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_rx_strategy(channel
);
187 efx_fast_push_rx_descriptors(&efx
->rx_queue
[channel
->channel
]);
192 /* Mark channel as finished processing
194 * Note that since we will not receive further interrupts for this
195 * channel before we finish processing and call the eventq_read_ack()
196 * method, there is no need to use the interrupt hold-off timers.
198 static inline void efx_channel_processed(struct efx_channel
*channel
)
200 /* The interrupt handler for this channel may set work_pending
201 * as soon as we acknowledge the events we've seen. Make sure
202 * it's cleared before then. */
203 channel
->work_pending
= false;
206 falcon_eventq_read_ack(channel
);
211 * NAPI guarantees serialisation of polls of the same device, which
212 * provides the guarantee required by efx_process_channel().
214 static int efx_poll(struct napi_struct
*napi
, int budget
)
216 struct efx_channel
*channel
=
217 container_of(napi
, struct efx_channel
, napi_str
);
220 EFX_TRACE(channel
->efx
, "channel %d NAPI poll executing on CPU %d\n",
221 channel
->channel
, raw_smp_processor_id());
223 rx_packets
= efx_process_channel(channel
, budget
);
225 if (rx_packets
< budget
) {
226 /* There is no race here; although napi_disable() will
227 * only wait for napi_complete(), this isn't a problem
228 * since efx_channel_processed() will have no effect if
229 * interrupts have already been disabled.
232 efx_channel_processed(channel
);
238 /* Process the eventq of the specified channel immediately on this CPU
240 * Disable hardware generated interrupts, wait for any existing
241 * processing to finish, then directly poll (and ack ) the eventq.
242 * Finally reenable NAPI and interrupts.
244 * Since we are touching interrupts the caller should hold the suspend lock
246 void efx_process_channel_now(struct efx_channel
*channel
)
248 struct efx_nic
*efx
= channel
->efx
;
250 BUG_ON(!channel
->used_flags
);
251 BUG_ON(!channel
->enabled
);
253 /* Disable interrupts and wait for ISRs to complete */
254 falcon_disable_interrupts(efx
);
256 synchronize_irq(efx
->legacy_irq
);
258 synchronize_irq(channel
->irq
);
260 /* Wait for any NAPI processing to complete */
261 napi_disable(&channel
->napi_str
);
263 /* Poll the channel */
264 efx_process_channel(channel
, efx
->type
->evq_size
);
266 /* Ack the eventq. This may cause an interrupt to be generated
267 * when they are reenabled */
268 efx_channel_processed(channel
);
270 napi_enable(&channel
->napi_str
);
271 falcon_enable_interrupts(efx
);
274 /* Create event queue
275 * Event queue memory allocations are done only once. If the channel
276 * is reset, the memory buffer will be reused; this guards against
277 * errors during channel reset and also simplifies interrupt handling.
279 static int efx_probe_eventq(struct efx_channel
*channel
)
281 EFX_LOG(channel
->efx
, "chan %d create event queue\n", channel
->channel
);
283 return falcon_probe_eventq(channel
);
286 /* Prepare channel's event queue */
287 static void efx_init_eventq(struct efx_channel
*channel
)
289 EFX_LOG(channel
->efx
, "chan %d init event queue\n", channel
->channel
);
291 channel
->eventq_read_ptr
= 0;
293 falcon_init_eventq(channel
);
296 static void efx_fini_eventq(struct efx_channel
*channel
)
298 EFX_LOG(channel
->efx
, "chan %d fini event queue\n", channel
->channel
);
300 falcon_fini_eventq(channel
);
303 static void efx_remove_eventq(struct efx_channel
*channel
)
305 EFX_LOG(channel
->efx
, "chan %d remove event queue\n", channel
->channel
);
307 falcon_remove_eventq(channel
);
310 /**************************************************************************
314 *************************************************************************/
316 static int efx_probe_channel(struct efx_channel
*channel
)
318 struct efx_tx_queue
*tx_queue
;
319 struct efx_rx_queue
*rx_queue
;
322 EFX_LOG(channel
->efx
, "creating channel %d\n", channel
->channel
);
324 rc
= efx_probe_eventq(channel
);
328 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
329 rc
= efx_probe_tx_queue(tx_queue
);
334 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
335 rc
= efx_probe_rx_queue(rx_queue
);
340 channel
->n_rx_frm_trunc
= 0;
345 efx_for_each_channel_rx_queue(rx_queue
, channel
)
346 efx_remove_rx_queue(rx_queue
);
348 efx_for_each_channel_tx_queue(tx_queue
, channel
)
349 efx_remove_tx_queue(tx_queue
);
355 static void efx_set_channel_names(struct efx_nic
*efx
)
357 struct efx_channel
*channel
;
358 const char *type
= "";
361 efx_for_each_channel(channel
, efx
) {
362 number
= channel
->channel
;
363 if (efx
->n_channels
> efx
->n_rx_queues
) {
364 if (channel
->channel
< efx
->n_rx_queues
) {
368 number
-= efx
->n_rx_queues
;
371 snprintf(channel
->name
, sizeof(channel
->name
),
372 "%s%s-%d", efx
->name
, type
, number
);
376 /* Channels are shutdown and reinitialised whilst the NIC is running
377 * to propagate configuration changes (mtu, checksum offload), or
378 * to clear hardware error conditions
380 static void efx_init_channels(struct efx_nic
*efx
)
382 struct efx_tx_queue
*tx_queue
;
383 struct efx_rx_queue
*rx_queue
;
384 struct efx_channel
*channel
;
386 /* Calculate the rx buffer allocation parameters required to
387 * support the current MTU, including padding for header
388 * alignment and overruns.
390 efx
->rx_buffer_len
= (max(EFX_PAGE_IP_ALIGN
, NET_IP_ALIGN
) +
391 EFX_MAX_FRAME_LEN(efx
->net_dev
->mtu
) +
392 efx
->type
->rx_buffer_padding
);
393 efx
->rx_buffer_order
= get_order(efx
->rx_buffer_len
);
395 /* Initialise the channels */
396 efx_for_each_channel(channel
, efx
) {
397 EFX_LOG(channel
->efx
, "init chan %d\n", channel
->channel
);
399 efx_init_eventq(channel
);
401 efx_for_each_channel_tx_queue(tx_queue
, channel
)
402 efx_init_tx_queue(tx_queue
);
404 /* The rx buffer allocation strategy is MTU dependent */
405 efx_rx_strategy(channel
);
407 efx_for_each_channel_rx_queue(rx_queue
, channel
)
408 efx_init_rx_queue(rx_queue
);
410 WARN_ON(channel
->rx_pkt
!= NULL
);
411 efx_rx_strategy(channel
);
415 /* This enables event queue processing and packet transmission.
417 * Note that this function is not allowed to fail, since that would
418 * introduce too much complexity into the suspend/resume path.
420 static void efx_start_channel(struct efx_channel
*channel
)
422 struct efx_rx_queue
*rx_queue
;
424 EFX_LOG(channel
->efx
, "starting chan %d\n", channel
->channel
);
426 if (!(channel
->efx
->net_dev
->flags
& IFF_UP
))
427 netif_napi_add(channel
->napi_dev
, &channel
->napi_str
,
428 efx_poll
, napi_weight
);
430 /* The interrupt handler for this channel may set work_pending
431 * as soon as we enable it. Make sure it's cleared before
432 * then. Similarly, make sure it sees the enabled flag set. */
433 channel
->work_pending
= false;
434 channel
->enabled
= true;
437 napi_enable(&channel
->napi_str
);
439 /* Load up RX descriptors */
440 efx_for_each_channel_rx_queue(rx_queue
, channel
)
441 efx_fast_push_rx_descriptors(rx_queue
);
444 /* This disables event queue processing and packet transmission.
445 * This function does not guarantee that all queue processing
446 * (e.g. RX refill) is complete.
448 static void efx_stop_channel(struct efx_channel
*channel
)
450 struct efx_rx_queue
*rx_queue
;
452 if (!channel
->enabled
)
455 EFX_LOG(channel
->efx
, "stop chan %d\n", channel
->channel
);
457 channel
->enabled
= false;
458 napi_disable(&channel
->napi_str
);
460 /* Ensure that any worker threads have exited or will be no-ops */
461 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
462 spin_lock_bh(&rx_queue
->add_lock
);
463 spin_unlock_bh(&rx_queue
->add_lock
);
467 static void efx_fini_channels(struct efx_nic
*efx
)
469 struct efx_channel
*channel
;
470 struct efx_tx_queue
*tx_queue
;
471 struct efx_rx_queue
*rx_queue
;
474 EFX_ASSERT_RESET_SERIALISED(efx
);
475 BUG_ON(efx
->port_enabled
);
477 rc
= falcon_flush_queues(efx
);
479 EFX_ERR(efx
, "failed to flush queues\n");
481 EFX_LOG(efx
, "successfully flushed all queues\n");
483 efx_for_each_channel(channel
, efx
) {
484 EFX_LOG(channel
->efx
, "shut down chan %d\n", channel
->channel
);
486 efx_for_each_channel_rx_queue(rx_queue
, channel
)
487 efx_fini_rx_queue(rx_queue
);
488 efx_for_each_channel_tx_queue(tx_queue
, channel
)
489 efx_fini_tx_queue(tx_queue
);
490 efx_fini_eventq(channel
);
494 static void efx_remove_channel(struct efx_channel
*channel
)
496 struct efx_tx_queue
*tx_queue
;
497 struct efx_rx_queue
*rx_queue
;
499 EFX_LOG(channel
->efx
, "destroy chan %d\n", channel
->channel
);
501 efx_for_each_channel_rx_queue(rx_queue
, channel
)
502 efx_remove_rx_queue(rx_queue
);
503 efx_for_each_channel_tx_queue(tx_queue
, channel
)
504 efx_remove_tx_queue(tx_queue
);
505 efx_remove_eventq(channel
);
507 channel
->used_flags
= 0;
510 void efx_schedule_slow_fill(struct efx_rx_queue
*rx_queue
, int delay
)
512 queue_delayed_work(refill_workqueue
, &rx_queue
->work
, delay
);
515 /**************************************************************************
519 **************************************************************************/
521 /* This ensures that the kernel is kept informed (via
522 * netif_carrier_on/off) of the link status, and also maintains the
523 * link status's stop on the port's TX queue.
525 static void efx_link_status_changed(struct efx_nic
*efx
)
527 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
528 * that no events are triggered between unregister_netdev() and the
529 * driver unloading. A more general condition is that NETDEV_CHANGE
530 * can only be generated between NETDEV_UP and NETDEV_DOWN */
531 if (!netif_running(efx
->net_dev
))
534 if (efx
->port_inhibited
) {
535 netif_carrier_off(efx
->net_dev
);
539 if (efx
->link_up
!= netif_carrier_ok(efx
->net_dev
)) {
540 efx
->n_link_state_changes
++;
543 netif_carrier_on(efx
->net_dev
);
545 netif_carrier_off(efx
->net_dev
);
548 /* Status message for kernel log */
550 EFX_INFO(efx
, "link up at %uMbps %s-duplex (MTU %d)%s\n",
551 efx
->link_speed
, efx
->link_fd
? "full" : "half",
553 (efx
->promiscuous
? " [PROMISC]" : ""));
555 EFX_INFO(efx
, "link down\n");
560 /* This call reinitialises the MAC to pick up new PHY settings. The
561 * caller must hold the mac_lock */
562 void __efx_reconfigure_port(struct efx_nic
*efx
)
564 WARN_ON(!mutex_is_locked(&efx
->mac_lock
));
566 EFX_LOG(efx
, "reconfiguring MAC from PHY settings on CPU %d\n",
567 raw_smp_processor_id());
569 /* Serialise the promiscuous flag with efx_set_multicast_list. */
570 if (efx_dev_registered(efx
)) {
571 netif_addr_lock_bh(efx
->net_dev
);
572 netif_addr_unlock_bh(efx
->net_dev
);
575 falcon_deconfigure_mac_wrapper(efx
);
577 /* Reconfigure the PHY, disabling transmit in mac level loopback. */
578 if (LOOPBACK_INTERNAL(efx
))
579 efx
->phy_mode
|= PHY_MODE_TX_DISABLED
;
581 efx
->phy_mode
&= ~PHY_MODE_TX_DISABLED
;
582 efx
->phy_op
->reconfigure(efx
);
584 if (falcon_switch_mac(efx
))
587 efx
->mac_op
->reconfigure(efx
);
589 /* Inform kernel of loss/gain of carrier */
590 efx_link_status_changed(efx
);
594 EFX_ERR(efx
, "failed to reconfigure MAC\n");
595 efx
->phy_op
->fini(efx
);
596 efx
->port_initialized
= false;
599 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
601 void efx_reconfigure_port(struct efx_nic
*efx
)
603 EFX_ASSERT_RESET_SERIALISED(efx
);
605 mutex_lock(&efx
->mac_lock
);
606 __efx_reconfigure_port(efx
);
607 mutex_unlock(&efx
->mac_lock
);
610 /* Asynchronous efx_reconfigure_port work item. To speed up efx_flush_all()
611 * we don't efx_reconfigure_port() if the port is disabled. Care is taken
612 * in efx_stop_all() and efx_start_port() to prevent PHY events being lost */
613 static void efx_phy_work(struct work_struct
*data
)
615 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, phy_work
);
617 mutex_lock(&efx
->mac_lock
);
618 if (efx
->port_enabled
)
619 __efx_reconfigure_port(efx
);
620 mutex_unlock(&efx
->mac_lock
);
623 static void efx_mac_work(struct work_struct
*data
)
625 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, mac_work
);
627 mutex_lock(&efx
->mac_lock
);
628 if (efx
->port_enabled
)
629 efx
->mac_op
->irq(efx
);
630 mutex_unlock(&efx
->mac_lock
);
633 static int efx_probe_port(struct efx_nic
*efx
)
637 EFX_LOG(efx
, "create port\n");
639 /* Connect up MAC/PHY operations table and read MAC address */
640 rc
= falcon_probe_port(efx
);
645 efx
->phy_mode
= PHY_MODE_SPECIAL
;
647 /* Sanity check MAC address */
648 if (is_valid_ether_addr(efx
->mac_address
)) {
649 memcpy(efx
->net_dev
->dev_addr
, efx
->mac_address
, ETH_ALEN
);
651 EFX_ERR(efx
, "invalid MAC address %pM\n",
653 if (!allow_bad_hwaddr
) {
657 random_ether_addr(efx
->net_dev
->dev_addr
);
658 EFX_INFO(efx
, "using locally-generated MAC %pM\n",
659 efx
->net_dev
->dev_addr
);
665 efx_remove_port(efx
);
669 static int efx_init_port(struct efx_nic
*efx
)
673 EFX_LOG(efx
, "init port\n");
675 rc
= efx
->phy_op
->init(efx
);
678 mutex_lock(&efx
->mac_lock
);
679 efx
->phy_op
->reconfigure(efx
);
680 rc
= falcon_switch_mac(efx
);
681 mutex_unlock(&efx
->mac_lock
);
684 efx
->mac_op
->reconfigure(efx
);
686 efx
->port_initialized
= true;
687 efx_stats_enable(efx
);
691 efx
->phy_op
->fini(efx
);
695 /* Allow efx_reconfigure_port() to be scheduled, and close the window
696 * between efx_stop_port and efx_flush_all whereby a previously scheduled
697 * efx_phy_work()/efx_mac_work() may have been cancelled */
698 static void efx_start_port(struct efx_nic
*efx
)
700 EFX_LOG(efx
, "start port\n");
701 BUG_ON(efx
->port_enabled
);
703 mutex_lock(&efx
->mac_lock
);
704 efx
->port_enabled
= true;
705 __efx_reconfigure_port(efx
);
706 efx
->mac_op
->irq(efx
);
707 mutex_unlock(&efx
->mac_lock
);
710 /* Prevent efx_phy_work, efx_mac_work, and efx_monitor() from executing,
711 * and efx_set_multicast_list() from scheduling efx_phy_work. efx_phy_work
712 * and efx_mac_work may still be scheduled via NAPI processing until
713 * efx_flush_all() is called */
714 static void efx_stop_port(struct efx_nic
*efx
)
716 EFX_LOG(efx
, "stop port\n");
718 mutex_lock(&efx
->mac_lock
);
719 efx
->port_enabled
= false;
720 mutex_unlock(&efx
->mac_lock
);
722 /* Serialise against efx_set_multicast_list() */
723 if (efx_dev_registered(efx
)) {
724 netif_addr_lock_bh(efx
->net_dev
);
725 netif_addr_unlock_bh(efx
->net_dev
);
729 static void efx_fini_port(struct efx_nic
*efx
)
731 EFX_LOG(efx
, "shut down port\n");
733 if (!efx
->port_initialized
)
736 efx_stats_disable(efx
);
737 efx
->phy_op
->fini(efx
);
738 efx
->port_initialized
= false;
740 efx
->link_up
= false;
741 efx_link_status_changed(efx
);
744 static void efx_remove_port(struct efx_nic
*efx
)
746 EFX_LOG(efx
, "destroying port\n");
748 falcon_remove_port(efx
);
751 /**************************************************************************
755 **************************************************************************/
757 /* This configures the PCI device to enable I/O and DMA. */
758 static int efx_init_io(struct efx_nic
*efx
)
760 struct pci_dev
*pci_dev
= efx
->pci_dev
;
761 dma_addr_t dma_mask
= efx
->type
->max_dma_mask
;
764 EFX_LOG(efx
, "initialising I/O\n");
766 rc
= pci_enable_device(pci_dev
);
768 EFX_ERR(efx
, "failed to enable PCI device\n");
772 pci_set_master(pci_dev
);
774 /* Set the PCI DMA mask. Try all possibilities from our
775 * genuine mask down to 32 bits, because some architectures
776 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
777 * masks event though they reject 46 bit masks.
779 while (dma_mask
> 0x7fffffffUL
) {
780 if (pci_dma_supported(pci_dev
, dma_mask
) &&
781 ((rc
= pci_set_dma_mask(pci_dev
, dma_mask
)) == 0))
786 EFX_ERR(efx
, "could not find a suitable DMA mask\n");
789 EFX_LOG(efx
, "using DMA mask %llx\n", (unsigned long long) dma_mask
);
790 rc
= pci_set_consistent_dma_mask(pci_dev
, dma_mask
);
792 /* pci_set_consistent_dma_mask() is not *allowed* to
793 * fail with a mask that pci_set_dma_mask() accepted,
794 * but just in case...
796 EFX_ERR(efx
, "failed to set consistent DMA mask\n");
800 efx
->membase_phys
= pci_resource_start(efx
->pci_dev
,
802 rc
= pci_request_region(pci_dev
, efx
->type
->mem_bar
, "sfc");
804 EFX_ERR(efx
, "request for memory BAR failed\n");
808 efx
->membase
= ioremap_nocache(efx
->membase_phys
,
809 efx
->type
->mem_map_size
);
811 EFX_ERR(efx
, "could not map memory BAR %d at %llx+%x\n",
813 (unsigned long long)efx
->membase_phys
,
814 efx
->type
->mem_map_size
);
818 EFX_LOG(efx
, "memory BAR %u at %llx+%x (virtual %p)\n",
819 efx
->type
->mem_bar
, (unsigned long long)efx
->membase_phys
,
820 efx
->type
->mem_map_size
, efx
->membase
);
825 pci_release_region(efx
->pci_dev
, efx
->type
->mem_bar
);
827 efx
->membase_phys
= 0;
829 pci_disable_device(efx
->pci_dev
);
834 static void efx_fini_io(struct efx_nic
*efx
)
836 EFX_LOG(efx
, "shutting down I/O\n");
839 iounmap(efx
->membase
);
843 if (efx
->membase_phys
) {
844 pci_release_region(efx
->pci_dev
, efx
->type
->mem_bar
);
845 efx
->membase_phys
= 0;
848 pci_disable_device(efx
->pci_dev
);
851 /* Get number of RX queues wanted. Return number of online CPU
852 * packages in the expectation that an IRQ balancer will spread
853 * interrupts across them. */
854 static int efx_wanted_rx_queues(void)
860 cpus_clear(core_mask
);
862 for_each_online_cpu(cpu
) {
863 if (!cpu_isset(cpu
, core_mask
)) {
865 cpus_or(core_mask
, core_mask
,
866 topology_core_siblings(cpu
));
873 /* Probe the number and type of interrupts we are able to obtain, and
874 * the resulting numbers of channels and RX queues.
876 static void efx_probe_interrupts(struct efx_nic
*efx
)
879 min_t(int, efx
->type
->phys_addr_channels
, EFX_MAX_CHANNELS
);
882 if (efx
->interrupt_mode
== EFX_INT_MODE_MSIX
) {
883 struct msix_entry xentries
[EFX_MAX_CHANNELS
];
887 /* We want one RX queue and interrupt per CPU package
888 * (or as specified by the rss_cpus module parameter).
889 * We will need one channel per interrupt.
891 rx_queues
= rss_cpus
? rss_cpus
: efx_wanted_rx_queues();
892 wanted_ints
= rx_queues
+ (separate_tx_channels
? 1 : 0);
893 wanted_ints
= min(wanted_ints
, max_channels
);
895 for (i
= 0; i
< wanted_ints
; i
++)
896 xentries
[i
].entry
= i
;
897 rc
= pci_enable_msix(efx
->pci_dev
, xentries
, wanted_ints
);
899 EFX_ERR(efx
, "WARNING: Insufficient MSI-X vectors"
900 " available (%d < %d).\n", rc
, wanted_ints
);
901 EFX_ERR(efx
, "WARNING: Performance may be reduced.\n");
902 EFX_BUG_ON_PARANOID(rc
>= wanted_ints
);
904 rc
= pci_enable_msix(efx
->pci_dev
, xentries
,
909 efx
->n_rx_queues
= min(rx_queues
, wanted_ints
);
910 efx
->n_channels
= wanted_ints
;
911 for (i
= 0; i
< wanted_ints
; i
++)
912 efx
->channel
[i
].irq
= xentries
[i
].vector
;
914 /* Fall back to single channel MSI */
915 efx
->interrupt_mode
= EFX_INT_MODE_MSI
;
916 EFX_ERR(efx
, "could not enable MSI-X\n");
920 /* Try single interrupt MSI */
921 if (efx
->interrupt_mode
== EFX_INT_MODE_MSI
) {
922 efx
->n_rx_queues
= 1;
924 rc
= pci_enable_msi(efx
->pci_dev
);
926 efx
->channel
[0].irq
= efx
->pci_dev
->irq
;
928 EFX_ERR(efx
, "could not enable MSI\n");
929 efx
->interrupt_mode
= EFX_INT_MODE_LEGACY
;
933 /* Assume legacy interrupts */
934 if (efx
->interrupt_mode
== EFX_INT_MODE_LEGACY
) {
935 efx
->n_rx_queues
= 1;
936 efx
->n_channels
= 1 + (separate_tx_channels
? 1 : 0);
937 efx
->legacy_irq
= efx
->pci_dev
->irq
;
941 static void efx_remove_interrupts(struct efx_nic
*efx
)
943 struct efx_channel
*channel
;
945 /* Remove MSI/MSI-X interrupts */
946 efx_for_each_channel(channel
, efx
)
948 pci_disable_msi(efx
->pci_dev
);
949 pci_disable_msix(efx
->pci_dev
);
951 /* Remove legacy interrupt */
955 static void efx_set_channels(struct efx_nic
*efx
)
957 struct efx_tx_queue
*tx_queue
;
958 struct efx_rx_queue
*rx_queue
;
960 efx_for_each_tx_queue(tx_queue
, efx
) {
961 if (separate_tx_channels
)
962 tx_queue
->channel
= &efx
->channel
[efx
->n_channels
-1];
964 tx_queue
->channel
= &efx
->channel
[0];
965 tx_queue
->channel
->used_flags
|= EFX_USED_BY_TX
;
968 efx_for_each_rx_queue(rx_queue
, efx
) {
969 rx_queue
->channel
= &efx
->channel
[rx_queue
->queue
];
970 rx_queue
->channel
->used_flags
|= EFX_USED_BY_RX
;
974 static int efx_probe_nic(struct efx_nic
*efx
)
978 EFX_LOG(efx
, "creating NIC\n");
980 /* Carry out hardware-type specific initialisation */
981 rc
= falcon_probe_nic(efx
);
985 /* Determine the number of channels and RX queues by trying to hook
986 * in MSI-X interrupts. */
987 efx_probe_interrupts(efx
);
989 efx_set_channels(efx
);
991 /* Initialise the interrupt moderation settings */
992 efx_init_irq_moderation(efx
, tx_irq_mod_usec
, rx_irq_mod_usec
);
997 static void efx_remove_nic(struct efx_nic
*efx
)
999 EFX_LOG(efx
, "destroying NIC\n");
1001 efx_remove_interrupts(efx
);
1002 falcon_remove_nic(efx
);
1005 /**************************************************************************
1007 * NIC startup/shutdown
1009 *************************************************************************/
1011 static int efx_probe_all(struct efx_nic
*efx
)
1013 struct efx_channel
*channel
;
1017 rc
= efx_probe_nic(efx
);
1019 EFX_ERR(efx
, "failed to create NIC\n");
1024 rc
= efx_probe_port(efx
);
1026 EFX_ERR(efx
, "failed to create port\n");
1030 /* Create channels */
1031 efx_for_each_channel(channel
, efx
) {
1032 rc
= efx_probe_channel(channel
);
1034 EFX_ERR(efx
, "failed to create channel %d\n",
1039 efx_set_channel_names(efx
);
1044 efx_for_each_channel(channel
, efx
)
1045 efx_remove_channel(channel
);
1046 efx_remove_port(efx
);
1048 efx_remove_nic(efx
);
1053 /* Called after previous invocation(s) of efx_stop_all, restarts the
1054 * port, kernel transmit queue, NAPI processing and hardware interrupts,
1055 * and ensures that the port is scheduled to be reconfigured.
1056 * This function is safe to call multiple times when the NIC is in any
1058 static void efx_start_all(struct efx_nic
*efx
)
1060 struct efx_channel
*channel
;
1062 EFX_ASSERT_RESET_SERIALISED(efx
);
1064 /* Check that it is appropriate to restart the interface. All
1065 * of these flags are safe to read under just the rtnl lock */
1066 if (efx
->port_enabled
)
1068 if ((efx
->state
!= STATE_RUNNING
) && (efx
->state
!= STATE_INIT
))
1070 if (efx_dev_registered(efx
) && !netif_running(efx
->net_dev
))
1073 /* Mark the port as enabled so port reconfigurations can start, then
1074 * restart the transmit interface early so the watchdog timer stops */
1075 efx_start_port(efx
);
1076 if (efx_dev_registered(efx
))
1077 efx_wake_queue(efx
);
1079 efx_for_each_channel(channel
, efx
)
1080 efx_start_channel(channel
);
1082 falcon_enable_interrupts(efx
);
1084 /* Start hardware monitor if we're in RUNNING */
1085 if (efx
->state
== STATE_RUNNING
)
1086 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1087 efx_monitor_interval
);
1090 /* Flush all delayed work. Should only be called when no more delayed work
1091 * will be scheduled. This doesn't flush pending online resets (efx_reset),
1092 * since we're holding the rtnl_lock at this point. */
1093 static void efx_flush_all(struct efx_nic
*efx
)
1095 struct efx_rx_queue
*rx_queue
;
1097 /* Make sure the hardware monitor is stopped */
1098 cancel_delayed_work_sync(&efx
->monitor_work
);
1100 /* Ensure that all RX slow refills are complete. */
1101 efx_for_each_rx_queue(rx_queue
, efx
)
1102 cancel_delayed_work_sync(&rx_queue
->work
);
1104 /* Stop scheduled port reconfigurations */
1105 cancel_work_sync(&efx
->mac_work
);
1106 cancel_work_sync(&efx
->phy_work
);
1110 /* Quiesce hardware and software without bringing the link down.
1111 * Safe to call multiple times, when the nic and interface is in any
1112 * state. The caller is guaranteed to subsequently be in a position
1113 * to modify any hardware and software state they see fit without
1115 static void efx_stop_all(struct efx_nic
*efx
)
1117 struct efx_channel
*channel
;
1119 EFX_ASSERT_RESET_SERIALISED(efx
);
1121 /* port_enabled can be read safely under the rtnl lock */
1122 if (!efx
->port_enabled
)
1125 /* Disable interrupts and wait for ISR to complete */
1126 falcon_disable_interrupts(efx
);
1127 if (efx
->legacy_irq
)
1128 synchronize_irq(efx
->legacy_irq
);
1129 efx_for_each_channel(channel
, efx
) {
1131 synchronize_irq(channel
->irq
);
1134 /* Stop all NAPI processing and synchronous rx refills */
1135 efx_for_each_channel(channel
, efx
)
1136 efx_stop_channel(channel
);
1138 /* Stop all asynchronous port reconfigurations. Since all
1139 * event processing has already been stopped, there is no
1140 * window to loose phy events */
1143 /* Flush efx_phy_work, efx_mac_work, refill_workqueue, monitor_work */
1146 /* Isolate the MAC from the TX and RX engines, so that queue
1147 * flushes will complete in a timely fashion. */
1148 falcon_drain_tx_fifo(efx
);
1150 /* Stop the kernel transmit interface late, so the watchdog
1151 * timer isn't ticking over the flush */
1152 if (efx_dev_registered(efx
)) {
1153 efx_stop_queue(efx
);
1154 netif_tx_lock_bh(efx
->net_dev
);
1155 netif_tx_unlock_bh(efx
->net_dev
);
1159 static void efx_remove_all(struct efx_nic
*efx
)
1161 struct efx_channel
*channel
;
1163 efx_for_each_channel(channel
, efx
)
1164 efx_remove_channel(channel
);
1165 efx_remove_port(efx
);
1166 efx_remove_nic(efx
);
1169 /* A convinience function to safely flush all the queues */
1170 void efx_flush_queues(struct efx_nic
*efx
)
1172 EFX_ASSERT_RESET_SERIALISED(efx
);
1176 efx_fini_channels(efx
);
1177 efx_init_channels(efx
);
1182 /**************************************************************************
1184 * Interrupt moderation
1186 **************************************************************************/
1188 /* Set interrupt moderation parameters */
1189 void efx_init_irq_moderation(struct efx_nic
*efx
, int tx_usecs
, int rx_usecs
)
1191 struct efx_tx_queue
*tx_queue
;
1192 struct efx_rx_queue
*rx_queue
;
1194 EFX_ASSERT_RESET_SERIALISED(efx
);
1196 efx_for_each_tx_queue(tx_queue
, efx
)
1197 tx_queue
->channel
->irq_moderation
= tx_usecs
;
1199 efx_for_each_rx_queue(rx_queue
, efx
)
1200 rx_queue
->channel
->irq_moderation
= rx_usecs
;
1203 /**************************************************************************
1207 **************************************************************************/
1209 /* Run periodically off the general workqueue. Serialised against
1210 * efx_reconfigure_port via the mac_lock */
1211 static void efx_monitor(struct work_struct
*data
)
1213 struct efx_nic
*efx
= container_of(data
, struct efx_nic
,
1217 EFX_TRACE(efx
, "hardware monitor executing on CPU %d\n",
1218 raw_smp_processor_id());
1220 /* If the mac_lock is already held then it is likely a port
1221 * reconfiguration is already in place, which will likely do
1222 * most of the work of check_hw() anyway. */
1223 if (!mutex_trylock(&efx
->mac_lock
))
1225 if (!efx
->port_enabled
)
1227 rc
= efx
->board_info
.monitor(efx
);
1229 EFX_ERR(efx
, "Board sensor %s; shutting down PHY\n",
1230 (rc
== -ERANGE
) ? "reported fault" : "failed");
1231 efx
->phy_mode
|= PHY_MODE_LOW_POWER
;
1232 falcon_sim_phy_event(efx
);
1234 efx
->phy_op
->poll(efx
);
1235 efx
->mac_op
->poll(efx
);
1238 mutex_unlock(&efx
->mac_lock
);
1240 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1241 efx_monitor_interval
);
1244 /**************************************************************************
1248 *************************************************************************/
1251 * Context: process, rtnl_lock() held.
1253 static int efx_ioctl(struct net_device
*net_dev
, struct ifreq
*ifr
, int cmd
)
1255 struct efx_nic
*efx
= netdev_priv(net_dev
);
1257 EFX_ASSERT_RESET_SERIALISED(efx
);
1259 return generic_mii_ioctl(&efx
->mii
, if_mii(ifr
), cmd
, NULL
);
1262 /**************************************************************************
1266 **************************************************************************/
1268 static int efx_init_napi(struct efx_nic
*efx
)
1270 struct efx_channel
*channel
;
1272 efx_for_each_channel(channel
, efx
) {
1273 channel
->napi_dev
= efx
->net_dev
;
1278 static void efx_fini_napi(struct efx_nic
*efx
)
1280 struct efx_channel
*channel
;
1282 efx_for_each_channel(channel
, efx
) {
1283 channel
->napi_dev
= NULL
;
1287 /**************************************************************************
1289 * Kernel netpoll interface
1291 *************************************************************************/
1293 #ifdef CONFIG_NET_POLL_CONTROLLER
1295 /* Although in the common case interrupts will be disabled, this is not
1296 * guaranteed. However, all our work happens inside the NAPI callback,
1297 * so no locking is required.
1299 static void efx_netpoll(struct net_device
*net_dev
)
1301 struct efx_nic
*efx
= netdev_priv(net_dev
);
1302 struct efx_channel
*channel
;
1304 efx_for_each_channel(channel
, efx
)
1305 efx_schedule_channel(channel
);
1310 /**************************************************************************
1312 * Kernel net device interface
1314 *************************************************************************/
1316 /* Context: process, rtnl_lock() held. */
1317 static int efx_net_open(struct net_device
*net_dev
)
1319 struct efx_nic
*efx
= netdev_priv(net_dev
);
1320 EFX_ASSERT_RESET_SERIALISED(efx
);
1322 EFX_LOG(efx
, "opening device %s on CPU %d\n", net_dev
->name
,
1323 raw_smp_processor_id());
1325 if (efx
->state
== STATE_DISABLED
)
1327 if (efx
->phy_mode
& PHY_MODE_SPECIAL
)
1334 /* Context: process, rtnl_lock() held.
1335 * Note that the kernel will ignore our return code; this method
1336 * should really be a void.
1338 static int efx_net_stop(struct net_device
*net_dev
)
1340 struct efx_nic
*efx
= netdev_priv(net_dev
);
1342 EFX_LOG(efx
, "closing %s on CPU %d\n", net_dev
->name
,
1343 raw_smp_processor_id());
1345 if (efx
->state
!= STATE_DISABLED
) {
1346 /* Stop the device and flush all the channels */
1348 efx_fini_channels(efx
);
1349 efx_init_channels(efx
);
1355 void efx_stats_disable(struct efx_nic
*efx
)
1357 spin_lock(&efx
->stats_lock
);
1358 ++efx
->stats_disable_count
;
1359 spin_unlock(&efx
->stats_lock
);
1362 void efx_stats_enable(struct efx_nic
*efx
)
1364 spin_lock(&efx
->stats_lock
);
1365 --efx
->stats_disable_count
;
1366 spin_unlock(&efx
->stats_lock
);
1369 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1370 static struct net_device_stats
*efx_net_stats(struct net_device
*net_dev
)
1372 struct efx_nic
*efx
= netdev_priv(net_dev
);
1373 struct efx_mac_stats
*mac_stats
= &efx
->mac_stats
;
1374 struct net_device_stats
*stats
= &net_dev
->stats
;
1376 /* Update stats if possible, but do not wait if another thread
1377 * is updating them or if MAC stats fetches are temporarily
1378 * disabled; slightly stale stats are acceptable.
1380 if (!spin_trylock(&efx
->stats_lock
))
1382 if (!efx
->stats_disable_count
) {
1383 efx
->mac_op
->update_stats(efx
);
1384 falcon_update_nic_stats(efx
);
1386 spin_unlock(&efx
->stats_lock
);
1388 stats
->rx_packets
= mac_stats
->rx_packets
;
1389 stats
->tx_packets
= mac_stats
->tx_packets
;
1390 stats
->rx_bytes
= mac_stats
->rx_bytes
;
1391 stats
->tx_bytes
= mac_stats
->tx_bytes
;
1392 stats
->multicast
= mac_stats
->rx_multicast
;
1393 stats
->collisions
= mac_stats
->tx_collision
;
1394 stats
->rx_length_errors
= (mac_stats
->rx_gtjumbo
+
1395 mac_stats
->rx_length_error
);
1396 stats
->rx_over_errors
= efx
->n_rx_nodesc_drop_cnt
;
1397 stats
->rx_crc_errors
= mac_stats
->rx_bad
;
1398 stats
->rx_frame_errors
= mac_stats
->rx_align_error
;
1399 stats
->rx_fifo_errors
= mac_stats
->rx_overflow
;
1400 stats
->rx_missed_errors
= mac_stats
->rx_missed
;
1401 stats
->tx_window_errors
= mac_stats
->tx_late_collision
;
1403 stats
->rx_errors
= (stats
->rx_length_errors
+
1404 stats
->rx_over_errors
+
1405 stats
->rx_crc_errors
+
1406 stats
->rx_frame_errors
+
1407 stats
->rx_fifo_errors
+
1408 stats
->rx_missed_errors
+
1409 mac_stats
->rx_symbol_error
);
1410 stats
->tx_errors
= (stats
->tx_window_errors
+
1416 /* Context: netif_tx_lock held, BHs disabled. */
1417 static void efx_watchdog(struct net_device
*net_dev
)
1419 struct efx_nic
*efx
= netdev_priv(net_dev
);
1421 EFX_ERR(efx
, "TX stuck with stop_count=%d port_enabled=%d:"
1422 " resetting channels\n",
1423 atomic_read(&efx
->netif_stop_count
), efx
->port_enabled
);
1425 efx_schedule_reset(efx
, RESET_TYPE_TX_WATCHDOG
);
1429 /* Context: process, rtnl_lock() held. */
1430 static int efx_change_mtu(struct net_device
*net_dev
, int new_mtu
)
1432 struct efx_nic
*efx
= netdev_priv(net_dev
);
1435 EFX_ASSERT_RESET_SERIALISED(efx
);
1437 if (new_mtu
> EFX_MAX_MTU
)
1442 EFX_LOG(efx
, "changing MTU to %d\n", new_mtu
);
1444 efx_fini_channels(efx
);
1445 net_dev
->mtu
= new_mtu
;
1446 efx_init_channels(efx
);
1452 static int efx_set_mac_address(struct net_device
*net_dev
, void *data
)
1454 struct efx_nic
*efx
= netdev_priv(net_dev
);
1455 struct sockaddr
*addr
= data
;
1456 char *new_addr
= addr
->sa_data
;
1458 EFX_ASSERT_RESET_SERIALISED(efx
);
1460 if (!is_valid_ether_addr(new_addr
)) {
1461 EFX_ERR(efx
, "invalid ethernet MAC address requested: %pM\n",
1466 memcpy(net_dev
->dev_addr
, new_addr
, net_dev
->addr_len
);
1468 /* Reconfigure the MAC */
1469 efx_reconfigure_port(efx
);
1474 /* Context: netif_addr_lock held, BHs disabled. */
1475 static void efx_set_multicast_list(struct net_device
*net_dev
)
1477 struct efx_nic
*efx
= netdev_priv(net_dev
);
1478 struct dev_mc_list
*mc_list
= net_dev
->mc_list
;
1479 union efx_multicast_hash
*mc_hash
= &efx
->multicast_hash
;
1480 bool promiscuous
= !!(net_dev
->flags
& IFF_PROMISC
);
1481 bool changed
= (efx
->promiscuous
!= promiscuous
);
1486 efx
->promiscuous
= promiscuous
;
1488 /* Build multicast hash table */
1489 if (promiscuous
|| (net_dev
->flags
& IFF_ALLMULTI
)) {
1490 memset(mc_hash
, 0xff, sizeof(*mc_hash
));
1492 memset(mc_hash
, 0x00, sizeof(*mc_hash
));
1493 for (i
= 0; i
< net_dev
->mc_count
; i
++) {
1494 crc
= ether_crc_le(ETH_ALEN
, mc_list
->dmi_addr
);
1495 bit
= crc
& (EFX_MCAST_HASH_ENTRIES
- 1);
1496 set_bit_le(bit
, mc_hash
->byte
);
1497 mc_list
= mc_list
->next
;
1501 if (!efx
->port_enabled
)
1502 /* Delay pushing settings until efx_start_port() */
1506 queue_work(efx
->workqueue
, &efx
->phy_work
);
1508 /* Create and activate new global multicast hash table */
1509 falcon_set_multicast_hash(efx
);
1512 static const struct net_device_ops efx_netdev_ops
= {
1513 .ndo_open
= efx_net_open
,
1514 .ndo_stop
= efx_net_stop
,
1515 .ndo_get_stats
= efx_net_stats
,
1516 .ndo_tx_timeout
= efx_watchdog
,
1517 .ndo_start_xmit
= efx_hard_start_xmit
,
1518 .ndo_validate_addr
= eth_validate_addr
,
1519 .ndo_do_ioctl
= efx_ioctl
,
1520 .ndo_change_mtu
= efx_change_mtu
,
1521 .ndo_set_mac_address
= efx_set_mac_address
,
1522 .ndo_set_multicast_list
= efx_set_multicast_list
,
1523 #ifdef CONFIG_NET_POLL_CONTROLLER
1524 .ndo_poll_controller
= efx_netpoll
,
1528 static void efx_update_name(struct efx_nic
*efx
)
1530 strcpy(efx
->name
, efx
->net_dev
->name
);
1531 efx_mtd_rename(efx
);
1532 efx_set_channel_names(efx
);
1535 static int efx_netdev_event(struct notifier_block
*this,
1536 unsigned long event
, void *ptr
)
1538 struct net_device
*net_dev
= ptr
;
1540 if (net_dev
->netdev_ops
== &efx_netdev_ops
&&
1541 event
== NETDEV_CHANGENAME
)
1542 efx_update_name(netdev_priv(net_dev
));
1547 static struct notifier_block efx_netdev_notifier
= {
1548 .notifier_call
= efx_netdev_event
,
1552 show_phy_type(struct device
*dev
, struct device_attribute
*attr
, char *buf
)
1554 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
1555 return sprintf(buf
, "%d\n", efx
->phy_type
);
1557 static DEVICE_ATTR(phy_type
, 0644, show_phy_type
, NULL
);
1559 static int efx_register_netdev(struct efx_nic
*efx
)
1561 struct net_device
*net_dev
= efx
->net_dev
;
1564 net_dev
->watchdog_timeo
= 5 * HZ
;
1565 net_dev
->irq
= efx
->pci_dev
->irq
;
1566 net_dev
->netdev_ops
= &efx_netdev_ops
;
1567 SET_NETDEV_DEV(net_dev
, &efx
->pci_dev
->dev
);
1568 SET_ETHTOOL_OPS(net_dev
, &efx_ethtool_ops
);
1570 /* Always start with carrier off; PHY events will detect the link */
1571 netif_carrier_off(efx
->net_dev
);
1573 /* Clear MAC statistics */
1574 efx
->mac_op
->update_stats(efx
);
1575 memset(&efx
->mac_stats
, 0, sizeof(efx
->mac_stats
));
1577 rc
= register_netdev(net_dev
);
1579 EFX_ERR(efx
, "could not register net dev\n");
1584 efx_update_name(efx
);
1587 rc
= device_create_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1589 EFX_ERR(efx
, "failed to init net dev attributes\n");
1590 goto fail_registered
;
1596 unregister_netdev(net_dev
);
1600 static void efx_unregister_netdev(struct efx_nic
*efx
)
1602 struct efx_tx_queue
*tx_queue
;
1607 BUG_ON(netdev_priv(efx
->net_dev
) != efx
);
1609 /* Free up any skbs still remaining. This has to happen before
1610 * we try to unregister the netdev as running their destructors
1611 * may be needed to get the device ref. count to 0. */
1612 efx_for_each_tx_queue(tx_queue
, efx
)
1613 efx_release_tx_buffers(tx_queue
);
1615 if (efx_dev_registered(efx
)) {
1616 strlcpy(efx
->name
, pci_name(efx
->pci_dev
), sizeof(efx
->name
));
1617 device_remove_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1618 unregister_netdev(efx
->net_dev
);
1622 /**************************************************************************
1624 * Device reset and suspend
1626 **************************************************************************/
1628 /* Tears down the entire software state and most of the hardware state
1630 void efx_reset_down(struct efx_nic
*efx
, enum reset_type method
,
1631 struct ethtool_cmd
*ecmd
)
1633 EFX_ASSERT_RESET_SERIALISED(efx
);
1635 efx_stats_disable(efx
);
1637 mutex_lock(&efx
->mac_lock
);
1638 mutex_lock(&efx
->spi_lock
);
1640 efx
->phy_op
->get_settings(efx
, ecmd
);
1642 efx_fini_channels(efx
);
1643 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
)
1644 efx
->phy_op
->fini(efx
);
1647 /* This function will always ensure that the locks acquired in
1648 * efx_reset_down() are released. A failure return code indicates
1649 * that we were unable to reinitialise the hardware, and the
1650 * driver should be disabled. If ok is false, then the rx and tx
1651 * engines are not restarted, pending a RESET_DISABLE. */
1652 int efx_reset_up(struct efx_nic
*efx
, enum reset_type method
,
1653 struct ethtool_cmd
*ecmd
, bool ok
)
1657 EFX_ASSERT_RESET_SERIALISED(efx
);
1659 rc
= falcon_init_nic(efx
);
1661 EFX_ERR(efx
, "failed to initialise NIC\n");
1665 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
) {
1667 rc
= efx
->phy_op
->init(efx
);
1671 efx
->port_initialized
= false;
1675 efx_init_channels(efx
);
1677 if (efx
->phy_op
->set_settings(efx
, ecmd
))
1678 EFX_ERR(efx
, "could not restore PHY settings\n");
1681 mutex_unlock(&efx
->spi_lock
);
1682 mutex_unlock(&efx
->mac_lock
);
1686 efx_stats_enable(efx
);
1691 /* Reset the NIC as transparently as possible. Do not reset the PHY
1692 * Note that the reset may fail, in which case the card will be left
1693 * in a most-probably-unusable state.
1695 * This function will sleep. You cannot reset from within an atomic
1696 * state; use efx_schedule_reset() instead.
1698 * Grabs the rtnl_lock.
1700 static int efx_reset(struct efx_nic
*efx
)
1702 struct ethtool_cmd ecmd
;
1703 enum reset_type method
= efx
->reset_pending
;
1706 /* Serialise with kernel interfaces */
1709 /* If we're not RUNNING then don't reset. Leave the reset_pending
1710 * flag set so that efx_pci_probe_main will be retried */
1711 if (efx
->state
!= STATE_RUNNING
) {
1712 EFX_INFO(efx
, "scheduled reset quenched. NIC not RUNNING\n");
1716 EFX_INFO(efx
, "resetting (%d)\n", method
);
1718 efx_reset_down(efx
, method
, &ecmd
);
1720 rc
= falcon_reset_hw(efx
, method
);
1722 EFX_ERR(efx
, "failed to reset hardware\n");
1726 /* Allow resets to be rescheduled. */
1727 efx
->reset_pending
= RESET_TYPE_NONE
;
1729 /* Reinitialise bus-mastering, which may have been turned off before
1730 * the reset was scheduled. This is still appropriate, even in the
1731 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
1732 * can respond to requests. */
1733 pci_set_master(efx
->pci_dev
);
1735 /* Leave device stopped if necessary */
1736 if (method
== RESET_TYPE_DISABLE
) {
1737 efx_reset_up(efx
, method
, &ecmd
, false);
1740 rc
= efx_reset_up(efx
, method
, &ecmd
, true);
1745 EFX_ERR(efx
, "has been disabled\n");
1746 efx
->state
= STATE_DISABLED
;
1747 dev_close(efx
->net_dev
);
1749 EFX_LOG(efx
, "reset complete\n");
1757 /* The worker thread exists so that code that cannot sleep can
1758 * schedule a reset for later.
1760 static void efx_reset_work(struct work_struct
*data
)
1762 struct efx_nic
*nic
= container_of(data
, struct efx_nic
, reset_work
);
1767 void efx_schedule_reset(struct efx_nic
*efx
, enum reset_type type
)
1769 enum reset_type method
;
1771 if (efx
->reset_pending
!= RESET_TYPE_NONE
) {
1772 EFX_INFO(efx
, "quenching already scheduled reset\n");
1777 case RESET_TYPE_INVISIBLE
:
1778 case RESET_TYPE_ALL
:
1779 case RESET_TYPE_WORLD
:
1780 case RESET_TYPE_DISABLE
:
1783 case RESET_TYPE_RX_RECOVERY
:
1784 case RESET_TYPE_RX_DESC_FETCH
:
1785 case RESET_TYPE_TX_DESC_FETCH
:
1786 case RESET_TYPE_TX_SKIP
:
1787 method
= RESET_TYPE_INVISIBLE
;
1790 method
= RESET_TYPE_ALL
;
1795 EFX_LOG(efx
, "scheduling reset (%d:%d)\n", type
, method
);
1797 EFX_LOG(efx
, "scheduling reset (%d)\n", method
);
1799 efx
->reset_pending
= method
;
1801 queue_work(reset_workqueue
, &efx
->reset_work
);
1804 /**************************************************************************
1806 * List of NICs we support
1808 **************************************************************************/
1810 /* PCI device ID table */
1811 static struct pci_device_id efx_pci_table
[] __devinitdata
= {
1812 {PCI_DEVICE(EFX_VENDID_SFC
, FALCON_A_P_DEVID
),
1813 .driver_data
= (unsigned long) &falcon_a_nic_type
},
1814 {PCI_DEVICE(EFX_VENDID_SFC
, FALCON_B_P_DEVID
),
1815 .driver_data
= (unsigned long) &falcon_b_nic_type
},
1816 {0} /* end of list */
1819 /**************************************************************************
1821 * Dummy PHY/MAC/Board operations
1823 * Can be used for some unimplemented operations
1824 * Needed so all function pointers are valid and do not have to be tested
1827 **************************************************************************/
1828 int efx_port_dummy_op_int(struct efx_nic
*efx
)
1832 void efx_port_dummy_op_void(struct efx_nic
*efx
) {}
1833 void efx_port_dummy_op_blink(struct efx_nic
*efx
, bool blink
) {}
1835 static struct efx_mac_operations efx_dummy_mac_operations
= {
1836 .reconfigure
= efx_port_dummy_op_void
,
1837 .poll
= efx_port_dummy_op_void
,
1838 .irq
= efx_port_dummy_op_void
,
1841 static struct efx_phy_operations efx_dummy_phy_operations
= {
1842 .init
= efx_port_dummy_op_int
,
1843 .reconfigure
= efx_port_dummy_op_void
,
1844 .poll
= efx_port_dummy_op_void
,
1845 .fini
= efx_port_dummy_op_void
,
1846 .clear_interrupt
= efx_port_dummy_op_void
,
1849 static struct efx_board efx_dummy_board_info
= {
1850 .init
= efx_port_dummy_op_int
,
1851 .init_leds
= efx_port_dummy_op_void
,
1852 .set_id_led
= efx_port_dummy_op_blink
,
1853 .monitor
= efx_port_dummy_op_int
,
1854 .blink
= efx_port_dummy_op_blink
,
1855 .fini
= efx_port_dummy_op_void
,
1858 /**************************************************************************
1862 **************************************************************************/
1864 /* This zeroes out and then fills in the invariants in a struct
1865 * efx_nic (including all sub-structures).
1867 static int efx_init_struct(struct efx_nic
*efx
, struct efx_nic_type
*type
,
1868 struct pci_dev
*pci_dev
, struct net_device
*net_dev
)
1870 struct efx_channel
*channel
;
1871 struct efx_tx_queue
*tx_queue
;
1872 struct efx_rx_queue
*rx_queue
;
1875 /* Initialise common structures */
1876 memset(efx
, 0, sizeof(*efx
));
1877 spin_lock_init(&efx
->biu_lock
);
1878 spin_lock_init(&efx
->phy_lock
);
1879 mutex_init(&efx
->spi_lock
);
1880 INIT_WORK(&efx
->reset_work
, efx_reset_work
);
1881 INIT_DELAYED_WORK(&efx
->monitor_work
, efx_monitor
);
1882 efx
->pci_dev
= pci_dev
;
1883 efx
->state
= STATE_INIT
;
1884 efx
->reset_pending
= RESET_TYPE_NONE
;
1885 strlcpy(efx
->name
, pci_name(pci_dev
), sizeof(efx
->name
));
1886 efx
->board_info
= efx_dummy_board_info
;
1888 efx
->net_dev
= net_dev
;
1889 efx
->rx_checksum_enabled
= true;
1890 spin_lock_init(&efx
->netif_stop_lock
);
1891 spin_lock_init(&efx
->stats_lock
);
1892 efx
->stats_disable_count
= 1;
1893 mutex_init(&efx
->mac_lock
);
1894 efx
->mac_op
= &efx_dummy_mac_operations
;
1895 efx
->phy_op
= &efx_dummy_phy_operations
;
1896 efx
->mii
.dev
= net_dev
;
1897 INIT_WORK(&efx
->phy_work
, efx_phy_work
);
1898 INIT_WORK(&efx
->mac_work
, efx_mac_work
);
1899 atomic_set(&efx
->netif_stop_count
, 1);
1901 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++) {
1902 channel
= &efx
->channel
[i
];
1904 channel
->channel
= i
;
1905 channel
->work_pending
= false;
1907 for (i
= 0; i
< EFX_TX_QUEUE_COUNT
; i
++) {
1908 tx_queue
= &efx
->tx_queue
[i
];
1909 tx_queue
->efx
= efx
;
1910 tx_queue
->queue
= i
;
1911 tx_queue
->buffer
= NULL
;
1912 tx_queue
->channel
= &efx
->channel
[0]; /* for safety */
1913 tx_queue
->tso_headers_free
= NULL
;
1915 for (i
= 0; i
< EFX_MAX_RX_QUEUES
; i
++) {
1916 rx_queue
= &efx
->rx_queue
[i
];
1917 rx_queue
->efx
= efx
;
1918 rx_queue
->queue
= i
;
1919 rx_queue
->channel
= &efx
->channel
[0]; /* for safety */
1920 rx_queue
->buffer
= NULL
;
1921 spin_lock_init(&rx_queue
->add_lock
);
1922 INIT_DELAYED_WORK(&rx_queue
->work
, efx_rx_work
);
1927 /* Sanity-check NIC type */
1928 EFX_BUG_ON_PARANOID(efx
->type
->txd_ring_mask
&
1929 (efx
->type
->txd_ring_mask
+ 1));
1930 EFX_BUG_ON_PARANOID(efx
->type
->rxd_ring_mask
&
1931 (efx
->type
->rxd_ring_mask
+ 1));
1932 EFX_BUG_ON_PARANOID(efx
->type
->evq_size
&
1933 (efx
->type
->evq_size
- 1));
1934 /* As close as we can get to guaranteeing that we don't overflow */
1935 EFX_BUG_ON_PARANOID(efx
->type
->evq_size
<
1936 (efx
->type
->txd_ring_mask
+ 1 +
1937 efx
->type
->rxd_ring_mask
+ 1));
1938 EFX_BUG_ON_PARANOID(efx
->type
->phys_addr_channels
> EFX_MAX_CHANNELS
);
1940 /* Higher numbered interrupt modes are less capable! */
1941 efx
->interrupt_mode
= max(efx
->type
->max_interrupt_mode
,
1944 /* Would be good to use the net_dev name, but we're too early */
1945 snprintf(efx
->workqueue_name
, sizeof(efx
->workqueue_name
), "sfc%s",
1947 efx
->workqueue
= create_singlethread_workqueue(efx
->workqueue_name
);
1948 if (!efx
->workqueue
)
1954 static void efx_fini_struct(struct efx_nic
*efx
)
1956 if (efx
->workqueue
) {
1957 destroy_workqueue(efx
->workqueue
);
1958 efx
->workqueue
= NULL
;
1962 /**************************************************************************
1966 **************************************************************************/
1968 /* Main body of final NIC shutdown code
1969 * This is called only at module unload (or hotplug removal).
1971 static void efx_pci_remove_main(struct efx_nic
*efx
)
1973 EFX_ASSERT_RESET_SERIALISED(efx
);
1975 /* Skip everything if we never obtained a valid membase */
1979 efx_fini_channels(efx
);
1982 /* Shutdown the board, then the NIC and board state */
1983 efx
->board_info
.fini(efx
);
1984 falcon_fini_interrupt(efx
);
1987 efx_remove_all(efx
);
1990 /* Final NIC shutdown
1991 * This is called only at module unload (or hotplug removal).
1993 static void efx_pci_remove(struct pci_dev
*pci_dev
)
1995 struct efx_nic
*efx
;
1997 efx
= pci_get_drvdata(pci_dev
);
2001 /* Mark the NIC as fini, then stop the interface */
2003 efx
->state
= STATE_FINI
;
2004 dev_close(efx
->net_dev
);
2006 /* Allow any queued efx_resets() to complete */
2009 if (efx
->membase
== NULL
)
2012 efx_unregister_netdev(efx
);
2014 efx_mtd_remove(efx
);
2016 /* Wait for any scheduled resets to complete. No more will be
2017 * scheduled from this point because efx_stop_all() has been
2018 * called, we are no longer registered with driverlink, and
2019 * the net_device's have been removed. */
2020 cancel_work_sync(&efx
->reset_work
);
2022 efx_pci_remove_main(efx
);
2026 EFX_LOG(efx
, "shutdown successful\n");
2028 pci_set_drvdata(pci_dev
, NULL
);
2029 efx_fini_struct(efx
);
2030 free_netdev(efx
->net_dev
);
2033 /* Main body of NIC initialisation
2034 * This is called at module load (or hotplug insertion, theoretically).
2036 static int efx_pci_probe_main(struct efx_nic
*efx
)
2040 /* Do start-of-day initialisation */
2041 rc
= efx_probe_all(efx
);
2045 rc
= efx_init_napi(efx
);
2049 /* Initialise the board */
2050 rc
= efx
->board_info
.init(efx
);
2052 EFX_ERR(efx
, "failed to initialise board\n");
2056 rc
= falcon_init_nic(efx
);
2058 EFX_ERR(efx
, "failed to initialise NIC\n");
2062 rc
= efx_init_port(efx
);
2064 EFX_ERR(efx
, "failed to initialise port\n");
2068 efx_init_channels(efx
);
2070 rc
= falcon_init_interrupt(efx
);
2077 efx_fini_channels(efx
);
2081 efx
->board_info
.fini(efx
);
2085 efx_remove_all(efx
);
2090 /* NIC initialisation
2092 * This is called at module load (or hotplug insertion,
2093 * theoretically). It sets up PCI mappings, tests and resets the NIC,
2094 * sets up and registers the network devices with the kernel and hooks
2095 * the interrupt service routine. It does not prepare the device for
2096 * transmission; this is left to the first time one of the network
2097 * interfaces is brought up (i.e. efx_net_open).
2099 static int __devinit
efx_pci_probe(struct pci_dev
*pci_dev
,
2100 const struct pci_device_id
*entry
)
2102 struct efx_nic_type
*type
= (struct efx_nic_type
*) entry
->driver_data
;
2103 struct net_device
*net_dev
;
2104 struct efx_nic
*efx
;
2107 /* Allocate and initialise a struct net_device and struct efx_nic */
2108 net_dev
= alloc_etherdev(sizeof(*efx
));
2111 net_dev
->features
|= (NETIF_F_IP_CSUM
| NETIF_F_SG
|
2112 NETIF_F_HIGHDMA
| NETIF_F_TSO
);
2114 net_dev
->features
|= NETIF_F_GRO
;
2115 /* Mask for features that also apply to VLAN devices */
2116 net_dev
->vlan_features
|= (NETIF_F_ALL_CSUM
| NETIF_F_SG
|
2117 NETIF_F_HIGHDMA
| NETIF_F_TSO
);
2118 efx
= netdev_priv(net_dev
);
2119 pci_set_drvdata(pci_dev
, efx
);
2120 rc
= efx_init_struct(efx
, type
, pci_dev
, net_dev
);
2124 EFX_INFO(efx
, "Solarflare Communications NIC detected\n");
2126 /* Set up basic I/O (BAR mappings etc) */
2127 rc
= efx_init_io(efx
);
2131 /* No serialisation is required with the reset path because
2132 * we're in STATE_INIT. */
2133 for (i
= 0; i
< 5; i
++) {
2134 rc
= efx_pci_probe_main(efx
);
2136 /* Serialise against efx_reset(). No more resets will be
2137 * scheduled since efx_stop_all() has been called, and we
2138 * have not and never have been registered with either
2139 * the rtnetlink or driverlink layers. */
2140 cancel_work_sync(&efx
->reset_work
);
2143 if (efx
->reset_pending
!= RESET_TYPE_NONE
) {
2144 /* If there was a scheduled reset during
2145 * probe, the NIC is probably hosed anyway */
2146 efx_pci_remove_main(efx
);
2153 /* Retry if a recoverably reset event has been scheduled */
2154 if ((efx
->reset_pending
!= RESET_TYPE_INVISIBLE
) &&
2155 (efx
->reset_pending
!= RESET_TYPE_ALL
))
2158 efx
->reset_pending
= RESET_TYPE_NONE
;
2162 EFX_ERR(efx
, "Could not reset NIC\n");
2166 /* Switch to the running state before we expose the device to
2167 * the OS. This is to ensure that the initial gathering of
2168 * MAC stats succeeds. */
2169 efx
->state
= STATE_RUNNING
;
2171 efx_mtd_probe(efx
); /* allowed to fail */
2173 rc
= efx_register_netdev(efx
);
2177 EFX_LOG(efx
, "initialisation successful\n");
2181 efx_pci_remove_main(efx
);
2186 efx_fini_struct(efx
);
2188 EFX_LOG(efx
, "initialisation failed. rc=%d\n", rc
);
2189 free_netdev(net_dev
);
2193 static struct pci_driver efx_pci_driver
= {
2194 .name
= EFX_DRIVER_NAME
,
2195 .id_table
= efx_pci_table
,
2196 .probe
= efx_pci_probe
,
2197 .remove
= efx_pci_remove
,
2200 /**************************************************************************
2202 * Kernel module interface
2204 *************************************************************************/
2206 module_param(interrupt_mode
, uint
, 0444);
2207 MODULE_PARM_DESC(interrupt_mode
,
2208 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2210 static int __init
efx_init_module(void)
2214 printk(KERN_INFO
"Solarflare NET driver v" EFX_DRIVER_VERSION
"\n");
2216 rc
= register_netdevice_notifier(&efx_netdev_notifier
);
2220 refill_workqueue
= create_workqueue("sfc_refill");
2221 if (!refill_workqueue
) {
2225 reset_workqueue
= create_singlethread_workqueue("sfc_reset");
2226 if (!reset_workqueue
) {
2231 rc
= pci_register_driver(&efx_pci_driver
);
2238 destroy_workqueue(reset_workqueue
);
2240 destroy_workqueue(refill_workqueue
);
2242 unregister_netdevice_notifier(&efx_netdev_notifier
);
2247 static void __exit
efx_exit_module(void)
2249 printk(KERN_INFO
"Solarflare NET driver unloading\n");
2251 pci_unregister_driver(&efx_pci_driver
);
2252 destroy_workqueue(reset_workqueue
);
2253 destroy_workqueue(refill_workqueue
);
2254 unregister_netdevice_notifier(&efx_netdev_notifier
);
2258 module_init(efx_init_module
);
2259 module_exit(efx_exit_module
);
2261 MODULE_AUTHOR("Michael Brown <mbrown@fensystems.co.uk> and "
2262 "Solarflare Communications");
2263 MODULE_DESCRIPTION("Solarflare Communications network driver");
2264 MODULE_LICENSE("GPL");
2265 MODULE_DEVICE_TABLE(pci
, efx_pci_table
);