1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2009 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 <linux/gfp.h>
24 #include "net_driver.h"
30 #include "workarounds.h"
32 /**************************************************************************
36 **************************************************************************
39 /* Loopback mode names (see LOOPBACK_MODE()) */
40 const unsigned int efx_loopback_mode_max
= LOOPBACK_MAX
;
41 const char *efx_loopback_mode_names
[] = {
42 [LOOPBACK_NONE
] = "NONE",
43 [LOOPBACK_DATA
] = "DATAPATH",
44 [LOOPBACK_GMAC
] = "GMAC",
45 [LOOPBACK_XGMII
] = "XGMII",
46 [LOOPBACK_XGXS
] = "XGXS",
47 [LOOPBACK_XAUI
] = "XAUI",
48 [LOOPBACK_GMII
] = "GMII",
49 [LOOPBACK_SGMII
] = "SGMII",
50 [LOOPBACK_XGBR
] = "XGBR",
51 [LOOPBACK_XFI
] = "XFI",
52 [LOOPBACK_XAUI_FAR
] = "XAUI_FAR",
53 [LOOPBACK_GMII_FAR
] = "GMII_FAR",
54 [LOOPBACK_SGMII_FAR
] = "SGMII_FAR",
55 [LOOPBACK_XFI_FAR
] = "XFI_FAR",
56 [LOOPBACK_GPHY
] = "GPHY",
57 [LOOPBACK_PHYXS
] = "PHYXS",
58 [LOOPBACK_PCS
] = "PCS",
59 [LOOPBACK_PMAPMD
] = "PMA/PMD",
60 [LOOPBACK_XPORT
] = "XPORT",
61 [LOOPBACK_XGMII_WS
] = "XGMII_WS",
62 [LOOPBACK_XAUI_WS
] = "XAUI_WS",
63 [LOOPBACK_XAUI_WS_FAR
] = "XAUI_WS_FAR",
64 [LOOPBACK_XAUI_WS_NEAR
] = "XAUI_WS_NEAR",
65 [LOOPBACK_GMII_WS
] = "GMII_WS",
66 [LOOPBACK_XFI_WS
] = "XFI_WS",
67 [LOOPBACK_XFI_WS_FAR
] = "XFI_WS_FAR",
68 [LOOPBACK_PHYXS_WS
] = "PHYXS_WS",
71 /* Interrupt mode names (see INT_MODE())) */
72 const unsigned int efx_interrupt_mode_max
= EFX_INT_MODE_MAX
;
73 const char *efx_interrupt_mode_names
[] = {
74 [EFX_INT_MODE_MSIX
] = "MSI-X",
75 [EFX_INT_MODE_MSI
] = "MSI",
76 [EFX_INT_MODE_LEGACY
] = "legacy",
79 const unsigned int efx_reset_type_max
= RESET_TYPE_MAX
;
80 const char *efx_reset_type_names
[] = {
81 [RESET_TYPE_INVISIBLE
] = "INVISIBLE",
82 [RESET_TYPE_ALL
] = "ALL",
83 [RESET_TYPE_WORLD
] = "WORLD",
84 [RESET_TYPE_DISABLE
] = "DISABLE",
85 [RESET_TYPE_TX_WATCHDOG
] = "TX_WATCHDOG",
86 [RESET_TYPE_INT_ERROR
] = "INT_ERROR",
87 [RESET_TYPE_RX_RECOVERY
] = "RX_RECOVERY",
88 [RESET_TYPE_RX_DESC_FETCH
] = "RX_DESC_FETCH",
89 [RESET_TYPE_TX_DESC_FETCH
] = "TX_DESC_FETCH",
90 [RESET_TYPE_TX_SKIP
] = "TX_SKIP",
91 [RESET_TYPE_MC_FAILURE
] = "MC_FAILURE",
94 #define EFX_MAX_MTU (9 * 1024)
96 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
97 * queued onto this work queue. This is not a per-nic work queue, because
98 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
100 static struct workqueue_struct
*reset_workqueue
;
102 /**************************************************************************
104 * Configurable values
106 *************************************************************************/
109 * Use separate channels for TX and RX events
111 * Set this to 1 to use separate channels for TX and RX. It allows us
112 * to control interrupt affinity separately for TX and RX.
114 * This is only used in MSI-X interrupt mode
116 static unsigned int separate_tx_channels
;
117 module_param(separate_tx_channels
, uint
, 0644);
118 MODULE_PARM_DESC(separate_tx_channels
,
119 "Use separate channels for TX and RX");
121 /* This is the weight assigned to each of the (per-channel) virtual
124 static int napi_weight
= 64;
126 /* This is the time (in jiffies) between invocations of the hardware
127 * monitor, which checks for known hardware bugs and resets the
128 * hardware and driver as necessary.
130 unsigned int efx_monitor_interval
= 1 * HZ
;
132 /* This controls whether or not the driver will initialise devices
133 * with invalid MAC addresses stored in the EEPROM or flash. If true,
134 * such devices will be initialised with a random locally-generated
135 * MAC address. This allows for loading the sfc_mtd driver to
136 * reprogram the flash, even if the flash contents (including the MAC
137 * address) have previously been erased.
139 static unsigned int allow_bad_hwaddr
;
141 /* Initial interrupt moderation settings. They can be modified after
142 * module load with ethtool.
144 * The default for RX should strike a balance between increasing the
145 * round-trip latency and reducing overhead.
147 static unsigned int rx_irq_mod_usec
= 60;
149 /* Initial interrupt moderation settings. They can be modified after
150 * module load with ethtool.
152 * This default is chosen to ensure that a 10G link does not go idle
153 * while a TX queue is stopped after it has become full. A queue is
154 * restarted when it drops below half full. The time this takes (assuming
155 * worst case 3 descriptors per packet and 1024 descriptors) is
156 * 512 / 3 * 1.2 = 205 usec.
158 static unsigned int tx_irq_mod_usec
= 150;
160 /* This is the first interrupt mode to try out of:
165 static unsigned int interrupt_mode
;
167 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
168 * i.e. the number of CPUs among which we may distribute simultaneous
169 * interrupt handling.
171 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
172 * The default (0) means to assign an interrupt to each package (level II cache)
174 static unsigned int rss_cpus
;
175 module_param(rss_cpus
, uint
, 0444);
176 MODULE_PARM_DESC(rss_cpus
, "Number of CPUs to use for Receive-Side Scaling");
178 static int phy_flash_cfg
;
179 module_param(phy_flash_cfg
, int, 0644);
180 MODULE_PARM_DESC(phy_flash_cfg
, "Set PHYs into reflash mode initially");
182 static unsigned irq_adapt_low_thresh
= 10000;
183 module_param(irq_adapt_low_thresh
, uint
, 0644);
184 MODULE_PARM_DESC(irq_adapt_low_thresh
,
185 "Threshold score for reducing IRQ moderation");
187 static unsigned irq_adapt_high_thresh
= 20000;
188 module_param(irq_adapt_high_thresh
, uint
, 0644);
189 MODULE_PARM_DESC(irq_adapt_high_thresh
,
190 "Threshold score for increasing IRQ moderation");
192 static unsigned debug
= (NETIF_MSG_DRV
| NETIF_MSG_PROBE
|
193 NETIF_MSG_LINK
| NETIF_MSG_IFDOWN
|
194 NETIF_MSG_IFUP
| NETIF_MSG_RX_ERR
|
195 NETIF_MSG_TX_ERR
| NETIF_MSG_HW
);
196 module_param(debug
, uint
, 0);
197 MODULE_PARM_DESC(debug
, "Bitmapped debugging message enable value");
199 /**************************************************************************
201 * Utility functions and prototypes
203 *************************************************************************/
204 static void efx_remove_channel(struct efx_channel
*channel
);
205 static void efx_remove_port(struct efx_nic
*efx
);
206 static void efx_fini_napi(struct efx_nic
*efx
);
207 static void efx_fini_channels(struct efx_nic
*efx
);
209 #define EFX_ASSERT_RESET_SERIALISED(efx) \
211 if ((efx->state == STATE_RUNNING) || \
212 (efx->state == STATE_DISABLED)) \
216 /**************************************************************************
218 * Event queue processing
220 *************************************************************************/
222 /* Process channel's event queue
224 * This function is responsible for processing the event queue of a
225 * single channel. The caller must guarantee that this function will
226 * never be concurrently called more than once on the same channel,
227 * though different channels may be being processed concurrently.
229 static int efx_process_channel(struct efx_channel
*channel
, int budget
)
231 struct efx_nic
*efx
= channel
->efx
;
234 if (unlikely(efx
->reset_pending
!= RESET_TYPE_NONE
||
238 spent
= efx_nic_process_eventq(channel
, budget
);
242 /* Deliver last RX packet. */
243 if (channel
->rx_pkt
) {
244 __efx_rx_packet(channel
, channel
->rx_pkt
,
245 channel
->rx_pkt_csummed
);
246 channel
->rx_pkt
= NULL
;
249 efx_rx_strategy(channel
);
251 efx_fast_push_rx_descriptors(&efx
->rx_queue
[channel
->channel
]);
256 /* Mark channel as finished processing
258 * Note that since we will not receive further interrupts for this
259 * channel before we finish processing and call the eventq_read_ack()
260 * method, there is no need to use the interrupt hold-off timers.
262 static inline void efx_channel_processed(struct efx_channel
*channel
)
264 /* The interrupt handler for this channel may set work_pending
265 * as soon as we acknowledge the events we've seen. Make sure
266 * it's cleared before then. */
267 channel
->work_pending
= false;
270 efx_nic_eventq_read_ack(channel
);
275 * NAPI guarantees serialisation of polls of the same device, which
276 * provides the guarantee required by efx_process_channel().
278 static int efx_poll(struct napi_struct
*napi
, int budget
)
280 struct efx_channel
*channel
=
281 container_of(napi
, struct efx_channel
, napi_str
);
282 struct efx_nic
*efx
= channel
->efx
;
285 netif_vdbg(efx
, intr
, efx
->net_dev
,
286 "channel %d NAPI poll executing on CPU %d\n",
287 channel
->channel
, raw_smp_processor_id());
289 spent
= efx_process_channel(channel
, budget
);
291 if (spent
< budget
) {
292 if (channel
->channel
< efx
->n_rx_channels
&&
293 efx
->irq_rx_adaptive
&&
294 unlikely(++channel
->irq_count
== 1000)) {
295 if (unlikely(channel
->irq_mod_score
<
296 irq_adapt_low_thresh
)) {
297 if (channel
->irq_moderation
> 1) {
298 channel
->irq_moderation
-= 1;
299 efx
->type
->push_irq_moderation(channel
);
301 } else if (unlikely(channel
->irq_mod_score
>
302 irq_adapt_high_thresh
)) {
303 if (channel
->irq_moderation
<
304 efx
->irq_rx_moderation
) {
305 channel
->irq_moderation
+= 1;
306 efx
->type
->push_irq_moderation(channel
);
309 channel
->irq_count
= 0;
310 channel
->irq_mod_score
= 0;
313 /* There is no race here; although napi_disable() will
314 * only wait for napi_complete(), this isn't a problem
315 * since efx_channel_processed() will have no effect if
316 * interrupts have already been disabled.
319 efx_channel_processed(channel
);
325 /* Process the eventq of the specified channel immediately on this CPU
327 * Disable hardware generated interrupts, wait for any existing
328 * processing to finish, then directly poll (and ack ) the eventq.
329 * Finally reenable NAPI and interrupts.
331 * Since we are touching interrupts the caller should hold the suspend lock
333 void efx_process_channel_now(struct efx_channel
*channel
)
335 struct efx_nic
*efx
= channel
->efx
;
337 BUG_ON(!channel
->enabled
);
339 /* Disable interrupts and wait for ISRs to complete */
340 efx_nic_disable_interrupts(efx
);
342 synchronize_irq(efx
->legacy_irq
);
344 synchronize_irq(channel
->irq
);
346 /* Wait for any NAPI processing to complete */
347 napi_disable(&channel
->napi_str
);
349 /* Poll the channel */
350 efx_process_channel(channel
, EFX_EVQ_SIZE
);
352 /* Ack the eventq. This may cause an interrupt to be generated
353 * when they are reenabled */
354 efx_channel_processed(channel
);
356 napi_enable(&channel
->napi_str
);
357 efx_nic_enable_interrupts(efx
);
360 /* Create event queue
361 * Event queue memory allocations are done only once. If the channel
362 * is reset, the memory buffer will be reused; this guards against
363 * errors during channel reset and also simplifies interrupt handling.
365 static int efx_probe_eventq(struct efx_channel
*channel
)
367 netif_dbg(channel
->efx
, probe
, channel
->efx
->net_dev
,
368 "chan %d create event queue\n", channel
->channel
);
370 return efx_nic_probe_eventq(channel
);
373 /* Prepare channel's event queue */
374 static void efx_init_eventq(struct efx_channel
*channel
)
376 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
377 "chan %d init event queue\n", channel
->channel
);
379 channel
->eventq_read_ptr
= 0;
381 efx_nic_init_eventq(channel
);
384 static void efx_fini_eventq(struct efx_channel
*channel
)
386 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
387 "chan %d fini event queue\n", channel
->channel
);
389 efx_nic_fini_eventq(channel
);
392 static void efx_remove_eventq(struct efx_channel
*channel
)
394 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
395 "chan %d remove event queue\n", channel
->channel
);
397 efx_nic_remove_eventq(channel
);
400 /**************************************************************************
404 *************************************************************************/
406 static int efx_probe_channel(struct efx_channel
*channel
)
408 struct efx_tx_queue
*tx_queue
;
409 struct efx_rx_queue
*rx_queue
;
412 netif_dbg(channel
->efx
, probe
, channel
->efx
->net_dev
,
413 "creating channel %d\n", channel
->channel
);
415 rc
= efx_probe_eventq(channel
);
419 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
420 rc
= efx_probe_tx_queue(tx_queue
);
425 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
426 rc
= efx_probe_rx_queue(rx_queue
);
431 channel
->n_rx_frm_trunc
= 0;
436 efx_for_each_channel_rx_queue(rx_queue
, channel
)
437 efx_remove_rx_queue(rx_queue
);
439 efx_for_each_channel_tx_queue(tx_queue
, channel
)
440 efx_remove_tx_queue(tx_queue
);
446 static void efx_set_channel_names(struct efx_nic
*efx
)
448 struct efx_channel
*channel
;
449 const char *type
= "";
452 efx_for_each_channel(channel
, efx
) {
453 number
= channel
->channel
;
454 if (efx
->n_channels
> efx
->n_rx_channels
) {
455 if (channel
->channel
< efx
->n_rx_channels
) {
459 number
-= efx
->n_rx_channels
;
462 snprintf(channel
->name
, sizeof(channel
->name
),
463 "%s%s-%d", efx
->name
, type
, number
);
467 /* Channels are shutdown and reinitialised whilst the NIC is running
468 * to propagate configuration changes (mtu, checksum offload), or
469 * to clear hardware error conditions
471 static void efx_init_channels(struct efx_nic
*efx
)
473 struct efx_tx_queue
*tx_queue
;
474 struct efx_rx_queue
*rx_queue
;
475 struct efx_channel
*channel
;
477 /* Calculate the rx buffer allocation parameters required to
478 * support the current MTU, including padding for header
479 * alignment and overruns.
481 efx
->rx_buffer_len
= (max(EFX_PAGE_IP_ALIGN
, NET_IP_ALIGN
) +
482 EFX_MAX_FRAME_LEN(efx
->net_dev
->mtu
) +
483 efx
->type
->rx_buffer_hash_size
+
484 efx
->type
->rx_buffer_padding
);
485 efx
->rx_buffer_order
= get_order(efx
->rx_buffer_len
+
486 sizeof(struct efx_rx_page_state
));
488 /* Initialise the channels */
489 efx_for_each_channel(channel
, efx
) {
490 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
491 "init chan %d\n", channel
->channel
);
493 efx_init_eventq(channel
);
495 efx_for_each_channel_tx_queue(tx_queue
, channel
)
496 efx_init_tx_queue(tx_queue
);
498 /* The rx buffer allocation strategy is MTU dependent */
499 efx_rx_strategy(channel
);
501 efx_for_each_channel_rx_queue(rx_queue
, channel
)
502 efx_init_rx_queue(rx_queue
);
504 WARN_ON(channel
->rx_pkt
!= NULL
);
505 efx_rx_strategy(channel
);
509 /* This enables event queue processing and packet transmission.
511 * Note that this function is not allowed to fail, since that would
512 * introduce too much complexity into the suspend/resume path.
514 static void efx_start_channel(struct efx_channel
*channel
)
516 struct efx_rx_queue
*rx_queue
;
518 netif_dbg(channel
->efx
, ifup
, channel
->efx
->net_dev
,
519 "starting chan %d\n", channel
->channel
);
521 /* The interrupt handler for this channel may set work_pending
522 * as soon as we enable it. Make sure it's cleared before
523 * then. Similarly, make sure it sees the enabled flag set. */
524 channel
->work_pending
= false;
525 channel
->enabled
= true;
528 /* Fill the queues before enabling NAPI */
529 efx_for_each_channel_rx_queue(rx_queue
, channel
)
530 efx_fast_push_rx_descriptors(rx_queue
);
532 napi_enable(&channel
->napi_str
);
535 /* This disables event queue processing and packet transmission.
536 * This function does not guarantee that all queue processing
537 * (e.g. RX refill) is complete.
539 static void efx_stop_channel(struct efx_channel
*channel
)
541 if (!channel
->enabled
)
544 netif_dbg(channel
->efx
, ifdown
, channel
->efx
->net_dev
,
545 "stop chan %d\n", channel
->channel
);
547 channel
->enabled
= false;
548 napi_disable(&channel
->napi_str
);
551 static void efx_fini_channels(struct efx_nic
*efx
)
553 struct efx_channel
*channel
;
554 struct efx_tx_queue
*tx_queue
;
555 struct efx_rx_queue
*rx_queue
;
558 EFX_ASSERT_RESET_SERIALISED(efx
);
559 BUG_ON(efx
->port_enabled
);
561 rc
= efx_nic_flush_queues(efx
);
562 if (rc
&& EFX_WORKAROUND_7803(efx
)) {
563 /* Schedule a reset to recover from the flush failure. The
564 * descriptor caches reference memory we're about to free,
565 * but falcon_reconfigure_mac_wrapper() won't reconnect
566 * the MACs because of the pending reset. */
567 netif_err(efx
, drv
, efx
->net_dev
,
568 "Resetting to recover from flush failure\n");
569 efx_schedule_reset(efx
, RESET_TYPE_ALL
);
571 netif_err(efx
, drv
, efx
->net_dev
, "failed to flush queues\n");
573 netif_dbg(efx
, drv
, efx
->net_dev
,
574 "successfully flushed all queues\n");
577 efx_for_each_channel(channel
, efx
) {
578 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
579 "shut down chan %d\n", channel
->channel
);
581 efx_for_each_channel_rx_queue(rx_queue
, channel
)
582 efx_fini_rx_queue(rx_queue
);
583 efx_for_each_channel_tx_queue(tx_queue
, channel
)
584 efx_fini_tx_queue(tx_queue
);
585 efx_fini_eventq(channel
);
589 static void efx_remove_channel(struct efx_channel
*channel
)
591 struct efx_tx_queue
*tx_queue
;
592 struct efx_rx_queue
*rx_queue
;
594 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
595 "destroy chan %d\n", channel
->channel
);
597 efx_for_each_channel_rx_queue(rx_queue
, channel
)
598 efx_remove_rx_queue(rx_queue
);
599 efx_for_each_channel_tx_queue(tx_queue
, channel
)
600 efx_remove_tx_queue(tx_queue
);
601 efx_remove_eventq(channel
);
604 void efx_schedule_slow_fill(struct efx_rx_queue
*rx_queue
)
606 mod_timer(&rx_queue
->slow_fill
, jiffies
+ msecs_to_jiffies(100));
609 /**************************************************************************
613 **************************************************************************/
615 /* This ensures that the kernel is kept informed (via
616 * netif_carrier_on/off) of the link status, and also maintains the
617 * link status's stop on the port's TX queue.
619 void efx_link_status_changed(struct efx_nic
*efx
)
621 struct efx_link_state
*link_state
= &efx
->link_state
;
623 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
624 * that no events are triggered between unregister_netdev() and the
625 * driver unloading. A more general condition is that NETDEV_CHANGE
626 * can only be generated between NETDEV_UP and NETDEV_DOWN */
627 if (!netif_running(efx
->net_dev
))
630 if (efx
->port_inhibited
) {
631 netif_carrier_off(efx
->net_dev
);
635 if (link_state
->up
!= netif_carrier_ok(efx
->net_dev
)) {
636 efx
->n_link_state_changes
++;
639 netif_carrier_on(efx
->net_dev
);
641 netif_carrier_off(efx
->net_dev
);
644 /* Status message for kernel log */
645 if (link_state
->up
) {
646 netif_info(efx
, link
, efx
->net_dev
,
647 "link up at %uMbps %s-duplex (MTU %d)%s\n",
648 link_state
->speed
, link_state
->fd
? "full" : "half",
650 (efx
->promiscuous
? " [PROMISC]" : ""));
652 netif_info(efx
, link
, efx
->net_dev
, "link down\n");
657 void efx_link_set_advertising(struct efx_nic
*efx
, u32 advertising
)
659 efx
->link_advertising
= advertising
;
661 if (advertising
& ADVERTISED_Pause
)
662 efx
->wanted_fc
|= (EFX_FC_TX
| EFX_FC_RX
);
664 efx
->wanted_fc
&= ~(EFX_FC_TX
| EFX_FC_RX
);
665 if (advertising
& ADVERTISED_Asym_Pause
)
666 efx
->wanted_fc
^= EFX_FC_TX
;
670 void efx_link_set_wanted_fc(struct efx_nic
*efx
, enum efx_fc_type wanted_fc
)
672 efx
->wanted_fc
= wanted_fc
;
673 if (efx
->link_advertising
) {
674 if (wanted_fc
& EFX_FC_RX
)
675 efx
->link_advertising
|= (ADVERTISED_Pause
|
676 ADVERTISED_Asym_Pause
);
678 efx
->link_advertising
&= ~(ADVERTISED_Pause
|
679 ADVERTISED_Asym_Pause
);
680 if (wanted_fc
& EFX_FC_TX
)
681 efx
->link_advertising
^= ADVERTISED_Asym_Pause
;
685 static void efx_fini_port(struct efx_nic
*efx
);
687 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
688 * the MAC appropriately. All other PHY configuration changes are pushed
689 * through phy_op->set_settings(), and pushed asynchronously to the MAC
690 * through efx_monitor().
692 * Callers must hold the mac_lock
694 int __efx_reconfigure_port(struct efx_nic
*efx
)
696 enum efx_phy_mode phy_mode
;
699 WARN_ON(!mutex_is_locked(&efx
->mac_lock
));
701 /* Serialise the promiscuous flag with efx_set_multicast_list. */
702 if (efx_dev_registered(efx
)) {
703 netif_addr_lock_bh(efx
->net_dev
);
704 netif_addr_unlock_bh(efx
->net_dev
);
707 /* Disable PHY transmit in mac level loopbacks */
708 phy_mode
= efx
->phy_mode
;
709 if (LOOPBACK_INTERNAL(efx
))
710 efx
->phy_mode
|= PHY_MODE_TX_DISABLED
;
712 efx
->phy_mode
&= ~PHY_MODE_TX_DISABLED
;
714 rc
= efx
->type
->reconfigure_port(efx
);
717 efx
->phy_mode
= phy_mode
;
722 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
724 int efx_reconfigure_port(struct efx_nic
*efx
)
728 EFX_ASSERT_RESET_SERIALISED(efx
);
730 mutex_lock(&efx
->mac_lock
);
731 rc
= __efx_reconfigure_port(efx
);
732 mutex_unlock(&efx
->mac_lock
);
737 /* Asynchronous work item for changing MAC promiscuity and multicast
738 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
740 static void efx_mac_work(struct work_struct
*data
)
742 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, mac_work
);
744 mutex_lock(&efx
->mac_lock
);
745 if (efx
->port_enabled
) {
746 efx
->type
->push_multicast_hash(efx
);
747 efx
->mac_op
->reconfigure(efx
);
749 mutex_unlock(&efx
->mac_lock
);
752 static int efx_probe_port(struct efx_nic
*efx
)
756 netif_dbg(efx
, probe
, efx
->net_dev
, "create port\n");
759 efx
->phy_mode
= PHY_MODE_SPECIAL
;
761 /* Connect up MAC/PHY operations table */
762 rc
= efx
->type
->probe_port(efx
);
766 /* Sanity check MAC address */
767 if (is_valid_ether_addr(efx
->mac_address
)) {
768 memcpy(efx
->net_dev
->dev_addr
, efx
->mac_address
, ETH_ALEN
);
770 netif_err(efx
, probe
, efx
->net_dev
, "invalid MAC address %pM\n",
772 if (!allow_bad_hwaddr
) {
776 random_ether_addr(efx
->net_dev
->dev_addr
);
777 netif_info(efx
, probe
, efx
->net_dev
,
778 "using locally-generated MAC %pM\n",
779 efx
->net_dev
->dev_addr
);
785 efx_remove_port(efx
);
789 static int efx_init_port(struct efx_nic
*efx
)
793 netif_dbg(efx
, drv
, efx
->net_dev
, "init port\n");
795 mutex_lock(&efx
->mac_lock
);
797 rc
= efx
->phy_op
->init(efx
);
801 efx
->port_initialized
= true;
803 /* Reconfigure the MAC before creating dma queues (required for
804 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
805 efx
->mac_op
->reconfigure(efx
);
807 /* Ensure the PHY advertises the correct flow control settings */
808 rc
= efx
->phy_op
->reconfigure(efx
);
812 mutex_unlock(&efx
->mac_lock
);
816 efx
->phy_op
->fini(efx
);
818 mutex_unlock(&efx
->mac_lock
);
822 static void efx_start_port(struct efx_nic
*efx
)
824 netif_dbg(efx
, ifup
, efx
->net_dev
, "start port\n");
825 BUG_ON(efx
->port_enabled
);
827 mutex_lock(&efx
->mac_lock
);
828 efx
->port_enabled
= true;
830 /* efx_mac_work() might have been scheduled after efx_stop_port(),
831 * and then cancelled by efx_flush_all() */
832 efx
->type
->push_multicast_hash(efx
);
833 efx
->mac_op
->reconfigure(efx
);
835 mutex_unlock(&efx
->mac_lock
);
838 /* Prevent efx_mac_work() and efx_monitor() from working */
839 static void efx_stop_port(struct efx_nic
*efx
)
841 netif_dbg(efx
, ifdown
, efx
->net_dev
, "stop port\n");
843 mutex_lock(&efx
->mac_lock
);
844 efx
->port_enabled
= false;
845 mutex_unlock(&efx
->mac_lock
);
847 /* Serialise against efx_set_multicast_list() */
848 if (efx_dev_registered(efx
)) {
849 netif_addr_lock_bh(efx
->net_dev
);
850 netif_addr_unlock_bh(efx
->net_dev
);
854 static void efx_fini_port(struct efx_nic
*efx
)
856 netif_dbg(efx
, drv
, efx
->net_dev
, "shut down port\n");
858 if (!efx
->port_initialized
)
861 efx
->phy_op
->fini(efx
);
862 efx
->port_initialized
= false;
864 efx
->link_state
.up
= false;
865 efx_link_status_changed(efx
);
868 static void efx_remove_port(struct efx_nic
*efx
)
870 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying port\n");
872 efx
->type
->remove_port(efx
);
875 /**************************************************************************
879 **************************************************************************/
881 /* This configures the PCI device to enable I/O and DMA. */
882 static int efx_init_io(struct efx_nic
*efx
)
884 struct pci_dev
*pci_dev
= efx
->pci_dev
;
885 dma_addr_t dma_mask
= efx
->type
->max_dma_mask
;
888 netif_dbg(efx
, probe
, efx
->net_dev
, "initialising I/O\n");
890 rc
= pci_enable_device(pci_dev
);
892 netif_err(efx
, probe
, efx
->net_dev
,
893 "failed to enable PCI device\n");
897 pci_set_master(pci_dev
);
899 /* Set the PCI DMA mask. Try all possibilities from our
900 * genuine mask down to 32 bits, because some architectures
901 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
902 * masks event though they reject 46 bit masks.
904 while (dma_mask
> 0x7fffffffUL
) {
905 if (pci_dma_supported(pci_dev
, dma_mask
) &&
906 ((rc
= pci_set_dma_mask(pci_dev
, dma_mask
)) == 0))
911 netif_err(efx
, probe
, efx
->net_dev
,
912 "could not find a suitable DMA mask\n");
915 netif_dbg(efx
, probe
, efx
->net_dev
,
916 "using DMA mask %llx\n", (unsigned long long) dma_mask
);
917 rc
= pci_set_consistent_dma_mask(pci_dev
, dma_mask
);
919 /* pci_set_consistent_dma_mask() is not *allowed* to
920 * fail with a mask that pci_set_dma_mask() accepted,
921 * but just in case...
923 netif_err(efx
, probe
, efx
->net_dev
,
924 "failed to set consistent DMA mask\n");
928 efx
->membase_phys
= pci_resource_start(efx
->pci_dev
, EFX_MEM_BAR
);
929 rc
= pci_request_region(pci_dev
, EFX_MEM_BAR
, "sfc");
931 netif_err(efx
, probe
, efx
->net_dev
,
932 "request for memory BAR failed\n");
936 efx
->membase
= ioremap_nocache(efx
->membase_phys
,
937 efx
->type
->mem_map_size
);
939 netif_err(efx
, probe
, efx
->net_dev
,
940 "could not map memory BAR at %llx+%x\n",
941 (unsigned long long)efx
->membase_phys
,
942 efx
->type
->mem_map_size
);
946 netif_dbg(efx
, probe
, efx
->net_dev
,
947 "memory BAR at %llx+%x (virtual %p)\n",
948 (unsigned long long)efx
->membase_phys
,
949 efx
->type
->mem_map_size
, efx
->membase
);
954 pci_release_region(efx
->pci_dev
, EFX_MEM_BAR
);
956 efx
->membase_phys
= 0;
958 pci_disable_device(efx
->pci_dev
);
963 static void efx_fini_io(struct efx_nic
*efx
)
965 netif_dbg(efx
, drv
, efx
->net_dev
, "shutting down I/O\n");
968 iounmap(efx
->membase
);
972 if (efx
->membase_phys
) {
973 pci_release_region(efx
->pci_dev
, EFX_MEM_BAR
);
974 efx
->membase_phys
= 0;
977 pci_disable_device(efx
->pci_dev
);
980 /* Get number of channels wanted. Each channel will have its own IRQ,
981 * 1 RX queue and/or 2 TX queues. */
982 static int efx_wanted_channels(void)
984 cpumask_var_t core_mask
;
988 if (unlikely(!zalloc_cpumask_var(&core_mask
, GFP_KERNEL
))) {
990 "sfc: RSS disabled due to allocation failure\n");
995 for_each_online_cpu(cpu
) {
996 if (!cpumask_test_cpu(cpu
, core_mask
)) {
998 cpumask_or(core_mask
, core_mask
,
999 topology_core_cpumask(cpu
));
1003 free_cpumask_var(core_mask
);
1007 /* Probe the number and type of interrupts we are able to obtain, and
1008 * the resulting numbers of channels and RX queues.
1010 static void efx_probe_interrupts(struct efx_nic
*efx
)
1013 min_t(int, efx
->type
->phys_addr_channels
, EFX_MAX_CHANNELS
);
1016 if (efx
->interrupt_mode
== EFX_INT_MODE_MSIX
) {
1017 struct msix_entry xentries
[EFX_MAX_CHANNELS
];
1020 n_channels
= efx_wanted_channels();
1021 if (separate_tx_channels
)
1023 n_channels
= min(n_channels
, max_channels
);
1025 for (i
= 0; i
< n_channels
; i
++)
1026 xentries
[i
].entry
= i
;
1027 rc
= pci_enable_msix(efx
->pci_dev
, xentries
, n_channels
);
1029 netif_err(efx
, drv
, efx
->net_dev
,
1030 "WARNING: Insufficient MSI-X vectors"
1031 " available (%d < %d).\n", rc
, n_channels
);
1032 netif_err(efx
, drv
, efx
->net_dev
,
1033 "WARNING: Performance may be reduced.\n");
1034 EFX_BUG_ON_PARANOID(rc
>= n_channels
);
1036 rc
= pci_enable_msix(efx
->pci_dev
, xentries
,
1041 efx
->n_channels
= n_channels
;
1042 if (separate_tx_channels
) {
1043 efx
->n_tx_channels
=
1044 max(efx
->n_channels
/ 2, 1U);
1045 efx
->n_rx_channels
=
1046 max(efx
->n_channels
-
1047 efx
->n_tx_channels
, 1U);
1049 efx
->n_tx_channels
= efx
->n_channels
;
1050 efx
->n_rx_channels
= efx
->n_channels
;
1052 for (i
= 0; i
< n_channels
; i
++)
1053 efx
->channel
[i
].irq
= xentries
[i
].vector
;
1055 /* Fall back to single channel MSI */
1056 efx
->interrupt_mode
= EFX_INT_MODE_MSI
;
1057 netif_err(efx
, drv
, efx
->net_dev
,
1058 "could not enable MSI-X\n");
1062 /* Try single interrupt MSI */
1063 if (efx
->interrupt_mode
== EFX_INT_MODE_MSI
) {
1064 efx
->n_channels
= 1;
1065 efx
->n_rx_channels
= 1;
1066 efx
->n_tx_channels
= 1;
1067 rc
= pci_enable_msi(efx
->pci_dev
);
1069 efx
->channel
[0].irq
= efx
->pci_dev
->irq
;
1071 netif_err(efx
, drv
, efx
->net_dev
,
1072 "could not enable MSI\n");
1073 efx
->interrupt_mode
= EFX_INT_MODE_LEGACY
;
1077 /* Assume legacy interrupts */
1078 if (efx
->interrupt_mode
== EFX_INT_MODE_LEGACY
) {
1079 efx
->n_channels
= 1 + (separate_tx_channels
? 1 : 0);
1080 efx
->n_rx_channels
= 1;
1081 efx
->n_tx_channels
= 1;
1082 efx
->legacy_irq
= efx
->pci_dev
->irq
;
1086 static void efx_remove_interrupts(struct efx_nic
*efx
)
1088 struct efx_channel
*channel
;
1090 /* Remove MSI/MSI-X interrupts */
1091 efx_for_each_channel(channel
, efx
)
1093 pci_disable_msi(efx
->pci_dev
);
1094 pci_disable_msix(efx
->pci_dev
);
1096 /* Remove legacy interrupt */
1097 efx
->legacy_irq
= 0;
1100 static void efx_set_channels(struct efx_nic
*efx
)
1102 struct efx_channel
*channel
;
1103 struct efx_tx_queue
*tx_queue
;
1104 struct efx_rx_queue
*rx_queue
;
1105 unsigned tx_channel_offset
=
1106 separate_tx_channels
? efx
->n_channels
- efx
->n_tx_channels
: 0;
1108 efx_for_each_channel(channel
, efx
) {
1109 if (channel
->channel
- tx_channel_offset
< efx
->n_tx_channels
) {
1110 channel
->tx_queue
= &efx
->tx_queue
[
1111 (channel
->channel
- tx_channel_offset
) *
1113 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1114 tx_queue
->channel
= channel
;
1118 efx_for_each_rx_queue(rx_queue
, efx
)
1119 rx_queue
->channel
= &efx
->channel
[rx_queue
->queue
];
1122 static int efx_probe_nic(struct efx_nic
*efx
)
1126 netif_dbg(efx
, probe
, efx
->net_dev
, "creating NIC\n");
1128 /* Carry out hardware-type specific initialisation */
1129 rc
= efx
->type
->probe(efx
);
1133 /* Determine the number of channels and queues by trying to hook
1134 * in MSI-X interrupts. */
1135 efx_probe_interrupts(efx
);
1137 efx_set_channels(efx
);
1138 efx
->net_dev
->real_num_tx_queues
= efx
->n_tx_channels
;
1140 /* Initialise the interrupt moderation settings */
1141 efx_init_irq_moderation(efx
, tx_irq_mod_usec
, rx_irq_mod_usec
, true);
1146 static void efx_remove_nic(struct efx_nic
*efx
)
1148 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying NIC\n");
1150 efx_remove_interrupts(efx
);
1151 efx
->type
->remove(efx
);
1154 /**************************************************************************
1156 * NIC startup/shutdown
1158 *************************************************************************/
1160 static int efx_probe_all(struct efx_nic
*efx
)
1162 struct efx_channel
*channel
;
1166 rc
= efx_probe_nic(efx
);
1168 netif_err(efx
, probe
, efx
->net_dev
, "failed to create NIC\n");
1173 rc
= efx_probe_port(efx
);
1175 netif_err(efx
, probe
, efx
->net_dev
, "failed to create port\n");
1179 /* Create channels */
1180 efx_for_each_channel(channel
, efx
) {
1181 rc
= efx_probe_channel(channel
);
1183 netif_err(efx
, probe
, efx
->net_dev
,
1184 "failed to create channel %d\n",
1189 efx_set_channel_names(efx
);
1194 efx_for_each_channel(channel
, efx
)
1195 efx_remove_channel(channel
);
1196 efx_remove_port(efx
);
1198 efx_remove_nic(efx
);
1203 /* Called after previous invocation(s) of efx_stop_all, restarts the
1204 * port, kernel transmit queue, NAPI processing and hardware interrupts,
1205 * and ensures that the port is scheduled to be reconfigured.
1206 * This function is safe to call multiple times when the NIC is in any
1208 static void efx_start_all(struct efx_nic
*efx
)
1210 struct efx_channel
*channel
;
1212 EFX_ASSERT_RESET_SERIALISED(efx
);
1214 /* Check that it is appropriate to restart the interface. All
1215 * of these flags are safe to read under just the rtnl lock */
1216 if (efx
->port_enabled
)
1218 if ((efx
->state
!= STATE_RUNNING
) && (efx
->state
!= STATE_INIT
))
1220 if (efx_dev_registered(efx
) && !netif_running(efx
->net_dev
))
1223 /* Mark the port as enabled so port reconfigurations can start, then
1224 * restart the transmit interface early so the watchdog timer stops */
1225 efx_start_port(efx
);
1227 efx_for_each_channel(channel
, efx
) {
1228 if (efx_dev_registered(efx
))
1229 efx_wake_queue(channel
);
1230 efx_start_channel(channel
);
1233 efx_nic_enable_interrupts(efx
);
1235 /* Switch to event based MCDI completions after enabling interrupts.
1236 * If a reset has been scheduled, then we need to stay in polled mode.
1237 * Rather than serialising efx_mcdi_mode_event() [which sleeps] and
1238 * reset_pending [modified from an atomic context], we instead guarantee
1239 * that efx_mcdi_mode_poll() isn't reverted erroneously */
1240 efx_mcdi_mode_event(efx
);
1241 if (efx
->reset_pending
!= RESET_TYPE_NONE
)
1242 efx_mcdi_mode_poll(efx
);
1244 /* Start the hardware monitor if there is one. Otherwise (we're link
1245 * event driven), we have to poll the PHY because after an event queue
1246 * flush, we could have a missed a link state change */
1247 if (efx
->type
->monitor
!= NULL
) {
1248 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1249 efx_monitor_interval
);
1251 mutex_lock(&efx
->mac_lock
);
1252 if (efx
->phy_op
->poll(efx
))
1253 efx_link_status_changed(efx
);
1254 mutex_unlock(&efx
->mac_lock
);
1257 efx
->type
->start_stats(efx
);
1260 /* Flush all delayed work. Should only be called when no more delayed work
1261 * will be scheduled. This doesn't flush pending online resets (efx_reset),
1262 * since we're holding the rtnl_lock at this point. */
1263 static void efx_flush_all(struct efx_nic
*efx
)
1265 /* Make sure the hardware monitor is stopped */
1266 cancel_delayed_work_sync(&efx
->monitor_work
);
1267 /* Stop scheduled port reconfigurations */
1268 cancel_work_sync(&efx
->mac_work
);
1271 /* Quiesce hardware and software without bringing the link down.
1272 * Safe to call multiple times, when the nic and interface is in any
1273 * state. The caller is guaranteed to subsequently be in a position
1274 * to modify any hardware and software state they see fit without
1276 static void efx_stop_all(struct efx_nic
*efx
)
1278 struct efx_channel
*channel
;
1280 EFX_ASSERT_RESET_SERIALISED(efx
);
1282 /* port_enabled can be read safely under the rtnl lock */
1283 if (!efx
->port_enabled
)
1286 efx
->type
->stop_stats(efx
);
1288 /* Switch to MCDI polling on Siena before disabling interrupts */
1289 efx_mcdi_mode_poll(efx
);
1291 /* Disable interrupts and wait for ISR to complete */
1292 efx_nic_disable_interrupts(efx
);
1293 if (efx
->legacy_irq
)
1294 synchronize_irq(efx
->legacy_irq
);
1295 efx_for_each_channel(channel
, efx
) {
1297 synchronize_irq(channel
->irq
);
1300 /* Stop all NAPI processing and synchronous rx refills */
1301 efx_for_each_channel(channel
, efx
)
1302 efx_stop_channel(channel
);
1304 /* Stop all asynchronous port reconfigurations. Since all
1305 * event processing has already been stopped, there is no
1306 * window to loose phy events */
1309 /* Flush efx_mac_work(), refill_workqueue, monitor_work */
1312 /* Stop the kernel transmit interface late, so the watchdog
1313 * timer isn't ticking over the flush */
1314 if (efx_dev_registered(efx
)) {
1315 struct efx_channel
*channel
;
1316 efx_for_each_channel(channel
, efx
)
1317 efx_stop_queue(channel
);
1318 netif_tx_lock_bh(efx
->net_dev
);
1319 netif_tx_unlock_bh(efx
->net_dev
);
1323 static void efx_remove_all(struct efx_nic
*efx
)
1325 struct efx_channel
*channel
;
1327 efx_for_each_channel(channel
, efx
)
1328 efx_remove_channel(channel
);
1329 efx_remove_port(efx
);
1330 efx_remove_nic(efx
);
1333 /**************************************************************************
1335 * Interrupt moderation
1337 **************************************************************************/
1339 static unsigned irq_mod_ticks(int usecs
, int resolution
)
1342 return 0; /* cannot receive interrupts ahead of time :-) */
1343 if (usecs
< resolution
)
1344 return 1; /* never round down to 0 */
1345 return usecs
/ resolution
;
1348 /* Set interrupt moderation parameters */
1349 void efx_init_irq_moderation(struct efx_nic
*efx
, int tx_usecs
, int rx_usecs
,
1352 struct efx_tx_queue
*tx_queue
;
1353 struct efx_rx_queue
*rx_queue
;
1354 unsigned tx_ticks
= irq_mod_ticks(tx_usecs
, EFX_IRQ_MOD_RESOLUTION
);
1355 unsigned rx_ticks
= irq_mod_ticks(rx_usecs
, EFX_IRQ_MOD_RESOLUTION
);
1357 EFX_ASSERT_RESET_SERIALISED(efx
);
1359 efx_for_each_tx_queue(tx_queue
, efx
)
1360 tx_queue
->channel
->irq_moderation
= tx_ticks
;
1362 efx
->irq_rx_adaptive
= rx_adaptive
;
1363 efx
->irq_rx_moderation
= rx_ticks
;
1364 efx_for_each_rx_queue(rx_queue
, efx
)
1365 rx_queue
->channel
->irq_moderation
= rx_ticks
;
1368 /**************************************************************************
1372 **************************************************************************/
1374 /* Run periodically off the general workqueue. Serialised against
1375 * efx_reconfigure_port via the mac_lock */
1376 static void efx_monitor(struct work_struct
*data
)
1378 struct efx_nic
*efx
= container_of(data
, struct efx_nic
,
1381 netif_vdbg(efx
, timer
, efx
->net_dev
,
1382 "hardware monitor executing on CPU %d\n",
1383 raw_smp_processor_id());
1384 BUG_ON(efx
->type
->monitor
== NULL
);
1386 /* If the mac_lock is already held then it is likely a port
1387 * reconfiguration is already in place, which will likely do
1388 * most of the work of check_hw() anyway. */
1389 if (!mutex_trylock(&efx
->mac_lock
))
1391 if (!efx
->port_enabled
)
1393 efx
->type
->monitor(efx
);
1396 mutex_unlock(&efx
->mac_lock
);
1398 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1399 efx_monitor_interval
);
1402 /**************************************************************************
1406 *************************************************************************/
1409 * Context: process, rtnl_lock() held.
1411 static int efx_ioctl(struct net_device
*net_dev
, struct ifreq
*ifr
, int cmd
)
1413 struct efx_nic
*efx
= netdev_priv(net_dev
);
1414 struct mii_ioctl_data
*data
= if_mii(ifr
);
1416 EFX_ASSERT_RESET_SERIALISED(efx
);
1418 /* Convert phy_id from older PRTAD/DEVAD format */
1419 if ((cmd
== SIOCGMIIREG
|| cmd
== SIOCSMIIREG
) &&
1420 (data
->phy_id
& 0xfc00) == 0x0400)
1421 data
->phy_id
^= MDIO_PHY_ID_C45
| 0x0400;
1423 return mdio_mii_ioctl(&efx
->mdio
, data
, cmd
);
1426 /**************************************************************************
1430 **************************************************************************/
1432 static int efx_init_napi(struct efx_nic
*efx
)
1434 struct efx_channel
*channel
;
1436 efx_for_each_channel(channel
, efx
) {
1437 channel
->napi_dev
= efx
->net_dev
;
1438 netif_napi_add(channel
->napi_dev
, &channel
->napi_str
,
1439 efx_poll
, napi_weight
);
1444 static void efx_fini_napi(struct efx_nic
*efx
)
1446 struct efx_channel
*channel
;
1448 efx_for_each_channel(channel
, efx
) {
1449 if (channel
->napi_dev
)
1450 netif_napi_del(&channel
->napi_str
);
1451 channel
->napi_dev
= NULL
;
1455 /**************************************************************************
1457 * Kernel netpoll interface
1459 *************************************************************************/
1461 #ifdef CONFIG_NET_POLL_CONTROLLER
1463 /* Although in the common case interrupts will be disabled, this is not
1464 * guaranteed. However, all our work happens inside the NAPI callback,
1465 * so no locking is required.
1467 static void efx_netpoll(struct net_device
*net_dev
)
1469 struct efx_nic
*efx
= netdev_priv(net_dev
);
1470 struct efx_channel
*channel
;
1472 efx_for_each_channel(channel
, efx
)
1473 efx_schedule_channel(channel
);
1478 /**************************************************************************
1480 * Kernel net device interface
1482 *************************************************************************/
1484 /* Context: process, rtnl_lock() held. */
1485 static int efx_net_open(struct net_device
*net_dev
)
1487 struct efx_nic
*efx
= netdev_priv(net_dev
);
1488 EFX_ASSERT_RESET_SERIALISED(efx
);
1490 netif_dbg(efx
, ifup
, efx
->net_dev
, "opening device on CPU %d\n",
1491 raw_smp_processor_id());
1493 if (efx
->state
== STATE_DISABLED
)
1495 if (efx
->phy_mode
& PHY_MODE_SPECIAL
)
1497 if (efx_mcdi_poll_reboot(efx
) && efx_reset(efx
, RESET_TYPE_ALL
))
1500 /* Notify the kernel of the link state polled during driver load,
1501 * before the monitor starts running */
1502 efx_link_status_changed(efx
);
1508 /* Context: process, rtnl_lock() held.
1509 * Note that the kernel will ignore our return code; this method
1510 * should really be a void.
1512 static int efx_net_stop(struct net_device
*net_dev
)
1514 struct efx_nic
*efx
= netdev_priv(net_dev
);
1516 netif_dbg(efx
, ifdown
, efx
->net_dev
, "closing on CPU %d\n",
1517 raw_smp_processor_id());
1519 if (efx
->state
!= STATE_DISABLED
) {
1520 /* Stop the device and flush all the channels */
1522 efx_fini_channels(efx
);
1523 efx_init_channels(efx
);
1529 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1530 static struct rtnl_link_stats64
*efx_net_stats(struct net_device
*net_dev
)
1532 struct efx_nic
*efx
= netdev_priv(net_dev
);
1533 struct efx_mac_stats
*mac_stats
= &efx
->mac_stats
;
1534 struct rtnl_link_stats64
*stats
= &net_dev
->stats64
;
1536 spin_lock_bh(&efx
->stats_lock
);
1537 efx
->type
->update_stats(efx
);
1538 spin_unlock_bh(&efx
->stats_lock
);
1540 stats
->rx_packets
= mac_stats
->rx_packets
;
1541 stats
->tx_packets
= mac_stats
->tx_packets
;
1542 stats
->rx_bytes
= mac_stats
->rx_bytes
;
1543 stats
->tx_bytes
= mac_stats
->tx_bytes
;
1544 stats
->multicast
= mac_stats
->rx_multicast
;
1545 stats
->collisions
= mac_stats
->tx_collision
;
1546 stats
->rx_length_errors
= (mac_stats
->rx_gtjumbo
+
1547 mac_stats
->rx_length_error
);
1548 stats
->rx_over_errors
= efx
->n_rx_nodesc_drop_cnt
;
1549 stats
->rx_crc_errors
= mac_stats
->rx_bad
;
1550 stats
->rx_frame_errors
= mac_stats
->rx_align_error
;
1551 stats
->rx_fifo_errors
= mac_stats
->rx_overflow
;
1552 stats
->rx_missed_errors
= mac_stats
->rx_missed
;
1553 stats
->tx_window_errors
= mac_stats
->tx_late_collision
;
1555 stats
->rx_errors
= (stats
->rx_length_errors
+
1556 stats
->rx_crc_errors
+
1557 stats
->rx_frame_errors
+
1558 mac_stats
->rx_symbol_error
);
1559 stats
->tx_errors
= (stats
->tx_window_errors
+
1565 /* Context: netif_tx_lock held, BHs disabled. */
1566 static void efx_watchdog(struct net_device
*net_dev
)
1568 struct efx_nic
*efx
= netdev_priv(net_dev
);
1570 netif_err(efx
, tx_err
, efx
->net_dev
,
1571 "TX stuck with port_enabled=%d: resetting channels\n",
1574 efx_schedule_reset(efx
, RESET_TYPE_TX_WATCHDOG
);
1578 /* Context: process, rtnl_lock() held. */
1579 static int efx_change_mtu(struct net_device
*net_dev
, int new_mtu
)
1581 struct efx_nic
*efx
= netdev_priv(net_dev
);
1584 EFX_ASSERT_RESET_SERIALISED(efx
);
1586 if (new_mtu
> EFX_MAX_MTU
)
1591 netif_dbg(efx
, drv
, efx
->net_dev
, "changing MTU to %d\n", new_mtu
);
1593 efx_fini_channels(efx
);
1595 mutex_lock(&efx
->mac_lock
);
1596 /* Reconfigure the MAC before enabling the dma queues so that
1597 * the RX buffers don't overflow */
1598 net_dev
->mtu
= new_mtu
;
1599 efx
->mac_op
->reconfigure(efx
);
1600 mutex_unlock(&efx
->mac_lock
);
1602 efx_init_channels(efx
);
1608 static int efx_set_mac_address(struct net_device
*net_dev
, void *data
)
1610 struct efx_nic
*efx
= netdev_priv(net_dev
);
1611 struct sockaddr
*addr
= data
;
1612 char *new_addr
= addr
->sa_data
;
1614 EFX_ASSERT_RESET_SERIALISED(efx
);
1616 if (!is_valid_ether_addr(new_addr
)) {
1617 netif_err(efx
, drv
, efx
->net_dev
,
1618 "invalid ethernet MAC address requested: %pM\n",
1623 memcpy(net_dev
->dev_addr
, new_addr
, net_dev
->addr_len
);
1625 /* Reconfigure the MAC */
1626 mutex_lock(&efx
->mac_lock
);
1627 efx
->mac_op
->reconfigure(efx
);
1628 mutex_unlock(&efx
->mac_lock
);
1633 /* Context: netif_addr_lock held, BHs disabled. */
1634 static void efx_set_multicast_list(struct net_device
*net_dev
)
1636 struct efx_nic
*efx
= netdev_priv(net_dev
);
1637 struct netdev_hw_addr
*ha
;
1638 union efx_multicast_hash
*mc_hash
= &efx
->multicast_hash
;
1642 efx
->promiscuous
= !!(net_dev
->flags
& IFF_PROMISC
);
1644 /* Build multicast hash table */
1645 if (efx
->promiscuous
|| (net_dev
->flags
& IFF_ALLMULTI
)) {
1646 memset(mc_hash
, 0xff, sizeof(*mc_hash
));
1648 memset(mc_hash
, 0x00, sizeof(*mc_hash
));
1649 netdev_for_each_mc_addr(ha
, net_dev
) {
1650 crc
= ether_crc_le(ETH_ALEN
, ha
->addr
);
1651 bit
= crc
& (EFX_MCAST_HASH_ENTRIES
- 1);
1652 set_bit_le(bit
, mc_hash
->byte
);
1655 /* Broadcast packets go through the multicast hash filter.
1656 * ether_crc_le() of the broadcast address is 0xbe2612ff
1657 * so we always add bit 0xff to the mask.
1659 set_bit_le(0xff, mc_hash
->byte
);
1662 if (efx
->port_enabled
)
1663 queue_work(efx
->workqueue
, &efx
->mac_work
);
1664 /* Otherwise efx_start_port() will do this */
1667 static const struct net_device_ops efx_netdev_ops
= {
1668 .ndo_open
= efx_net_open
,
1669 .ndo_stop
= efx_net_stop
,
1670 .ndo_get_stats64
= efx_net_stats
,
1671 .ndo_tx_timeout
= efx_watchdog
,
1672 .ndo_start_xmit
= efx_hard_start_xmit
,
1673 .ndo_validate_addr
= eth_validate_addr
,
1674 .ndo_do_ioctl
= efx_ioctl
,
1675 .ndo_change_mtu
= efx_change_mtu
,
1676 .ndo_set_mac_address
= efx_set_mac_address
,
1677 .ndo_set_multicast_list
= efx_set_multicast_list
,
1678 #ifdef CONFIG_NET_POLL_CONTROLLER
1679 .ndo_poll_controller
= efx_netpoll
,
1683 static void efx_update_name(struct efx_nic
*efx
)
1685 strcpy(efx
->name
, efx
->net_dev
->name
);
1686 efx_mtd_rename(efx
);
1687 efx_set_channel_names(efx
);
1690 static int efx_netdev_event(struct notifier_block
*this,
1691 unsigned long event
, void *ptr
)
1693 struct net_device
*net_dev
= ptr
;
1695 if (net_dev
->netdev_ops
== &efx_netdev_ops
&&
1696 event
== NETDEV_CHANGENAME
)
1697 efx_update_name(netdev_priv(net_dev
));
1702 static struct notifier_block efx_netdev_notifier
= {
1703 .notifier_call
= efx_netdev_event
,
1707 show_phy_type(struct device
*dev
, struct device_attribute
*attr
, char *buf
)
1709 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
1710 return sprintf(buf
, "%d\n", efx
->phy_type
);
1712 static DEVICE_ATTR(phy_type
, 0644, show_phy_type
, NULL
);
1714 static int efx_register_netdev(struct efx_nic
*efx
)
1716 struct net_device
*net_dev
= efx
->net_dev
;
1719 net_dev
->watchdog_timeo
= 5 * HZ
;
1720 net_dev
->irq
= efx
->pci_dev
->irq
;
1721 net_dev
->netdev_ops
= &efx_netdev_ops
;
1722 SET_ETHTOOL_OPS(net_dev
, &efx_ethtool_ops
);
1724 /* Clear MAC statistics */
1725 efx
->mac_op
->update_stats(efx
);
1726 memset(&efx
->mac_stats
, 0, sizeof(efx
->mac_stats
));
1730 rc
= dev_alloc_name(net_dev
, net_dev
->name
);
1733 efx_update_name(efx
);
1735 rc
= register_netdevice(net_dev
);
1739 /* Always start with carrier off; PHY events will detect the link */
1740 netif_carrier_off(efx
->net_dev
);
1744 rc
= device_create_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1746 netif_err(efx
, drv
, efx
->net_dev
,
1747 "failed to init net dev attributes\n");
1748 goto fail_registered
;
1755 netif_err(efx
, drv
, efx
->net_dev
, "could not register net dev\n");
1759 unregister_netdev(net_dev
);
1763 static void efx_unregister_netdev(struct efx_nic
*efx
)
1765 struct efx_tx_queue
*tx_queue
;
1770 BUG_ON(netdev_priv(efx
->net_dev
) != efx
);
1772 /* Free up any skbs still remaining. This has to happen before
1773 * we try to unregister the netdev as running their destructors
1774 * may be needed to get the device ref. count to 0. */
1775 efx_for_each_tx_queue(tx_queue
, efx
)
1776 efx_release_tx_buffers(tx_queue
);
1778 if (efx_dev_registered(efx
)) {
1779 strlcpy(efx
->name
, pci_name(efx
->pci_dev
), sizeof(efx
->name
));
1780 device_remove_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1781 unregister_netdev(efx
->net_dev
);
1785 /**************************************************************************
1787 * Device reset and suspend
1789 **************************************************************************/
1791 /* Tears down the entire software state and most of the hardware state
1793 void efx_reset_down(struct efx_nic
*efx
, enum reset_type method
)
1795 EFX_ASSERT_RESET_SERIALISED(efx
);
1798 mutex_lock(&efx
->mac_lock
);
1799 mutex_lock(&efx
->spi_lock
);
1801 efx_fini_channels(efx
);
1802 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
)
1803 efx
->phy_op
->fini(efx
);
1804 efx
->type
->fini(efx
);
1807 /* This function will always ensure that the locks acquired in
1808 * efx_reset_down() are released. A failure return code indicates
1809 * that we were unable to reinitialise the hardware, and the
1810 * driver should be disabled. If ok is false, then the rx and tx
1811 * engines are not restarted, pending a RESET_DISABLE. */
1812 int efx_reset_up(struct efx_nic
*efx
, enum reset_type method
, bool ok
)
1816 EFX_ASSERT_RESET_SERIALISED(efx
);
1818 rc
= efx
->type
->init(efx
);
1820 netif_err(efx
, drv
, efx
->net_dev
, "failed to initialise NIC\n");
1827 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
) {
1828 rc
= efx
->phy_op
->init(efx
);
1831 if (efx
->phy_op
->reconfigure(efx
))
1832 netif_err(efx
, drv
, efx
->net_dev
,
1833 "could not restore PHY settings\n");
1836 efx
->mac_op
->reconfigure(efx
);
1838 efx_init_channels(efx
);
1840 mutex_unlock(&efx
->spi_lock
);
1841 mutex_unlock(&efx
->mac_lock
);
1848 efx
->port_initialized
= false;
1850 mutex_unlock(&efx
->spi_lock
);
1851 mutex_unlock(&efx
->mac_lock
);
1856 /* Reset the NIC using the specified method. Note that the reset may
1857 * fail, in which case the card will be left in an unusable state.
1859 * Caller must hold the rtnl_lock.
1861 int efx_reset(struct efx_nic
*efx
, enum reset_type method
)
1866 netif_info(efx
, drv
, efx
->net_dev
, "resetting (%s)\n",
1867 RESET_TYPE(method
));
1869 efx_reset_down(efx
, method
);
1871 rc
= efx
->type
->reset(efx
, method
);
1873 netif_err(efx
, drv
, efx
->net_dev
, "failed to reset hardware\n");
1877 /* Allow resets to be rescheduled. */
1878 efx
->reset_pending
= RESET_TYPE_NONE
;
1880 /* Reinitialise bus-mastering, which may have been turned off before
1881 * the reset was scheduled. This is still appropriate, even in the
1882 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
1883 * can respond to requests. */
1884 pci_set_master(efx
->pci_dev
);
1887 /* Leave device stopped if necessary */
1888 disabled
= rc
|| method
== RESET_TYPE_DISABLE
;
1889 rc2
= efx_reset_up(efx
, method
, !disabled
);
1897 dev_close(efx
->net_dev
);
1898 netif_err(efx
, drv
, efx
->net_dev
, "has been disabled\n");
1899 efx
->state
= STATE_DISABLED
;
1901 netif_dbg(efx
, drv
, efx
->net_dev
, "reset complete\n");
1906 /* The worker thread exists so that code that cannot sleep can
1907 * schedule a reset for later.
1909 static void efx_reset_work(struct work_struct
*data
)
1911 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, reset_work
);
1913 if (efx
->reset_pending
== RESET_TYPE_NONE
)
1916 /* If we're not RUNNING then don't reset. Leave the reset_pending
1917 * flag set so that efx_pci_probe_main will be retried */
1918 if (efx
->state
!= STATE_RUNNING
) {
1919 netif_info(efx
, drv
, efx
->net_dev
,
1920 "scheduled reset quenched. NIC not RUNNING\n");
1925 (void)efx_reset(efx
, efx
->reset_pending
);
1929 void efx_schedule_reset(struct efx_nic
*efx
, enum reset_type type
)
1931 enum reset_type method
;
1933 if (efx
->reset_pending
!= RESET_TYPE_NONE
) {
1934 netif_info(efx
, drv
, efx
->net_dev
,
1935 "quenching already scheduled reset\n");
1940 case RESET_TYPE_INVISIBLE
:
1941 case RESET_TYPE_ALL
:
1942 case RESET_TYPE_WORLD
:
1943 case RESET_TYPE_DISABLE
:
1946 case RESET_TYPE_RX_RECOVERY
:
1947 case RESET_TYPE_RX_DESC_FETCH
:
1948 case RESET_TYPE_TX_DESC_FETCH
:
1949 case RESET_TYPE_TX_SKIP
:
1950 method
= RESET_TYPE_INVISIBLE
;
1952 case RESET_TYPE_MC_FAILURE
:
1954 method
= RESET_TYPE_ALL
;
1959 netif_dbg(efx
, drv
, efx
->net_dev
,
1960 "scheduling %s reset for %s\n",
1961 RESET_TYPE(method
), RESET_TYPE(type
));
1963 netif_dbg(efx
, drv
, efx
->net_dev
, "scheduling %s reset\n",
1964 RESET_TYPE(method
));
1966 efx
->reset_pending
= method
;
1968 /* efx_process_channel() will no longer read events once a
1969 * reset is scheduled. So switch back to poll'd MCDI completions. */
1970 efx_mcdi_mode_poll(efx
);
1972 queue_work(reset_workqueue
, &efx
->reset_work
);
1975 /**************************************************************************
1977 * List of NICs we support
1979 **************************************************************************/
1981 /* PCI device ID table */
1982 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table
) = {
1983 {PCI_DEVICE(EFX_VENDID_SFC
, FALCON_A_P_DEVID
),
1984 .driver_data
= (unsigned long) &falcon_a1_nic_type
},
1985 {PCI_DEVICE(EFX_VENDID_SFC
, FALCON_B_P_DEVID
),
1986 .driver_data
= (unsigned long) &falcon_b0_nic_type
},
1987 {PCI_DEVICE(EFX_VENDID_SFC
, BETHPAGE_A_P_DEVID
),
1988 .driver_data
= (unsigned long) &siena_a0_nic_type
},
1989 {PCI_DEVICE(EFX_VENDID_SFC
, SIENA_A_P_DEVID
),
1990 .driver_data
= (unsigned long) &siena_a0_nic_type
},
1991 {0} /* end of list */
1994 /**************************************************************************
1996 * Dummy PHY/MAC operations
1998 * Can be used for some unimplemented operations
1999 * Needed so all function pointers are valid and do not have to be tested
2002 **************************************************************************/
2003 int efx_port_dummy_op_int(struct efx_nic
*efx
)
2007 void efx_port_dummy_op_void(struct efx_nic
*efx
) {}
2008 void efx_port_dummy_op_set_id_led(struct efx_nic
*efx
, enum efx_led_mode mode
)
2011 bool efx_port_dummy_op_poll(struct efx_nic
*efx
)
2016 static struct efx_phy_operations efx_dummy_phy_operations
= {
2017 .init
= efx_port_dummy_op_int
,
2018 .reconfigure
= efx_port_dummy_op_int
,
2019 .poll
= efx_port_dummy_op_poll
,
2020 .fini
= efx_port_dummy_op_void
,
2023 /**************************************************************************
2027 **************************************************************************/
2029 /* This zeroes out and then fills in the invariants in a struct
2030 * efx_nic (including all sub-structures).
2032 static int efx_init_struct(struct efx_nic
*efx
, struct efx_nic_type
*type
,
2033 struct pci_dev
*pci_dev
, struct net_device
*net_dev
)
2035 struct efx_channel
*channel
;
2036 struct efx_tx_queue
*tx_queue
;
2037 struct efx_rx_queue
*rx_queue
;
2040 /* Initialise common structures */
2041 memset(efx
, 0, sizeof(*efx
));
2042 spin_lock_init(&efx
->biu_lock
);
2043 mutex_init(&efx
->mdio_lock
);
2044 mutex_init(&efx
->spi_lock
);
2045 #ifdef CONFIG_SFC_MTD
2046 INIT_LIST_HEAD(&efx
->mtd_list
);
2048 INIT_WORK(&efx
->reset_work
, efx_reset_work
);
2049 INIT_DELAYED_WORK(&efx
->monitor_work
, efx_monitor
);
2050 efx
->pci_dev
= pci_dev
;
2051 efx
->msg_enable
= debug
;
2052 efx
->state
= STATE_INIT
;
2053 efx
->reset_pending
= RESET_TYPE_NONE
;
2054 strlcpy(efx
->name
, pci_name(pci_dev
), sizeof(efx
->name
));
2056 efx
->net_dev
= net_dev
;
2057 efx
->rx_checksum_enabled
= true;
2058 spin_lock_init(&efx
->stats_lock
);
2059 mutex_init(&efx
->mac_lock
);
2060 efx
->mac_op
= type
->default_mac_ops
;
2061 efx
->phy_op
= &efx_dummy_phy_operations
;
2062 efx
->mdio
.dev
= net_dev
;
2063 INIT_WORK(&efx
->mac_work
, efx_mac_work
);
2065 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++) {
2066 channel
= &efx
->channel
[i
];
2068 channel
->channel
= i
;
2069 channel
->work_pending
= false;
2070 spin_lock_init(&channel
->tx_stop_lock
);
2071 atomic_set(&channel
->tx_stop_count
, 1);
2073 for (i
= 0; i
< EFX_MAX_TX_QUEUES
; i
++) {
2074 tx_queue
= &efx
->tx_queue
[i
];
2075 tx_queue
->efx
= efx
;
2076 tx_queue
->queue
= i
;
2077 tx_queue
->buffer
= NULL
;
2078 tx_queue
->channel
= &efx
->channel
[0]; /* for safety */
2079 tx_queue
->tso_headers_free
= NULL
;
2081 for (i
= 0; i
< EFX_MAX_RX_QUEUES
; i
++) {
2082 rx_queue
= &efx
->rx_queue
[i
];
2083 rx_queue
->efx
= efx
;
2084 rx_queue
->queue
= i
;
2085 rx_queue
->channel
= &efx
->channel
[0]; /* for safety */
2086 rx_queue
->buffer
= NULL
;
2087 setup_timer(&rx_queue
->slow_fill
, efx_rx_slow_fill
,
2088 (unsigned long)rx_queue
);
2093 /* As close as we can get to guaranteeing that we don't overflow */
2094 BUILD_BUG_ON(EFX_EVQ_SIZE
< EFX_TXQ_SIZE
+ EFX_RXQ_SIZE
);
2096 EFX_BUG_ON_PARANOID(efx
->type
->phys_addr_channels
> EFX_MAX_CHANNELS
);
2098 /* Higher numbered interrupt modes are less capable! */
2099 efx
->interrupt_mode
= max(efx
->type
->max_interrupt_mode
,
2102 /* Would be good to use the net_dev name, but we're too early */
2103 snprintf(efx
->workqueue_name
, sizeof(efx
->workqueue_name
), "sfc%s",
2105 efx
->workqueue
= create_singlethread_workqueue(efx
->workqueue_name
);
2106 if (!efx
->workqueue
)
2112 static void efx_fini_struct(struct efx_nic
*efx
)
2114 if (efx
->workqueue
) {
2115 destroy_workqueue(efx
->workqueue
);
2116 efx
->workqueue
= NULL
;
2120 /**************************************************************************
2124 **************************************************************************/
2126 /* Main body of final NIC shutdown code
2127 * This is called only at module unload (or hotplug removal).
2129 static void efx_pci_remove_main(struct efx_nic
*efx
)
2131 efx_nic_fini_interrupt(efx
);
2132 efx_fini_channels(efx
);
2134 efx
->type
->fini(efx
);
2136 efx_remove_all(efx
);
2139 /* Final NIC shutdown
2140 * This is called only at module unload (or hotplug removal).
2142 static void efx_pci_remove(struct pci_dev
*pci_dev
)
2144 struct efx_nic
*efx
;
2146 efx
= pci_get_drvdata(pci_dev
);
2150 /* Mark the NIC as fini, then stop the interface */
2152 efx
->state
= STATE_FINI
;
2153 dev_close(efx
->net_dev
);
2155 /* Allow any queued efx_resets() to complete */
2158 efx_unregister_netdev(efx
);
2160 efx_mtd_remove(efx
);
2162 /* Wait for any scheduled resets to complete. No more will be
2163 * scheduled from this point because efx_stop_all() has been
2164 * called, we are no longer registered with driverlink, and
2165 * the net_device's have been removed. */
2166 cancel_work_sync(&efx
->reset_work
);
2168 efx_pci_remove_main(efx
);
2171 netif_dbg(efx
, drv
, efx
->net_dev
, "shutdown successful\n");
2173 pci_set_drvdata(pci_dev
, NULL
);
2174 efx_fini_struct(efx
);
2175 free_netdev(efx
->net_dev
);
2178 /* Main body of NIC initialisation
2179 * This is called at module load (or hotplug insertion, theoretically).
2181 static int efx_pci_probe_main(struct efx_nic
*efx
)
2185 /* Do start-of-day initialisation */
2186 rc
= efx_probe_all(efx
);
2190 rc
= efx_init_napi(efx
);
2194 rc
= efx
->type
->init(efx
);
2196 netif_err(efx
, probe
, efx
->net_dev
,
2197 "failed to initialise NIC\n");
2201 rc
= efx_init_port(efx
);
2203 netif_err(efx
, probe
, efx
->net_dev
,
2204 "failed to initialise port\n");
2208 efx_init_channels(efx
);
2210 rc
= efx_nic_init_interrupt(efx
);
2217 efx_fini_channels(efx
);
2220 efx
->type
->fini(efx
);
2224 efx_remove_all(efx
);
2229 /* NIC initialisation
2231 * This is called at module load (or hotplug insertion,
2232 * theoretically). It sets up PCI mappings, tests and resets the NIC,
2233 * sets up and registers the network devices with the kernel and hooks
2234 * the interrupt service routine. It does not prepare the device for
2235 * transmission; this is left to the first time one of the network
2236 * interfaces is brought up (i.e. efx_net_open).
2238 static int __devinit
efx_pci_probe(struct pci_dev
*pci_dev
,
2239 const struct pci_device_id
*entry
)
2241 struct efx_nic_type
*type
= (struct efx_nic_type
*) entry
->driver_data
;
2242 struct net_device
*net_dev
;
2243 struct efx_nic
*efx
;
2246 /* Allocate and initialise a struct net_device and struct efx_nic */
2247 net_dev
= alloc_etherdev_mq(sizeof(*efx
), EFX_MAX_CORE_TX_QUEUES
);
2250 net_dev
->features
|= (type
->offload_features
| NETIF_F_SG
|
2251 NETIF_F_HIGHDMA
| NETIF_F_TSO
|
2253 if (type
->offload_features
& NETIF_F_V6_CSUM
)
2254 net_dev
->features
|= NETIF_F_TSO6
;
2255 /* Mask for features that also apply to VLAN devices */
2256 net_dev
->vlan_features
|= (NETIF_F_ALL_CSUM
| NETIF_F_SG
|
2257 NETIF_F_HIGHDMA
| NETIF_F_TSO
);
2258 efx
= netdev_priv(net_dev
);
2259 pci_set_drvdata(pci_dev
, efx
);
2260 SET_NETDEV_DEV(net_dev
, &pci_dev
->dev
);
2261 rc
= efx_init_struct(efx
, type
, pci_dev
, net_dev
);
2265 netif_info(efx
, probe
, efx
->net_dev
,
2266 "Solarflare Communications NIC detected\n");
2268 /* Set up basic I/O (BAR mappings etc) */
2269 rc
= efx_init_io(efx
);
2273 /* No serialisation is required with the reset path because
2274 * we're in STATE_INIT. */
2275 for (i
= 0; i
< 5; i
++) {
2276 rc
= efx_pci_probe_main(efx
);
2278 /* Serialise against efx_reset(). No more resets will be
2279 * scheduled since efx_stop_all() has been called, and we
2280 * have not and never have been registered with either
2281 * the rtnetlink or driverlink layers. */
2282 cancel_work_sync(&efx
->reset_work
);
2285 if (efx
->reset_pending
!= RESET_TYPE_NONE
) {
2286 /* If there was a scheduled reset during
2287 * probe, the NIC is probably hosed anyway */
2288 efx_pci_remove_main(efx
);
2295 /* Retry if a recoverably reset event has been scheduled */
2296 if ((efx
->reset_pending
!= RESET_TYPE_INVISIBLE
) &&
2297 (efx
->reset_pending
!= RESET_TYPE_ALL
))
2300 efx
->reset_pending
= RESET_TYPE_NONE
;
2304 netif_err(efx
, probe
, efx
->net_dev
, "Could not reset NIC\n");
2308 /* Switch to the running state before we expose the device to the OS,
2309 * so that dev_open()|efx_start_all() will actually start the device */
2310 efx
->state
= STATE_RUNNING
;
2312 rc
= efx_register_netdev(efx
);
2316 netif_dbg(efx
, probe
, efx
->net_dev
, "initialisation successful\n");
2319 efx_mtd_probe(efx
); /* allowed to fail */
2324 efx_pci_remove_main(efx
);
2329 efx_fini_struct(efx
);
2332 netif_dbg(efx
, drv
, efx
->net_dev
, "initialisation failed. rc=%d\n", rc
);
2333 free_netdev(net_dev
);
2337 static int efx_pm_freeze(struct device
*dev
)
2339 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2341 efx
->state
= STATE_FINI
;
2343 netif_device_detach(efx
->net_dev
);
2346 efx_fini_channels(efx
);
2351 static int efx_pm_thaw(struct device
*dev
)
2353 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2355 efx
->state
= STATE_INIT
;
2357 efx_init_channels(efx
);
2359 mutex_lock(&efx
->mac_lock
);
2360 efx
->phy_op
->reconfigure(efx
);
2361 mutex_unlock(&efx
->mac_lock
);
2365 netif_device_attach(efx
->net_dev
);
2367 efx
->state
= STATE_RUNNING
;
2369 efx
->type
->resume_wol(efx
);
2371 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2372 queue_work(reset_workqueue
, &efx
->reset_work
);
2377 static int efx_pm_poweroff(struct device
*dev
)
2379 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
2380 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
2382 efx
->type
->fini(efx
);
2384 efx
->reset_pending
= RESET_TYPE_NONE
;
2386 pci_save_state(pci_dev
);
2387 return pci_set_power_state(pci_dev
, PCI_D3hot
);
2390 /* Used for both resume and restore */
2391 static int efx_pm_resume(struct device
*dev
)
2393 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
2394 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
2397 rc
= pci_set_power_state(pci_dev
, PCI_D0
);
2400 pci_restore_state(pci_dev
);
2401 rc
= pci_enable_device(pci_dev
);
2404 pci_set_master(efx
->pci_dev
);
2405 rc
= efx
->type
->reset(efx
, RESET_TYPE_ALL
);
2408 rc
= efx
->type
->init(efx
);
2415 static int efx_pm_suspend(struct device
*dev
)
2420 rc
= efx_pm_poweroff(dev
);
2426 static struct dev_pm_ops efx_pm_ops
= {
2427 .suspend
= efx_pm_suspend
,
2428 .resume
= efx_pm_resume
,
2429 .freeze
= efx_pm_freeze
,
2430 .thaw
= efx_pm_thaw
,
2431 .poweroff
= efx_pm_poweroff
,
2432 .restore
= efx_pm_resume
,
2435 static struct pci_driver efx_pci_driver
= {
2436 .name
= KBUILD_MODNAME
,
2437 .id_table
= efx_pci_table
,
2438 .probe
= efx_pci_probe
,
2439 .remove
= efx_pci_remove
,
2440 .driver
.pm
= &efx_pm_ops
,
2443 /**************************************************************************
2445 * Kernel module interface
2447 *************************************************************************/
2449 module_param(interrupt_mode
, uint
, 0444);
2450 MODULE_PARM_DESC(interrupt_mode
,
2451 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2453 static int __init
efx_init_module(void)
2457 printk(KERN_INFO
"Solarflare NET driver v" EFX_DRIVER_VERSION
"\n");
2459 rc
= register_netdevice_notifier(&efx_netdev_notifier
);
2463 reset_workqueue
= create_singlethread_workqueue("sfc_reset");
2464 if (!reset_workqueue
) {
2469 rc
= pci_register_driver(&efx_pci_driver
);
2476 destroy_workqueue(reset_workqueue
);
2478 unregister_netdevice_notifier(&efx_netdev_notifier
);
2483 static void __exit
efx_exit_module(void)
2485 printk(KERN_INFO
"Solarflare NET driver unloading\n");
2487 pci_unregister_driver(&efx_pci_driver
);
2488 destroy_workqueue(reset_workqueue
);
2489 unregister_netdevice_notifier(&efx_netdev_notifier
);
2493 module_init(efx_init_module
);
2494 module_exit(efx_exit_module
);
2496 MODULE_AUTHOR("Solarflare Communications and "
2497 "Michael Brown <mbrown@fensystems.co.uk>");
2498 MODULE_DESCRIPTION("Solarflare Communications network driver");
2499 MODULE_LICENSE("GPL");
2500 MODULE_DEVICE_TABLE(pci
, efx_pci_table
);