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 /**************************************************************************
138 * Utility functions and prototypes
140 *************************************************************************/
141 static void efx_remove_channel(struct efx_channel
*channel
);
142 static void efx_remove_port(struct efx_nic
*efx
);
143 static void efx_fini_napi(struct efx_nic
*efx
);
144 static void efx_fini_channels(struct efx_nic
*efx
);
146 #define EFX_ASSERT_RESET_SERIALISED(efx) \
148 if (efx->state == STATE_RUNNING) \
152 /**************************************************************************
154 * Event queue processing
156 *************************************************************************/
158 /* Process channel's event queue
160 * This function is responsible for processing the event queue of a
161 * single channel. The caller must guarantee that this function will
162 * never be concurrently called more than once on the same channel,
163 * though different channels may be being processed concurrently.
165 static int efx_process_channel(struct efx_channel
*channel
, int rx_quota
)
167 struct efx_nic
*efx
= channel
->efx
;
170 if (unlikely(efx
->reset_pending
!= RESET_TYPE_NONE
||
174 rx_packets
= falcon_process_eventq(channel
, rx_quota
);
178 /* Deliver last RX packet. */
179 if (channel
->rx_pkt
) {
180 __efx_rx_packet(channel
, channel
->rx_pkt
,
181 channel
->rx_pkt_csummed
);
182 channel
->rx_pkt
= NULL
;
185 efx_flush_lro(channel
);
186 efx_rx_strategy(channel
);
188 efx_fast_push_rx_descriptors(&efx
->rx_queue
[channel
->channel
]);
193 /* Mark channel as finished processing
195 * Note that since we will not receive further interrupts for this
196 * channel before we finish processing and call the eventq_read_ack()
197 * method, there is no need to use the interrupt hold-off timers.
199 static inline void efx_channel_processed(struct efx_channel
*channel
)
201 /* The interrupt handler for this channel may set work_pending
202 * as soon as we acknowledge the events we've seen. Make sure
203 * it's cleared before then. */
204 channel
->work_pending
= false;
207 falcon_eventq_read_ack(channel
);
212 * NAPI guarantees serialisation of polls of the same device, which
213 * provides the guarantee required by efx_process_channel().
215 static int efx_poll(struct napi_struct
*napi
, int budget
)
217 struct efx_channel
*channel
=
218 container_of(napi
, struct efx_channel
, napi_str
);
221 EFX_TRACE(channel
->efx
, "channel %d NAPI poll executing on CPU %d\n",
222 channel
->channel
, raw_smp_processor_id());
224 rx_packets
= efx_process_channel(channel
, budget
);
226 if (rx_packets
< budget
) {
227 /* There is no race here; although napi_disable() will
228 * only wait for netif_rx_complete(), this isn't a problem
229 * since efx_channel_processed() will have no effect if
230 * interrupts have already been disabled.
232 netif_rx_complete(napi
);
233 efx_channel_processed(channel
);
239 /* Process the eventq of the specified channel immediately on this CPU
241 * Disable hardware generated interrupts, wait for any existing
242 * processing to finish, then directly poll (and ack ) the eventq.
243 * Finally reenable NAPI and interrupts.
245 * Since we are touching interrupts the caller should hold the suspend lock
247 void efx_process_channel_now(struct efx_channel
*channel
)
249 struct efx_nic
*efx
= channel
->efx
;
251 BUG_ON(!channel
->used_flags
);
252 BUG_ON(!channel
->enabled
);
254 /* Disable interrupts and wait for ISRs to complete */
255 falcon_disable_interrupts(efx
);
257 synchronize_irq(efx
->legacy_irq
);
259 synchronize_irq(channel
->irq
);
261 /* Wait for any NAPI processing to complete */
262 napi_disable(&channel
->napi_str
);
264 /* Poll the channel */
265 efx_process_channel(channel
, efx
->type
->evq_size
);
267 /* Ack the eventq. This may cause an interrupt to be generated
268 * when they are reenabled */
269 efx_channel_processed(channel
);
271 napi_enable(&channel
->napi_str
);
272 falcon_enable_interrupts(efx
);
275 /* Create event queue
276 * Event queue memory allocations are done only once. If the channel
277 * is reset, the memory buffer will be reused; this guards against
278 * errors during channel reset and also simplifies interrupt handling.
280 static int efx_probe_eventq(struct efx_channel
*channel
)
282 EFX_LOG(channel
->efx
, "chan %d create event queue\n", channel
->channel
);
284 return falcon_probe_eventq(channel
);
287 /* Prepare channel's event queue */
288 static void efx_init_eventq(struct efx_channel
*channel
)
290 EFX_LOG(channel
->efx
, "chan %d init event queue\n", channel
->channel
);
292 channel
->eventq_read_ptr
= 0;
294 falcon_init_eventq(channel
);
297 static void efx_fini_eventq(struct efx_channel
*channel
)
299 EFX_LOG(channel
->efx
, "chan %d fini event queue\n", channel
->channel
);
301 falcon_fini_eventq(channel
);
304 static void efx_remove_eventq(struct efx_channel
*channel
)
306 EFX_LOG(channel
->efx
, "chan %d remove event queue\n", channel
->channel
);
308 falcon_remove_eventq(channel
);
311 /**************************************************************************
315 *************************************************************************/
317 static int efx_probe_channel(struct efx_channel
*channel
)
319 struct efx_tx_queue
*tx_queue
;
320 struct efx_rx_queue
*rx_queue
;
323 EFX_LOG(channel
->efx
, "creating channel %d\n", channel
->channel
);
325 rc
= efx_probe_eventq(channel
);
329 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
330 rc
= efx_probe_tx_queue(tx_queue
);
335 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
336 rc
= efx_probe_rx_queue(rx_queue
);
341 channel
->n_rx_frm_trunc
= 0;
346 efx_for_each_channel_rx_queue(rx_queue
, channel
)
347 efx_remove_rx_queue(rx_queue
);
349 efx_for_each_channel_tx_queue(tx_queue
, channel
)
350 efx_remove_tx_queue(tx_queue
);
356 static void efx_set_channel_names(struct efx_nic
*efx
)
358 struct efx_channel
*channel
;
359 const char *type
= "";
362 efx_for_each_channel(channel
, efx
) {
363 number
= channel
->channel
;
364 if (efx
->n_channels
> efx
->n_rx_queues
) {
365 if (channel
->channel
< efx
->n_rx_queues
) {
369 number
-= efx
->n_rx_queues
;
372 snprintf(channel
->name
, sizeof(channel
->name
),
373 "%s%s-%d", efx
->name
, type
, number
);
377 /* Channels are shutdown and reinitialised whilst the NIC is running
378 * to propagate configuration changes (mtu, checksum offload), or
379 * to clear hardware error conditions
381 static void efx_init_channels(struct efx_nic
*efx
)
383 struct efx_tx_queue
*tx_queue
;
384 struct efx_rx_queue
*rx_queue
;
385 struct efx_channel
*channel
;
387 /* Calculate the rx buffer allocation parameters required to
388 * support the current MTU, including padding for header
389 * alignment and overruns.
391 efx
->rx_buffer_len
= (max(EFX_PAGE_IP_ALIGN
, NET_IP_ALIGN
) +
392 EFX_MAX_FRAME_LEN(efx
->net_dev
->mtu
) +
393 efx
->type
->rx_buffer_padding
);
394 efx
->rx_buffer_order
= get_order(efx
->rx_buffer_len
);
396 /* Initialise the channels */
397 efx_for_each_channel(channel
, efx
) {
398 EFX_LOG(channel
->efx
, "init chan %d\n", channel
->channel
);
400 efx_init_eventq(channel
);
402 efx_for_each_channel_tx_queue(tx_queue
, channel
)
403 efx_init_tx_queue(tx_queue
);
405 /* The rx buffer allocation strategy is MTU dependent */
406 efx_rx_strategy(channel
);
408 efx_for_each_channel_rx_queue(rx_queue
, channel
)
409 efx_init_rx_queue(rx_queue
);
411 WARN_ON(channel
->rx_pkt
!= NULL
);
412 efx_rx_strategy(channel
);
416 /* This enables event queue processing and packet transmission.
418 * Note that this function is not allowed to fail, since that would
419 * introduce too much complexity into the suspend/resume path.
421 static void efx_start_channel(struct efx_channel
*channel
)
423 struct efx_rx_queue
*rx_queue
;
425 EFX_LOG(channel
->efx
, "starting chan %d\n", channel
->channel
);
427 if (!(channel
->efx
->net_dev
->flags
& IFF_UP
))
428 netif_napi_add(channel
->napi_dev
, &channel
->napi_str
,
429 efx_poll
, napi_weight
);
431 /* The interrupt handler for this channel may set work_pending
432 * as soon as we enable it. Make sure it's cleared before
433 * then. Similarly, make sure it sees the enabled flag set. */
434 channel
->work_pending
= false;
435 channel
->enabled
= true;
438 napi_enable(&channel
->napi_str
);
440 /* Load up RX descriptors */
441 efx_for_each_channel_rx_queue(rx_queue
, channel
)
442 efx_fast_push_rx_descriptors(rx_queue
);
445 /* This disables event queue processing and packet transmission.
446 * This function does not guarantee that all queue processing
447 * (e.g. RX refill) is complete.
449 static void efx_stop_channel(struct efx_channel
*channel
)
451 struct efx_rx_queue
*rx_queue
;
453 if (!channel
->enabled
)
456 EFX_LOG(channel
->efx
, "stop chan %d\n", channel
->channel
);
458 channel
->enabled
= false;
459 napi_disable(&channel
->napi_str
);
461 /* Ensure that any worker threads have exited or will be no-ops */
462 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
463 spin_lock_bh(&rx_queue
->add_lock
);
464 spin_unlock_bh(&rx_queue
->add_lock
);
468 static void efx_fini_channels(struct efx_nic
*efx
)
470 struct efx_channel
*channel
;
471 struct efx_tx_queue
*tx_queue
;
472 struct efx_rx_queue
*rx_queue
;
475 EFX_ASSERT_RESET_SERIALISED(efx
);
476 BUG_ON(efx
->port_enabled
);
478 rc
= falcon_flush_queues(efx
);
480 EFX_ERR(efx
, "failed to flush queues\n");
482 EFX_LOG(efx
, "successfully flushed all queues\n");
484 efx_for_each_channel(channel
, efx
) {
485 EFX_LOG(channel
->efx
, "shut down chan %d\n", channel
->channel
);
487 efx_for_each_channel_rx_queue(rx_queue
, channel
)
488 efx_fini_rx_queue(rx_queue
);
489 efx_for_each_channel_tx_queue(tx_queue
, channel
)
490 efx_fini_tx_queue(tx_queue
);
491 efx_fini_eventq(channel
);
495 static void efx_remove_channel(struct efx_channel
*channel
)
497 struct efx_tx_queue
*tx_queue
;
498 struct efx_rx_queue
*rx_queue
;
500 EFX_LOG(channel
->efx
, "destroy chan %d\n", channel
->channel
);
502 efx_for_each_channel_rx_queue(rx_queue
, channel
)
503 efx_remove_rx_queue(rx_queue
);
504 efx_for_each_channel_tx_queue(tx_queue
, channel
)
505 efx_remove_tx_queue(tx_queue
);
506 efx_remove_eventq(channel
);
508 channel
->used_flags
= 0;
511 void efx_schedule_slow_fill(struct efx_rx_queue
*rx_queue
, int delay
)
513 queue_delayed_work(refill_workqueue
, &rx_queue
->work
, delay
);
516 /**************************************************************************
520 **************************************************************************/
522 /* This ensures that the kernel is kept informed (via
523 * netif_carrier_on/off) of the link status, and also maintains the
524 * link status's stop on the port's TX queue.
526 static void efx_link_status_changed(struct efx_nic
*efx
)
528 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
529 * that no events are triggered between unregister_netdev() and the
530 * driver unloading. A more general condition is that NETDEV_CHANGE
531 * can only be generated between NETDEV_UP and NETDEV_DOWN */
532 if (!netif_running(efx
->net_dev
))
535 if (efx
->port_inhibited
) {
536 netif_carrier_off(efx
->net_dev
);
540 if (efx
->link_up
!= netif_carrier_ok(efx
->net_dev
)) {
541 efx
->n_link_state_changes
++;
544 netif_carrier_on(efx
->net_dev
);
546 netif_carrier_off(efx
->net_dev
);
549 /* Status message for kernel log */
551 EFX_INFO(efx
, "link up at %uMbps %s-duplex (MTU %d)%s\n",
552 efx
->link_speed
, efx
->link_fd
? "full" : "half",
554 (efx
->promiscuous
? " [PROMISC]" : ""));
556 EFX_INFO(efx
, "link down\n");
561 /* This call reinitialises the MAC to pick up new PHY settings. The
562 * caller must hold the mac_lock */
563 void __efx_reconfigure_port(struct efx_nic
*efx
)
565 WARN_ON(!mutex_is_locked(&efx
->mac_lock
));
567 EFX_LOG(efx
, "reconfiguring MAC from PHY settings on CPU %d\n",
568 raw_smp_processor_id());
570 /* Serialise the promiscuous flag with efx_set_multicast_list. */
571 if (efx_dev_registered(efx
)) {
572 netif_addr_lock_bh(efx
->net_dev
);
573 netif_addr_unlock_bh(efx
->net_dev
);
576 falcon_deconfigure_mac_wrapper(efx
);
578 /* Reconfigure the PHY, disabling transmit in mac level loopback. */
579 if (LOOPBACK_INTERNAL(efx
))
580 efx
->phy_mode
|= PHY_MODE_TX_DISABLED
;
582 efx
->phy_mode
&= ~PHY_MODE_TX_DISABLED
;
583 efx
->phy_op
->reconfigure(efx
);
585 if (falcon_switch_mac(efx
))
588 efx
->mac_op
->reconfigure(efx
);
590 /* Inform kernel of loss/gain of carrier */
591 efx_link_status_changed(efx
);
595 EFX_ERR(efx
, "failed to reconfigure MAC\n");
596 efx
->phy_op
->fini(efx
);
597 efx
->port_initialized
= false;
600 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
602 void efx_reconfigure_port(struct efx_nic
*efx
)
604 EFX_ASSERT_RESET_SERIALISED(efx
);
606 mutex_lock(&efx
->mac_lock
);
607 __efx_reconfigure_port(efx
);
608 mutex_unlock(&efx
->mac_lock
);
611 /* Asynchronous efx_reconfigure_port work item. To speed up efx_flush_all()
612 * we don't efx_reconfigure_port() if the port is disabled. Care is taken
613 * in efx_stop_all() and efx_start_port() to prevent PHY events being lost */
614 static void efx_phy_work(struct work_struct
*data
)
616 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, phy_work
);
618 mutex_lock(&efx
->mac_lock
);
619 if (efx
->port_enabled
)
620 __efx_reconfigure_port(efx
);
621 mutex_unlock(&efx
->mac_lock
);
624 static void efx_mac_work(struct work_struct
*data
)
626 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, mac_work
);
628 mutex_lock(&efx
->mac_lock
);
629 if (efx
->port_enabled
)
630 efx
->mac_op
->irq(efx
);
631 mutex_unlock(&efx
->mac_lock
);
634 static int efx_probe_port(struct efx_nic
*efx
)
638 EFX_LOG(efx
, "create port\n");
640 /* Connect up MAC/PHY operations table and read MAC address */
641 rc
= falcon_probe_port(efx
);
646 efx
->phy_mode
= PHY_MODE_SPECIAL
;
648 /* Sanity check MAC address */
649 if (is_valid_ether_addr(efx
->mac_address
)) {
650 memcpy(efx
->net_dev
->dev_addr
, efx
->mac_address
, ETH_ALEN
);
652 EFX_ERR(efx
, "invalid MAC address %pM\n",
654 if (!allow_bad_hwaddr
) {
658 random_ether_addr(efx
->net_dev
->dev_addr
);
659 EFX_INFO(efx
, "using locally-generated MAC %pM\n",
660 efx
->net_dev
->dev_addr
);
666 efx_remove_port(efx
);
670 static int efx_init_port(struct efx_nic
*efx
)
674 EFX_LOG(efx
, "init port\n");
676 rc
= efx
->phy_op
->init(efx
);
679 efx
->phy_op
->reconfigure(efx
);
681 mutex_lock(&efx
->mac_lock
);
682 rc
= falcon_switch_mac(efx
);
683 mutex_unlock(&efx
->mac_lock
);
686 efx
->mac_op
->reconfigure(efx
);
688 efx
->port_initialized
= true;
689 efx
->stats_enabled
= true;
693 efx
->phy_op
->fini(efx
);
697 /* Allow efx_reconfigure_port() to be scheduled, and close the window
698 * between efx_stop_port and efx_flush_all whereby a previously scheduled
699 * efx_phy_work()/efx_mac_work() may have been cancelled */
700 static void efx_start_port(struct efx_nic
*efx
)
702 EFX_LOG(efx
, "start port\n");
703 BUG_ON(efx
->port_enabled
);
705 mutex_lock(&efx
->mac_lock
);
706 efx
->port_enabled
= true;
707 __efx_reconfigure_port(efx
);
708 efx
->mac_op
->irq(efx
);
709 mutex_unlock(&efx
->mac_lock
);
712 /* Prevent efx_phy_work, efx_mac_work, and efx_monitor() from executing,
713 * and efx_set_multicast_list() from scheduling efx_phy_work. efx_phy_work
714 * and efx_mac_work may still be scheduled via NAPI processing until
715 * efx_flush_all() is called */
716 static void efx_stop_port(struct efx_nic
*efx
)
718 EFX_LOG(efx
, "stop port\n");
720 mutex_lock(&efx
->mac_lock
);
721 efx
->port_enabled
= false;
722 mutex_unlock(&efx
->mac_lock
);
724 /* Serialise against efx_set_multicast_list() */
725 if (efx_dev_registered(efx
)) {
726 netif_addr_lock_bh(efx
->net_dev
);
727 netif_addr_unlock_bh(efx
->net_dev
);
731 static void efx_fini_port(struct efx_nic
*efx
)
733 EFX_LOG(efx
, "shut down port\n");
735 if (!efx
->port_initialized
)
738 efx
->phy_op
->fini(efx
);
739 efx
->port_initialized
= false;
741 efx
->link_up
= false;
742 efx_link_status_changed(efx
);
745 static void efx_remove_port(struct efx_nic
*efx
)
747 EFX_LOG(efx
, "destroying port\n");
749 falcon_remove_port(efx
);
752 /**************************************************************************
756 **************************************************************************/
758 /* This configures the PCI device to enable I/O and DMA. */
759 static int efx_init_io(struct efx_nic
*efx
)
761 struct pci_dev
*pci_dev
= efx
->pci_dev
;
762 dma_addr_t dma_mask
= efx
->type
->max_dma_mask
;
765 EFX_LOG(efx
, "initialising I/O\n");
767 rc
= pci_enable_device(pci_dev
);
769 EFX_ERR(efx
, "failed to enable PCI device\n");
773 pci_set_master(pci_dev
);
775 /* Set the PCI DMA mask. Try all possibilities from our
776 * genuine mask down to 32 bits, because some architectures
777 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
778 * masks event though they reject 46 bit masks.
780 while (dma_mask
> 0x7fffffffUL
) {
781 if (pci_dma_supported(pci_dev
, dma_mask
) &&
782 ((rc
= pci_set_dma_mask(pci_dev
, dma_mask
)) == 0))
787 EFX_ERR(efx
, "could not find a suitable DMA mask\n");
790 EFX_LOG(efx
, "using DMA mask %llx\n", (unsigned long long) dma_mask
);
791 rc
= pci_set_consistent_dma_mask(pci_dev
, dma_mask
);
793 /* pci_set_consistent_dma_mask() is not *allowed* to
794 * fail with a mask that pci_set_dma_mask() accepted,
795 * but just in case...
797 EFX_ERR(efx
, "failed to set consistent DMA mask\n");
801 efx
->membase_phys
= pci_resource_start(efx
->pci_dev
,
803 rc
= pci_request_region(pci_dev
, efx
->type
->mem_bar
, "sfc");
805 EFX_ERR(efx
, "request for memory BAR failed\n");
809 efx
->membase
= ioremap_nocache(efx
->membase_phys
,
810 efx
->type
->mem_map_size
);
812 EFX_ERR(efx
, "could not map memory BAR %d at %llx+%x\n",
814 (unsigned long long)efx
->membase_phys
,
815 efx
->type
->mem_map_size
);
819 EFX_LOG(efx
, "memory BAR %u at %llx+%x (virtual %p)\n",
820 efx
->type
->mem_bar
, (unsigned long long)efx
->membase_phys
,
821 efx
->type
->mem_map_size
, efx
->membase
);
826 pci_release_region(efx
->pci_dev
, efx
->type
->mem_bar
);
828 efx
->membase_phys
= 0;
830 pci_disable_device(efx
->pci_dev
);
835 static void efx_fini_io(struct efx_nic
*efx
)
837 EFX_LOG(efx
, "shutting down I/O\n");
840 iounmap(efx
->membase
);
844 if (efx
->membase_phys
) {
845 pci_release_region(efx
->pci_dev
, efx
->type
->mem_bar
);
846 efx
->membase_phys
= 0;
849 pci_disable_device(efx
->pci_dev
);
852 /* Get number of RX queues wanted. Return number of online CPU
853 * packages in the expectation that an IRQ balancer will spread
854 * interrupts across them. */
855 static int efx_wanted_rx_queues(void)
861 cpus_clear(core_mask
);
863 for_each_online_cpu(cpu
) {
864 if (!cpu_isset(cpu
, core_mask
)) {
866 cpus_or(core_mask
, core_mask
,
867 topology_core_siblings(cpu
));
874 /* Probe the number and type of interrupts we are able to obtain, and
875 * the resulting numbers of channels and RX queues.
877 static void efx_probe_interrupts(struct efx_nic
*efx
)
880 min_t(int, efx
->type
->phys_addr_channels
, EFX_MAX_CHANNELS
);
883 if (efx
->interrupt_mode
== EFX_INT_MODE_MSIX
) {
884 struct msix_entry xentries
[EFX_MAX_CHANNELS
];
888 /* We want one RX queue and interrupt per CPU package
889 * (or as specified by the rss_cpus module parameter).
890 * We will need one channel per interrupt.
892 rx_queues
= rss_cpus
? rss_cpus
: efx_wanted_rx_queues();
893 wanted_ints
= rx_queues
+ (separate_tx_channels
? 1 : 0);
894 wanted_ints
= min(wanted_ints
, max_channels
);
896 for (i
= 0; i
< wanted_ints
; i
++)
897 xentries
[i
].entry
= i
;
898 rc
= pci_enable_msix(efx
->pci_dev
, xentries
, wanted_ints
);
900 EFX_ERR(efx
, "WARNING: Insufficient MSI-X vectors"
901 " available (%d < %d).\n", rc
, wanted_ints
);
902 EFX_ERR(efx
, "WARNING: Performance may be reduced.\n");
903 EFX_BUG_ON_PARANOID(rc
>= wanted_ints
);
905 rc
= pci_enable_msix(efx
->pci_dev
, xentries
,
910 efx
->n_rx_queues
= min(rx_queues
, wanted_ints
);
911 efx
->n_channels
= wanted_ints
;
912 for (i
= 0; i
< wanted_ints
; i
++)
913 efx
->channel
[i
].irq
= xentries
[i
].vector
;
915 /* Fall back to single channel MSI */
916 efx
->interrupt_mode
= EFX_INT_MODE_MSI
;
917 EFX_ERR(efx
, "could not enable MSI-X\n");
921 /* Try single interrupt MSI */
922 if (efx
->interrupt_mode
== EFX_INT_MODE_MSI
) {
923 efx
->n_rx_queues
= 1;
925 rc
= pci_enable_msi(efx
->pci_dev
);
927 efx
->channel
[0].irq
= efx
->pci_dev
->irq
;
929 EFX_ERR(efx
, "could not enable MSI\n");
930 efx
->interrupt_mode
= EFX_INT_MODE_LEGACY
;
934 /* Assume legacy interrupts */
935 if (efx
->interrupt_mode
== EFX_INT_MODE_LEGACY
) {
936 efx
->n_rx_queues
= 1;
937 efx
->n_channels
= 1 + (separate_tx_channels
? 1 : 0);
938 efx
->legacy_irq
= efx
->pci_dev
->irq
;
942 static void efx_remove_interrupts(struct efx_nic
*efx
)
944 struct efx_channel
*channel
;
946 /* Remove MSI/MSI-X interrupts */
947 efx_for_each_channel(channel
, efx
)
949 pci_disable_msi(efx
->pci_dev
);
950 pci_disable_msix(efx
->pci_dev
);
952 /* Remove legacy interrupt */
956 static void efx_set_channels(struct efx_nic
*efx
)
958 struct efx_tx_queue
*tx_queue
;
959 struct efx_rx_queue
*rx_queue
;
961 efx_for_each_tx_queue(tx_queue
, efx
) {
962 if (separate_tx_channels
)
963 tx_queue
->channel
= &efx
->channel
[efx
->n_channels
-1];
965 tx_queue
->channel
= &efx
->channel
[0];
966 tx_queue
->channel
->used_flags
|= EFX_USED_BY_TX
;
969 efx_for_each_rx_queue(rx_queue
, efx
) {
970 rx_queue
->channel
= &efx
->channel
[rx_queue
->queue
];
971 rx_queue
->channel
->used_flags
|= EFX_USED_BY_RX
;
975 static int efx_probe_nic(struct efx_nic
*efx
)
979 EFX_LOG(efx
, "creating NIC\n");
981 /* Carry out hardware-type specific initialisation */
982 rc
= falcon_probe_nic(efx
);
986 /* Determine the number of channels and RX queues by trying to hook
987 * in MSI-X interrupts. */
988 efx_probe_interrupts(efx
);
990 efx_set_channels(efx
);
992 /* Initialise the interrupt moderation settings */
993 efx_init_irq_moderation(efx
, tx_irq_mod_usec
, rx_irq_mod_usec
);
998 static void efx_remove_nic(struct efx_nic
*efx
)
1000 EFX_LOG(efx
, "destroying NIC\n");
1002 efx_remove_interrupts(efx
);
1003 falcon_remove_nic(efx
);
1006 /**************************************************************************
1008 * NIC startup/shutdown
1010 *************************************************************************/
1012 static int efx_probe_all(struct efx_nic
*efx
)
1014 struct efx_channel
*channel
;
1018 rc
= efx_probe_nic(efx
);
1020 EFX_ERR(efx
, "failed to create NIC\n");
1025 rc
= efx_probe_port(efx
);
1027 EFX_ERR(efx
, "failed to create port\n");
1031 /* Create channels */
1032 efx_for_each_channel(channel
, efx
) {
1033 rc
= efx_probe_channel(channel
);
1035 EFX_ERR(efx
, "failed to create channel %d\n",
1040 efx_set_channel_names(efx
);
1045 efx_for_each_channel(channel
, efx
)
1046 efx_remove_channel(channel
);
1047 efx_remove_port(efx
);
1049 efx_remove_nic(efx
);
1054 /* Called after previous invocation(s) of efx_stop_all, restarts the
1055 * port, kernel transmit queue, NAPI processing and hardware interrupts,
1056 * and ensures that the port is scheduled to be reconfigured.
1057 * This function is safe to call multiple times when the NIC is in any
1059 static void efx_start_all(struct efx_nic
*efx
)
1061 struct efx_channel
*channel
;
1063 EFX_ASSERT_RESET_SERIALISED(efx
);
1065 /* Check that it is appropriate to restart the interface. All
1066 * of these flags are safe to read under just the rtnl lock */
1067 if (efx
->port_enabled
)
1069 if ((efx
->state
!= STATE_RUNNING
) && (efx
->state
!= STATE_INIT
))
1071 if (efx_dev_registered(efx
) && !netif_running(efx
->net_dev
))
1074 /* Mark the port as enabled so port reconfigurations can start, then
1075 * restart the transmit interface early so the watchdog timer stops */
1076 efx_start_port(efx
);
1077 if (efx_dev_registered(efx
))
1078 efx_wake_queue(efx
);
1080 efx_for_each_channel(channel
, efx
)
1081 efx_start_channel(channel
);
1083 falcon_enable_interrupts(efx
);
1085 /* Start hardware monitor if we're in RUNNING */
1086 if (efx
->state
== STATE_RUNNING
)
1087 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1088 efx_monitor_interval
);
1091 /* Flush all delayed work. Should only be called when no more delayed work
1092 * will be scheduled. This doesn't flush pending online resets (efx_reset),
1093 * since we're holding the rtnl_lock at this point. */
1094 static void efx_flush_all(struct efx_nic
*efx
)
1096 struct efx_rx_queue
*rx_queue
;
1098 /* Make sure the hardware monitor is stopped */
1099 cancel_delayed_work_sync(&efx
->monitor_work
);
1101 /* Ensure that all RX slow refills are complete. */
1102 efx_for_each_rx_queue(rx_queue
, efx
)
1103 cancel_delayed_work_sync(&rx_queue
->work
);
1105 /* Stop scheduled port reconfigurations */
1106 cancel_work_sync(&efx
->mac_work
);
1107 cancel_work_sync(&efx
->phy_work
);
1111 /* Quiesce hardware and software without bringing the link down.
1112 * Safe to call multiple times, when the nic and interface is in any
1113 * state. The caller is guaranteed to subsequently be in a position
1114 * to modify any hardware and software state they see fit without
1116 static void efx_stop_all(struct efx_nic
*efx
)
1118 struct efx_channel
*channel
;
1120 EFX_ASSERT_RESET_SERIALISED(efx
);
1122 /* port_enabled can be read safely under the rtnl lock */
1123 if (!efx
->port_enabled
)
1126 /* Disable interrupts and wait for ISR to complete */
1127 falcon_disable_interrupts(efx
);
1128 if (efx
->legacy_irq
)
1129 synchronize_irq(efx
->legacy_irq
);
1130 efx_for_each_channel(channel
, efx
) {
1132 synchronize_irq(channel
->irq
);
1135 /* Stop all NAPI processing and synchronous rx refills */
1136 efx_for_each_channel(channel
, efx
)
1137 efx_stop_channel(channel
);
1139 /* Stop all asynchronous port reconfigurations. Since all
1140 * event processing has already been stopped, there is no
1141 * window to loose phy events */
1144 /* Flush efx_phy_work, efx_mac_work, refill_workqueue, monitor_work */
1147 /* Isolate the MAC from the TX and RX engines, so that queue
1148 * flushes will complete in a timely fashion. */
1149 falcon_drain_tx_fifo(efx
);
1151 /* Stop the kernel transmit interface late, so the watchdog
1152 * timer isn't ticking over the flush */
1153 if (efx_dev_registered(efx
)) {
1154 efx_stop_queue(efx
);
1155 netif_tx_lock_bh(efx
->net_dev
);
1156 netif_tx_unlock_bh(efx
->net_dev
);
1160 static void efx_remove_all(struct efx_nic
*efx
)
1162 struct efx_channel
*channel
;
1164 efx_for_each_channel(channel
, efx
)
1165 efx_remove_channel(channel
);
1166 efx_remove_port(efx
);
1167 efx_remove_nic(efx
);
1170 /* A convinience function to safely flush all the queues */
1171 void efx_flush_queues(struct efx_nic
*efx
)
1173 EFX_ASSERT_RESET_SERIALISED(efx
);
1177 efx_fini_channels(efx
);
1178 efx_init_channels(efx
);
1183 /**************************************************************************
1185 * Interrupt moderation
1187 **************************************************************************/
1189 /* Set interrupt moderation parameters */
1190 void efx_init_irq_moderation(struct efx_nic
*efx
, int tx_usecs
, int rx_usecs
)
1192 struct efx_tx_queue
*tx_queue
;
1193 struct efx_rx_queue
*rx_queue
;
1195 EFX_ASSERT_RESET_SERIALISED(efx
);
1197 efx_for_each_tx_queue(tx_queue
, efx
)
1198 tx_queue
->channel
->irq_moderation
= tx_usecs
;
1200 efx_for_each_rx_queue(rx_queue
, efx
)
1201 rx_queue
->channel
->irq_moderation
= rx_usecs
;
1204 /**************************************************************************
1208 **************************************************************************/
1210 /* Run periodically off the general workqueue. Serialised against
1211 * efx_reconfigure_port via the mac_lock */
1212 static void efx_monitor(struct work_struct
*data
)
1214 struct efx_nic
*efx
= container_of(data
, struct efx_nic
,
1218 EFX_TRACE(efx
, "hardware monitor executing on CPU %d\n",
1219 raw_smp_processor_id());
1221 /* If the mac_lock is already held then it is likely a port
1222 * reconfiguration is already in place, which will likely do
1223 * most of the work of check_hw() anyway. */
1224 if (!mutex_trylock(&efx
->mac_lock
))
1226 if (!efx
->port_enabled
)
1228 rc
= efx
->board_info
.monitor(efx
);
1230 EFX_ERR(efx
, "Board sensor %s; shutting down PHY\n",
1231 (rc
== -ERANGE
) ? "reported fault" : "failed");
1232 efx
->phy_mode
|= PHY_MODE_LOW_POWER
;
1233 falcon_sim_phy_event(efx
);
1235 efx
->phy_op
->poll(efx
);
1236 efx
->mac_op
->poll(efx
);
1239 mutex_unlock(&efx
->mac_lock
);
1241 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1242 efx_monitor_interval
);
1245 /**************************************************************************
1249 *************************************************************************/
1252 * Context: process, rtnl_lock() held.
1254 static int efx_ioctl(struct net_device
*net_dev
, struct ifreq
*ifr
, int cmd
)
1256 struct efx_nic
*efx
= netdev_priv(net_dev
);
1258 EFX_ASSERT_RESET_SERIALISED(efx
);
1260 return generic_mii_ioctl(&efx
->mii
, if_mii(ifr
), cmd
, NULL
);
1263 /**************************************************************************
1267 **************************************************************************/
1269 static int efx_init_napi(struct efx_nic
*efx
)
1271 struct efx_channel
*channel
;
1274 efx_for_each_channel(channel
, efx
) {
1275 channel
->napi_dev
= efx
->net_dev
;
1276 rc
= efx_lro_init(&channel
->lro_mgr
, efx
);
1286 static void efx_fini_napi(struct efx_nic
*efx
)
1288 struct efx_channel
*channel
;
1290 efx_for_each_channel(channel
, efx
) {
1291 efx_lro_fini(&channel
->lro_mgr
);
1292 channel
->napi_dev
= NULL
;
1296 /**************************************************************************
1298 * Kernel netpoll interface
1300 *************************************************************************/
1302 #ifdef CONFIG_NET_POLL_CONTROLLER
1304 /* Although in the common case interrupts will be disabled, this is not
1305 * guaranteed. However, all our work happens inside the NAPI callback,
1306 * so no locking is required.
1308 static void efx_netpoll(struct net_device
*net_dev
)
1310 struct efx_nic
*efx
= netdev_priv(net_dev
);
1311 struct efx_channel
*channel
;
1313 efx_for_each_channel(channel
, efx
)
1314 efx_schedule_channel(channel
);
1319 /**************************************************************************
1321 * Kernel net device interface
1323 *************************************************************************/
1325 /* Context: process, rtnl_lock() held. */
1326 static int efx_net_open(struct net_device
*net_dev
)
1328 struct efx_nic
*efx
= netdev_priv(net_dev
);
1329 EFX_ASSERT_RESET_SERIALISED(efx
);
1331 EFX_LOG(efx
, "opening device %s on CPU %d\n", net_dev
->name
,
1332 raw_smp_processor_id());
1334 if (efx
->state
== STATE_DISABLED
)
1336 if (efx
->phy_mode
& PHY_MODE_SPECIAL
)
1343 /* Context: process, rtnl_lock() held.
1344 * Note that the kernel will ignore our return code; this method
1345 * should really be a void.
1347 static int efx_net_stop(struct net_device
*net_dev
)
1349 struct efx_nic
*efx
= netdev_priv(net_dev
);
1351 EFX_LOG(efx
, "closing %s on CPU %d\n", net_dev
->name
,
1352 raw_smp_processor_id());
1354 if (efx
->state
!= STATE_DISABLED
) {
1355 /* Stop the device and flush all the channels */
1357 efx_fini_channels(efx
);
1358 efx_init_channels(efx
);
1364 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1365 static struct net_device_stats
*efx_net_stats(struct net_device
*net_dev
)
1367 struct efx_nic
*efx
= netdev_priv(net_dev
);
1368 struct efx_mac_stats
*mac_stats
= &efx
->mac_stats
;
1369 struct net_device_stats
*stats
= &net_dev
->stats
;
1371 /* Update stats if possible, but do not wait if another thread
1372 * is updating them (or resetting the NIC); slightly stale
1373 * stats are acceptable.
1375 if (!spin_trylock(&efx
->stats_lock
))
1377 if (efx
->stats_enabled
) {
1378 efx
->mac_op
->update_stats(efx
);
1379 falcon_update_nic_stats(efx
);
1381 spin_unlock(&efx
->stats_lock
);
1383 stats
->rx_packets
= mac_stats
->rx_packets
;
1384 stats
->tx_packets
= mac_stats
->tx_packets
;
1385 stats
->rx_bytes
= mac_stats
->rx_bytes
;
1386 stats
->tx_bytes
= mac_stats
->tx_bytes
;
1387 stats
->multicast
= mac_stats
->rx_multicast
;
1388 stats
->collisions
= mac_stats
->tx_collision
;
1389 stats
->rx_length_errors
= (mac_stats
->rx_gtjumbo
+
1390 mac_stats
->rx_length_error
);
1391 stats
->rx_over_errors
= efx
->n_rx_nodesc_drop_cnt
;
1392 stats
->rx_crc_errors
= mac_stats
->rx_bad
;
1393 stats
->rx_frame_errors
= mac_stats
->rx_align_error
;
1394 stats
->rx_fifo_errors
= mac_stats
->rx_overflow
;
1395 stats
->rx_missed_errors
= mac_stats
->rx_missed
;
1396 stats
->tx_window_errors
= mac_stats
->tx_late_collision
;
1398 stats
->rx_errors
= (stats
->rx_length_errors
+
1399 stats
->rx_over_errors
+
1400 stats
->rx_crc_errors
+
1401 stats
->rx_frame_errors
+
1402 stats
->rx_fifo_errors
+
1403 stats
->rx_missed_errors
+
1404 mac_stats
->rx_symbol_error
);
1405 stats
->tx_errors
= (stats
->tx_window_errors
+
1411 /* Context: netif_tx_lock held, BHs disabled. */
1412 static void efx_watchdog(struct net_device
*net_dev
)
1414 struct efx_nic
*efx
= netdev_priv(net_dev
);
1416 EFX_ERR(efx
, "TX stuck with stop_count=%d port_enabled=%d:"
1417 " resetting channels\n",
1418 atomic_read(&efx
->netif_stop_count
), efx
->port_enabled
);
1420 efx_schedule_reset(efx
, RESET_TYPE_TX_WATCHDOG
);
1424 /* Context: process, rtnl_lock() held. */
1425 static int efx_change_mtu(struct net_device
*net_dev
, int new_mtu
)
1427 struct efx_nic
*efx
= netdev_priv(net_dev
);
1430 EFX_ASSERT_RESET_SERIALISED(efx
);
1432 if (new_mtu
> EFX_MAX_MTU
)
1437 EFX_LOG(efx
, "changing MTU to %d\n", new_mtu
);
1439 efx_fini_channels(efx
);
1440 net_dev
->mtu
= new_mtu
;
1441 efx_init_channels(efx
);
1447 static int efx_set_mac_address(struct net_device
*net_dev
, void *data
)
1449 struct efx_nic
*efx
= netdev_priv(net_dev
);
1450 struct sockaddr
*addr
= data
;
1451 char *new_addr
= addr
->sa_data
;
1453 EFX_ASSERT_RESET_SERIALISED(efx
);
1455 if (!is_valid_ether_addr(new_addr
)) {
1456 EFX_ERR(efx
, "invalid ethernet MAC address requested: %pM\n",
1461 memcpy(net_dev
->dev_addr
, new_addr
, net_dev
->addr_len
);
1463 /* Reconfigure the MAC */
1464 efx_reconfigure_port(efx
);
1469 /* Context: netif_addr_lock held, BHs disabled. */
1470 static void efx_set_multicast_list(struct net_device
*net_dev
)
1472 struct efx_nic
*efx
= netdev_priv(net_dev
);
1473 struct dev_mc_list
*mc_list
= net_dev
->mc_list
;
1474 union efx_multicast_hash
*mc_hash
= &efx
->multicast_hash
;
1475 bool promiscuous
= !!(net_dev
->flags
& IFF_PROMISC
);
1476 bool changed
= (efx
->promiscuous
!= promiscuous
);
1481 efx
->promiscuous
= promiscuous
;
1483 /* Build multicast hash table */
1484 if (promiscuous
|| (net_dev
->flags
& IFF_ALLMULTI
)) {
1485 memset(mc_hash
, 0xff, sizeof(*mc_hash
));
1487 memset(mc_hash
, 0x00, sizeof(*mc_hash
));
1488 for (i
= 0; i
< net_dev
->mc_count
; i
++) {
1489 crc
= ether_crc_le(ETH_ALEN
, mc_list
->dmi_addr
);
1490 bit
= crc
& (EFX_MCAST_HASH_ENTRIES
- 1);
1491 set_bit_le(bit
, mc_hash
->byte
);
1492 mc_list
= mc_list
->next
;
1496 if (!efx
->port_enabled
)
1497 /* Delay pushing settings until efx_start_port() */
1501 queue_work(efx
->workqueue
, &efx
->phy_work
);
1503 /* Create and activate new global multicast hash table */
1504 falcon_set_multicast_hash(efx
);
1507 static const struct net_device_ops efx_netdev_ops
= {
1508 .ndo_open
= efx_net_open
,
1509 .ndo_stop
= efx_net_stop
,
1510 .ndo_get_stats
= efx_net_stats
,
1511 .ndo_tx_timeout
= efx_watchdog
,
1512 .ndo_start_xmit
= efx_hard_start_xmit
,
1513 .ndo_validate_addr
= eth_validate_addr
,
1514 .ndo_do_ioctl
= efx_ioctl
,
1515 .ndo_change_mtu
= efx_change_mtu
,
1516 .ndo_set_mac_address
= efx_set_mac_address
,
1517 .ndo_set_multicast_list
= efx_set_multicast_list
,
1518 #ifdef CONFIG_NET_POLL_CONTROLLER
1519 .ndo_poll_controller
= efx_netpoll
,
1523 static void efx_update_name(struct efx_nic
*efx
)
1525 strcpy(efx
->name
, efx
->net_dev
->name
);
1526 efx_mtd_rename(efx
);
1527 efx_set_channel_names(efx
);
1530 static int efx_netdev_event(struct notifier_block
*this,
1531 unsigned long event
, void *ptr
)
1533 struct net_device
*net_dev
= ptr
;
1535 if (net_dev
->netdev_ops
== &efx_netdev_ops
&&
1536 event
== NETDEV_CHANGENAME
)
1537 efx_update_name(netdev_priv(net_dev
));
1542 static struct notifier_block efx_netdev_notifier
= {
1543 .notifier_call
= efx_netdev_event
,
1547 show_phy_type(struct device
*dev
, struct device_attribute
*attr
, char *buf
)
1549 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
1550 return sprintf(buf
, "%d\n", efx
->phy_type
);
1552 static DEVICE_ATTR(phy_type
, 0644, show_phy_type
, NULL
);
1554 static int efx_register_netdev(struct efx_nic
*efx
)
1556 struct net_device
*net_dev
= efx
->net_dev
;
1559 net_dev
->watchdog_timeo
= 5 * HZ
;
1560 net_dev
->irq
= efx
->pci_dev
->irq
;
1561 net_dev
->netdev_ops
= &efx_netdev_ops
;
1562 SET_NETDEV_DEV(net_dev
, &efx
->pci_dev
->dev
);
1563 SET_ETHTOOL_OPS(net_dev
, &efx_ethtool_ops
);
1565 /* Always start with carrier off; PHY events will detect the link */
1566 netif_carrier_off(efx
->net_dev
);
1568 /* Clear MAC statistics */
1569 efx
->mac_op
->update_stats(efx
);
1570 memset(&efx
->mac_stats
, 0, sizeof(efx
->mac_stats
));
1572 rc
= register_netdev(net_dev
);
1574 EFX_ERR(efx
, "could not register net dev\n");
1579 efx_update_name(efx
);
1582 rc
= device_create_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1584 EFX_ERR(efx
, "failed to init net dev attributes\n");
1585 goto fail_registered
;
1591 unregister_netdev(net_dev
);
1595 static void efx_unregister_netdev(struct efx_nic
*efx
)
1597 struct efx_tx_queue
*tx_queue
;
1602 BUG_ON(netdev_priv(efx
->net_dev
) != efx
);
1604 /* Free up any skbs still remaining. This has to happen before
1605 * we try to unregister the netdev as running their destructors
1606 * may be needed to get the device ref. count to 0. */
1607 efx_for_each_tx_queue(tx_queue
, efx
)
1608 efx_release_tx_buffers(tx_queue
);
1610 if (efx_dev_registered(efx
)) {
1611 strlcpy(efx
->name
, pci_name(efx
->pci_dev
), sizeof(efx
->name
));
1612 device_remove_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1613 unregister_netdev(efx
->net_dev
);
1617 /**************************************************************************
1619 * Device reset and suspend
1621 **************************************************************************/
1623 /* Tears down the entire software state and most of the hardware state
1625 void efx_reset_down(struct efx_nic
*efx
, struct ethtool_cmd
*ecmd
)
1627 EFX_ASSERT_RESET_SERIALISED(efx
);
1629 /* The net_dev->get_stats handler is quite slow, and will fail
1630 * if a fetch is pending over reset. Serialise against it. */
1631 spin_lock(&efx
->stats_lock
);
1632 efx
->stats_enabled
= false;
1633 spin_unlock(&efx
->stats_lock
);
1636 mutex_lock(&efx
->mac_lock
);
1637 mutex_lock(&efx
->spi_lock
);
1639 efx
->phy_op
->get_settings(efx
, ecmd
);
1641 efx_fini_channels(efx
);
1644 /* This function will always ensure that the locks acquired in
1645 * efx_reset_down() are released. A failure return code indicates
1646 * that we were unable to reinitialise the hardware, and the
1647 * driver should be disabled. If ok is false, then the rx and tx
1648 * engines are not restarted, pending a RESET_DISABLE. */
1649 int efx_reset_up(struct efx_nic
*efx
, struct ethtool_cmd
*ecmd
, bool ok
)
1653 EFX_ASSERT_RESET_SERIALISED(efx
);
1655 rc
= falcon_init_nic(efx
);
1657 EFX_ERR(efx
, "failed to initialise NIC\n");
1662 efx_init_channels(efx
);
1664 if (efx
->phy_op
->set_settings(efx
, ecmd
))
1665 EFX_ERR(efx
, "could not restore PHY settings\n");
1668 mutex_unlock(&efx
->spi_lock
);
1669 mutex_unlock(&efx
->mac_lock
);
1673 efx
->stats_enabled
= true;
1678 /* Reset the NIC as transparently as possible. Do not reset the PHY
1679 * Note that the reset may fail, in which case the card will be left
1680 * in a most-probably-unusable state.
1682 * This function will sleep. You cannot reset from within an atomic
1683 * state; use efx_schedule_reset() instead.
1685 * Grabs the rtnl_lock.
1687 static int efx_reset(struct efx_nic
*efx
)
1689 struct ethtool_cmd ecmd
;
1690 enum reset_type method
= efx
->reset_pending
;
1693 /* Serialise with kernel interfaces */
1696 /* If we're not RUNNING then don't reset. Leave the reset_pending
1697 * flag set so that efx_pci_probe_main will be retried */
1698 if (efx
->state
!= STATE_RUNNING
) {
1699 EFX_INFO(efx
, "scheduled reset quenched. NIC not RUNNING\n");
1703 EFX_INFO(efx
, "resetting (%d)\n", method
);
1705 efx_reset_down(efx
, &ecmd
);
1707 rc
= falcon_reset_hw(efx
, method
);
1709 EFX_ERR(efx
, "failed to reset hardware\n");
1713 /* Allow resets to be rescheduled. */
1714 efx
->reset_pending
= RESET_TYPE_NONE
;
1716 /* Reinitialise bus-mastering, which may have been turned off before
1717 * the reset was scheduled. This is still appropriate, even in the
1718 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
1719 * can respond to requests. */
1720 pci_set_master(efx
->pci_dev
);
1722 /* Leave device stopped if necessary */
1723 if (method
== RESET_TYPE_DISABLE
) {
1724 efx_reset_up(efx
, &ecmd
, false);
1727 rc
= efx_reset_up(efx
, &ecmd
, true);
1732 EFX_ERR(efx
, "has been disabled\n");
1733 efx
->state
= STATE_DISABLED
;
1734 dev_close(efx
->net_dev
);
1736 EFX_LOG(efx
, "reset complete\n");
1744 /* The worker thread exists so that code that cannot sleep can
1745 * schedule a reset for later.
1747 static void efx_reset_work(struct work_struct
*data
)
1749 struct efx_nic
*nic
= container_of(data
, struct efx_nic
, reset_work
);
1754 void efx_schedule_reset(struct efx_nic
*efx
, enum reset_type type
)
1756 enum reset_type method
;
1758 if (efx
->reset_pending
!= RESET_TYPE_NONE
) {
1759 EFX_INFO(efx
, "quenching already scheduled reset\n");
1764 case RESET_TYPE_INVISIBLE
:
1765 case RESET_TYPE_ALL
:
1766 case RESET_TYPE_WORLD
:
1767 case RESET_TYPE_DISABLE
:
1770 case RESET_TYPE_RX_RECOVERY
:
1771 case RESET_TYPE_RX_DESC_FETCH
:
1772 case RESET_TYPE_TX_DESC_FETCH
:
1773 case RESET_TYPE_TX_SKIP
:
1774 method
= RESET_TYPE_INVISIBLE
;
1777 method
= RESET_TYPE_ALL
;
1782 EFX_LOG(efx
, "scheduling reset (%d:%d)\n", type
, method
);
1784 EFX_LOG(efx
, "scheduling reset (%d)\n", method
);
1786 efx
->reset_pending
= method
;
1788 queue_work(reset_workqueue
, &efx
->reset_work
);
1791 /**************************************************************************
1793 * List of NICs we support
1795 **************************************************************************/
1797 /* PCI device ID table */
1798 static struct pci_device_id efx_pci_table
[] __devinitdata
= {
1799 {PCI_DEVICE(EFX_VENDID_SFC
, FALCON_A_P_DEVID
),
1800 .driver_data
= (unsigned long) &falcon_a_nic_type
},
1801 {PCI_DEVICE(EFX_VENDID_SFC
, FALCON_B_P_DEVID
),
1802 .driver_data
= (unsigned long) &falcon_b_nic_type
},
1803 {0} /* end of list */
1806 /**************************************************************************
1808 * Dummy PHY/MAC/Board operations
1810 * Can be used for some unimplemented operations
1811 * Needed so all function pointers are valid and do not have to be tested
1814 **************************************************************************/
1815 int efx_port_dummy_op_int(struct efx_nic
*efx
)
1819 void efx_port_dummy_op_void(struct efx_nic
*efx
) {}
1820 void efx_port_dummy_op_blink(struct efx_nic
*efx
, bool blink
) {}
1822 static struct efx_mac_operations efx_dummy_mac_operations
= {
1823 .reconfigure
= efx_port_dummy_op_void
,
1824 .poll
= efx_port_dummy_op_void
,
1825 .irq
= efx_port_dummy_op_void
,
1828 static struct efx_phy_operations efx_dummy_phy_operations
= {
1829 .init
= efx_port_dummy_op_int
,
1830 .reconfigure
= efx_port_dummy_op_void
,
1831 .poll
= efx_port_dummy_op_void
,
1832 .fini
= efx_port_dummy_op_void
,
1833 .clear_interrupt
= efx_port_dummy_op_void
,
1836 static struct efx_board efx_dummy_board_info
= {
1837 .init
= efx_port_dummy_op_int
,
1838 .init_leds
= efx_port_dummy_op_int
,
1839 .set_fault_led
= efx_port_dummy_op_blink
,
1840 .monitor
= efx_port_dummy_op_int
,
1841 .blink
= efx_port_dummy_op_blink
,
1842 .fini
= efx_port_dummy_op_void
,
1845 /**************************************************************************
1849 **************************************************************************/
1851 /* This zeroes out and then fills in the invariants in a struct
1852 * efx_nic (including all sub-structures).
1854 static int efx_init_struct(struct efx_nic
*efx
, struct efx_nic_type
*type
,
1855 struct pci_dev
*pci_dev
, struct net_device
*net_dev
)
1857 struct efx_channel
*channel
;
1858 struct efx_tx_queue
*tx_queue
;
1859 struct efx_rx_queue
*rx_queue
;
1862 /* Initialise common structures */
1863 memset(efx
, 0, sizeof(*efx
));
1864 spin_lock_init(&efx
->biu_lock
);
1865 spin_lock_init(&efx
->phy_lock
);
1866 mutex_init(&efx
->spi_lock
);
1867 INIT_WORK(&efx
->reset_work
, efx_reset_work
);
1868 INIT_DELAYED_WORK(&efx
->monitor_work
, efx_monitor
);
1869 efx
->pci_dev
= pci_dev
;
1870 efx
->state
= STATE_INIT
;
1871 efx
->reset_pending
= RESET_TYPE_NONE
;
1872 strlcpy(efx
->name
, pci_name(pci_dev
), sizeof(efx
->name
));
1873 efx
->board_info
= efx_dummy_board_info
;
1875 efx
->net_dev
= net_dev
;
1876 efx
->rx_checksum_enabled
= true;
1877 spin_lock_init(&efx
->netif_stop_lock
);
1878 spin_lock_init(&efx
->stats_lock
);
1879 mutex_init(&efx
->mac_lock
);
1880 efx
->mac_op
= &efx_dummy_mac_operations
;
1881 efx
->phy_op
= &efx_dummy_phy_operations
;
1882 efx
->mii
.dev
= net_dev
;
1883 INIT_WORK(&efx
->phy_work
, efx_phy_work
);
1884 INIT_WORK(&efx
->mac_work
, efx_mac_work
);
1885 atomic_set(&efx
->netif_stop_count
, 1);
1887 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++) {
1888 channel
= &efx
->channel
[i
];
1890 channel
->channel
= i
;
1891 channel
->work_pending
= false;
1893 for (i
= 0; i
< EFX_TX_QUEUE_COUNT
; i
++) {
1894 tx_queue
= &efx
->tx_queue
[i
];
1895 tx_queue
->efx
= efx
;
1896 tx_queue
->queue
= i
;
1897 tx_queue
->buffer
= NULL
;
1898 tx_queue
->channel
= &efx
->channel
[0]; /* for safety */
1899 tx_queue
->tso_headers_free
= NULL
;
1901 for (i
= 0; i
< EFX_MAX_RX_QUEUES
; i
++) {
1902 rx_queue
= &efx
->rx_queue
[i
];
1903 rx_queue
->efx
= efx
;
1904 rx_queue
->queue
= i
;
1905 rx_queue
->channel
= &efx
->channel
[0]; /* for safety */
1906 rx_queue
->buffer
= NULL
;
1907 spin_lock_init(&rx_queue
->add_lock
);
1908 INIT_DELAYED_WORK(&rx_queue
->work
, efx_rx_work
);
1913 /* Sanity-check NIC type */
1914 EFX_BUG_ON_PARANOID(efx
->type
->txd_ring_mask
&
1915 (efx
->type
->txd_ring_mask
+ 1));
1916 EFX_BUG_ON_PARANOID(efx
->type
->rxd_ring_mask
&
1917 (efx
->type
->rxd_ring_mask
+ 1));
1918 EFX_BUG_ON_PARANOID(efx
->type
->evq_size
&
1919 (efx
->type
->evq_size
- 1));
1920 /* As close as we can get to guaranteeing that we don't overflow */
1921 EFX_BUG_ON_PARANOID(efx
->type
->evq_size
<
1922 (efx
->type
->txd_ring_mask
+ 1 +
1923 efx
->type
->rxd_ring_mask
+ 1));
1924 EFX_BUG_ON_PARANOID(efx
->type
->phys_addr_channels
> EFX_MAX_CHANNELS
);
1926 /* Higher numbered interrupt modes are less capable! */
1927 efx
->interrupt_mode
= max(efx
->type
->max_interrupt_mode
,
1930 /* Would be good to use the net_dev name, but we're too early */
1931 snprintf(efx
->workqueue_name
, sizeof(efx
->workqueue_name
), "sfc%s",
1933 efx
->workqueue
= create_singlethread_workqueue(efx
->workqueue_name
);
1934 if (!efx
->workqueue
)
1940 static void efx_fini_struct(struct efx_nic
*efx
)
1942 if (efx
->workqueue
) {
1943 destroy_workqueue(efx
->workqueue
);
1944 efx
->workqueue
= NULL
;
1948 /**************************************************************************
1952 **************************************************************************/
1954 /* Main body of final NIC shutdown code
1955 * This is called only at module unload (or hotplug removal).
1957 static void efx_pci_remove_main(struct efx_nic
*efx
)
1959 EFX_ASSERT_RESET_SERIALISED(efx
);
1961 /* Skip everything if we never obtained a valid membase */
1965 efx_fini_channels(efx
);
1968 /* Shutdown the board, then the NIC and board state */
1969 efx
->board_info
.fini(efx
);
1970 falcon_fini_interrupt(efx
);
1973 efx_remove_all(efx
);
1976 /* Final NIC shutdown
1977 * This is called only at module unload (or hotplug removal).
1979 static void efx_pci_remove(struct pci_dev
*pci_dev
)
1981 struct efx_nic
*efx
;
1983 efx
= pci_get_drvdata(pci_dev
);
1987 /* Mark the NIC as fini, then stop the interface */
1989 efx
->state
= STATE_FINI
;
1990 dev_close(efx
->net_dev
);
1992 /* Allow any queued efx_resets() to complete */
1995 if (efx
->membase
== NULL
)
1998 efx_unregister_netdev(efx
);
2000 efx_mtd_remove(efx
);
2002 /* Wait for any scheduled resets to complete. No more will be
2003 * scheduled from this point because efx_stop_all() has been
2004 * called, we are no longer registered with driverlink, and
2005 * the net_device's have been removed. */
2006 cancel_work_sync(&efx
->reset_work
);
2008 efx_pci_remove_main(efx
);
2012 EFX_LOG(efx
, "shutdown successful\n");
2014 pci_set_drvdata(pci_dev
, NULL
);
2015 efx_fini_struct(efx
);
2016 free_netdev(efx
->net_dev
);
2019 /* Main body of NIC initialisation
2020 * This is called at module load (or hotplug insertion, theoretically).
2022 static int efx_pci_probe_main(struct efx_nic
*efx
)
2026 /* Do start-of-day initialisation */
2027 rc
= efx_probe_all(efx
);
2031 rc
= efx_init_napi(efx
);
2035 /* Initialise the board */
2036 rc
= efx
->board_info
.init(efx
);
2038 EFX_ERR(efx
, "failed to initialise board\n");
2042 rc
= falcon_init_nic(efx
);
2044 EFX_ERR(efx
, "failed to initialise NIC\n");
2048 rc
= efx_init_port(efx
);
2050 EFX_ERR(efx
, "failed to initialise port\n");
2054 efx_init_channels(efx
);
2056 rc
= falcon_init_interrupt(efx
);
2063 efx_fini_channels(efx
);
2067 efx
->board_info
.fini(efx
);
2071 efx_remove_all(efx
);
2076 /* NIC initialisation
2078 * This is called at module load (or hotplug insertion,
2079 * theoretically). It sets up PCI mappings, tests and resets the NIC,
2080 * sets up and registers the network devices with the kernel and hooks
2081 * the interrupt service routine. It does not prepare the device for
2082 * transmission; this is left to the first time one of the network
2083 * interfaces is brought up (i.e. efx_net_open).
2085 static int __devinit
efx_pci_probe(struct pci_dev
*pci_dev
,
2086 const struct pci_device_id
*entry
)
2088 struct efx_nic_type
*type
= (struct efx_nic_type
*) entry
->driver_data
;
2089 struct net_device
*net_dev
;
2090 struct efx_nic
*efx
;
2093 /* Allocate and initialise a struct net_device and struct efx_nic */
2094 net_dev
= alloc_etherdev(sizeof(*efx
));
2097 net_dev
->features
|= (NETIF_F_IP_CSUM
| NETIF_F_SG
|
2098 NETIF_F_HIGHDMA
| NETIF_F_TSO
);
2100 net_dev
->features
|= NETIF_F_LRO
;
2101 /* Mask for features that also apply to VLAN devices */
2102 net_dev
->vlan_features
|= (NETIF_F_ALL_CSUM
| NETIF_F_SG
|
2103 NETIF_F_HIGHDMA
| NETIF_F_TSO
);
2104 efx
= netdev_priv(net_dev
);
2105 pci_set_drvdata(pci_dev
, efx
);
2106 rc
= efx_init_struct(efx
, type
, pci_dev
, net_dev
);
2110 EFX_INFO(efx
, "Solarflare Communications NIC detected\n");
2112 /* Set up basic I/O (BAR mappings etc) */
2113 rc
= efx_init_io(efx
);
2117 /* No serialisation is required with the reset path because
2118 * we're in STATE_INIT. */
2119 for (i
= 0; i
< 5; i
++) {
2120 rc
= efx_pci_probe_main(efx
);
2122 /* Serialise against efx_reset(). No more resets will be
2123 * scheduled since efx_stop_all() has been called, and we
2124 * have not and never have been registered with either
2125 * the rtnetlink or driverlink layers. */
2126 cancel_work_sync(&efx
->reset_work
);
2129 if (efx
->reset_pending
!= RESET_TYPE_NONE
) {
2130 /* If there was a scheduled reset during
2131 * probe, the NIC is probably hosed anyway */
2132 efx_pci_remove_main(efx
);
2139 /* Retry if a recoverably reset event has been scheduled */
2140 if ((efx
->reset_pending
!= RESET_TYPE_INVISIBLE
) &&
2141 (efx
->reset_pending
!= RESET_TYPE_ALL
))
2144 efx
->reset_pending
= RESET_TYPE_NONE
;
2148 EFX_ERR(efx
, "Could not reset NIC\n");
2152 /* Switch to the running state before we expose the device to
2153 * the OS. This is to ensure that the initial gathering of
2154 * MAC stats succeeds. */
2155 efx
->state
= STATE_RUNNING
;
2157 efx_mtd_probe(efx
); /* allowed to fail */
2159 rc
= efx_register_netdev(efx
);
2163 EFX_LOG(efx
, "initialisation successful\n");
2167 efx_pci_remove_main(efx
);
2172 efx_fini_struct(efx
);
2174 EFX_LOG(efx
, "initialisation failed. rc=%d\n", rc
);
2175 free_netdev(net_dev
);
2179 static struct pci_driver efx_pci_driver
= {
2180 .name
= EFX_DRIVER_NAME
,
2181 .id_table
= efx_pci_table
,
2182 .probe
= efx_pci_probe
,
2183 .remove
= efx_pci_remove
,
2186 /**************************************************************************
2188 * Kernel module interface
2190 *************************************************************************/
2192 module_param(interrupt_mode
, uint
, 0444);
2193 MODULE_PARM_DESC(interrupt_mode
,
2194 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2196 static int __init
efx_init_module(void)
2200 printk(KERN_INFO
"Solarflare NET driver v" EFX_DRIVER_VERSION
"\n");
2202 rc
= register_netdevice_notifier(&efx_netdev_notifier
);
2206 refill_workqueue
= create_workqueue("sfc_refill");
2207 if (!refill_workqueue
) {
2211 reset_workqueue
= create_singlethread_workqueue("sfc_reset");
2212 if (!reset_workqueue
) {
2217 rc
= pci_register_driver(&efx_pci_driver
);
2224 destroy_workqueue(reset_workqueue
);
2226 destroy_workqueue(refill_workqueue
);
2228 unregister_netdevice_notifier(&efx_netdev_notifier
);
2233 static void __exit
efx_exit_module(void)
2235 printk(KERN_INFO
"Solarflare NET driver unloading\n");
2237 pci_unregister_driver(&efx_pci_driver
);
2238 destroy_workqueue(reset_workqueue
);
2239 destroy_workqueue(refill_workqueue
);
2240 unregister_netdevice_notifier(&efx_netdev_notifier
);
2244 module_init(efx_init_module
);
2245 module_exit(efx_exit_module
);
2247 MODULE_AUTHOR("Michael Brown <mbrown@fensystems.co.uk> and "
2248 "Solarflare Communications");
2249 MODULE_DESCRIPTION("Solarflare Communications network driver");
2250 MODULE_LICENSE("GPL");
2251 MODULE_DEVICE_TABLE(pci
, efx_pci_table
);