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"
31 #include "workarounds.h"
33 /**************************************************************************
37 **************************************************************************
40 /* Loopback mode names (see LOOPBACK_MODE()) */
41 const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
42 const char *const efx_loopback_mode_names[] = {
43 [LOOPBACK_NONE] = "NONE",
44 [LOOPBACK_DATA] = "DATAPATH",
45 [LOOPBACK_GMAC] = "GMAC",
46 [LOOPBACK_XGMII] = "XGMII",
47 [LOOPBACK_XGXS] = "XGXS",
48 [LOOPBACK_XAUI] = "XAUI",
49 [LOOPBACK_GMII] = "GMII",
50 [LOOPBACK_SGMII] = "SGMII",
51 [LOOPBACK_XGBR] = "XGBR",
52 [LOOPBACK_XFI] = "XFI",
53 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
54 [LOOPBACK_GMII_FAR] = "GMII_FAR",
55 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
56 [LOOPBACK_XFI_FAR] = "XFI_FAR",
57 [LOOPBACK_GPHY] = "GPHY",
58 [LOOPBACK_PHYXS] = "PHYXS",
59 [LOOPBACK_PCS] = "PCS",
60 [LOOPBACK_PMAPMD] = "PMA/PMD",
61 [LOOPBACK_XPORT] = "XPORT",
62 [LOOPBACK_XGMII_WS] = "XGMII_WS",
63 [LOOPBACK_XAUI_WS] = "XAUI_WS",
64 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
65 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
66 [LOOPBACK_GMII_WS] = "GMII_WS",
67 [LOOPBACK_XFI_WS] = "XFI_WS",
68 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
69 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
72 const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
73 const char *const efx_reset_type_names[] = {
74 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
75 [RESET_TYPE_ALL] = "ALL",
76 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
77 [RESET_TYPE_WORLD] = "WORLD",
78 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
79 [RESET_TYPE_DISABLE] = "DISABLE",
80 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
81 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
82 [RESET_TYPE_RX_RECOVERY] = "RX_RECOVERY",
83 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
84 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
85 [RESET_TYPE_MC_FAILURE] = "MC_FAILURE",
86 [RESET_TYPE_MC_BIST] = "MC_BIST",
89 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
90 * queued onto this work queue. This is not a per-nic work queue, because
91 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
93 static struct workqueue_struct *reset_workqueue;
95 /* How often and how many times to poll for a reset while waiting for a
96 * BIST that another function started to complete.
98 #define BIST_WAIT_DELAY_MS 100
99 #define BIST_WAIT_DELAY_COUNT 100
101 /**************************************************************************
103 * Configurable values
105 *************************************************************************/
108 * Use separate channels for TX and RX events
110 * Set this to 1 to use separate channels for TX and RX. It allows us
111 * to control interrupt affinity separately for TX and RX.
113 * This is only used in MSI-X interrupt mode
115 static bool separate_tx_channels;
116 module_param(separate_tx_channels, bool, 0444);
117 MODULE_PARM_DESC(separate_tx_channels,
118 "Use separate channels for TX and RX");
120 /* This is the weight assigned to each of the (per-channel) virtual
123 static int napi_weight = 64;
125 /* This is the time (in jiffies) between invocations of the hardware
127 * On Falcon-based NICs, this will:
128 * - Check the on-board hardware monitor;
129 * - Poll the link state and reconfigure the hardware as necessary.
130 * On Siena-based NICs for power systems with EEH support, this will give EEH a
133 static unsigned int efx_monitor_interval = 1 * HZ;
135 /* Initial interrupt moderation settings. They can be modified after
136 * module load with ethtool.
138 * The default for RX should strike a balance between increasing the
139 * round-trip latency and reducing overhead.
141 static unsigned int rx_irq_mod_usec = 60;
143 /* Initial interrupt moderation settings. They can be modified after
144 * module load with ethtool.
146 * This default is chosen to ensure that a 10G link does not go idle
147 * while a TX queue is stopped after it has become full. A queue is
148 * restarted when it drops below half full. The time this takes (assuming
149 * worst case 3 descriptors per packet and 1024 descriptors) is
150 * 512 / 3 * 1.2 = 205 usec.
152 static unsigned int tx_irq_mod_usec = 150;
154 /* This is the first interrupt mode to try out of:
159 static unsigned int interrupt_mode;
161 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
162 * i.e. the number of CPUs among which we may distribute simultaneous
163 * interrupt handling.
165 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
166 * The default (0) means to assign an interrupt to each core.
168 static unsigned int rss_cpus;
169 module_param(rss_cpus, uint, 0444);
170 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
172 static bool phy_flash_cfg;
173 module_param(phy_flash_cfg, bool, 0644);
174 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
176 static unsigned irq_adapt_low_thresh = 8000;
177 module_param(irq_adapt_low_thresh, uint, 0644);
178 MODULE_PARM_DESC(irq_adapt_low_thresh,
179 "Threshold score for reducing IRQ moderation");
181 static unsigned irq_adapt_high_thresh = 16000;
182 module_param(irq_adapt_high_thresh, uint, 0644);
183 MODULE_PARM_DESC(irq_adapt_high_thresh,
184 "Threshold score for increasing IRQ moderation");
186 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
187 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
188 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
189 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
190 module_param(debug, uint, 0);
191 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
193 /**************************************************************************
195 * Utility functions and prototypes
197 *************************************************************************/
199 static int efx_soft_enable_interrupts(struct efx_nic *efx);
200 static void efx_soft_disable_interrupts(struct efx_nic *efx);
201 static void efx_remove_channel(struct efx_channel *channel);
202 static void efx_remove_channels(struct efx_nic *efx);
203 static const struct efx_channel_type efx_default_channel_type;
204 static void efx_remove_port(struct efx_nic *efx);
205 static void efx_init_napi_channel(struct efx_channel *channel);
206 static void efx_fini_napi(struct efx_nic *efx);
207 static void efx_fini_napi_channel(struct efx_channel *channel);
208 static void efx_fini_struct(struct efx_nic *efx);
209 static void efx_start_all(struct efx_nic *efx);
210 static void efx_stop_all(struct efx_nic *efx);
212 #define EFX_ASSERT_RESET_SERIALISED(efx) \
214 if ((efx->state == STATE_READY) || \
215 (efx->state == STATE_RECOVERY) || \
216 (efx->state == STATE_DISABLED)) \
220 static int efx_check_disabled(struct efx_nic *efx)
222 if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
223 netif_err(efx, drv, efx->net_dev,
224 "device is disabled due to earlier errors\n");
230 /**************************************************************************
232 * Event queue processing
234 *************************************************************************/
236 /* Process channel's event queue
238 * This function is responsible for processing the event queue of a
239 * single channel. The caller must guarantee that this function will
240 * never be concurrently called more than once on the same channel,
241 * though different channels may be being processed concurrently.
243 static int efx_process_channel(struct efx_channel *channel, int budget)
247 if (unlikely(!channel->enabled))
250 spent = efx_nic_process_eventq(channel, budget);
251 if (spent && efx_channel_has_rx_queue(channel)) {
252 struct efx_rx_queue *rx_queue =
253 efx_channel_get_rx_queue(channel);
255 efx_rx_flush_packet(channel);
256 efx_fast_push_rx_descriptors(rx_queue, true);
264 * NAPI guarantees serialisation of polls of the same device, which
265 * provides the guarantee required by efx_process_channel().
267 static int efx_poll(struct napi_struct *napi, int budget)
269 struct efx_channel *channel =
270 container_of(napi, struct efx_channel, napi_str);
271 struct efx_nic *efx = channel->efx;
274 netif_vdbg(efx, intr, efx->net_dev,
275 "channel %d NAPI poll executing on CPU %d\n",
276 channel->channel, raw_smp_processor_id());
278 spent = efx_process_channel(channel, budget);
280 if (spent < budget) {
281 if (efx_channel_has_rx_queue(channel) &&
282 efx->irq_rx_adaptive &&
283 unlikely(++channel->irq_count == 1000)) {
284 if (unlikely(channel->irq_mod_score <
285 irq_adapt_low_thresh)) {
286 if (channel->irq_moderation > 1) {
287 channel->irq_moderation -= 1;
288 efx->type->push_irq_moderation(channel);
290 } else if (unlikely(channel->irq_mod_score >
291 irq_adapt_high_thresh)) {
292 if (channel->irq_moderation <
293 efx->irq_rx_moderation) {
294 channel->irq_moderation += 1;
295 efx->type->push_irq_moderation(channel);
298 channel->irq_count = 0;
299 channel->irq_mod_score = 0;
302 efx_filter_rfs_expire(channel);
304 /* There is no race here; although napi_disable() will
305 * only wait for napi_complete(), this isn't a problem
306 * since efx_nic_eventq_read_ack() will have no effect if
307 * interrupts have already been disabled.
310 efx_nic_eventq_read_ack(channel);
316 /* Create event queue
317 * Event queue memory allocations are done only once. If the channel
318 * is reset, the memory buffer will be reused; this guards against
319 * errors during channel reset and also simplifies interrupt handling.
321 static int efx_probe_eventq(struct efx_channel *channel)
323 struct efx_nic *efx = channel->efx;
324 unsigned long entries;
326 netif_dbg(efx, probe, efx->net_dev,
327 "chan %d create event queue\n", channel->channel);
329 /* Build an event queue with room for one event per tx and rx buffer,
330 * plus some extra for link state events and MCDI completions. */
331 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
332 EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
333 channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
335 return efx_nic_probe_eventq(channel);
338 /* Prepare channel's event queue */
339 static int efx_init_eventq(struct efx_channel *channel)
341 struct efx_nic *efx = channel->efx;
344 EFX_WARN_ON_PARANOID(channel->eventq_init);
346 netif_dbg(efx, drv, efx->net_dev,
347 "chan %d init event queue\n", channel->channel);
349 rc = efx_nic_init_eventq(channel);
351 efx->type->push_irq_moderation(channel);
352 channel->eventq_read_ptr = 0;
353 channel->eventq_init = true;
358 /* Enable event queue processing and NAPI */
359 static void efx_start_eventq(struct efx_channel *channel)
361 netif_dbg(channel->efx, ifup, channel->efx->net_dev,
362 "chan %d start event queue\n", channel->channel);
364 /* Make sure the NAPI handler sees the enabled flag set */
365 channel->enabled = true;
368 napi_enable(&channel->napi_str);
369 efx_nic_eventq_read_ack(channel);
372 /* Disable event queue processing and NAPI */
373 static void efx_stop_eventq(struct efx_channel *channel)
375 if (!channel->enabled)
378 napi_disable(&channel->napi_str);
379 channel->enabled = false;
382 static void efx_fini_eventq(struct efx_channel *channel)
384 if (!channel->eventq_init)
387 netif_dbg(channel->efx, drv, channel->efx->net_dev,
388 "chan %d fini event queue\n", channel->channel);
390 efx_nic_fini_eventq(channel);
391 channel->eventq_init = false;
394 static void efx_remove_eventq(struct efx_channel *channel)
396 netif_dbg(channel->efx, drv, channel->efx->net_dev,
397 "chan %d remove event queue\n", channel->channel);
399 efx_nic_remove_eventq(channel);
402 /**************************************************************************
406 *************************************************************************/
408 /* Allocate and initialise a channel structure. */
409 static struct efx_channel *
410 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
412 struct efx_channel *channel;
413 struct efx_rx_queue *rx_queue;
414 struct efx_tx_queue *tx_queue;
417 channel = kzalloc(sizeof(*channel), GFP_KERNEL);
422 channel->channel = i;
423 channel->type = &efx_default_channel_type;
425 for (j = 0; j < EFX_TXQ_TYPES; j++) {
426 tx_queue = &channel->tx_queue[j];
428 tx_queue->queue = i * EFX_TXQ_TYPES + j;
429 tx_queue->channel = channel;
432 rx_queue = &channel->rx_queue;
434 setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
435 (unsigned long)rx_queue);
440 /* Allocate and initialise a channel structure, copying parameters
441 * (but not resources) from an old channel structure.
443 static struct efx_channel *
444 efx_copy_channel(const struct efx_channel *old_channel)
446 struct efx_channel *channel;
447 struct efx_rx_queue *rx_queue;
448 struct efx_tx_queue *tx_queue;
451 channel = kmalloc(sizeof(*channel), GFP_KERNEL);
455 *channel = *old_channel;
457 channel->napi_dev = NULL;
458 memset(&channel->eventq, 0, sizeof(channel->eventq));
460 for (j = 0; j < EFX_TXQ_TYPES; j++) {
461 tx_queue = &channel->tx_queue[j];
462 if (tx_queue->channel)
463 tx_queue->channel = channel;
464 tx_queue->buffer = NULL;
465 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
468 rx_queue = &channel->rx_queue;
469 rx_queue->buffer = NULL;
470 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
471 setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
472 (unsigned long)rx_queue);
477 static int efx_probe_channel(struct efx_channel *channel)
479 struct efx_tx_queue *tx_queue;
480 struct efx_rx_queue *rx_queue;
483 netif_dbg(channel->efx, probe, channel->efx->net_dev,
484 "creating channel %d\n", channel->channel);
486 rc = channel->type->pre_probe(channel);
490 rc = efx_probe_eventq(channel);
494 efx_for_each_channel_tx_queue(tx_queue, channel) {
495 rc = efx_probe_tx_queue(tx_queue);
500 efx_for_each_channel_rx_queue(rx_queue, channel) {
501 rc = efx_probe_rx_queue(rx_queue);
509 efx_remove_channel(channel);
514 efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
516 struct efx_nic *efx = channel->efx;
520 number = channel->channel;
521 if (efx->tx_channel_offset == 0) {
523 } else if (channel->channel < efx->tx_channel_offset) {
527 number -= efx->tx_channel_offset;
529 snprintf(buf, len, "%s%s-%d", efx->name, type, number);
532 static void efx_set_channel_names(struct efx_nic *efx)
534 struct efx_channel *channel;
536 efx_for_each_channel(channel, efx)
537 channel->type->get_name(channel,
538 efx->msi_context[channel->channel].name,
539 sizeof(efx->msi_context[0].name));
542 static int efx_probe_channels(struct efx_nic *efx)
544 struct efx_channel *channel;
547 /* Restart special buffer allocation */
548 efx->next_buffer_table = 0;
550 /* Probe channels in reverse, so that any 'extra' channels
551 * use the start of the buffer table. This allows the traffic
552 * channels to be resized without moving them or wasting the
553 * entries before them.
555 efx_for_each_channel_rev(channel, efx) {
556 rc = efx_probe_channel(channel);
558 netif_err(efx, probe, efx->net_dev,
559 "failed to create channel %d\n",
564 efx_set_channel_names(efx);
569 efx_remove_channels(efx);
573 /* Channels are shutdown and reinitialised whilst the NIC is running
574 * to propagate configuration changes (mtu, checksum offload), or
575 * to clear hardware error conditions
577 static void efx_start_datapath(struct efx_nic *efx)
579 bool old_rx_scatter = efx->rx_scatter;
580 struct efx_tx_queue *tx_queue;
581 struct efx_rx_queue *rx_queue;
582 struct efx_channel *channel;
585 /* Calculate the rx buffer allocation parameters required to
586 * support the current MTU, including padding for header
587 * alignment and overruns.
589 efx->rx_dma_len = (efx->rx_prefix_size +
590 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
591 efx->type->rx_buffer_padding);
592 rx_buf_len = (sizeof(struct efx_rx_page_state) +
593 efx->rx_ip_align + efx->rx_dma_len);
594 if (rx_buf_len <= PAGE_SIZE) {
595 efx->rx_scatter = efx->type->always_rx_scatter;
596 efx->rx_buffer_order = 0;
597 } else if (efx->type->can_rx_scatter) {
598 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
599 BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
600 2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
601 EFX_RX_BUF_ALIGNMENT) >
603 efx->rx_scatter = true;
604 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
605 efx->rx_buffer_order = 0;
607 efx->rx_scatter = false;
608 efx->rx_buffer_order = get_order(rx_buf_len);
611 efx_rx_config_page_split(efx);
612 if (efx->rx_buffer_order)
613 netif_dbg(efx, drv, efx->net_dev,
614 "RX buf len=%u; page order=%u batch=%u\n",
615 efx->rx_dma_len, efx->rx_buffer_order,
616 efx->rx_pages_per_batch);
618 netif_dbg(efx, drv, efx->net_dev,
619 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
620 efx->rx_dma_len, efx->rx_page_buf_step,
621 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
623 /* RX filters may also have scatter-enabled flags */
624 if (efx->rx_scatter != old_rx_scatter)
625 efx->type->filter_update_rx_scatter(efx);
627 /* We must keep at least one descriptor in a TX ring empty.
628 * We could avoid this when the queue size does not exactly
629 * match the hardware ring size, but it's not that important.
630 * Therefore we stop the queue when one more skb might fill
631 * the ring completely. We wake it when half way back to
634 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
635 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
637 /* Initialise the channels */
638 efx_for_each_channel(channel, efx) {
639 efx_for_each_channel_tx_queue(tx_queue, channel) {
640 efx_init_tx_queue(tx_queue);
641 atomic_inc(&efx->active_queues);
644 efx_for_each_channel_rx_queue(rx_queue, channel) {
645 efx_init_rx_queue(rx_queue);
646 atomic_inc(&efx->active_queues);
647 efx_stop_eventq(channel);
648 efx_fast_push_rx_descriptors(rx_queue, false);
649 efx_start_eventq(channel);
652 WARN_ON(channel->rx_pkt_n_frags);
655 efx_ptp_start_datapath(efx);
657 if (netif_device_present(efx->net_dev))
658 netif_tx_wake_all_queues(efx->net_dev);
661 static void efx_stop_datapath(struct efx_nic *efx)
663 struct efx_channel *channel;
664 struct efx_tx_queue *tx_queue;
665 struct efx_rx_queue *rx_queue;
668 EFX_ASSERT_RESET_SERIALISED(efx);
669 BUG_ON(efx->port_enabled);
671 efx_ptp_stop_datapath(efx);
674 efx_for_each_channel(channel, efx) {
675 efx_for_each_channel_rx_queue(rx_queue, channel)
676 rx_queue->refill_enabled = false;
679 efx_for_each_channel(channel, efx) {
680 /* RX packet processing is pipelined, so wait for the
681 * NAPI handler to complete. At least event queue 0
682 * might be kept active by non-data events, so don't
683 * use napi_synchronize() but actually disable NAPI
686 if (efx_channel_has_rx_queue(channel)) {
687 efx_stop_eventq(channel);
688 efx_start_eventq(channel);
692 rc = efx->type->fini_dmaq(efx);
693 if (rc && EFX_WORKAROUND_7803(efx)) {
694 /* Schedule a reset to recover from the flush failure. The
695 * descriptor caches reference memory we're about to free,
696 * but falcon_reconfigure_mac_wrapper() won't reconnect
697 * the MACs because of the pending reset.
699 netif_err(efx, drv, efx->net_dev,
700 "Resetting to recover from flush failure\n");
701 efx_schedule_reset(efx, RESET_TYPE_ALL);
703 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
705 netif_dbg(efx, drv, efx->net_dev,
706 "successfully flushed all queues\n");
709 efx_for_each_channel(channel, efx) {
710 efx_for_each_channel_rx_queue(rx_queue, channel)
711 efx_fini_rx_queue(rx_queue);
712 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
713 efx_fini_tx_queue(tx_queue);
717 static void efx_remove_channel(struct efx_channel *channel)
719 struct efx_tx_queue *tx_queue;
720 struct efx_rx_queue *rx_queue;
722 netif_dbg(channel->efx, drv, channel->efx->net_dev,
723 "destroy chan %d\n", channel->channel);
725 efx_for_each_channel_rx_queue(rx_queue, channel)
726 efx_remove_rx_queue(rx_queue);
727 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
728 efx_remove_tx_queue(tx_queue);
729 efx_remove_eventq(channel);
730 channel->type->post_remove(channel);
733 static void efx_remove_channels(struct efx_nic *efx)
735 struct efx_channel *channel;
737 efx_for_each_channel(channel, efx)
738 efx_remove_channel(channel);
742 efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
744 struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
745 u32 old_rxq_entries, old_txq_entries;
746 unsigned i, next_buffer_table = 0;
749 rc = efx_check_disabled(efx);
753 /* Not all channels should be reallocated. We must avoid
754 * reallocating their buffer table entries.
756 efx_for_each_channel(channel, efx) {
757 struct efx_rx_queue *rx_queue;
758 struct efx_tx_queue *tx_queue;
760 if (channel->type->copy)
762 next_buffer_table = max(next_buffer_table,
763 channel->eventq.index +
764 channel->eventq.entries);
765 efx_for_each_channel_rx_queue(rx_queue, channel)
766 next_buffer_table = max(next_buffer_table,
767 rx_queue->rxd.index +
768 rx_queue->rxd.entries);
769 efx_for_each_channel_tx_queue(tx_queue, channel)
770 next_buffer_table = max(next_buffer_table,
771 tx_queue->txd.index +
772 tx_queue->txd.entries);
775 efx_device_detach_sync(efx);
777 efx_soft_disable_interrupts(efx);
779 /* Clone channels (where possible) */
780 memset(other_channel, 0, sizeof(other_channel));
781 for (i = 0; i < efx->n_channels; i++) {
782 channel = efx->channel[i];
783 if (channel->type->copy)
784 channel = channel->type->copy(channel);
789 other_channel[i] = channel;
792 /* Swap entry counts and channel pointers */
793 old_rxq_entries = efx->rxq_entries;
794 old_txq_entries = efx->txq_entries;
795 efx->rxq_entries = rxq_entries;
796 efx->txq_entries = txq_entries;
797 for (i = 0; i < efx->n_channels; i++) {
798 channel = efx->channel[i];
799 efx->channel[i] = other_channel[i];
800 other_channel[i] = channel;
803 /* Restart buffer table allocation */
804 efx->next_buffer_table = next_buffer_table;
806 for (i = 0; i < efx->n_channels; i++) {
807 channel = efx->channel[i];
808 if (!channel->type->copy)
810 rc = efx_probe_channel(channel);
813 efx_init_napi_channel(efx->channel[i]);
817 /* Destroy unused channel structures */
818 for (i = 0; i < efx->n_channels; i++) {
819 channel = other_channel[i];
820 if (channel && channel->type->copy) {
821 efx_fini_napi_channel(channel);
822 efx_remove_channel(channel);
827 rc2 = efx_soft_enable_interrupts(efx);
830 netif_err(efx, drv, efx->net_dev,
831 "unable to restart interrupts on channel reallocation\n");
832 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
835 netif_device_attach(efx->net_dev);
841 efx->rxq_entries = old_rxq_entries;
842 efx->txq_entries = old_txq_entries;
843 for (i = 0; i < efx->n_channels; i++) {
844 channel = efx->channel[i];
845 efx->channel[i] = other_channel[i];
846 other_channel[i] = channel;
851 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
853 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
856 static const struct efx_channel_type efx_default_channel_type = {
857 .pre_probe = efx_channel_dummy_op_int,
858 .post_remove = efx_channel_dummy_op_void,
859 .get_name = efx_get_channel_name,
860 .copy = efx_copy_channel,
861 .keep_eventq = false,
864 int efx_channel_dummy_op_int(struct efx_channel *channel)
869 void efx_channel_dummy_op_void(struct efx_channel *channel)
873 /**************************************************************************
877 **************************************************************************/
879 /* This ensures that the kernel is kept informed (via
880 * netif_carrier_on/off) of the link status, and also maintains the
881 * link status's stop on the port's TX queue.
883 void efx_link_status_changed(struct efx_nic *efx)
885 struct efx_link_state *link_state = &efx->link_state;
887 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
888 * that no events are triggered between unregister_netdev() and the
889 * driver unloading. A more general condition is that NETDEV_CHANGE
890 * can only be generated between NETDEV_UP and NETDEV_DOWN */
891 if (!netif_running(efx->net_dev))
894 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
895 efx->n_link_state_changes++;
898 netif_carrier_on(efx->net_dev);
900 netif_carrier_off(efx->net_dev);
903 /* Status message for kernel log */
905 netif_info(efx, link, efx->net_dev,
906 "link up at %uMbps %s-duplex (MTU %d)\n",
907 link_state->speed, link_state->fd ? "full" : "half",
910 netif_info(efx, link, efx->net_dev, "link down\n");
913 void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
915 efx->link_advertising = advertising;
917 if (advertising & ADVERTISED_Pause)
918 efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
920 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
921 if (advertising & ADVERTISED_Asym_Pause)
922 efx->wanted_fc ^= EFX_FC_TX;
926 void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
928 efx->wanted_fc = wanted_fc;
929 if (efx->link_advertising) {
930 if (wanted_fc & EFX_FC_RX)
931 efx->link_advertising |= (ADVERTISED_Pause |
932 ADVERTISED_Asym_Pause);
934 efx->link_advertising &= ~(ADVERTISED_Pause |
935 ADVERTISED_Asym_Pause);
936 if (wanted_fc & EFX_FC_TX)
937 efx->link_advertising ^= ADVERTISED_Asym_Pause;
941 static void efx_fini_port(struct efx_nic *efx);
943 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
944 * the MAC appropriately. All other PHY configuration changes are pushed
945 * through phy_op->set_settings(), and pushed asynchronously to the MAC
946 * through efx_monitor().
948 * Callers must hold the mac_lock
950 int __efx_reconfigure_port(struct efx_nic *efx)
952 enum efx_phy_mode phy_mode;
955 WARN_ON(!mutex_is_locked(&efx->mac_lock));
957 /* Disable PHY transmit in mac level loopbacks */
958 phy_mode = efx->phy_mode;
959 if (LOOPBACK_INTERNAL(efx))
960 efx->phy_mode |= PHY_MODE_TX_DISABLED;
962 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
964 rc = efx->type->reconfigure_port(efx);
967 efx->phy_mode = phy_mode;
972 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
974 int efx_reconfigure_port(struct efx_nic *efx)
978 EFX_ASSERT_RESET_SERIALISED(efx);
980 mutex_lock(&efx->mac_lock);
981 rc = __efx_reconfigure_port(efx);
982 mutex_unlock(&efx->mac_lock);
987 /* Asynchronous work item for changing MAC promiscuity and multicast
988 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
990 static void efx_mac_work(struct work_struct *data)
992 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
994 mutex_lock(&efx->mac_lock);
995 if (efx->port_enabled)
996 efx->type->reconfigure_mac(efx);
997 mutex_unlock(&efx->mac_lock);
1000 static int efx_probe_port(struct efx_nic *efx)
1004 netif_dbg(efx, probe, efx->net_dev, "create port\n");
1007 efx->phy_mode = PHY_MODE_SPECIAL;
1009 /* Connect up MAC/PHY operations table */
1010 rc = efx->type->probe_port(efx);
1014 /* Initialise MAC address to permanent address */
1015 memcpy(efx->net_dev->dev_addr, efx->net_dev->perm_addr, ETH_ALEN);
1020 static int efx_init_port(struct efx_nic *efx)
1024 netif_dbg(efx, drv, efx->net_dev, "init port\n");
1026 mutex_lock(&efx->mac_lock);
1028 rc = efx->phy_op->init(efx);
1032 efx->port_initialized = true;
1034 /* Reconfigure the MAC before creating dma queues (required for
1035 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1036 efx->type->reconfigure_mac(efx);
1038 /* Ensure the PHY advertises the correct flow control settings */
1039 rc = efx->phy_op->reconfigure(efx);
1043 mutex_unlock(&efx->mac_lock);
1047 efx->phy_op->fini(efx);
1049 mutex_unlock(&efx->mac_lock);
1053 static void efx_start_port(struct efx_nic *efx)
1055 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1056 BUG_ON(efx->port_enabled);
1058 mutex_lock(&efx->mac_lock);
1059 efx->port_enabled = true;
1061 /* Ensure MAC ingress/egress is enabled */
1062 efx->type->reconfigure_mac(efx);
1064 mutex_unlock(&efx->mac_lock);
1067 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1068 * and the async self-test, wait for them to finish and prevent them
1069 * being scheduled again. This doesn't cover online resets, which
1070 * should only be cancelled when removing the device.
1072 static void efx_stop_port(struct efx_nic *efx)
1074 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1076 EFX_ASSERT_RESET_SERIALISED(efx);
1078 mutex_lock(&efx->mac_lock);
1079 efx->port_enabled = false;
1080 mutex_unlock(&efx->mac_lock);
1082 /* Serialise against efx_set_multicast_list() */
1083 netif_addr_lock_bh(efx->net_dev);
1084 netif_addr_unlock_bh(efx->net_dev);
1086 cancel_delayed_work_sync(&efx->monitor_work);
1087 efx_selftest_async_cancel(efx);
1088 cancel_work_sync(&efx->mac_work);
1091 static void efx_fini_port(struct efx_nic *efx)
1093 netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1095 if (!efx->port_initialized)
1098 efx->phy_op->fini(efx);
1099 efx->port_initialized = false;
1101 efx->link_state.up = false;
1102 efx_link_status_changed(efx);
1105 static void efx_remove_port(struct efx_nic *efx)
1107 netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1109 efx->type->remove_port(efx);
1112 /**************************************************************************
1116 **************************************************************************/
1118 static LIST_HEAD(efx_primary_list);
1119 static LIST_HEAD(efx_unassociated_list);
1121 static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1123 return left->type == right->type &&
1124 left->vpd_sn && right->vpd_sn &&
1125 !strcmp(left->vpd_sn, right->vpd_sn);
1128 static void efx_associate(struct efx_nic *efx)
1130 struct efx_nic *other, *next;
1132 if (efx->primary == efx) {
1133 /* Adding primary function; look for secondaries */
1135 netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1136 list_add_tail(&efx->node, &efx_primary_list);
1138 list_for_each_entry_safe(other, next, &efx_unassociated_list,
1140 if (efx_same_controller(efx, other)) {
1141 list_del(&other->node);
1142 netif_dbg(other, probe, other->net_dev,
1143 "moving to secondary list of %s %s\n",
1144 pci_name(efx->pci_dev),
1145 efx->net_dev->name);
1146 list_add_tail(&other->node,
1147 &efx->secondary_list);
1148 other->primary = efx;
1152 /* Adding secondary function; look for primary */
1154 list_for_each_entry(other, &efx_primary_list, node) {
1155 if (efx_same_controller(efx, other)) {
1156 netif_dbg(efx, probe, efx->net_dev,
1157 "adding to secondary list of %s %s\n",
1158 pci_name(other->pci_dev),
1159 other->net_dev->name);
1160 list_add_tail(&efx->node,
1161 &other->secondary_list);
1162 efx->primary = other;
1167 netif_dbg(efx, probe, efx->net_dev,
1168 "adding to unassociated list\n");
1169 list_add_tail(&efx->node, &efx_unassociated_list);
1173 static void efx_dissociate(struct efx_nic *efx)
1175 struct efx_nic *other, *next;
1177 list_del(&efx->node);
1178 efx->primary = NULL;
1180 list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1181 list_del(&other->node);
1182 netif_dbg(other, probe, other->net_dev,
1183 "moving to unassociated list\n");
1184 list_add_tail(&other->node, &efx_unassociated_list);
1185 other->primary = NULL;
1189 /* This configures the PCI device to enable I/O and DMA. */
1190 static int efx_init_io(struct efx_nic *efx)
1192 struct pci_dev *pci_dev = efx->pci_dev;
1193 dma_addr_t dma_mask = efx->type->max_dma_mask;
1194 unsigned int mem_map_size = efx->type->mem_map_size(efx);
1197 netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1199 rc = pci_enable_device(pci_dev);
1201 netif_err(efx, probe, efx->net_dev,
1202 "failed to enable PCI device\n");
1206 pci_set_master(pci_dev);
1208 /* Set the PCI DMA mask. Try all possibilities from our
1209 * genuine mask down to 32 bits, because some architectures
1210 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1211 * masks event though they reject 46 bit masks.
1213 while (dma_mask > 0x7fffffffUL) {
1214 if (dma_supported(&pci_dev->dev, dma_mask)) {
1215 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1222 netif_err(efx, probe, efx->net_dev,
1223 "could not find a suitable DMA mask\n");
1226 netif_dbg(efx, probe, efx->net_dev,
1227 "using DMA mask %llx\n", (unsigned long long) dma_mask);
1229 efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1230 rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1232 netif_err(efx, probe, efx->net_dev,
1233 "request for memory BAR failed\n");
1237 efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1238 if (!efx->membase) {
1239 netif_err(efx, probe, efx->net_dev,
1240 "could not map memory BAR at %llx+%x\n",
1241 (unsigned long long)efx->membase_phys, mem_map_size);
1245 netif_dbg(efx, probe, efx->net_dev,
1246 "memory BAR at %llx+%x (virtual %p)\n",
1247 (unsigned long long)efx->membase_phys, mem_map_size,
1253 pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1255 efx->membase_phys = 0;
1257 pci_disable_device(efx->pci_dev);
1262 static void efx_fini_io(struct efx_nic *efx)
1264 netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1267 iounmap(efx->membase);
1268 efx->membase = NULL;
1271 if (efx->membase_phys) {
1272 pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1273 efx->membase_phys = 0;
1276 pci_disable_device(efx->pci_dev);
1279 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1281 cpumask_var_t thread_mask;
1288 if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1289 netif_warn(efx, probe, efx->net_dev,
1290 "RSS disabled due to allocation failure\n");
1295 for_each_online_cpu(cpu) {
1296 if (!cpumask_test_cpu(cpu, thread_mask)) {
1298 cpumask_or(thread_mask, thread_mask,
1299 topology_thread_cpumask(cpu));
1303 free_cpumask_var(thread_mask);
1306 /* If RSS is requested for the PF *and* VFs then we can't write RSS
1307 * table entries that are inaccessible to VFs
1309 if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1310 count > efx_vf_size(efx)) {
1311 netif_warn(efx, probe, efx->net_dev,
1312 "Reducing number of RSS channels from %u to %u for "
1313 "VF support. Increase vf-msix-limit to use more "
1314 "channels on the PF.\n",
1315 count, efx_vf_size(efx));
1316 count = efx_vf_size(efx);
1322 /* Probe the number and type of interrupts we are able to obtain, and
1323 * the resulting numbers of channels and RX queues.
1325 static int efx_probe_interrupts(struct efx_nic *efx)
1327 unsigned int extra_channels = 0;
1331 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1332 if (efx->extra_channel_type[i])
1335 if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1336 struct msix_entry xentries[EFX_MAX_CHANNELS];
1337 unsigned int n_channels;
1339 n_channels = efx_wanted_parallelism(efx);
1340 if (separate_tx_channels)
1342 n_channels += extra_channels;
1343 n_channels = min(n_channels, efx->max_channels);
1345 for (i = 0; i < n_channels; i++)
1346 xentries[i].entry = i;
1347 rc = pci_enable_msix(efx->pci_dev, xentries, n_channels);
1349 netif_err(efx, drv, efx->net_dev,
1350 "WARNING: Insufficient MSI-X vectors"
1351 " available (%d < %u).\n", rc, n_channels);
1352 netif_err(efx, drv, efx->net_dev,
1353 "WARNING: Performance may be reduced.\n");
1354 EFX_BUG_ON_PARANOID(rc >= n_channels);
1356 rc = pci_enable_msix(efx->pci_dev, xentries,
1361 efx->n_channels = n_channels;
1362 if (n_channels > extra_channels)
1363 n_channels -= extra_channels;
1364 if (separate_tx_channels) {
1365 efx->n_tx_channels = max(n_channels / 2, 1U);
1366 efx->n_rx_channels = max(n_channels -
1370 efx->n_tx_channels = n_channels;
1371 efx->n_rx_channels = n_channels;
1373 for (i = 0; i < efx->n_channels; i++)
1374 efx_get_channel(efx, i)->irq =
1377 /* Fall back to single channel MSI */
1378 efx->interrupt_mode = EFX_INT_MODE_MSI;
1379 netif_err(efx, drv, efx->net_dev,
1380 "could not enable MSI-X\n");
1384 /* Try single interrupt MSI */
1385 if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1386 efx->n_channels = 1;
1387 efx->n_rx_channels = 1;
1388 efx->n_tx_channels = 1;
1389 rc = pci_enable_msi(efx->pci_dev);
1391 efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1393 netif_err(efx, drv, efx->net_dev,
1394 "could not enable MSI\n");
1395 efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1399 /* Assume legacy interrupts */
1400 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1401 efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1402 efx->n_rx_channels = 1;
1403 efx->n_tx_channels = 1;
1404 efx->legacy_irq = efx->pci_dev->irq;
1407 /* Assign extra channels if possible */
1408 j = efx->n_channels;
1409 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1410 if (!efx->extra_channel_type[i])
1412 if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1413 efx->n_channels <= extra_channels) {
1414 efx->extra_channel_type[i]->handle_no_channel(efx);
1417 efx_get_channel(efx, j)->type =
1418 efx->extra_channel_type[i];
1422 /* RSS might be usable on VFs even if it is disabled on the PF */
1423 efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1424 efx->n_rx_channels : efx_vf_size(efx));
1429 static int efx_soft_enable_interrupts(struct efx_nic *efx)
1431 struct efx_channel *channel, *end_channel;
1434 BUG_ON(efx->state == STATE_DISABLED);
1436 efx->irq_soft_enabled = true;
1439 efx_for_each_channel(channel, efx) {
1440 if (!channel->type->keep_eventq) {
1441 rc = efx_init_eventq(channel);
1445 efx_start_eventq(channel);
1448 efx_mcdi_mode_event(efx);
1452 end_channel = channel;
1453 efx_for_each_channel(channel, efx) {
1454 if (channel == end_channel)
1456 efx_stop_eventq(channel);
1457 if (!channel->type->keep_eventq)
1458 efx_fini_eventq(channel);
1464 static void efx_soft_disable_interrupts(struct efx_nic *efx)
1466 struct efx_channel *channel;
1468 if (efx->state == STATE_DISABLED)
1471 efx_mcdi_mode_poll(efx);
1473 efx->irq_soft_enabled = false;
1476 if (efx->legacy_irq)
1477 synchronize_irq(efx->legacy_irq);
1479 efx_for_each_channel(channel, efx) {
1481 synchronize_irq(channel->irq);
1483 efx_stop_eventq(channel);
1484 if (!channel->type->keep_eventq)
1485 efx_fini_eventq(channel);
1488 /* Flush the asynchronous MCDI request queue */
1489 efx_mcdi_flush_async(efx);
1492 static int efx_enable_interrupts(struct efx_nic *efx)
1494 struct efx_channel *channel, *end_channel;
1497 BUG_ON(efx->state == STATE_DISABLED);
1499 if (efx->eeh_disabled_legacy_irq) {
1500 enable_irq(efx->legacy_irq);
1501 efx->eeh_disabled_legacy_irq = false;
1504 efx->type->irq_enable_master(efx);
1506 efx_for_each_channel(channel, efx) {
1507 if (channel->type->keep_eventq) {
1508 rc = efx_init_eventq(channel);
1514 rc = efx_soft_enable_interrupts(efx);
1521 end_channel = channel;
1522 efx_for_each_channel(channel, efx) {
1523 if (channel == end_channel)
1525 if (channel->type->keep_eventq)
1526 efx_fini_eventq(channel);
1529 efx->type->irq_disable_non_ev(efx);
1534 static void efx_disable_interrupts(struct efx_nic *efx)
1536 struct efx_channel *channel;
1538 efx_soft_disable_interrupts(efx);
1540 efx_for_each_channel(channel, efx) {
1541 if (channel->type->keep_eventq)
1542 efx_fini_eventq(channel);
1545 efx->type->irq_disable_non_ev(efx);
1548 static void efx_remove_interrupts(struct efx_nic *efx)
1550 struct efx_channel *channel;
1552 /* Remove MSI/MSI-X interrupts */
1553 efx_for_each_channel(channel, efx)
1555 pci_disable_msi(efx->pci_dev);
1556 pci_disable_msix(efx->pci_dev);
1558 /* Remove legacy interrupt */
1559 efx->legacy_irq = 0;
1562 static void efx_set_channels(struct efx_nic *efx)
1564 struct efx_channel *channel;
1565 struct efx_tx_queue *tx_queue;
1567 efx->tx_channel_offset =
1568 separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1570 /* We need to mark which channels really have RX and TX
1571 * queues, and adjust the TX queue numbers if we have separate
1572 * RX-only and TX-only channels.
1574 efx_for_each_channel(channel, efx) {
1575 if (channel->channel < efx->n_rx_channels)
1576 channel->rx_queue.core_index = channel->channel;
1578 channel->rx_queue.core_index = -1;
1580 efx_for_each_channel_tx_queue(tx_queue, channel)
1581 tx_queue->queue -= (efx->tx_channel_offset *
1586 static int efx_probe_nic(struct efx_nic *efx)
1591 netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1593 /* Carry out hardware-type specific initialisation */
1594 rc = efx->type->probe(efx);
1598 /* Determine the number of channels and queues by trying to hook
1599 * in MSI-X interrupts. */
1600 rc = efx_probe_interrupts(efx);
1604 rc = efx->type->dimension_resources(efx);
1608 if (efx->n_channels > 1)
1609 get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1610 for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1611 efx->rx_indir_table[i] =
1612 ethtool_rxfh_indir_default(i, efx->rss_spread);
1614 efx_set_channels(efx);
1615 netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1616 netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1618 /* Initialise the interrupt moderation settings */
1619 efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1625 efx_remove_interrupts(efx);
1627 efx->type->remove(efx);
1631 static void efx_remove_nic(struct efx_nic *efx)
1633 netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1635 efx_remove_interrupts(efx);
1636 efx->type->remove(efx);
1639 static int efx_probe_filters(struct efx_nic *efx)
1643 spin_lock_init(&efx->filter_lock);
1645 rc = efx->type->filter_table_probe(efx);
1649 #ifdef CONFIG_RFS_ACCEL
1650 if (efx->type->offload_features & NETIF_F_NTUPLE) {
1651 efx->rps_flow_id = kcalloc(efx->type->max_rx_ip_filters,
1652 sizeof(*efx->rps_flow_id),
1654 if (!efx->rps_flow_id) {
1655 efx->type->filter_table_remove(efx);
1664 static void efx_remove_filters(struct efx_nic *efx)
1666 #ifdef CONFIG_RFS_ACCEL
1667 kfree(efx->rps_flow_id);
1669 efx->type->filter_table_remove(efx);
1672 static void efx_restore_filters(struct efx_nic *efx)
1674 efx->type->filter_table_restore(efx);
1677 /**************************************************************************
1679 * NIC startup/shutdown
1681 *************************************************************************/
1683 static int efx_probe_all(struct efx_nic *efx)
1687 rc = efx_probe_nic(efx);
1689 netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1693 rc = efx_probe_port(efx);
1695 netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1699 BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1700 if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1704 efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1706 rc = efx_probe_filters(efx);
1708 netif_err(efx, probe, efx->net_dev,
1709 "failed to create filter tables\n");
1713 rc = efx_probe_channels(efx);
1720 efx_remove_filters(efx);
1722 efx_remove_port(efx);
1724 efx_remove_nic(efx);
1729 /* If the interface is supposed to be running but is not, start
1730 * the hardware and software data path, regular activity for the port
1731 * (MAC statistics, link polling, etc.) and schedule the port to be
1732 * reconfigured. Interrupts must already be enabled. This function
1733 * is safe to call multiple times, so long as the NIC is not disabled.
1734 * Requires the RTNL lock.
1736 static void efx_start_all(struct efx_nic *efx)
1738 EFX_ASSERT_RESET_SERIALISED(efx);
1739 BUG_ON(efx->state == STATE_DISABLED);
1741 /* Check that it is appropriate to restart the interface. All
1742 * of these flags are safe to read under just the rtnl lock */
1743 if (efx->port_enabled || !netif_running(efx->net_dev))
1746 efx_start_port(efx);
1747 efx_start_datapath(efx);
1749 /* Start the hardware monitor if there is one */
1750 if (efx->type->monitor != NULL)
1751 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1752 efx_monitor_interval);
1754 /* If link state detection is normally event-driven, we have
1755 * to poll now because we could have missed a change
1757 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1758 mutex_lock(&efx->mac_lock);
1759 if (efx->phy_op->poll(efx))
1760 efx_link_status_changed(efx);
1761 mutex_unlock(&efx->mac_lock);
1764 efx->type->start_stats(efx);
1765 efx->type->pull_stats(efx);
1766 spin_lock_bh(&efx->stats_lock);
1767 efx->type->update_stats(efx, NULL, NULL);
1768 spin_unlock_bh(&efx->stats_lock);
1771 /* Quiesce the hardware and software data path, and regular activity
1772 * for the port without bringing the link down. Safe to call multiple
1773 * times with the NIC in almost any state, but interrupts should be
1774 * enabled. Requires the RTNL lock.
1776 static void efx_stop_all(struct efx_nic *efx)
1778 EFX_ASSERT_RESET_SERIALISED(efx);
1780 /* port_enabled can be read safely under the rtnl lock */
1781 if (!efx->port_enabled)
1784 /* update stats before we go down so we can accurately count
1787 efx->type->pull_stats(efx);
1788 spin_lock_bh(&efx->stats_lock);
1789 efx->type->update_stats(efx, NULL, NULL);
1790 spin_unlock_bh(&efx->stats_lock);
1791 efx->type->stop_stats(efx);
1794 /* Stop the kernel transmit interface. This is only valid if
1795 * the device is stopped or detached; otherwise the watchdog
1796 * may fire immediately.
1798 WARN_ON(netif_running(efx->net_dev) &&
1799 netif_device_present(efx->net_dev));
1800 netif_tx_disable(efx->net_dev);
1802 efx_stop_datapath(efx);
1805 static void efx_remove_all(struct efx_nic *efx)
1807 efx_remove_channels(efx);
1808 efx_remove_filters(efx);
1809 efx_remove_port(efx);
1810 efx_remove_nic(efx);
1813 /**************************************************************************
1815 * Interrupt moderation
1817 **************************************************************************/
1819 static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1823 if (usecs * 1000 < quantum_ns)
1824 return 1; /* never round down to 0 */
1825 return usecs * 1000 / quantum_ns;
1828 /* Set interrupt moderation parameters */
1829 int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1830 unsigned int rx_usecs, bool rx_adaptive,
1831 bool rx_may_override_tx)
1833 struct efx_channel *channel;
1834 unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1835 efx->timer_quantum_ns,
1837 unsigned int tx_ticks;
1838 unsigned int rx_ticks;
1840 EFX_ASSERT_RESET_SERIALISED(efx);
1842 if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1845 tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1846 rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1848 if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1849 !rx_may_override_tx) {
1850 netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1851 "RX and TX IRQ moderation must be equal\n");
1855 efx->irq_rx_adaptive = rx_adaptive;
1856 efx->irq_rx_moderation = rx_ticks;
1857 efx_for_each_channel(channel, efx) {
1858 if (efx_channel_has_rx_queue(channel))
1859 channel->irq_moderation = rx_ticks;
1860 else if (efx_channel_has_tx_queues(channel))
1861 channel->irq_moderation = tx_ticks;
1867 void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1868 unsigned int *rx_usecs, bool *rx_adaptive)
1870 /* We must round up when converting ticks to microseconds
1871 * because we round down when converting the other way.
1874 *rx_adaptive = efx->irq_rx_adaptive;
1875 *rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1876 efx->timer_quantum_ns,
1879 /* If channels are shared between RX and TX, so is IRQ
1880 * moderation. Otherwise, IRQ moderation is the same for all
1881 * TX channels and is not adaptive.
1883 if (efx->tx_channel_offset == 0)
1884 *tx_usecs = *rx_usecs;
1886 *tx_usecs = DIV_ROUND_UP(
1887 efx->channel[efx->tx_channel_offset]->irq_moderation *
1888 efx->timer_quantum_ns,
1892 /**************************************************************************
1896 **************************************************************************/
1898 /* Run periodically off the general workqueue */
1899 static void efx_monitor(struct work_struct *data)
1901 struct efx_nic *efx = container_of(data, struct efx_nic,
1904 netif_vdbg(efx, timer, efx->net_dev,
1905 "hardware monitor executing on CPU %d\n",
1906 raw_smp_processor_id());
1907 BUG_ON(efx->type->monitor == NULL);
1909 /* If the mac_lock is already held then it is likely a port
1910 * reconfiguration is already in place, which will likely do
1911 * most of the work of monitor() anyway. */
1912 if (mutex_trylock(&efx->mac_lock)) {
1913 if (efx->port_enabled)
1914 efx->type->monitor(efx);
1915 mutex_unlock(&efx->mac_lock);
1918 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1919 efx_monitor_interval);
1922 /**************************************************************************
1926 *************************************************************************/
1929 * Context: process, rtnl_lock() held.
1931 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1933 struct efx_nic *efx = netdev_priv(net_dev);
1934 struct mii_ioctl_data *data = if_mii(ifr);
1936 if (cmd == SIOCSHWTSTAMP)
1937 return efx_ptp_set_ts_config(efx, ifr);
1938 if (cmd == SIOCGHWTSTAMP)
1939 return efx_ptp_get_ts_config(efx, ifr);
1941 /* Convert phy_id from older PRTAD/DEVAD format */
1942 if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1943 (data->phy_id & 0xfc00) == 0x0400)
1944 data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1946 return mdio_mii_ioctl(&efx->mdio, data, cmd);
1949 /**************************************************************************
1953 **************************************************************************/
1955 static void efx_init_napi_channel(struct efx_channel *channel)
1957 struct efx_nic *efx = channel->efx;
1959 channel->napi_dev = efx->net_dev;
1960 netif_napi_add(channel->napi_dev, &channel->napi_str,
1961 efx_poll, napi_weight);
1964 static void efx_init_napi(struct efx_nic *efx)
1966 struct efx_channel *channel;
1968 efx_for_each_channel(channel, efx)
1969 efx_init_napi_channel(channel);
1972 static void efx_fini_napi_channel(struct efx_channel *channel)
1974 if (channel->napi_dev)
1975 netif_napi_del(&channel->napi_str);
1976 channel->napi_dev = NULL;
1979 static void efx_fini_napi(struct efx_nic *efx)
1981 struct efx_channel *channel;
1983 efx_for_each_channel(channel, efx)
1984 efx_fini_napi_channel(channel);
1987 /**************************************************************************
1989 * Kernel netpoll interface
1991 *************************************************************************/
1993 #ifdef CONFIG_NET_POLL_CONTROLLER
1995 /* Although in the common case interrupts will be disabled, this is not
1996 * guaranteed. However, all our work happens inside the NAPI callback,
1997 * so no locking is required.
1999 static void efx_netpoll(struct net_device *net_dev)
2001 struct efx_nic *efx = netdev_priv(net_dev);
2002 struct efx_channel *channel;
2004 efx_for_each_channel(channel, efx)
2005 efx_schedule_channel(channel);
2010 /**************************************************************************
2012 * Kernel net device interface
2014 *************************************************************************/
2016 /* Context: process, rtnl_lock() held. */
2017 static int efx_net_open(struct net_device *net_dev)
2019 struct efx_nic *efx = netdev_priv(net_dev);
2022 netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2023 raw_smp_processor_id());
2025 rc = efx_check_disabled(efx);
2028 if (efx->phy_mode & PHY_MODE_SPECIAL)
2030 if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2033 /* Notify the kernel of the link state polled during driver load,
2034 * before the monitor starts running */
2035 efx_link_status_changed(efx);
2038 efx_selftest_async_start(efx);
2042 /* Context: process, rtnl_lock() held.
2043 * Note that the kernel will ignore our return code; this method
2044 * should really be a void.
2046 static int efx_net_stop(struct net_device *net_dev)
2048 struct efx_nic *efx = netdev_priv(net_dev);
2050 netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2051 raw_smp_processor_id());
2053 /* Stop the device and flush all the channels */
2059 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2060 static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
2061 struct rtnl_link_stats64 *stats)
2063 struct efx_nic *efx = netdev_priv(net_dev);
2065 spin_lock_bh(&efx->stats_lock);
2066 efx->type->update_stats(efx, NULL, stats);
2067 spin_unlock_bh(&efx->stats_lock);
2072 /* Context: netif_tx_lock held, BHs disabled. */
2073 static void efx_watchdog(struct net_device *net_dev)
2075 struct efx_nic *efx = netdev_priv(net_dev);
2077 netif_err(efx, tx_err, efx->net_dev,
2078 "TX stuck with port_enabled=%d: resetting channels\n",
2081 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2085 /* Context: process, rtnl_lock() held. */
2086 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2088 struct efx_nic *efx = netdev_priv(net_dev);
2091 rc = efx_check_disabled(efx);
2094 if (new_mtu > EFX_MAX_MTU)
2097 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2099 efx_device_detach_sync(efx);
2102 mutex_lock(&efx->mac_lock);
2103 net_dev->mtu = new_mtu;
2104 efx->type->reconfigure_mac(efx);
2105 mutex_unlock(&efx->mac_lock);
2108 netif_device_attach(efx->net_dev);
2112 static int efx_set_mac_address(struct net_device *net_dev, void *data)
2114 struct efx_nic *efx = netdev_priv(net_dev);
2115 struct sockaddr *addr = data;
2116 u8 *new_addr = addr->sa_data;
2118 if (!is_valid_ether_addr(new_addr)) {
2119 netif_err(efx, drv, efx->net_dev,
2120 "invalid ethernet MAC address requested: %pM\n",
2122 return -EADDRNOTAVAIL;
2125 memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
2126 efx_sriov_mac_address_changed(efx);
2128 /* Reconfigure the MAC */
2129 mutex_lock(&efx->mac_lock);
2130 efx->type->reconfigure_mac(efx);
2131 mutex_unlock(&efx->mac_lock);
2136 /* Context: netif_addr_lock held, BHs disabled. */
2137 static void efx_set_rx_mode(struct net_device *net_dev)
2139 struct efx_nic *efx = netdev_priv(net_dev);
2141 if (efx->port_enabled)
2142 queue_work(efx->workqueue, &efx->mac_work);
2143 /* Otherwise efx_start_port() will do this */
2146 static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2148 struct efx_nic *efx = netdev_priv(net_dev);
2150 /* If disabling RX n-tuple filtering, clear existing filters */
2151 if (net_dev->features & ~data & NETIF_F_NTUPLE)
2152 return efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2157 static const struct net_device_ops efx_farch_netdev_ops = {
2158 .ndo_open = efx_net_open,
2159 .ndo_stop = efx_net_stop,
2160 .ndo_get_stats64 = efx_net_stats,
2161 .ndo_tx_timeout = efx_watchdog,
2162 .ndo_start_xmit = efx_hard_start_xmit,
2163 .ndo_validate_addr = eth_validate_addr,
2164 .ndo_do_ioctl = efx_ioctl,
2165 .ndo_change_mtu = efx_change_mtu,
2166 .ndo_set_mac_address = efx_set_mac_address,
2167 .ndo_set_rx_mode = efx_set_rx_mode,
2168 .ndo_set_features = efx_set_features,
2169 #ifdef CONFIG_SFC_SRIOV
2170 .ndo_set_vf_mac = efx_sriov_set_vf_mac,
2171 .ndo_set_vf_vlan = efx_sriov_set_vf_vlan,
2172 .ndo_set_vf_spoofchk = efx_sriov_set_vf_spoofchk,
2173 .ndo_get_vf_config = efx_sriov_get_vf_config,
2175 #ifdef CONFIG_NET_POLL_CONTROLLER
2176 .ndo_poll_controller = efx_netpoll,
2178 .ndo_setup_tc = efx_setup_tc,
2179 #ifdef CONFIG_RFS_ACCEL
2180 .ndo_rx_flow_steer = efx_filter_rfs,
2184 static const struct net_device_ops efx_ef10_netdev_ops = {
2185 .ndo_open = efx_net_open,
2186 .ndo_stop = efx_net_stop,
2187 .ndo_get_stats64 = efx_net_stats,
2188 .ndo_tx_timeout = efx_watchdog,
2189 .ndo_start_xmit = efx_hard_start_xmit,
2190 .ndo_validate_addr = eth_validate_addr,
2191 .ndo_do_ioctl = efx_ioctl,
2192 .ndo_change_mtu = efx_change_mtu,
2193 .ndo_set_mac_address = efx_set_mac_address,
2194 .ndo_set_rx_mode = efx_set_rx_mode,
2195 .ndo_set_features = efx_set_features,
2196 #ifdef CONFIG_NET_POLL_CONTROLLER
2197 .ndo_poll_controller = efx_netpoll,
2199 #ifdef CONFIG_RFS_ACCEL
2200 .ndo_rx_flow_steer = efx_filter_rfs,
2204 static void efx_update_name(struct efx_nic *efx)
2206 strcpy(efx->name, efx->net_dev->name);
2207 efx_mtd_rename(efx);
2208 efx_set_channel_names(efx);
2211 static int efx_netdev_event(struct notifier_block *this,
2212 unsigned long event, void *ptr)
2214 struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2216 if ((net_dev->netdev_ops == &efx_farch_netdev_ops ||
2217 net_dev->netdev_ops == &efx_ef10_netdev_ops) &&
2218 event == NETDEV_CHANGENAME)
2219 efx_update_name(netdev_priv(net_dev));
2224 static struct notifier_block efx_netdev_notifier = {
2225 .notifier_call = efx_netdev_event,
2229 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2231 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2232 return sprintf(buf, "%d\n", efx->phy_type);
2234 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2236 static int efx_register_netdev(struct efx_nic *efx)
2238 struct net_device *net_dev = efx->net_dev;
2239 struct efx_channel *channel;
2242 net_dev->watchdog_timeo = 5 * HZ;
2243 net_dev->irq = efx->pci_dev->irq;
2244 if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0) {
2245 net_dev->netdev_ops = &efx_ef10_netdev_ops;
2246 net_dev->priv_flags |= IFF_UNICAST_FLT;
2248 net_dev->netdev_ops = &efx_farch_netdev_ops;
2250 SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2251 net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2255 /* Enable resets to be scheduled and check whether any were
2256 * already requested. If so, the NIC is probably hosed so we
2259 efx->state = STATE_READY;
2260 smp_mb(); /* ensure we change state before checking reset_pending */
2261 if (efx->reset_pending) {
2262 netif_err(efx, probe, efx->net_dev,
2263 "aborting probe due to scheduled reset\n");
2268 rc = dev_alloc_name(net_dev, net_dev->name);
2271 efx_update_name(efx);
2273 /* Always start with carrier off; PHY events will detect the link */
2274 netif_carrier_off(net_dev);
2276 rc = register_netdevice(net_dev);
2280 efx_for_each_channel(channel, efx) {
2281 struct efx_tx_queue *tx_queue;
2282 efx_for_each_channel_tx_queue(tx_queue, channel)
2283 efx_init_tx_queue_core_txq(tx_queue);
2290 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2292 netif_err(efx, drv, efx->net_dev,
2293 "failed to init net dev attributes\n");
2294 goto fail_registered;
2301 efx_dissociate(efx);
2302 unregister_netdevice(net_dev);
2304 efx->state = STATE_UNINIT;
2306 netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2310 static void efx_unregister_netdev(struct efx_nic *efx)
2315 BUG_ON(netdev_priv(efx->net_dev) != efx);
2317 strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2318 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2321 unregister_netdevice(efx->net_dev);
2322 efx->state = STATE_UNINIT;
2326 /**************************************************************************
2328 * Device reset and suspend
2330 **************************************************************************/
2332 /* Tears down the entire software state and most of the hardware state
2334 void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2336 EFX_ASSERT_RESET_SERIALISED(efx);
2339 efx_disable_interrupts(efx);
2341 mutex_lock(&efx->mac_lock);
2342 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2343 efx->phy_op->fini(efx);
2344 efx->type->fini(efx);
2347 /* This function will always ensure that the locks acquired in
2348 * efx_reset_down() are released. A failure return code indicates
2349 * that we were unable to reinitialise the hardware, and the
2350 * driver should be disabled. If ok is false, then the rx and tx
2351 * engines are not restarted, pending a RESET_DISABLE. */
2352 int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2356 EFX_ASSERT_RESET_SERIALISED(efx);
2358 rc = efx->type->init(efx);
2360 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2367 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2368 rc = efx->phy_op->init(efx);
2371 if (efx->phy_op->reconfigure(efx))
2372 netif_err(efx, drv, efx->net_dev,
2373 "could not restore PHY settings\n");
2376 rc = efx_enable_interrupts(efx);
2379 efx_restore_filters(efx);
2380 efx_sriov_reset(efx);
2382 mutex_unlock(&efx->mac_lock);
2389 efx->port_initialized = false;
2391 mutex_unlock(&efx->mac_lock);
2396 /* Reset the NIC using the specified method. Note that the reset may
2397 * fail, in which case the card will be left in an unusable state.
2399 * Caller must hold the rtnl_lock.
2401 int efx_reset(struct efx_nic *efx, enum reset_type method)
2406 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2407 RESET_TYPE(method));
2409 efx_device_detach_sync(efx);
2410 efx_reset_down(efx, method);
2412 rc = efx->type->reset(efx, method);
2414 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2418 /* Clear flags for the scopes we covered. We assume the NIC and
2419 * driver are now quiescent so that there is no race here.
2421 efx->reset_pending &= -(1 << (method + 1));
2423 /* Reinitialise bus-mastering, which may have been turned off before
2424 * the reset was scheduled. This is still appropriate, even in the
2425 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2426 * can respond to requests. */
2427 pci_set_master(efx->pci_dev);
2430 /* Leave device stopped if necessary */
2432 method == RESET_TYPE_DISABLE ||
2433 method == RESET_TYPE_RECOVER_OR_DISABLE;
2434 rc2 = efx_reset_up(efx, method, !disabled);
2442 dev_close(efx->net_dev);
2443 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2444 efx->state = STATE_DISABLED;
2446 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2447 netif_device_attach(efx->net_dev);
2452 /* Try recovery mechanisms.
2453 * For now only EEH is supported.
2454 * Returns 0 if the recovery mechanisms are unsuccessful.
2455 * Returns a non-zero value otherwise.
2457 int efx_try_recovery(struct efx_nic *efx)
2460 /* A PCI error can occur and not be seen by EEH because nothing
2461 * happens on the PCI bus. In this case the driver may fail and
2462 * schedule a 'recover or reset', leading to this recovery handler.
2463 * Manually call the eeh failure check function.
2465 struct eeh_dev *eehdev =
2466 of_node_to_eeh_dev(pci_device_to_OF_node(efx->pci_dev));
2468 if (eeh_dev_check_failure(eehdev)) {
2469 /* The EEH mechanisms will handle the error and reset the
2470 * device if necessary.
2478 static void efx_wait_for_bist_end(struct efx_nic *efx)
2482 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2483 if (efx_mcdi_poll_reboot(efx))
2485 msleep(BIST_WAIT_DELAY_MS);
2488 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2490 /* Either way unset the BIST flag. If we found no reboot we probably
2491 * won't recover, but we should try.
2493 efx->mc_bist_for_other_fn = false;
2496 /* The worker thread exists so that code that cannot sleep can
2497 * schedule a reset for later.
2499 static void efx_reset_work(struct work_struct *data)
2501 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2502 unsigned long pending;
2503 enum reset_type method;
2505 pending = ACCESS_ONCE(efx->reset_pending);
2506 method = fls(pending) - 1;
2508 if (method == RESET_TYPE_MC_BIST)
2509 efx_wait_for_bist_end(efx);
2511 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2512 method == RESET_TYPE_RECOVER_OR_ALL) &&
2513 efx_try_recovery(efx))
2521 /* We checked the state in efx_schedule_reset() but it may
2522 * have changed by now. Now that we have the RTNL lock,
2523 * it cannot change again.
2525 if (efx->state == STATE_READY)
2526 (void)efx_reset(efx, method);
2531 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2533 enum reset_type method;
2535 if (efx->state == STATE_RECOVERY) {
2536 netif_dbg(efx, drv, efx->net_dev,
2537 "recovering: skip scheduling %s reset\n",
2543 case RESET_TYPE_INVISIBLE:
2544 case RESET_TYPE_ALL:
2545 case RESET_TYPE_RECOVER_OR_ALL:
2546 case RESET_TYPE_WORLD:
2547 case RESET_TYPE_DISABLE:
2548 case RESET_TYPE_RECOVER_OR_DISABLE:
2549 case RESET_TYPE_MC_BIST:
2551 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2552 RESET_TYPE(method));
2555 method = efx->type->map_reset_reason(type);
2556 netif_dbg(efx, drv, efx->net_dev,
2557 "scheduling %s reset for %s\n",
2558 RESET_TYPE(method), RESET_TYPE(type));
2562 set_bit(method, &efx->reset_pending);
2563 smp_mb(); /* ensure we change reset_pending before checking state */
2565 /* If we're not READY then just leave the flags set as the cue
2566 * to abort probing or reschedule the reset later.
2568 if (ACCESS_ONCE(efx->state) != STATE_READY)
2571 /* efx_process_channel() will no longer read events once a
2572 * reset is scheduled. So switch back to poll'd MCDI completions. */
2573 efx_mcdi_mode_poll(efx);
2575 queue_work(reset_workqueue, &efx->reset_work);
2578 /**************************************************************************
2580 * List of NICs we support
2582 **************************************************************************/
2584 /* PCI device ID table */
2585 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2586 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2587 PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2588 .driver_data = (unsigned long) &falcon_a1_nic_type},
2589 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2590 PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2591 .driver_data = (unsigned long) &falcon_b0_nic_type},
2592 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803), /* SFC9020 */
2593 .driver_data = (unsigned long) &siena_a0_nic_type},
2594 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813), /* SFL9021 */
2595 .driver_data = (unsigned long) &siena_a0_nic_type},
2596 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903), /* SFC9120 PF */
2597 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2598 {0} /* end of list */
2601 /**************************************************************************
2603 * Dummy PHY/MAC operations
2605 * Can be used for some unimplemented operations
2606 * Needed so all function pointers are valid and do not have to be tested
2609 **************************************************************************/
2610 int efx_port_dummy_op_int(struct efx_nic *efx)
2614 void efx_port_dummy_op_void(struct efx_nic *efx) {}
2616 static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2621 static const struct efx_phy_operations efx_dummy_phy_operations = {
2622 .init = efx_port_dummy_op_int,
2623 .reconfigure = efx_port_dummy_op_int,
2624 .poll = efx_port_dummy_op_poll,
2625 .fini = efx_port_dummy_op_void,
2628 /**************************************************************************
2632 **************************************************************************/
2634 /* This zeroes out and then fills in the invariants in a struct
2635 * efx_nic (including all sub-structures).
2637 static int efx_init_struct(struct efx_nic *efx,
2638 struct pci_dev *pci_dev, struct net_device *net_dev)
2642 /* Initialise common structures */
2643 INIT_LIST_HEAD(&efx->node);
2644 INIT_LIST_HEAD(&efx->secondary_list);
2645 spin_lock_init(&efx->biu_lock);
2646 #ifdef CONFIG_SFC_MTD
2647 INIT_LIST_HEAD(&efx->mtd_list);
2649 INIT_WORK(&efx->reset_work, efx_reset_work);
2650 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2651 INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2652 efx->pci_dev = pci_dev;
2653 efx->msg_enable = debug;
2654 efx->state = STATE_UNINIT;
2655 strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2657 efx->net_dev = net_dev;
2658 efx->rx_prefix_size = efx->type->rx_prefix_size;
2660 NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2661 efx->rx_packet_hash_offset =
2662 efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2663 efx->rx_packet_ts_offset =
2664 efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2665 spin_lock_init(&efx->stats_lock);
2666 mutex_init(&efx->mac_lock);
2667 efx->phy_op = &efx_dummy_phy_operations;
2668 efx->mdio.dev = net_dev;
2669 INIT_WORK(&efx->mac_work, efx_mac_work);
2670 init_waitqueue_head(&efx->flush_wq);
2672 for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2673 efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2674 if (!efx->channel[i])
2676 efx->msi_context[i].efx = efx;
2677 efx->msi_context[i].index = i;
2680 /* Higher numbered interrupt modes are less capable! */
2681 efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2684 /* Would be good to use the net_dev name, but we're too early */
2685 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2687 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2688 if (!efx->workqueue)
2694 efx_fini_struct(efx);
2698 static void efx_fini_struct(struct efx_nic *efx)
2702 for (i = 0; i < EFX_MAX_CHANNELS; i++)
2703 kfree(efx->channel[i]);
2707 if (efx->workqueue) {
2708 destroy_workqueue(efx->workqueue);
2709 efx->workqueue = NULL;
2713 /**************************************************************************
2717 **************************************************************************/
2719 /* Main body of final NIC shutdown code
2720 * This is called only at module unload (or hotplug removal).
2722 static void efx_pci_remove_main(struct efx_nic *efx)
2724 /* Flush reset_work. It can no longer be scheduled since we
2727 BUG_ON(efx->state == STATE_READY);
2728 cancel_work_sync(&efx->reset_work);
2730 efx_disable_interrupts(efx);
2731 efx_nic_fini_interrupt(efx);
2733 efx->type->fini(efx);
2735 efx_remove_all(efx);
2738 /* Final NIC shutdown
2739 * This is called only at module unload (or hotplug removal).
2741 static void efx_pci_remove(struct pci_dev *pci_dev)
2743 struct efx_nic *efx;
2745 efx = pci_get_drvdata(pci_dev);
2749 /* Mark the NIC as fini, then stop the interface */
2751 efx_dissociate(efx);
2752 dev_close(efx->net_dev);
2753 efx_disable_interrupts(efx);
2756 efx_sriov_fini(efx);
2757 efx_unregister_netdev(efx);
2759 efx_mtd_remove(efx);
2761 efx_pci_remove_main(efx);
2764 netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2766 efx_fini_struct(efx);
2767 free_netdev(efx->net_dev);
2769 pci_disable_pcie_error_reporting(pci_dev);
2772 /* NIC VPD information
2773 * Called during probe to display the part number of the
2774 * installed NIC. VPD is potentially very large but this should
2775 * always appear within the first 512 bytes.
2777 #define SFC_VPD_LEN 512
2778 static void efx_probe_vpd_strings(struct efx_nic *efx)
2780 struct pci_dev *dev = efx->pci_dev;
2781 char vpd_data[SFC_VPD_LEN];
2783 int ro_start, ro_size, i, j;
2785 /* Get the vpd data from the device */
2786 vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2787 if (vpd_size <= 0) {
2788 netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2792 /* Get the Read only section */
2793 ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2795 netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2799 ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
2801 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2802 if (i + j > vpd_size)
2805 /* Get the Part number */
2806 i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2808 netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2812 j = pci_vpd_info_field_size(&vpd_data[i]);
2813 i += PCI_VPD_INFO_FLD_HDR_SIZE;
2814 if (i + j > vpd_size) {
2815 netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2819 netif_info(efx, drv, efx->net_dev,
2820 "Part Number : %.*s\n", j, &vpd_data[i]);
2822 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2824 i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
2826 netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
2830 j = pci_vpd_info_field_size(&vpd_data[i]);
2831 i += PCI_VPD_INFO_FLD_HDR_SIZE;
2832 if (i + j > vpd_size) {
2833 netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
2837 efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
2841 snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
2845 /* Main body of NIC initialisation
2846 * This is called at module load (or hotplug insertion, theoretically).
2848 static int efx_pci_probe_main(struct efx_nic *efx)
2852 /* Do start-of-day initialisation */
2853 rc = efx_probe_all(efx);
2859 rc = efx->type->init(efx);
2861 netif_err(efx, probe, efx->net_dev,
2862 "failed to initialise NIC\n");
2866 rc = efx_init_port(efx);
2868 netif_err(efx, probe, efx->net_dev,
2869 "failed to initialise port\n");
2873 rc = efx_nic_init_interrupt(efx);
2876 rc = efx_enable_interrupts(efx);
2883 efx_nic_fini_interrupt(efx);
2887 efx->type->fini(efx);
2890 efx_remove_all(efx);
2895 /* NIC initialisation
2897 * This is called at module load (or hotplug insertion,
2898 * theoretically). It sets up PCI mappings, resets the NIC,
2899 * sets up and registers the network devices with the kernel and hooks
2900 * the interrupt service routine. It does not prepare the device for
2901 * transmission; this is left to the first time one of the network
2902 * interfaces is brought up (i.e. efx_net_open).
2904 static int efx_pci_probe(struct pci_dev *pci_dev,
2905 const struct pci_device_id *entry)
2907 struct net_device *net_dev;
2908 struct efx_nic *efx;
2911 /* Allocate and initialise a struct net_device and struct efx_nic */
2912 net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2916 efx = netdev_priv(net_dev);
2917 efx->type = (const struct efx_nic_type *) entry->driver_data;
2918 net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2919 NETIF_F_HIGHDMA | NETIF_F_TSO |
2921 if (efx->type->offload_features & NETIF_F_V6_CSUM)
2922 net_dev->features |= NETIF_F_TSO6;
2923 /* Mask for features that also apply to VLAN devices */
2924 net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2925 NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2927 /* All offloads can be toggled */
2928 net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2929 pci_set_drvdata(pci_dev, efx);
2930 SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2931 rc = efx_init_struct(efx, pci_dev, net_dev);
2935 netif_info(efx, probe, efx->net_dev,
2936 "Solarflare NIC detected\n");
2938 efx_probe_vpd_strings(efx);
2940 /* Set up basic I/O (BAR mappings etc) */
2941 rc = efx_init_io(efx);
2945 rc = efx_pci_probe_main(efx);
2949 rc = efx_register_netdev(efx);
2953 rc = efx_sriov_init(efx);
2955 netif_err(efx, probe, efx->net_dev,
2956 "SR-IOV can't be enabled rc %d\n", rc);
2958 netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2960 /* Try to create MTDs, but allow this to fail */
2962 rc = efx_mtd_probe(efx);
2965 netif_warn(efx, probe, efx->net_dev,
2966 "failed to create MTDs (%d)\n", rc);
2968 rc = pci_enable_pcie_error_reporting(pci_dev);
2969 if (rc && rc != -EINVAL)
2970 netif_warn(efx, probe, efx->net_dev,
2971 "pci_enable_pcie_error_reporting failed (%d)\n", rc);
2976 efx_pci_remove_main(efx);
2980 efx_fini_struct(efx);
2983 netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2984 free_netdev(net_dev);
2988 static int efx_pm_freeze(struct device *dev)
2990 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2994 if (efx->state != STATE_DISABLED) {
2995 efx->state = STATE_UNINIT;
2997 efx_device_detach_sync(efx);
3000 efx_disable_interrupts(efx);
3008 static int efx_pm_thaw(struct device *dev)
3011 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3015 if (efx->state != STATE_DISABLED) {
3016 rc = efx_enable_interrupts(efx);
3020 mutex_lock(&efx->mac_lock);
3021 efx->phy_op->reconfigure(efx);
3022 mutex_unlock(&efx->mac_lock);
3026 netif_device_attach(efx->net_dev);
3028 efx->state = STATE_READY;
3030 efx->type->resume_wol(efx);
3035 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3036 queue_work(reset_workqueue, &efx->reset_work);
3046 static int efx_pm_poweroff(struct device *dev)
3048 struct pci_dev *pci_dev = to_pci_dev(dev);
3049 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3051 efx->type->fini(efx);
3053 efx->reset_pending = 0;
3055 pci_save_state(pci_dev);
3056 return pci_set_power_state(pci_dev, PCI_D3hot);
3059 /* Used for both resume and restore */
3060 static int efx_pm_resume(struct device *dev)
3062 struct pci_dev *pci_dev = to_pci_dev(dev);
3063 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3066 rc = pci_set_power_state(pci_dev, PCI_D0);
3069 pci_restore_state(pci_dev);
3070 rc = pci_enable_device(pci_dev);
3073 pci_set_master(efx->pci_dev);
3074 rc = efx->type->reset(efx, RESET_TYPE_ALL);
3077 rc = efx->type->init(efx);
3080 rc = efx_pm_thaw(dev);
3084 static int efx_pm_suspend(struct device *dev)
3089 rc = efx_pm_poweroff(dev);
3095 static const struct dev_pm_ops efx_pm_ops = {
3096 .suspend = efx_pm_suspend,
3097 .resume = efx_pm_resume,
3098 .freeze = efx_pm_freeze,
3099 .thaw = efx_pm_thaw,
3100 .poweroff = efx_pm_poweroff,
3101 .restore = efx_pm_resume,
3104 /* A PCI error affecting this device was detected.
3105 * At this point MMIO and DMA may be disabled.
3106 * Stop the software path and request a slot reset.
3108 static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3109 enum pci_channel_state state)
3111 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3112 struct efx_nic *efx = pci_get_drvdata(pdev);
3114 if (state == pci_channel_io_perm_failure)
3115 return PCI_ERS_RESULT_DISCONNECT;
3119 if (efx->state != STATE_DISABLED) {
3120 efx->state = STATE_RECOVERY;
3121 efx->reset_pending = 0;
3123 efx_device_detach_sync(efx);
3126 efx_disable_interrupts(efx);
3128 status = PCI_ERS_RESULT_NEED_RESET;
3130 /* If the interface is disabled we don't want to do anything
3133 status = PCI_ERS_RESULT_RECOVERED;
3138 pci_disable_device(pdev);
3143 /* Fake a successfull reset, which will be performed later in efx_io_resume. */
3144 static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3146 struct efx_nic *efx = pci_get_drvdata(pdev);
3147 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3150 if (pci_enable_device(pdev)) {
3151 netif_err(efx, hw, efx->net_dev,
3152 "Cannot re-enable PCI device after reset.\n");
3153 status = PCI_ERS_RESULT_DISCONNECT;
3156 rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3158 netif_err(efx, hw, efx->net_dev,
3159 "pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3160 /* Non-fatal error. Continue. */
3166 /* Perform the actual reset and resume I/O operations. */
3167 static void efx_io_resume(struct pci_dev *pdev)
3169 struct efx_nic *efx = pci_get_drvdata(pdev);
3174 if (efx->state == STATE_DISABLED)
3177 rc = efx_reset(efx, RESET_TYPE_ALL);
3179 netif_err(efx, hw, efx->net_dev,
3180 "efx_reset failed after PCI error (%d)\n", rc);
3182 efx->state = STATE_READY;
3183 netif_dbg(efx, hw, efx->net_dev,
3184 "Done resetting and resuming IO after PCI error.\n");
3191 /* For simplicity and reliability, we always require a slot reset and try to
3192 * reset the hardware when a pci error affecting the device is detected.
3193 * We leave both the link_reset and mmio_enabled callback unimplemented:
3194 * with our request for slot reset the mmio_enabled callback will never be
3195 * called, and the link_reset callback is not used by AER or EEH mechanisms.
3197 static struct pci_error_handlers efx_err_handlers = {
3198 .error_detected = efx_io_error_detected,
3199 .slot_reset = efx_io_slot_reset,
3200 .resume = efx_io_resume,
3203 static struct pci_driver efx_pci_driver = {
3204 .name = KBUILD_MODNAME,
3205 .id_table = efx_pci_table,
3206 .probe = efx_pci_probe,
3207 .remove = efx_pci_remove,
3208 .driver.pm = &efx_pm_ops,
3209 .err_handler = &efx_err_handlers,
3212 /**************************************************************************
3214 * Kernel module interface
3216 *************************************************************************/
3218 module_param(interrupt_mode, uint, 0444);
3219 MODULE_PARM_DESC(interrupt_mode,
3220 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3222 static int __init efx_init_module(void)
3226 printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3228 rc = register_netdevice_notifier(&efx_netdev_notifier);
3232 rc = efx_init_sriov();
3236 reset_workqueue = create_singlethread_workqueue("sfc_reset");
3237 if (!reset_workqueue) {
3242 rc = pci_register_driver(&efx_pci_driver);
3249 destroy_workqueue(reset_workqueue);
3253 unregister_netdevice_notifier(&efx_netdev_notifier);
3258 static void __exit efx_exit_module(void)
3260 printk(KERN_INFO "Solarflare NET driver unloading\n");
3262 pci_unregister_driver(&efx_pci_driver);
3263 destroy_workqueue(reset_workqueue);
3265 unregister_netdevice_notifier(&efx_netdev_notifier);
3269 module_init(efx_init_module);
3270 module_exit(efx_exit_module);
3272 MODULE_AUTHOR("Solarflare Communications and "
3273 "Michael Brown <mbrown@fensystems.co.uk>");
3274 MODULE_DESCRIPTION("Solarflare Communications network driver");
3275 MODULE_LICENSE("GPL");
3276 MODULE_DEVICE_TABLE(pci, efx_pci_table);