1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2008 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/crc32.h>
21 #include <linux/ethtool.h>
22 #include <linux/topology.h>
23 #include "net_driver.h"
31 #define EFX_MAX_MTU (9 * 1024)
33 /* RX slow fill workqueue. If memory allocation fails in the fast path,
34 * a work item is pushed onto this work queue to retry the allocation later,
35 * to avoid the NIC being starved of RX buffers. Since this is a per cpu
36 * workqueue, there is nothing to be gained in making it per NIC
38 static struct workqueue_struct *refill_workqueue;
40 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
41 * queued onto this work queue. This is not a per-nic work queue, because
42 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
44 static struct workqueue_struct *reset_workqueue;
46 /**************************************************************************
50 *************************************************************************/
53 * Enable large receive offload (LRO) aka soft segment reassembly (SSR)
55 * This sets the default for new devices. It can be controlled later
58 static int lro = true;
59 module_param(lro, int, 0644);
60 MODULE_PARM_DESC(lro, "Large receive offload acceleration");
63 * Use separate channels for TX and RX events
65 * Set this to 1 to use separate channels for TX and RX. It allows us
66 * to control interrupt affinity separately for TX and RX.
68 * This is only used in MSI-X interrupt mode
70 static unsigned int separate_tx_channels;
71 module_param(separate_tx_channels, uint, 0644);
72 MODULE_PARM_DESC(separate_tx_channels,
73 "Use separate channels for TX and RX");
75 /* This is the weight assigned to each of the (per-channel) virtual
78 static int napi_weight = 64;
80 /* This is the time (in jiffies) between invocations of the hardware
81 * monitor, which checks for known hardware bugs and resets the
82 * hardware and driver as necessary.
84 unsigned int efx_monitor_interval = 1 * HZ;
86 /* This controls whether or not the driver will initialise devices
87 * with invalid MAC addresses stored in the EEPROM or flash. If true,
88 * such devices will be initialised with a random locally-generated
89 * MAC address. This allows for loading the sfc_mtd driver to
90 * reprogram the flash, even if the flash contents (including the MAC
91 * address) have previously been erased.
93 static unsigned int allow_bad_hwaddr;
95 /* Initial interrupt moderation settings. They can be modified after
96 * module load with ethtool.
98 * The default for RX should strike a balance between increasing the
99 * round-trip latency and reducing overhead.
101 static unsigned int rx_irq_mod_usec = 60;
103 /* Initial interrupt moderation settings. They can be modified after
104 * module load with ethtool.
106 * This default is chosen to ensure that a 10G link does not go idle
107 * while a TX queue is stopped after it has become full. A queue is
108 * restarted when it drops below half full. The time this takes (assuming
109 * worst case 3 descriptors per packet and 1024 descriptors) is
110 * 512 / 3 * 1.2 = 205 usec.
112 static unsigned int tx_irq_mod_usec = 150;
114 /* This is the first interrupt mode to try out of:
119 static unsigned int interrupt_mode;
121 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
122 * i.e. the number of CPUs among which we may distribute simultaneous
123 * interrupt handling.
125 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
126 * The default (0) means to assign an interrupt to each package (level II cache)
128 static unsigned int rss_cpus;
129 module_param(rss_cpus, uint, 0444);
130 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
132 static int phy_flash_cfg;
133 module_param(phy_flash_cfg, int, 0644);
134 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
136 static unsigned irq_adapt_low_thresh = 10000;
137 module_param(irq_adapt_low_thresh, uint, 0644);
138 MODULE_PARM_DESC(irq_adapt_low_thresh,
139 "Threshold score for reducing IRQ moderation");
141 static unsigned irq_adapt_high_thresh = 20000;
142 module_param(irq_adapt_high_thresh, uint, 0644);
143 MODULE_PARM_DESC(irq_adapt_high_thresh,
144 "Threshold score for increasing IRQ moderation");
146 /**************************************************************************
148 * Utility functions and prototypes
150 *************************************************************************/
151 static void efx_remove_channel(struct efx_channel *channel);
152 static void efx_remove_port(struct efx_nic *efx);
153 static void efx_fini_napi(struct efx_nic *efx);
154 static void efx_fini_channels(struct efx_nic *efx);
156 #define EFX_ASSERT_RESET_SERIALISED(efx) \
158 if (efx->state == STATE_RUNNING) \
162 /**************************************************************************
164 * Event queue processing
166 *************************************************************************/
168 /* Process channel's event queue
170 * This function is responsible for processing the event queue of a
171 * single channel. The caller must guarantee that this function will
172 * never be concurrently called more than once on the same channel,
173 * though different channels may be being processed concurrently.
175 static int efx_process_channel(struct efx_channel *channel, int rx_quota)
177 struct efx_nic *efx = channel->efx;
180 if (unlikely(efx->reset_pending != RESET_TYPE_NONE ||
184 rx_packets = falcon_process_eventq(channel, rx_quota);
188 /* Deliver last RX packet. */
189 if (channel->rx_pkt) {
190 __efx_rx_packet(channel, channel->rx_pkt,
191 channel->rx_pkt_csummed);
192 channel->rx_pkt = NULL;
195 efx_rx_strategy(channel);
197 efx_fast_push_rx_descriptors(&efx->rx_queue[channel->channel]);
202 /* Mark channel as finished processing
204 * Note that since we will not receive further interrupts for this
205 * channel before we finish processing and call the eventq_read_ack()
206 * method, there is no need to use the interrupt hold-off timers.
208 static inline void efx_channel_processed(struct efx_channel *channel)
210 /* The interrupt handler for this channel may set work_pending
211 * as soon as we acknowledge the events we've seen. Make sure
212 * it's cleared before then. */
213 channel->work_pending = false;
216 falcon_eventq_read_ack(channel);
221 * NAPI guarantees serialisation of polls of the same device, which
222 * provides the guarantee required by efx_process_channel().
224 static int efx_poll(struct napi_struct *napi, int budget)
226 struct efx_channel *channel =
227 container_of(napi, struct efx_channel, napi_str);
230 EFX_TRACE(channel->efx, "channel %d NAPI poll executing on CPU %d\n",
231 channel->channel, raw_smp_processor_id());
233 rx_packets = efx_process_channel(channel, budget);
235 if (rx_packets < budget) {
236 struct efx_nic *efx = channel->efx;
238 if (channel->used_flags & EFX_USED_BY_RX &&
239 efx->irq_rx_adaptive &&
240 unlikely(++channel->irq_count == 1000)) {
241 unsigned old_irq_moderation = channel->irq_moderation;
243 if (unlikely(channel->irq_mod_score <
244 irq_adapt_low_thresh)) {
245 channel->irq_moderation =
247 channel->irq_moderation -
248 FALCON_IRQ_MOD_RESOLUTION,
249 FALCON_IRQ_MOD_RESOLUTION);
250 } else if (unlikely(channel->irq_mod_score >
251 irq_adapt_high_thresh)) {
252 channel->irq_moderation =
253 min(channel->irq_moderation +
254 FALCON_IRQ_MOD_RESOLUTION,
255 efx->irq_rx_moderation);
258 if (channel->irq_moderation != old_irq_moderation)
259 falcon_set_int_moderation(channel);
261 channel->irq_count = 0;
262 channel->irq_mod_score = 0;
265 /* There is no race here; although napi_disable() will
266 * only wait for napi_complete(), this isn't a problem
267 * since efx_channel_processed() will have no effect if
268 * interrupts have already been disabled.
271 efx_channel_processed(channel);
277 /* Process the eventq of the specified channel immediately on this CPU
279 * Disable hardware generated interrupts, wait for any existing
280 * processing to finish, then directly poll (and ack ) the eventq.
281 * Finally reenable NAPI and interrupts.
283 * Since we are touching interrupts the caller should hold the suspend lock
285 void efx_process_channel_now(struct efx_channel *channel)
287 struct efx_nic *efx = channel->efx;
289 BUG_ON(!channel->used_flags);
290 BUG_ON(!channel->enabled);
292 /* Disable interrupts and wait for ISRs to complete */
293 falcon_disable_interrupts(efx);
295 synchronize_irq(efx->legacy_irq);
297 synchronize_irq(channel->irq);
299 /* Wait for any NAPI processing to complete */
300 napi_disable(&channel->napi_str);
302 /* Poll the channel */
303 efx_process_channel(channel, efx->type->evq_size);
305 /* Ack the eventq. This may cause an interrupt to be generated
306 * when they are reenabled */
307 efx_channel_processed(channel);
309 napi_enable(&channel->napi_str);
310 falcon_enable_interrupts(efx);
313 /* Create event queue
314 * Event queue memory allocations are done only once. If the channel
315 * is reset, the memory buffer will be reused; this guards against
316 * errors during channel reset and also simplifies interrupt handling.
318 static int efx_probe_eventq(struct efx_channel *channel)
320 EFX_LOG(channel->efx, "chan %d create event queue\n", channel->channel);
322 return falcon_probe_eventq(channel);
325 /* Prepare channel's event queue */
326 static void efx_init_eventq(struct efx_channel *channel)
328 EFX_LOG(channel->efx, "chan %d init event queue\n", channel->channel);
330 channel->eventq_read_ptr = 0;
332 falcon_init_eventq(channel);
335 static void efx_fini_eventq(struct efx_channel *channel)
337 EFX_LOG(channel->efx, "chan %d fini event queue\n", channel->channel);
339 falcon_fini_eventq(channel);
342 static void efx_remove_eventq(struct efx_channel *channel)
344 EFX_LOG(channel->efx, "chan %d remove event queue\n", channel->channel);
346 falcon_remove_eventq(channel);
349 /**************************************************************************
353 *************************************************************************/
355 static int efx_probe_channel(struct efx_channel *channel)
357 struct efx_tx_queue *tx_queue;
358 struct efx_rx_queue *rx_queue;
361 EFX_LOG(channel->efx, "creating channel %d\n", channel->channel);
363 rc = efx_probe_eventq(channel);
367 efx_for_each_channel_tx_queue(tx_queue, channel) {
368 rc = efx_probe_tx_queue(tx_queue);
373 efx_for_each_channel_rx_queue(rx_queue, channel) {
374 rc = efx_probe_rx_queue(rx_queue);
379 channel->n_rx_frm_trunc = 0;
384 efx_for_each_channel_rx_queue(rx_queue, channel)
385 efx_remove_rx_queue(rx_queue);
387 efx_for_each_channel_tx_queue(tx_queue, channel)
388 efx_remove_tx_queue(tx_queue);
394 static void efx_set_channel_names(struct efx_nic *efx)
396 struct efx_channel *channel;
397 const char *type = "";
400 efx_for_each_channel(channel, efx) {
401 number = channel->channel;
402 if (efx->n_channels > efx->n_rx_queues) {
403 if (channel->channel < efx->n_rx_queues) {
407 number -= efx->n_rx_queues;
410 snprintf(channel->name, sizeof(channel->name),
411 "%s%s-%d", efx->name, type, number);
415 /* Channels are shutdown and reinitialised whilst the NIC is running
416 * to propagate configuration changes (mtu, checksum offload), or
417 * to clear hardware error conditions
419 static void efx_init_channels(struct efx_nic *efx)
421 struct efx_tx_queue *tx_queue;
422 struct efx_rx_queue *rx_queue;
423 struct efx_channel *channel;
425 /* Calculate the rx buffer allocation parameters required to
426 * support the current MTU, including padding for header
427 * alignment and overruns.
429 efx->rx_buffer_len = (max(EFX_PAGE_IP_ALIGN, NET_IP_ALIGN) +
430 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
431 efx->type->rx_buffer_padding);
432 efx->rx_buffer_order = get_order(efx->rx_buffer_len);
434 /* Initialise the channels */
435 efx_for_each_channel(channel, efx) {
436 EFX_LOG(channel->efx, "init chan %d\n", channel->channel);
438 efx_init_eventq(channel);
440 efx_for_each_channel_tx_queue(tx_queue, channel)
441 efx_init_tx_queue(tx_queue);
443 /* The rx buffer allocation strategy is MTU dependent */
444 efx_rx_strategy(channel);
446 efx_for_each_channel_rx_queue(rx_queue, channel)
447 efx_init_rx_queue(rx_queue);
449 WARN_ON(channel->rx_pkt != NULL);
450 efx_rx_strategy(channel);
452 netif_napi_add(channel->napi_dev, &channel->napi_str,
453 efx_poll, napi_weight);
457 /* This enables event queue processing and packet transmission.
459 * Note that this function is not allowed to fail, since that would
460 * introduce too much complexity into the suspend/resume path.
462 static void efx_start_channel(struct efx_channel *channel)
464 struct efx_rx_queue *rx_queue;
466 EFX_LOG(channel->efx, "starting chan %d\n", channel->channel);
468 /* The interrupt handler for this channel may set work_pending
469 * as soon as we enable it. Make sure it's cleared before
470 * then. Similarly, make sure it sees the enabled flag set. */
471 channel->work_pending = false;
472 channel->enabled = true;
475 napi_enable(&channel->napi_str);
477 /* Load up RX descriptors */
478 efx_for_each_channel_rx_queue(rx_queue, channel)
479 efx_fast_push_rx_descriptors(rx_queue);
482 /* This disables event queue processing and packet transmission.
483 * This function does not guarantee that all queue processing
484 * (e.g. RX refill) is complete.
486 static void efx_stop_channel(struct efx_channel *channel)
488 struct efx_rx_queue *rx_queue;
490 if (!channel->enabled)
493 EFX_LOG(channel->efx, "stop chan %d\n", channel->channel);
495 channel->enabled = false;
496 napi_disable(&channel->napi_str);
498 /* Ensure that any worker threads have exited or will be no-ops */
499 efx_for_each_channel_rx_queue(rx_queue, channel) {
500 spin_lock_bh(&rx_queue->add_lock);
501 spin_unlock_bh(&rx_queue->add_lock);
505 static void efx_fini_channels(struct efx_nic *efx)
507 struct efx_channel *channel;
508 struct efx_tx_queue *tx_queue;
509 struct efx_rx_queue *rx_queue;
512 EFX_ASSERT_RESET_SERIALISED(efx);
513 BUG_ON(efx->port_enabled);
515 rc = falcon_flush_queues(efx);
517 EFX_ERR(efx, "failed to flush queues\n");
519 EFX_LOG(efx, "successfully flushed all queues\n");
521 efx_for_each_channel(channel, efx) {
522 EFX_LOG(channel->efx, "shut down chan %d\n", channel->channel);
524 efx_for_each_channel_rx_queue(rx_queue, channel)
525 efx_fini_rx_queue(rx_queue);
526 efx_for_each_channel_tx_queue(tx_queue, channel)
527 efx_fini_tx_queue(tx_queue);
528 efx_fini_eventq(channel);
532 static void efx_remove_channel(struct efx_channel *channel)
534 struct efx_tx_queue *tx_queue;
535 struct efx_rx_queue *rx_queue;
537 EFX_LOG(channel->efx, "destroy chan %d\n", channel->channel);
539 efx_for_each_channel_rx_queue(rx_queue, channel)
540 efx_remove_rx_queue(rx_queue);
541 efx_for_each_channel_tx_queue(tx_queue, channel)
542 efx_remove_tx_queue(tx_queue);
543 efx_remove_eventq(channel);
545 channel->used_flags = 0;
548 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue, int delay)
550 queue_delayed_work(refill_workqueue, &rx_queue->work, delay);
553 /**************************************************************************
557 **************************************************************************/
559 /* This ensures that the kernel is kept informed (via
560 * netif_carrier_on/off) of the link status, and also maintains the
561 * link status's stop on the port's TX queue.
563 static void efx_link_status_changed(struct efx_nic *efx)
565 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
566 * that no events are triggered between unregister_netdev() and the
567 * driver unloading. A more general condition is that NETDEV_CHANGE
568 * can only be generated between NETDEV_UP and NETDEV_DOWN */
569 if (!netif_running(efx->net_dev))
572 if (efx->port_inhibited) {
573 netif_carrier_off(efx->net_dev);
577 if (efx->link_up != netif_carrier_ok(efx->net_dev)) {
578 efx->n_link_state_changes++;
581 netif_carrier_on(efx->net_dev);
583 netif_carrier_off(efx->net_dev);
586 /* Status message for kernel log */
588 EFX_INFO(efx, "link up at %uMbps %s-duplex (MTU %d)%s\n",
589 efx->link_speed, efx->link_fd ? "full" : "half",
591 (efx->promiscuous ? " [PROMISC]" : ""));
593 EFX_INFO(efx, "link down\n");
598 static void efx_fini_port(struct efx_nic *efx);
600 /* This call reinitialises the MAC to pick up new PHY settings. The
601 * caller must hold the mac_lock */
602 void __efx_reconfigure_port(struct efx_nic *efx)
604 WARN_ON(!mutex_is_locked(&efx->mac_lock));
606 EFX_LOG(efx, "reconfiguring MAC from PHY settings on CPU %d\n",
607 raw_smp_processor_id());
609 /* Serialise the promiscuous flag with efx_set_multicast_list. */
610 if (efx_dev_registered(efx)) {
611 netif_addr_lock_bh(efx->net_dev);
612 netif_addr_unlock_bh(efx->net_dev);
615 falcon_deconfigure_mac_wrapper(efx);
617 /* Reconfigure the PHY, disabling transmit in mac level loopback. */
618 if (LOOPBACK_INTERNAL(efx))
619 efx->phy_mode |= PHY_MODE_TX_DISABLED;
621 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
622 efx->phy_op->reconfigure(efx);
624 if (falcon_switch_mac(efx))
627 efx->mac_op->reconfigure(efx);
629 /* Inform kernel of loss/gain of carrier */
630 efx_link_status_changed(efx);
634 EFX_ERR(efx, "failed to reconfigure MAC\n");
635 efx->port_enabled = false;
639 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
641 void efx_reconfigure_port(struct efx_nic *efx)
643 EFX_ASSERT_RESET_SERIALISED(efx);
645 mutex_lock(&efx->mac_lock);
646 __efx_reconfigure_port(efx);
647 mutex_unlock(&efx->mac_lock);
650 /* Asynchronous efx_reconfigure_port work item. To speed up efx_flush_all()
651 * we don't efx_reconfigure_port() if the port is disabled. Care is taken
652 * in efx_stop_all() and efx_start_port() to prevent PHY events being lost */
653 static void efx_phy_work(struct work_struct *data)
655 struct efx_nic *efx = container_of(data, struct efx_nic, phy_work);
657 mutex_lock(&efx->mac_lock);
658 if (efx->port_enabled)
659 __efx_reconfigure_port(efx);
660 mutex_unlock(&efx->mac_lock);
663 static void efx_mac_work(struct work_struct *data)
665 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
667 mutex_lock(&efx->mac_lock);
668 if (efx->port_enabled)
669 efx->mac_op->irq(efx);
670 mutex_unlock(&efx->mac_lock);
673 static int efx_probe_port(struct efx_nic *efx)
677 EFX_LOG(efx, "create port\n");
679 /* Connect up MAC/PHY operations table and read MAC address */
680 rc = falcon_probe_port(efx);
685 efx->phy_mode = PHY_MODE_SPECIAL;
687 /* Sanity check MAC address */
688 if (is_valid_ether_addr(efx->mac_address)) {
689 memcpy(efx->net_dev->dev_addr, efx->mac_address, ETH_ALEN);
691 EFX_ERR(efx, "invalid MAC address %pM\n",
693 if (!allow_bad_hwaddr) {
697 random_ether_addr(efx->net_dev->dev_addr);
698 EFX_INFO(efx, "using locally-generated MAC %pM\n",
699 efx->net_dev->dev_addr);
705 efx_remove_port(efx);
709 static int efx_init_port(struct efx_nic *efx)
713 EFX_LOG(efx, "init port\n");
715 rc = efx->phy_op->init(efx);
718 mutex_lock(&efx->mac_lock);
719 efx->phy_op->reconfigure(efx);
720 rc = falcon_switch_mac(efx);
721 mutex_unlock(&efx->mac_lock);
724 efx->mac_op->reconfigure(efx);
726 efx->port_initialized = true;
727 efx_stats_enable(efx);
731 efx->phy_op->fini(efx);
735 /* Allow efx_reconfigure_port() to be scheduled, and close the window
736 * between efx_stop_port and efx_flush_all whereby a previously scheduled
737 * efx_phy_work()/efx_mac_work() may have been cancelled */
738 static void efx_start_port(struct efx_nic *efx)
740 EFX_LOG(efx, "start port\n");
741 BUG_ON(efx->port_enabled);
743 mutex_lock(&efx->mac_lock);
744 efx->port_enabled = true;
745 __efx_reconfigure_port(efx);
746 efx->mac_op->irq(efx);
747 mutex_unlock(&efx->mac_lock);
750 /* Prevent efx_phy_work, efx_mac_work, and efx_monitor() from executing,
751 * and efx_set_multicast_list() from scheduling efx_phy_work. efx_phy_work
752 * and efx_mac_work may still be scheduled via NAPI processing until
753 * efx_flush_all() is called */
754 static void efx_stop_port(struct efx_nic *efx)
756 EFX_LOG(efx, "stop port\n");
758 mutex_lock(&efx->mac_lock);
759 efx->port_enabled = false;
760 mutex_unlock(&efx->mac_lock);
762 /* Serialise against efx_set_multicast_list() */
763 if (efx_dev_registered(efx)) {
764 netif_addr_lock_bh(efx->net_dev);
765 netif_addr_unlock_bh(efx->net_dev);
769 static void efx_fini_port(struct efx_nic *efx)
771 EFX_LOG(efx, "shut down port\n");
773 if (!efx->port_initialized)
776 efx_stats_disable(efx);
777 efx->phy_op->fini(efx);
778 efx->port_initialized = false;
780 efx->link_up = false;
781 efx_link_status_changed(efx);
784 static void efx_remove_port(struct efx_nic *efx)
786 EFX_LOG(efx, "destroying port\n");
788 falcon_remove_port(efx);
791 /**************************************************************************
795 **************************************************************************/
797 /* This configures the PCI device to enable I/O and DMA. */
798 static int efx_init_io(struct efx_nic *efx)
800 struct pci_dev *pci_dev = efx->pci_dev;
801 dma_addr_t dma_mask = efx->type->max_dma_mask;
804 EFX_LOG(efx, "initialising I/O\n");
806 rc = pci_enable_device(pci_dev);
808 EFX_ERR(efx, "failed to enable PCI device\n");
812 pci_set_master(pci_dev);
814 /* Set the PCI DMA mask. Try all possibilities from our
815 * genuine mask down to 32 bits, because some architectures
816 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
817 * masks event though they reject 46 bit masks.
819 while (dma_mask > 0x7fffffffUL) {
820 if (pci_dma_supported(pci_dev, dma_mask) &&
821 ((rc = pci_set_dma_mask(pci_dev, dma_mask)) == 0))
826 EFX_ERR(efx, "could not find a suitable DMA mask\n");
829 EFX_LOG(efx, "using DMA mask %llx\n", (unsigned long long) dma_mask);
830 rc = pci_set_consistent_dma_mask(pci_dev, dma_mask);
832 /* pci_set_consistent_dma_mask() is not *allowed* to
833 * fail with a mask that pci_set_dma_mask() accepted,
834 * but just in case...
836 EFX_ERR(efx, "failed to set consistent DMA mask\n");
840 efx->membase_phys = pci_resource_start(efx->pci_dev,
842 rc = pci_request_region(pci_dev, efx->type->mem_bar, "sfc");
844 EFX_ERR(efx, "request for memory BAR failed\n");
848 efx->membase = ioremap_nocache(efx->membase_phys,
849 efx->type->mem_map_size);
851 EFX_ERR(efx, "could not map memory BAR %d at %llx+%x\n",
853 (unsigned long long)efx->membase_phys,
854 efx->type->mem_map_size);
858 EFX_LOG(efx, "memory BAR %u at %llx+%x (virtual %p)\n",
859 efx->type->mem_bar, (unsigned long long)efx->membase_phys,
860 efx->type->mem_map_size, efx->membase);
865 pci_release_region(efx->pci_dev, efx->type->mem_bar);
867 efx->membase_phys = 0;
869 pci_disable_device(efx->pci_dev);
874 static void efx_fini_io(struct efx_nic *efx)
876 EFX_LOG(efx, "shutting down I/O\n");
879 iounmap(efx->membase);
883 if (efx->membase_phys) {
884 pci_release_region(efx->pci_dev, efx->type->mem_bar);
885 efx->membase_phys = 0;
888 pci_disable_device(efx->pci_dev);
891 /* Get number of RX queues wanted. Return number of online CPU
892 * packages in the expectation that an IRQ balancer will spread
893 * interrupts across them. */
894 static int efx_wanted_rx_queues(void)
896 cpumask_var_t core_mask;
900 if (!alloc_cpumask_var(&core_mask, GFP_KERNEL)) {
902 "efx.c: allocation failure, irq balancing hobbled\n");
906 cpumask_clear(core_mask);
908 for_each_online_cpu(cpu) {
909 if (!cpumask_test_cpu(cpu, core_mask)) {
911 cpumask_or(core_mask, core_mask,
912 topology_core_cpumask(cpu));
916 free_cpumask_var(core_mask);
920 /* Probe the number and type of interrupts we are able to obtain, and
921 * the resulting numbers of channels and RX queues.
923 static void efx_probe_interrupts(struct efx_nic *efx)
926 min_t(int, efx->type->phys_addr_channels, EFX_MAX_CHANNELS);
929 if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
930 struct msix_entry xentries[EFX_MAX_CHANNELS];
934 /* We want one RX queue and interrupt per CPU package
935 * (or as specified by the rss_cpus module parameter).
936 * We will need one channel per interrupt.
938 rx_queues = rss_cpus ? rss_cpus : efx_wanted_rx_queues();
939 wanted_ints = rx_queues + (separate_tx_channels ? 1 : 0);
940 wanted_ints = min(wanted_ints, max_channels);
942 for (i = 0; i < wanted_ints; i++)
943 xentries[i].entry = i;
944 rc = pci_enable_msix(efx->pci_dev, xentries, wanted_ints);
946 EFX_ERR(efx, "WARNING: Insufficient MSI-X vectors"
947 " available (%d < %d).\n", rc, wanted_ints);
948 EFX_ERR(efx, "WARNING: Performance may be reduced.\n");
949 EFX_BUG_ON_PARANOID(rc >= wanted_ints);
951 rc = pci_enable_msix(efx->pci_dev, xentries,
956 efx->n_rx_queues = min(rx_queues, wanted_ints);
957 efx->n_channels = wanted_ints;
958 for (i = 0; i < wanted_ints; i++)
959 efx->channel[i].irq = xentries[i].vector;
961 /* Fall back to single channel MSI */
962 efx->interrupt_mode = EFX_INT_MODE_MSI;
963 EFX_ERR(efx, "could not enable MSI-X\n");
967 /* Try single interrupt MSI */
968 if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
969 efx->n_rx_queues = 1;
971 rc = pci_enable_msi(efx->pci_dev);
973 efx->channel[0].irq = efx->pci_dev->irq;
975 EFX_ERR(efx, "could not enable MSI\n");
976 efx->interrupt_mode = EFX_INT_MODE_LEGACY;
980 /* Assume legacy interrupts */
981 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
982 efx->n_rx_queues = 1;
983 efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
984 efx->legacy_irq = efx->pci_dev->irq;
988 static void efx_remove_interrupts(struct efx_nic *efx)
990 struct efx_channel *channel;
992 /* Remove MSI/MSI-X interrupts */
993 efx_for_each_channel(channel, efx)
995 pci_disable_msi(efx->pci_dev);
996 pci_disable_msix(efx->pci_dev);
998 /* Remove legacy interrupt */
1002 static void efx_set_channels(struct efx_nic *efx)
1004 struct efx_tx_queue *tx_queue;
1005 struct efx_rx_queue *rx_queue;
1007 efx_for_each_tx_queue(tx_queue, efx) {
1008 if (separate_tx_channels)
1009 tx_queue->channel = &efx->channel[efx->n_channels-1];
1011 tx_queue->channel = &efx->channel[0];
1012 tx_queue->channel->used_flags |= EFX_USED_BY_TX;
1015 efx_for_each_rx_queue(rx_queue, efx) {
1016 rx_queue->channel = &efx->channel[rx_queue->queue];
1017 rx_queue->channel->used_flags |= EFX_USED_BY_RX;
1021 static int efx_probe_nic(struct efx_nic *efx)
1025 EFX_LOG(efx, "creating NIC\n");
1027 /* Carry out hardware-type specific initialisation */
1028 rc = falcon_probe_nic(efx);
1032 /* Determine the number of channels and RX queues by trying to hook
1033 * in MSI-X interrupts. */
1034 efx_probe_interrupts(efx);
1036 efx_set_channels(efx);
1038 /* Initialise the interrupt moderation settings */
1039 efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true);
1044 static void efx_remove_nic(struct efx_nic *efx)
1046 EFX_LOG(efx, "destroying NIC\n");
1048 efx_remove_interrupts(efx);
1049 falcon_remove_nic(efx);
1052 /**************************************************************************
1054 * NIC startup/shutdown
1056 *************************************************************************/
1058 static int efx_probe_all(struct efx_nic *efx)
1060 struct efx_channel *channel;
1064 rc = efx_probe_nic(efx);
1066 EFX_ERR(efx, "failed to create NIC\n");
1071 rc = efx_probe_port(efx);
1073 EFX_ERR(efx, "failed to create port\n");
1077 /* Create channels */
1078 efx_for_each_channel(channel, efx) {
1079 rc = efx_probe_channel(channel);
1081 EFX_ERR(efx, "failed to create channel %d\n",
1086 efx_set_channel_names(efx);
1091 efx_for_each_channel(channel, efx)
1092 efx_remove_channel(channel);
1093 efx_remove_port(efx);
1095 efx_remove_nic(efx);
1100 /* Called after previous invocation(s) of efx_stop_all, restarts the
1101 * port, kernel transmit queue, NAPI processing and hardware interrupts,
1102 * and ensures that the port is scheduled to be reconfigured.
1103 * This function is safe to call multiple times when the NIC is in any
1105 static void efx_start_all(struct efx_nic *efx)
1107 struct efx_channel *channel;
1109 EFX_ASSERT_RESET_SERIALISED(efx);
1111 /* Check that it is appropriate to restart the interface. All
1112 * of these flags are safe to read under just the rtnl lock */
1113 if (efx->port_enabled)
1115 if ((efx->state != STATE_RUNNING) && (efx->state != STATE_INIT))
1117 if (efx_dev_registered(efx) && !netif_running(efx->net_dev))
1120 /* Mark the port as enabled so port reconfigurations can start, then
1121 * restart the transmit interface early so the watchdog timer stops */
1122 efx_start_port(efx);
1123 if (efx_dev_registered(efx))
1124 efx_wake_queue(efx);
1126 efx_for_each_channel(channel, efx)
1127 efx_start_channel(channel);
1129 falcon_enable_interrupts(efx);
1131 /* Start hardware monitor if we're in RUNNING */
1132 if (efx->state == STATE_RUNNING)
1133 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1134 efx_monitor_interval);
1137 /* Flush all delayed work. Should only be called when no more delayed work
1138 * will be scheduled. This doesn't flush pending online resets (efx_reset),
1139 * since we're holding the rtnl_lock at this point. */
1140 static void efx_flush_all(struct efx_nic *efx)
1142 struct efx_rx_queue *rx_queue;
1144 /* Make sure the hardware monitor is stopped */
1145 cancel_delayed_work_sync(&efx->monitor_work);
1147 /* Ensure that all RX slow refills are complete. */
1148 efx_for_each_rx_queue(rx_queue, efx)
1149 cancel_delayed_work_sync(&rx_queue->work);
1151 /* Stop scheduled port reconfigurations */
1152 cancel_work_sync(&efx->mac_work);
1153 cancel_work_sync(&efx->phy_work);
1157 /* Quiesce hardware and software without bringing the link down.
1158 * Safe to call multiple times, when the nic and interface is in any
1159 * state. The caller is guaranteed to subsequently be in a position
1160 * to modify any hardware and software state they see fit without
1162 static void efx_stop_all(struct efx_nic *efx)
1164 struct efx_channel *channel;
1166 EFX_ASSERT_RESET_SERIALISED(efx);
1168 /* port_enabled can be read safely under the rtnl lock */
1169 if (!efx->port_enabled)
1172 /* Disable interrupts and wait for ISR to complete */
1173 falcon_disable_interrupts(efx);
1174 if (efx->legacy_irq)
1175 synchronize_irq(efx->legacy_irq);
1176 efx_for_each_channel(channel, efx) {
1178 synchronize_irq(channel->irq);
1181 /* Stop all NAPI processing and synchronous rx refills */
1182 efx_for_each_channel(channel, efx)
1183 efx_stop_channel(channel);
1185 /* Stop all asynchronous port reconfigurations. Since all
1186 * event processing has already been stopped, there is no
1187 * window to loose phy events */
1190 /* Flush efx_phy_work, efx_mac_work, refill_workqueue, monitor_work */
1193 /* Isolate the MAC from the TX and RX engines, so that queue
1194 * flushes will complete in a timely fashion. */
1195 falcon_drain_tx_fifo(efx);
1197 /* Stop the kernel transmit interface late, so the watchdog
1198 * timer isn't ticking over the flush */
1199 if (efx_dev_registered(efx)) {
1200 efx_stop_queue(efx);
1201 netif_tx_lock_bh(efx->net_dev);
1202 netif_tx_unlock_bh(efx->net_dev);
1206 static void efx_remove_all(struct efx_nic *efx)
1208 struct efx_channel *channel;
1210 efx_for_each_channel(channel, efx)
1211 efx_remove_channel(channel);
1212 efx_remove_port(efx);
1213 efx_remove_nic(efx);
1216 /* A convinience function to safely flush all the queues */
1217 void efx_flush_queues(struct efx_nic *efx)
1219 EFX_ASSERT_RESET_SERIALISED(efx);
1223 efx_fini_channels(efx);
1224 efx_init_channels(efx);
1229 /**************************************************************************
1231 * Interrupt moderation
1233 **************************************************************************/
1235 /* Set interrupt moderation parameters */
1236 void efx_init_irq_moderation(struct efx_nic *efx, int tx_usecs, int rx_usecs,
1239 struct efx_tx_queue *tx_queue;
1240 struct efx_rx_queue *rx_queue;
1242 EFX_ASSERT_RESET_SERIALISED(efx);
1244 efx_for_each_tx_queue(tx_queue, efx)
1245 tx_queue->channel->irq_moderation = tx_usecs;
1247 efx->irq_rx_adaptive = rx_adaptive;
1248 efx->irq_rx_moderation = rx_usecs;
1249 efx_for_each_rx_queue(rx_queue, efx)
1250 rx_queue->channel->irq_moderation = rx_usecs;
1253 /**************************************************************************
1257 **************************************************************************/
1259 /* Run periodically off the general workqueue. Serialised against
1260 * efx_reconfigure_port via the mac_lock */
1261 static void efx_monitor(struct work_struct *data)
1263 struct efx_nic *efx = container_of(data, struct efx_nic,
1267 EFX_TRACE(efx, "hardware monitor executing on CPU %d\n",
1268 raw_smp_processor_id());
1270 /* If the mac_lock is already held then it is likely a port
1271 * reconfiguration is already in place, which will likely do
1272 * most of the work of check_hw() anyway. */
1273 if (!mutex_trylock(&efx->mac_lock))
1275 if (!efx->port_enabled)
1277 rc = efx->board_info.monitor(efx);
1279 EFX_ERR(efx, "Board sensor %s; shutting down PHY\n",
1280 (rc == -ERANGE) ? "reported fault" : "failed");
1281 efx->phy_mode |= PHY_MODE_LOW_POWER;
1282 falcon_sim_phy_event(efx);
1284 efx->phy_op->poll(efx);
1285 efx->mac_op->poll(efx);
1288 mutex_unlock(&efx->mac_lock);
1290 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1291 efx_monitor_interval);
1294 /**************************************************************************
1298 *************************************************************************/
1301 * Context: process, rtnl_lock() held.
1303 static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1305 struct efx_nic *efx = netdev_priv(net_dev);
1307 EFX_ASSERT_RESET_SERIALISED(efx);
1309 return generic_mii_ioctl(&efx->mii, if_mii(ifr), cmd, NULL);
1312 /**************************************************************************
1316 **************************************************************************/
1318 static int efx_init_napi(struct efx_nic *efx)
1320 struct efx_channel *channel;
1322 efx_for_each_channel(channel, efx) {
1323 channel->napi_dev = efx->net_dev;
1328 static void efx_fini_napi(struct efx_nic *efx)
1330 struct efx_channel *channel;
1332 efx_for_each_channel(channel, efx) {
1333 channel->napi_dev = NULL;
1337 /**************************************************************************
1339 * Kernel netpoll interface
1341 *************************************************************************/
1343 #ifdef CONFIG_NET_POLL_CONTROLLER
1345 /* Although in the common case interrupts will be disabled, this is not
1346 * guaranteed. However, all our work happens inside the NAPI callback,
1347 * so no locking is required.
1349 static void efx_netpoll(struct net_device *net_dev)
1351 struct efx_nic *efx = netdev_priv(net_dev);
1352 struct efx_channel *channel;
1354 efx_for_each_channel(channel, efx)
1355 efx_schedule_channel(channel);
1360 /**************************************************************************
1362 * Kernel net device interface
1364 *************************************************************************/
1366 /* Context: process, rtnl_lock() held. */
1367 static int efx_net_open(struct net_device *net_dev)
1369 struct efx_nic *efx = netdev_priv(net_dev);
1370 EFX_ASSERT_RESET_SERIALISED(efx);
1372 EFX_LOG(efx, "opening device %s on CPU %d\n", net_dev->name,
1373 raw_smp_processor_id());
1375 if (efx->state == STATE_DISABLED)
1377 if (efx->phy_mode & PHY_MODE_SPECIAL)
1384 /* Context: process, rtnl_lock() held.
1385 * Note that the kernel will ignore our return code; this method
1386 * should really be a void.
1388 static int efx_net_stop(struct net_device *net_dev)
1390 struct efx_nic *efx = netdev_priv(net_dev);
1392 EFX_LOG(efx, "closing %s on CPU %d\n", net_dev->name,
1393 raw_smp_processor_id());
1395 if (efx->state != STATE_DISABLED) {
1396 /* Stop the device and flush all the channels */
1398 efx_fini_channels(efx);
1399 efx_init_channels(efx);
1405 void efx_stats_disable(struct efx_nic *efx)
1407 spin_lock(&efx->stats_lock);
1408 ++efx->stats_disable_count;
1409 spin_unlock(&efx->stats_lock);
1412 void efx_stats_enable(struct efx_nic *efx)
1414 spin_lock(&efx->stats_lock);
1415 --efx->stats_disable_count;
1416 spin_unlock(&efx->stats_lock);
1419 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1420 static struct net_device_stats *efx_net_stats(struct net_device *net_dev)
1422 struct efx_nic *efx = netdev_priv(net_dev);
1423 struct efx_mac_stats *mac_stats = &efx->mac_stats;
1424 struct net_device_stats *stats = &net_dev->stats;
1426 /* Update stats if possible, but do not wait if another thread
1427 * is updating them or if MAC stats fetches are temporarily
1428 * disabled; slightly stale stats are acceptable.
1430 if (!spin_trylock(&efx->stats_lock))
1432 if (!efx->stats_disable_count) {
1433 efx->mac_op->update_stats(efx);
1434 falcon_update_nic_stats(efx);
1436 spin_unlock(&efx->stats_lock);
1438 stats->rx_packets = mac_stats->rx_packets;
1439 stats->tx_packets = mac_stats->tx_packets;
1440 stats->rx_bytes = mac_stats->rx_bytes;
1441 stats->tx_bytes = mac_stats->tx_bytes;
1442 stats->multicast = mac_stats->rx_multicast;
1443 stats->collisions = mac_stats->tx_collision;
1444 stats->rx_length_errors = (mac_stats->rx_gtjumbo +
1445 mac_stats->rx_length_error);
1446 stats->rx_over_errors = efx->n_rx_nodesc_drop_cnt;
1447 stats->rx_crc_errors = mac_stats->rx_bad;
1448 stats->rx_frame_errors = mac_stats->rx_align_error;
1449 stats->rx_fifo_errors = mac_stats->rx_overflow;
1450 stats->rx_missed_errors = mac_stats->rx_missed;
1451 stats->tx_window_errors = mac_stats->tx_late_collision;
1453 stats->rx_errors = (stats->rx_length_errors +
1454 stats->rx_over_errors +
1455 stats->rx_crc_errors +
1456 stats->rx_frame_errors +
1457 stats->rx_fifo_errors +
1458 stats->rx_missed_errors +
1459 mac_stats->rx_symbol_error);
1460 stats->tx_errors = (stats->tx_window_errors +
1466 /* Context: netif_tx_lock held, BHs disabled. */
1467 static void efx_watchdog(struct net_device *net_dev)
1469 struct efx_nic *efx = netdev_priv(net_dev);
1471 EFX_ERR(efx, "TX stuck with stop_count=%d port_enabled=%d:"
1472 " resetting channels\n",
1473 atomic_read(&efx->netif_stop_count), efx->port_enabled);
1475 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
1479 /* Context: process, rtnl_lock() held. */
1480 static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
1482 struct efx_nic *efx = netdev_priv(net_dev);
1485 EFX_ASSERT_RESET_SERIALISED(efx);
1487 if (new_mtu > EFX_MAX_MTU)
1492 EFX_LOG(efx, "changing MTU to %d\n", new_mtu);
1494 efx_fini_channels(efx);
1495 net_dev->mtu = new_mtu;
1496 efx_init_channels(efx);
1502 static int efx_set_mac_address(struct net_device *net_dev, void *data)
1504 struct efx_nic *efx = netdev_priv(net_dev);
1505 struct sockaddr *addr = data;
1506 char *new_addr = addr->sa_data;
1508 EFX_ASSERT_RESET_SERIALISED(efx);
1510 if (!is_valid_ether_addr(new_addr)) {
1511 EFX_ERR(efx, "invalid ethernet MAC address requested: %pM\n",
1516 memcpy(net_dev->dev_addr, new_addr, net_dev->addr_len);
1518 /* Reconfigure the MAC */
1519 efx_reconfigure_port(efx);
1524 /* Context: netif_addr_lock held, BHs disabled. */
1525 static void efx_set_multicast_list(struct net_device *net_dev)
1527 struct efx_nic *efx = netdev_priv(net_dev);
1528 struct dev_mc_list *mc_list = net_dev->mc_list;
1529 union efx_multicast_hash *mc_hash = &efx->multicast_hash;
1530 bool promiscuous = !!(net_dev->flags & IFF_PROMISC);
1531 bool changed = (efx->promiscuous != promiscuous);
1536 efx->promiscuous = promiscuous;
1538 /* Build multicast hash table */
1539 if (promiscuous || (net_dev->flags & IFF_ALLMULTI)) {
1540 memset(mc_hash, 0xff, sizeof(*mc_hash));
1542 memset(mc_hash, 0x00, sizeof(*mc_hash));
1543 for (i = 0; i < net_dev->mc_count; i++) {
1544 crc = ether_crc_le(ETH_ALEN, mc_list->dmi_addr);
1545 bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
1546 set_bit_le(bit, mc_hash->byte);
1547 mc_list = mc_list->next;
1551 if (!efx->port_enabled)
1552 /* Delay pushing settings until efx_start_port() */
1556 queue_work(efx->workqueue, &efx->phy_work);
1558 /* Create and activate new global multicast hash table */
1559 falcon_set_multicast_hash(efx);
1562 static const struct net_device_ops efx_netdev_ops = {
1563 .ndo_open = efx_net_open,
1564 .ndo_stop = efx_net_stop,
1565 .ndo_get_stats = efx_net_stats,
1566 .ndo_tx_timeout = efx_watchdog,
1567 .ndo_start_xmit = efx_hard_start_xmit,
1568 .ndo_validate_addr = eth_validate_addr,
1569 .ndo_do_ioctl = efx_ioctl,
1570 .ndo_change_mtu = efx_change_mtu,
1571 .ndo_set_mac_address = efx_set_mac_address,
1572 .ndo_set_multicast_list = efx_set_multicast_list,
1573 #ifdef CONFIG_NET_POLL_CONTROLLER
1574 .ndo_poll_controller = efx_netpoll,
1578 static void efx_update_name(struct efx_nic *efx)
1580 strcpy(efx->name, efx->net_dev->name);
1581 efx_mtd_rename(efx);
1582 efx_set_channel_names(efx);
1585 static int efx_netdev_event(struct notifier_block *this,
1586 unsigned long event, void *ptr)
1588 struct net_device *net_dev = ptr;
1590 if (net_dev->netdev_ops == &efx_netdev_ops &&
1591 event == NETDEV_CHANGENAME)
1592 efx_update_name(netdev_priv(net_dev));
1597 static struct notifier_block efx_netdev_notifier = {
1598 .notifier_call = efx_netdev_event,
1602 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
1604 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
1605 return sprintf(buf, "%d\n", efx->phy_type);
1607 static DEVICE_ATTR(phy_type, 0644, show_phy_type, NULL);
1609 static int efx_register_netdev(struct efx_nic *efx)
1611 struct net_device *net_dev = efx->net_dev;
1614 net_dev->watchdog_timeo = 5 * HZ;
1615 net_dev->irq = efx->pci_dev->irq;
1616 net_dev->netdev_ops = &efx_netdev_ops;
1617 SET_NETDEV_DEV(net_dev, &efx->pci_dev->dev);
1618 SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
1620 /* Always start with carrier off; PHY events will detect the link */
1621 netif_carrier_off(efx->net_dev);
1623 /* Clear MAC statistics */
1624 efx->mac_op->update_stats(efx);
1625 memset(&efx->mac_stats, 0, sizeof(efx->mac_stats));
1627 rc = register_netdev(net_dev);
1629 EFX_ERR(efx, "could not register net dev\n");
1634 efx_update_name(efx);
1637 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
1639 EFX_ERR(efx, "failed to init net dev attributes\n");
1640 goto fail_registered;
1646 unregister_netdev(net_dev);
1650 static void efx_unregister_netdev(struct efx_nic *efx)
1652 struct efx_tx_queue *tx_queue;
1657 BUG_ON(netdev_priv(efx->net_dev) != efx);
1659 /* Free up any skbs still remaining. This has to happen before
1660 * we try to unregister the netdev as running their destructors
1661 * may be needed to get the device ref. count to 0. */
1662 efx_for_each_tx_queue(tx_queue, efx)
1663 efx_release_tx_buffers(tx_queue);
1665 if (efx_dev_registered(efx)) {
1666 strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
1667 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
1668 unregister_netdev(efx->net_dev);
1672 /**************************************************************************
1674 * Device reset and suspend
1676 **************************************************************************/
1678 /* Tears down the entire software state and most of the hardware state
1680 void efx_reset_down(struct efx_nic *efx, enum reset_type method,
1681 struct ethtool_cmd *ecmd)
1683 EFX_ASSERT_RESET_SERIALISED(efx);
1685 efx_stats_disable(efx);
1687 mutex_lock(&efx->mac_lock);
1688 mutex_lock(&efx->spi_lock);
1690 efx->phy_op->get_settings(efx, ecmd);
1692 efx_fini_channels(efx);
1693 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
1694 efx->phy_op->fini(efx);
1697 /* This function will always ensure that the locks acquired in
1698 * efx_reset_down() are released. A failure return code indicates
1699 * that we were unable to reinitialise the hardware, and the
1700 * driver should be disabled. If ok is false, then the rx and tx
1701 * engines are not restarted, pending a RESET_DISABLE. */
1702 int efx_reset_up(struct efx_nic *efx, enum reset_type method,
1703 struct ethtool_cmd *ecmd, bool ok)
1707 EFX_ASSERT_RESET_SERIALISED(efx);
1709 rc = falcon_init_nic(efx);
1711 EFX_ERR(efx, "failed to initialise NIC\n");
1715 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
1717 rc = efx->phy_op->init(efx);
1722 efx->port_initialized = false;
1726 efx_init_channels(efx);
1728 if (efx->phy_op->set_settings(efx, ecmd))
1729 EFX_ERR(efx, "could not restore PHY settings\n");
1732 mutex_unlock(&efx->spi_lock);
1733 mutex_unlock(&efx->mac_lock);
1737 efx_stats_enable(efx);
1742 /* Reset the NIC as transparently as possible. Do not reset the PHY
1743 * Note that the reset may fail, in which case the card will be left
1744 * in a most-probably-unusable state.
1746 * This function will sleep. You cannot reset from within an atomic
1747 * state; use efx_schedule_reset() instead.
1749 * Grabs the rtnl_lock.
1751 static int efx_reset(struct efx_nic *efx)
1753 struct ethtool_cmd ecmd;
1754 enum reset_type method = efx->reset_pending;
1757 /* Serialise with kernel interfaces */
1760 /* If we're not RUNNING then don't reset. Leave the reset_pending
1761 * flag set so that efx_pci_probe_main will be retried */
1762 if (efx->state != STATE_RUNNING) {
1763 EFX_INFO(efx, "scheduled reset quenched. NIC not RUNNING\n");
1767 EFX_INFO(efx, "resetting (%d)\n", method);
1769 efx_reset_down(efx, method, &ecmd);
1771 rc = falcon_reset_hw(efx, method);
1773 EFX_ERR(efx, "failed to reset hardware\n");
1777 /* Allow resets to be rescheduled. */
1778 efx->reset_pending = RESET_TYPE_NONE;
1780 /* Reinitialise bus-mastering, which may have been turned off before
1781 * the reset was scheduled. This is still appropriate, even in the
1782 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
1783 * can respond to requests. */
1784 pci_set_master(efx->pci_dev);
1786 /* Leave device stopped if necessary */
1787 if (method == RESET_TYPE_DISABLE) {
1788 efx_reset_up(efx, method, &ecmd, false);
1791 rc = efx_reset_up(efx, method, &ecmd, true);
1796 EFX_ERR(efx, "has been disabled\n");
1797 efx->state = STATE_DISABLED;
1798 dev_close(efx->net_dev);
1800 EFX_LOG(efx, "reset complete\n");
1808 /* The worker thread exists so that code that cannot sleep can
1809 * schedule a reset for later.
1811 static void efx_reset_work(struct work_struct *data)
1813 struct efx_nic *nic = container_of(data, struct efx_nic, reset_work);
1818 void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
1820 enum reset_type method;
1822 if (efx->reset_pending != RESET_TYPE_NONE) {
1823 EFX_INFO(efx, "quenching already scheduled reset\n");
1828 case RESET_TYPE_INVISIBLE:
1829 case RESET_TYPE_ALL:
1830 case RESET_TYPE_WORLD:
1831 case RESET_TYPE_DISABLE:
1834 case RESET_TYPE_RX_RECOVERY:
1835 case RESET_TYPE_RX_DESC_FETCH:
1836 case RESET_TYPE_TX_DESC_FETCH:
1837 case RESET_TYPE_TX_SKIP:
1838 method = RESET_TYPE_INVISIBLE;
1841 method = RESET_TYPE_ALL;
1846 EFX_LOG(efx, "scheduling reset (%d:%d)\n", type, method);
1848 EFX_LOG(efx, "scheduling reset (%d)\n", method);
1850 efx->reset_pending = method;
1852 queue_work(reset_workqueue, &efx->reset_work);
1855 /**************************************************************************
1857 * List of NICs we support
1859 **************************************************************************/
1861 /* PCI device ID table */
1862 static struct pci_device_id efx_pci_table[] __devinitdata = {
1863 {PCI_DEVICE(EFX_VENDID_SFC, FALCON_A_P_DEVID),
1864 .driver_data = (unsigned long) &falcon_a_nic_type},
1865 {PCI_DEVICE(EFX_VENDID_SFC, FALCON_B_P_DEVID),
1866 .driver_data = (unsigned long) &falcon_b_nic_type},
1867 {0} /* end of list */
1870 /**************************************************************************
1872 * Dummy PHY/MAC/Board operations
1874 * Can be used for some unimplemented operations
1875 * Needed so all function pointers are valid and do not have to be tested
1878 **************************************************************************/
1879 int efx_port_dummy_op_int(struct efx_nic *efx)
1883 void efx_port_dummy_op_void(struct efx_nic *efx) {}
1884 void efx_port_dummy_op_blink(struct efx_nic *efx, bool blink) {}
1886 static struct efx_mac_operations efx_dummy_mac_operations = {
1887 .reconfigure = efx_port_dummy_op_void,
1888 .poll = efx_port_dummy_op_void,
1889 .irq = efx_port_dummy_op_void,
1892 static struct efx_phy_operations efx_dummy_phy_operations = {
1893 .init = efx_port_dummy_op_int,
1894 .reconfigure = efx_port_dummy_op_void,
1895 .poll = efx_port_dummy_op_void,
1896 .fini = efx_port_dummy_op_void,
1897 .clear_interrupt = efx_port_dummy_op_void,
1900 static struct efx_board efx_dummy_board_info = {
1901 .init = efx_port_dummy_op_int,
1902 .init_leds = efx_port_dummy_op_void,
1903 .set_id_led = efx_port_dummy_op_blink,
1904 .monitor = efx_port_dummy_op_int,
1905 .blink = efx_port_dummy_op_blink,
1906 .fini = efx_port_dummy_op_void,
1909 /**************************************************************************
1913 **************************************************************************/
1915 /* This zeroes out and then fills in the invariants in a struct
1916 * efx_nic (including all sub-structures).
1918 static int efx_init_struct(struct efx_nic *efx, struct efx_nic_type *type,
1919 struct pci_dev *pci_dev, struct net_device *net_dev)
1921 struct efx_channel *channel;
1922 struct efx_tx_queue *tx_queue;
1923 struct efx_rx_queue *rx_queue;
1926 /* Initialise common structures */
1927 memset(efx, 0, sizeof(*efx));
1928 spin_lock_init(&efx->biu_lock);
1929 spin_lock_init(&efx->phy_lock);
1930 mutex_init(&efx->spi_lock);
1931 INIT_WORK(&efx->reset_work, efx_reset_work);
1932 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
1933 efx->pci_dev = pci_dev;
1934 efx->state = STATE_INIT;
1935 efx->reset_pending = RESET_TYPE_NONE;
1936 strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
1937 efx->board_info = efx_dummy_board_info;
1939 efx->net_dev = net_dev;
1940 efx->rx_checksum_enabled = true;
1941 spin_lock_init(&efx->netif_stop_lock);
1942 spin_lock_init(&efx->stats_lock);
1943 efx->stats_disable_count = 1;
1944 mutex_init(&efx->mac_lock);
1945 efx->mac_op = &efx_dummy_mac_operations;
1946 efx->phy_op = &efx_dummy_phy_operations;
1947 efx->mii.dev = net_dev;
1948 INIT_WORK(&efx->phy_work, efx_phy_work);
1949 INIT_WORK(&efx->mac_work, efx_mac_work);
1950 atomic_set(&efx->netif_stop_count, 1);
1952 for (i = 0; i < EFX_MAX_CHANNELS; i++) {
1953 channel = &efx->channel[i];
1955 channel->channel = i;
1956 channel->work_pending = false;
1958 for (i = 0; i < EFX_TX_QUEUE_COUNT; i++) {
1959 tx_queue = &efx->tx_queue[i];
1960 tx_queue->efx = efx;
1961 tx_queue->queue = i;
1962 tx_queue->buffer = NULL;
1963 tx_queue->channel = &efx->channel[0]; /* for safety */
1964 tx_queue->tso_headers_free = NULL;
1966 for (i = 0; i < EFX_MAX_RX_QUEUES; i++) {
1967 rx_queue = &efx->rx_queue[i];
1968 rx_queue->efx = efx;
1969 rx_queue->queue = i;
1970 rx_queue->channel = &efx->channel[0]; /* for safety */
1971 rx_queue->buffer = NULL;
1972 spin_lock_init(&rx_queue->add_lock);
1973 INIT_DELAYED_WORK(&rx_queue->work, efx_rx_work);
1978 /* Sanity-check NIC type */
1979 EFX_BUG_ON_PARANOID(efx->type->txd_ring_mask &
1980 (efx->type->txd_ring_mask + 1));
1981 EFX_BUG_ON_PARANOID(efx->type->rxd_ring_mask &
1982 (efx->type->rxd_ring_mask + 1));
1983 EFX_BUG_ON_PARANOID(efx->type->evq_size &
1984 (efx->type->evq_size - 1));
1985 /* As close as we can get to guaranteeing that we don't overflow */
1986 EFX_BUG_ON_PARANOID(efx->type->evq_size <
1987 (efx->type->txd_ring_mask + 1 +
1988 efx->type->rxd_ring_mask + 1));
1989 EFX_BUG_ON_PARANOID(efx->type->phys_addr_channels > EFX_MAX_CHANNELS);
1991 /* Higher numbered interrupt modes are less capable! */
1992 efx->interrupt_mode = max(efx->type->max_interrupt_mode,
1995 /* Would be good to use the net_dev name, but we're too early */
1996 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
1998 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
1999 if (!efx->workqueue)
2005 static void efx_fini_struct(struct efx_nic *efx)
2007 if (efx->workqueue) {
2008 destroy_workqueue(efx->workqueue);
2009 efx->workqueue = NULL;
2013 /**************************************************************************
2017 **************************************************************************/
2019 /* Main body of final NIC shutdown code
2020 * This is called only at module unload (or hotplug removal).
2022 static void efx_pci_remove_main(struct efx_nic *efx)
2024 EFX_ASSERT_RESET_SERIALISED(efx);
2026 /* Skip everything if we never obtained a valid membase */
2030 efx_fini_channels(efx);
2033 /* Shutdown the board, then the NIC and board state */
2034 efx->board_info.fini(efx);
2035 falcon_fini_interrupt(efx);
2038 efx_remove_all(efx);
2041 /* Final NIC shutdown
2042 * This is called only at module unload (or hotplug removal).
2044 static void efx_pci_remove(struct pci_dev *pci_dev)
2046 struct efx_nic *efx;
2048 efx = pci_get_drvdata(pci_dev);
2052 /* Mark the NIC as fini, then stop the interface */
2054 efx->state = STATE_FINI;
2055 dev_close(efx->net_dev);
2057 /* Allow any queued efx_resets() to complete */
2060 if (efx->membase == NULL)
2063 efx_unregister_netdev(efx);
2065 efx_mtd_remove(efx);
2067 /* Wait for any scheduled resets to complete. No more will be
2068 * scheduled from this point because efx_stop_all() has been
2069 * called, we are no longer registered with driverlink, and
2070 * the net_device's have been removed. */
2071 cancel_work_sync(&efx->reset_work);
2073 efx_pci_remove_main(efx);
2077 EFX_LOG(efx, "shutdown successful\n");
2079 pci_set_drvdata(pci_dev, NULL);
2080 efx_fini_struct(efx);
2081 free_netdev(efx->net_dev);
2084 /* Main body of NIC initialisation
2085 * This is called at module load (or hotplug insertion, theoretically).
2087 static int efx_pci_probe_main(struct efx_nic *efx)
2091 /* Do start-of-day initialisation */
2092 rc = efx_probe_all(efx);
2096 rc = efx_init_napi(efx);
2100 /* Initialise the board */
2101 rc = efx->board_info.init(efx);
2103 EFX_ERR(efx, "failed to initialise board\n");
2107 rc = falcon_init_nic(efx);
2109 EFX_ERR(efx, "failed to initialise NIC\n");
2113 rc = efx_init_port(efx);
2115 EFX_ERR(efx, "failed to initialise port\n");
2119 efx_init_channels(efx);
2121 rc = falcon_init_interrupt(efx);
2128 efx_fini_channels(efx);
2132 efx->board_info.fini(efx);
2136 efx_remove_all(efx);
2141 /* NIC initialisation
2143 * This is called at module load (or hotplug insertion,
2144 * theoretically). It sets up PCI mappings, tests and resets the NIC,
2145 * sets up and registers the network devices with the kernel and hooks
2146 * the interrupt service routine. It does not prepare the device for
2147 * transmission; this is left to the first time one of the network
2148 * interfaces is brought up (i.e. efx_net_open).
2150 static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
2151 const struct pci_device_id *entry)
2153 struct efx_nic_type *type = (struct efx_nic_type *) entry->driver_data;
2154 struct net_device *net_dev;
2155 struct efx_nic *efx;
2158 /* Allocate and initialise a struct net_device and struct efx_nic */
2159 net_dev = alloc_etherdev(sizeof(*efx));
2162 net_dev->features |= (NETIF_F_IP_CSUM | NETIF_F_SG |
2163 NETIF_F_HIGHDMA | NETIF_F_TSO);
2165 net_dev->features |= NETIF_F_GRO;
2166 /* Mask for features that also apply to VLAN devices */
2167 net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2168 NETIF_F_HIGHDMA | NETIF_F_TSO);
2169 efx = netdev_priv(net_dev);
2170 pci_set_drvdata(pci_dev, efx);
2171 rc = efx_init_struct(efx, type, pci_dev, net_dev);
2175 EFX_INFO(efx, "Solarflare Communications NIC detected\n");
2177 /* Set up basic I/O (BAR mappings etc) */
2178 rc = efx_init_io(efx);
2182 /* No serialisation is required with the reset path because
2183 * we're in STATE_INIT. */
2184 for (i = 0; i < 5; i++) {
2185 rc = efx_pci_probe_main(efx);
2187 /* Serialise against efx_reset(). No more resets will be
2188 * scheduled since efx_stop_all() has been called, and we
2189 * have not and never have been registered with either
2190 * the rtnetlink or driverlink layers. */
2191 cancel_work_sync(&efx->reset_work);
2194 if (efx->reset_pending != RESET_TYPE_NONE) {
2195 /* If there was a scheduled reset during
2196 * probe, the NIC is probably hosed anyway */
2197 efx_pci_remove_main(efx);
2204 /* Retry if a recoverably reset event has been scheduled */
2205 if ((efx->reset_pending != RESET_TYPE_INVISIBLE) &&
2206 (efx->reset_pending != RESET_TYPE_ALL))
2209 efx->reset_pending = RESET_TYPE_NONE;
2213 EFX_ERR(efx, "Could not reset NIC\n");
2217 /* Switch to the running state before we expose the device to
2218 * the OS. This is to ensure that the initial gathering of
2219 * MAC stats succeeds. */
2220 efx->state = STATE_RUNNING;
2222 efx_mtd_probe(efx); /* allowed to fail */
2224 rc = efx_register_netdev(efx);
2228 EFX_LOG(efx, "initialisation successful\n");
2232 efx_pci_remove_main(efx);
2237 efx_fini_struct(efx);
2239 EFX_LOG(efx, "initialisation failed. rc=%d\n", rc);
2240 free_netdev(net_dev);
2244 static struct pci_driver efx_pci_driver = {
2245 .name = EFX_DRIVER_NAME,
2246 .id_table = efx_pci_table,
2247 .probe = efx_pci_probe,
2248 .remove = efx_pci_remove,
2251 /**************************************************************************
2253 * Kernel module interface
2255 *************************************************************************/
2257 module_param(interrupt_mode, uint, 0444);
2258 MODULE_PARM_DESC(interrupt_mode,
2259 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2261 static int __init efx_init_module(void)
2265 printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
2267 rc = register_netdevice_notifier(&efx_netdev_notifier);
2271 refill_workqueue = create_workqueue("sfc_refill");
2272 if (!refill_workqueue) {
2276 reset_workqueue = create_singlethread_workqueue("sfc_reset");
2277 if (!reset_workqueue) {
2282 rc = pci_register_driver(&efx_pci_driver);
2289 destroy_workqueue(reset_workqueue);
2291 destroy_workqueue(refill_workqueue);
2293 unregister_netdevice_notifier(&efx_netdev_notifier);
2298 static void __exit efx_exit_module(void)
2300 printk(KERN_INFO "Solarflare NET driver unloading\n");
2302 pci_unregister_driver(&efx_pci_driver);
2303 destroy_workqueue(reset_workqueue);
2304 destroy_workqueue(refill_workqueue);
2305 unregister_netdevice_notifier(&efx_netdev_notifier);
2309 module_init(efx_init_module);
2310 module_exit(efx_exit_module);
2312 MODULE_AUTHOR("Michael Brown <mbrown@fensystems.co.uk> and "
2313 "Solarflare Communications");
2314 MODULE_DESCRIPTION("Solarflare Communications network driver");
2315 MODULE_LICENSE("GPL");
2316 MODULE_DEVICE_TABLE(pci, efx_pci_table);