1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2011 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 <linux/cpu_rmap.h>
25 #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 *const 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 const unsigned int efx_reset_type_max
= RESET_TYPE_MAX
;
72 const char *const efx_reset_type_names
[] = {
73 [RESET_TYPE_INVISIBLE
] = "INVISIBLE",
74 [RESET_TYPE_ALL
] = "ALL",
75 [RESET_TYPE_WORLD
] = "WORLD",
76 [RESET_TYPE_DISABLE
] = "DISABLE",
77 [RESET_TYPE_TX_WATCHDOG
] = "TX_WATCHDOG",
78 [RESET_TYPE_INT_ERROR
] = "INT_ERROR",
79 [RESET_TYPE_RX_RECOVERY
] = "RX_RECOVERY",
80 [RESET_TYPE_RX_DESC_FETCH
] = "RX_DESC_FETCH",
81 [RESET_TYPE_TX_DESC_FETCH
] = "TX_DESC_FETCH",
82 [RESET_TYPE_TX_SKIP
] = "TX_SKIP",
83 [RESET_TYPE_MC_FAILURE
] = "MC_FAILURE",
86 #define EFX_MAX_MTU (9 * 1024)
88 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
89 * queued onto this work queue. This is not a per-nic work queue, because
90 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
92 static struct workqueue_struct
*reset_workqueue
;
94 /**************************************************************************
98 *************************************************************************/
101 * Use separate channels for TX and RX events
103 * Set this to 1 to use separate channels for TX and RX. It allows us
104 * to control interrupt affinity separately for TX and RX.
106 * This is only used in MSI-X interrupt mode
108 static unsigned int separate_tx_channels
;
109 module_param(separate_tx_channels
, uint
, 0444);
110 MODULE_PARM_DESC(separate_tx_channels
,
111 "Use separate channels for TX and RX");
113 /* This is the weight assigned to each of the (per-channel) virtual
116 static int napi_weight
= 64;
118 /* This is the time (in jiffies) between invocations of the hardware
119 * monitor. On Falcon-based NICs, this will:
120 * - Check the on-board hardware monitor;
121 * - Poll the link state and reconfigure the hardware as necessary.
123 static unsigned int efx_monitor_interval
= 1 * HZ
;
125 /* Initial interrupt moderation settings. They can be modified after
126 * module load with ethtool.
128 * The default for RX should strike a balance between increasing the
129 * round-trip latency and reducing overhead.
131 static unsigned int rx_irq_mod_usec
= 60;
133 /* Initial interrupt moderation settings. They can be modified after
134 * module load with ethtool.
136 * This default is chosen to ensure that a 10G link does not go idle
137 * while a TX queue is stopped after it has become full. A queue is
138 * restarted when it drops below half full. The time this takes (assuming
139 * worst case 3 descriptors per packet and 1024 descriptors) is
140 * 512 / 3 * 1.2 = 205 usec.
142 static unsigned int tx_irq_mod_usec
= 150;
144 /* This is the first interrupt mode to try out of:
149 static unsigned int interrupt_mode
;
151 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
152 * i.e. the number of CPUs among which we may distribute simultaneous
153 * interrupt handling.
155 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
156 * The default (0) means to assign an interrupt to each core.
158 static unsigned int rss_cpus
;
159 module_param(rss_cpus
, uint
, 0444);
160 MODULE_PARM_DESC(rss_cpus
, "Number of CPUs to use for Receive-Side Scaling");
162 static int phy_flash_cfg
;
163 module_param(phy_flash_cfg
, int, 0644);
164 MODULE_PARM_DESC(phy_flash_cfg
, "Set PHYs into reflash mode initially");
166 static unsigned irq_adapt_low_thresh
= 10000;
167 module_param(irq_adapt_low_thresh
, uint
, 0644);
168 MODULE_PARM_DESC(irq_adapt_low_thresh
,
169 "Threshold score for reducing IRQ moderation");
171 static unsigned irq_adapt_high_thresh
= 20000;
172 module_param(irq_adapt_high_thresh
, uint
, 0644);
173 MODULE_PARM_DESC(irq_adapt_high_thresh
,
174 "Threshold score for increasing IRQ moderation");
176 static unsigned debug
= (NETIF_MSG_DRV
| NETIF_MSG_PROBE
|
177 NETIF_MSG_LINK
| NETIF_MSG_IFDOWN
|
178 NETIF_MSG_IFUP
| NETIF_MSG_RX_ERR
|
179 NETIF_MSG_TX_ERR
| NETIF_MSG_HW
);
180 module_param(debug
, uint
, 0);
181 MODULE_PARM_DESC(debug
, "Bitmapped debugging message enable value");
183 /**************************************************************************
185 * Utility functions and prototypes
187 *************************************************************************/
189 static void efx_start_interrupts(struct efx_nic
*efx
);
190 static void efx_stop_interrupts(struct efx_nic
*efx
);
191 static void efx_remove_channels(struct efx_nic
*efx
);
192 static void efx_remove_port(struct efx_nic
*efx
);
193 static void efx_init_napi(struct efx_nic
*efx
);
194 static void efx_fini_napi(struct efx_nic
*efx
);
195 static void efx_fini_napi_channel(struct efx_channel
*channel
);
196 static void efx_fini_struct(struct efx_nic
*efx
);
197 static void efx_start_all(struct efx_nic
*efx
);
198 static void efx_stop_all(struct efx_nic
*efx
);
200 #define EFX_ASSERT_RESET_SERIALISED(efx) \
202 if ((efx->state == STATE_RUNNING) || \
203 (efx->state == STATE_DISABLED)) \
207 /**************************************************************************
209 * Event queue processing
211 *************************************************************************/
213 /* Process channel's event queue
215 * This function is responsible for processing the event queue of a
216 * single channel. The caller must guarantee that this function will
217 * never be concurrently called more than once on the same channel,
218 * though different channels may be being processed concurrently.
220 static int efx_process_channel(struct efx_channel
*channel
, int budget
)
224 if (unlikely(!channel
->enabled
))
227 spent
= efx_nic_process_eventq(channel
, budget
);
228 if (spent
&& efx_channel_has_rx_queue(channel
)) {
229 struct efx_rx_queue
*rx_queue
=
230 efx_channel_get_rx_queue(channel
);
232 /* Deliver last RX packet. */
233 if (channel
->rx_pkt
) {
234 __efx_rx_packet(channel
, channel
->rx_pkt
);
235 channel
->rx_pkt
= NULL
;
237 if (rx_queue
->enabled
) {
238 efx_rx_strategy(channel
);
239 efx_fast_push_rx_descriptors(rx_queue
);
246 /* Mark channel as finished processing
248 * Note that since we will not receive further interrupts for this
249 * channel before we finish processing and call the eventq_read_ack()
250 * method, there is no need to use the interrupt hold-off timers.
252 static inline void efx_channel_processed(struct efx_channel
*channel
)
254 /* The interrupt handler for this channel may set work_pending
255 * as soon as we acknowledge the events we've seen. Make sure
256 * it's cleared before then. */
257 channel
->work_pending
= false;
260 efx_nic_eventq_read_ack(channel
);
265 * NAPI guarantees serialisation of polls of the same device, which
266 * provides the guarantee required by efx_process_channel().
268 static int efx_poll(struct napi_struct
*napi
, int budget
)
270 struct efx_channel
*channel
=
271 container_of(napi
, struct efx_channel
, napi_str
);
272 struct efx_nic
*efx
= channel
->efx
;
275 netif_vdbg(efx
, intr
, efx
->net_dev
,
276 "channel %d NAPI poll executing on CPU %d\n",
277 channel
->channel
, raw_smp_processor_id());
279 spent
= efx_process_channel(channel
, budget
);
281 if (spent
< budget
) {
282 if (efx_channel_has_rx_queue(channel
) &&
283 efx
->irq_rx_adaptive
&&
284 unlikely(++channel
->irq_count
== 1000)) {
285 if (unlikely(channel
->irq_mod_score
<
286 irq_adapt_low_thresh
)) {
287 if (channel
->irq_moderation
> 1) {
288 channel
->irq_moderation
-= 1;
289 efx
->type
->push_irq_moderation(channel
);
291 } else if (unlikely(channel
->irq_mod_score
>
292 irq_adapt_high_thresh
)) {
293 if (channel
->irq_moderation
<
294 efx
->irq_rx_moderation
) {
295 channel
->irq_moderation
+= 1;
296 efx
->type
->push_irq_moderation(channel
);
299 channel
->irq_count
= 0;
300 channel
->irq_mod_score
= 0;
303 efx_filter_rfs_expire(channel
);
305 /* There is no race here; although napi_disable() will
306 * only wait for napi_complete(), this isn't a problem
307 * since efx_channel_processed() will have no effect if
308 * interrupts have already been disabled.
311 efx_channel_processed(channel
);
317 /* Process the eventq of the specified channel immediately on this CPU
319 * Disable hardware generated interrupts, wait for any existing
320 * processing to finish, then directly poll (and ack ) the eventq.
321 * Finally reenable NAPI and interrupts.
323 * This is for use only during a loopback self-test. It must not
324 * deliver any packets up the stack as this can result in deadlock.
326 void efx_process_channel_now(struct efx_channel
*channel
)
328 struct efx_nic
*efx
= channel
->efx
;
330 BUG_ON(channel
->channel
>= efx
->n_channels
);
331 BUG_ON(!channel
->enabled
);
332 BUG_ON(!efx
->loopback_selftest
);
334 /* Disable interrupts and wait for ISRs to complete */
335 efx_nic_disable_interrupts(efx
);
336 if (efx
->legacy_irq
) {
337 synchronize_irq(efx
->legacy_irq
);
338 efx
->legacy_irq_enabled
= false;
341 synchronize_irq(channel
->irq
);
343 /* Wait for any NAPI processing to complete */
344 napi_disable(&channel
->napi_str
);
346 /* Poll the channel */
347 efx_process_channel(channel
, channel
->eventq_mask
+ 1);
349 /* Ack the eventq. This may cause an interrupt to be generated
350 * when they are reenabled */
351 efx_channel_processed(channel
);
353 napi_enable(&channel
->napi_str
);
355 efx
->legacy_irq_enabled
= true;
356 efx_nic_enable_interrupts(efx
);
359 /* Create event queue
360 * Event queue memory allocations are done only once. If the channel
361 * is reset, the memory buffer will be reused; this guards against
362 * errors during channel reset and also simplifies interrupt handling.
364 static int efx_probe_eventq(struct efx_channel
*channel
)
366 struct efx_nic
*efx
= channel
->efx
;
367 unsigned long entries
;
369 netif_dbg(efx
, probe
, efx
->net_dev
,
370 "chan %d create event queue\n", channel
->channel
);
372 /* Build an event queue with room for one event per tx and rx buffer,
373 * plus some extra for link state events and MCDI completions. */
374 entries
= roundup_pow_of_two(efx
->rxq_entries
+ efx
->txq_entries
+ 128);
375 EFX_BUG_ON_PARANOID(entries
> EFX_MAX_EVQ_SIZE
);
376 channel
->eventq_mask
= max(entries
, EFX_MIN_EVQ_SIZE
) - 1;
378 return efx_nic_probe_eventq(channel
);
381 /* Prepare channel's event queue */
382 static void efx_init_eventq(struct efx_channel
*channel
)
384 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
385 "chan %d init event queue\n", channel
->channel
);
387 channel
->eventq_read_ptr
= 0;
389 efx_nic_init_eventq(channel
);
392 /* Enable event queue processing and NAPI */
393 static void efx_start_eventq(struct efx_channel
*channel
)
395 netif_dbg(channel
->efx
, ifup
, channel
->efx
->net_dev
,
396 "chan %d start event queue\n", channel
->channel
);
398 /* The interrupt handler for this channel may set work_pending
399 * as soon as we enable it. Make sure it's cleared before
400 * then. Similarly, make sure it sees the enabled flag set.
402 channel
->work_pending
= false;
403 channel
->enabled
= true;
406 napi_enable(&channel
->napi_str
);
407 efx_nic_eventq_read_ack(channel
);
410 /* Disable event queue processing and NAPI */
411 static void efx_stop_eventq(struct efx_channel
*channel
)
413 if (!channel
->enabled
)
416 napi_disable(&channel
->napi_str
);
417 channel
->enabled
= false;
420 static void efx_fini_eventq(struct efx_channel
*channel
)
422 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
423 "chan %d fini event queue\n", channel
->channel
);
425 efx_nic_fini_eventq(channel
);
428 static void efx_remove_eventq(struct efx_channel
*channel
)
430 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
431 "chan %d remove event queue\n", channel
->channel
);
433 efx_nic_remove_eventq(channel
);
436 /**************************************************************************
440 *************************************************************************/
442 /* Allocate and initialise a channel structure, optionally copying
443 * parameters (but not resources) from an old channel structure. */
444 static struct efx_channel
*
445 efx_alloc_channel(struct efx_nic
*efx
, int i
, struct efx_channel
*old_channel
)
447 struct efx_channel
*channel
;
448 struct efx_rx_queue
*rx_queue
;
449 struct efx_tx_queue
*tx_queue
;
453 channel
= kmalloc(sizeof(*channel
), GFP_KERNEL
);
457 *channel
= *old_channel
;
459 channel
->napi_dev
= NULL
;
460 memset(&channel
->eventq
, 0, sizeof(channel
->eventq
));
462 rx_queue
= &channel
->rx_queue
;
463 rx_queue
->buffer
= NULL
;
464 memset(&rx_queue
->rxd
, 0, sizeof(rx_queue
->rxd
));
466 for (j
= 0; j
< EFX_TXQ_TYPES
; j
++) {
467 tx_queue
= &channel
->tx_queue
[j
];
468 if (tx_queue
->channel
)
469 tx_queue
->channel
= channel
;
470 tx_queue
->buffer
= NULL
;
471 memset(&tx_queue
->txd
, 0, sizeof(tx_queue
->txd
));
474 channel
= kzalloc(sizeof(*channel
), GFP_KERNEL
);
479 channel
->channel
= i
;
481 for (j
= 0; j
< EFX_TXQ_TYPES
; j
++) {
482 tx_queue
= &channel
->tx_queue
[j
];
484 tx_queue
->queue
= i
* EFX_TXQ_TYPES
+ j
;
485 tx_queue
->channel
= channel
;
489 rx_queue
= &channel
->rx_queue
;
491 setup_timer(&rx_queue
->slow_fill
, efx_rx_slow_fill
,
492 (unsigned long)rx_queue
);
497 static int efx_probe_channel(struct efx_channel
*channel
)
499 struct efx_tx_queue
*tx_queue
;
500 struct efx_rx_queue
*rx_queue
;
503 netif_dbg(channel
->efx
, probe
, channel
->efx
->net_dev
,
504 "creating channel %d\n", channel
->channel
);
506 rc
= efx_probe_eventq(channel
);
510 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
511 rc
= efx_probe_tx_queue(tx_queue
);
516 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
517 rc
= efx_probe_rx_queue(rx_queue
);
522 channel
->n_rx_frm_trunc
= 0;
527 efx_for_each_channel_rx_queue(rx_queue
, channel
)
528 efx_remove_rx_queue(rx_queue
);
530 efx_for_each_channel_tx_queue(tx_queue
, channel
)
531 efx_remove_tx_queue(tx_queue
);
537 static void efx_set_channel_names(struct efx_nic
*efx
)
539 struct efx_channel
*channel
;
540 const char *type
= "";
543 efx_for_each_channel(channel
, efx
) {
544 number
= channel
->channel
;
545 if (efx
->n_channels
> efx
->n_rx_channels
) {
546 if (channel
->channel
< efx
->n_rx_channels
) {
550 number
-= efx
->n_rx_channels
;
553 snprintf(efx
->channel_name
[channel
->channel
],
554 sizeof(efx
->channel_name
[0]),
555 "%s%s-%d", efx
->name
, type
, number
);
559 static int efx_probe_channels(struct efx_nic
*efx
)
561 struct efx_channel
*channel
;
564 /* Restart special buffer allocation */
565 efx
->next_buffer_table
= 0;
567 efx_for_each_channel(channel
, efx
) {
568 rc
= efx_probe_channel(channel
);
570 netif_err(efx
, probe
, efx
->net_dev
,
571 "failed to create channel %d\n",
576 efx_set_channel_names(efx
);
581 efx_remove_channels(efx
);
585 /* Channels are shutdown and reinitialised whilst the NIC is running
586 * to propagate configuration changes (mtu, checksum offload), or
587 * to clear hardware error conditions
589 static void efx_start_datapath(struct efx_nic
*efx
)
591 struct efx_tx_queue
*tx_queue
;
592 struct efx_rx_queue
*rx_queue
;
593 struct efx_channel
*channel
;
595 /* Calculate the rx buffer allocation parameters required to
596 * support the current MTU, including padding for header
597 * alignment and overruns.
599 efx
->rx_buffer_len
= (max(EFX_PAGE_IP_ALIGN
, NET_IP_ALIGN
) +
600 EFX_MAX_FRAME_LEN(efx
->net_dev
->mtu
) +
601 efx
->type
->rx_buffer_hash_size
+
602 efx
->type
->rx_buffer_padding
);
603 efx
->rx_buffer_order
= get_order(efx
->rx_buffer_len
+
604 sizeof(struct efx_rx_page_state
));
606 /* Initialise the channels */
607 efx_for_each_channel(channel
, efx
) {
608 efx_for_each_channel_tx_queue(tx_queue
, channel
)
609 efx_init_tx_queue(tx_queue
);
611 /* The rx buffer allocation strategy is MTU dependent */
612 efx_rx_strategy(channel
);
614 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
615 efx_init_rx_queue(rx_queue
);
616 efx_nic_generate_fill_event(rx_queue
);
619 WARN_ON(channel
->rx_pkt
!= NULL
);
620 efx_rx_strategy(channel
);
623 if (netif_device_present(efx
->net_dev
))
624 netif_tx_wake_all_queues(efx
->net_dev
);
627 static void efx_stop_datapath(struct efx_nic
*efx
)
629 struct efx_channel
*channel
;
630 struct efx_tx_queue
*tx_queue
;
631 struct efx_rx_queue
*rx_queue
;
634 EFX_ASSERT_RESET_SERIALISED(efx
);
635 BUG_ON(efx
->port_enabled
);
637 rc
= efx_nic_flush_queues(efx
);
638 if (rc
&& EFX_WORKAROUND_7803(efx
)) {
639 /* Schedule a reset to recover from the flush failure. The
640 * descriptor caches reference memory we're about to free,
641 * but falcon_reconfigure_mac_wrapper() won't reconnect
642 * the MACs because of the pending reset. */
643 netif_err(efx
, drv
, efx
->net_dev
,
644 "Resetting to recover from flush failure\n");
645 efx_schedule_reset(efx
, RESET_TYPE_ALL
);
647 netif_err(efx
, drv
, efx
->net_dev
, "failed to flush queues\n");
649 netif_dbg(efx
, drv
, efx
->net_dev
,
650 "successfully flushed all queues\n");
653 efx_for_each_channel(channel
, efx
) {
654 /* RX packet processing is pipelined, so wait for the
655 * NAPI handler to complete. At least event queue 0
656 * might be kept active by non-data events, so don't
657 * use napi_synchronize() but actually disable NAPI
660 if (efx_channel_has_rx_queue(channel
)) {
661 efx_stop_eventq(channel
);
662 efx_start_eventq(channel
);
665 efx_for_each_channel_rx_queue(rx_queue
, channel
)
666 efx_fini_rx_queue(rx_queue
);
667 efx_for_each_possible_channel_tx_queue(tx_queue
, channel
)
668 efx_fini_tx_queue(tx_queue
);
672 static void efx_remove_channel(struct efx_channel
*channel
)
674 struct efx_tx_queue
*tx_queue
;
675 struct efx_rx_queue
*rx_queue
;
677 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
678 "destroy chan %d\n", channel
->channel
);
680 efx_for_each_channel_rx_queue(rx_queue
, channel
)
681 efx_remove_rx_queue(rx_queue
);
682 efx_for_each_possible_channel_tx_queue(tx_queue
, channel
)
683 efx_remove_tx_queue(tx_queue
);
684 efx_remove_eventq(channel
);
687 static void efx_remove_channels(struct efx_nic
*efx
)
689 struct efx_channel
*channel
;
691 efx_for_each_channel(channel
, efx
)
692 efx_remove_channel(channel
);
696 efx_realloc_channels(struct efx_nic
*efx
, u32 rxq_entries
, u32 txq_entries
)
698 struct efx_channel
*other_channel
[EFX_MAX_CHANNELS
], *channel
;
699 u32 old_rxq_entries
, old_txq_entries
;
704 efx_stop_interrupts(efx
);
707 memset(other_channel
, 0, sizeof(other_channel
));
708 for (i
= 0; i
< efx
->n_channels
; i
++) {
709 channel
= efx_alloc_channel(efx
, i
, efx
->channel
[i
]);
714 other_channel
[i
] = channel
;
717 /* Swap entry counts and channel pointers */
718 old_rxq_entries
= efx
->rxq_entries
;
719 old_txq_entries
= efx
->txq_entries
;
720 efx
->rxq_entries
= rxq_entries
;
721 efx
->txq_entries
= txq_entries
;
722 for (i
= 0; i
< efx
->n_channels
; i
++) {
723 channel
= efx
->channel
[i
];
724 efx
->channel
[i
] = other_channel
[i
];
725 other_channel
[i
] = channel
;
728 rc
= efx_probe_channels(efx
);
734 /* Destroy old channels */
735 for (i
= 0; i
< efx
->n_channels
; i
++) {
736 efx_fini_napi_channel(other_channel
[i
]);
737 efx_remove_channel(other_channel
[i
]);
740 /* Free unused channel structures */
741 for (i
= 0; i
< efx
->n_channels
; i
++)
742 kfree(other_channel
[i
]);
744 efx_start_interrupts(efx
);
750 efx
->rxq_entries
= old_rxq_entries
;
751 efx
->txq_entries
= old_txq_entries
;
752 for (i
= 0; i
< efx
->n_channels
; i
++) {
753 channel
= efx
->channel
[i
];
754 efx
->channel
[i
] = other_channel
[i
];
755 other_channel
[i
] = channel
;
760 void efx_schedule_slow_fill(struct efx_rx_queue
*rx_queue
)
762 mod_timer(&rx_queue
->slow_fill
, jiffies
+ msecs_to_jiffies(100));
765 /**************************************************************************
769 **************************************************************************/
771 /* This ensures that the kernel is kept informed (via
772 * netif_carrier_on/off) of the link status, and also maintains the
773 * link status's stop on the port's TX queue.
775 void efx_link_status_changed(struct efx_nic
*efx
)
777 struct efx_link_state
*link_state
= &efx
->link_state
;
779 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
780 * that no events are triggered between unregister_netdev() and the
781 * driver unloading. A more general condition is that NETDEV_CHANGE
782 * can only be generated between NETDEV_UP and NETDEV_DOWN */
783 if (!netif_running(efx
->net_dev
))
786 if (link_state
->up
!= netif_carrier_ok(efx
->net_dev
)) {
787 efx
->n_link_state_changes
++;
790 netif_carrier_on(efx
->net_dev
);
792 netif_carrier_off(efx
->net_dev
);
795 /* Status message for kernel log */
797 netif_info(efx
, link
, efx
->net_dev
,
798 "link up at %uMbps %s-duplex (MTU %d)%s\n",
799 link_state
->speed
, link_state
->fd
? "full" : "half",
801 (efx
->promiscuous
? " [PROMISC]" : ""));
803 netif_info(efx
, link
, efx
->net_dev
, "link down\n");
806 void efx_link_set_advertising(struct efx_nic
*efx
, u32 advertising
)
808 efx
->link_advertising
= advertising
;
810 if (advertising
& ADVERTISED_Pause
)
811 efx
->wanted_fc
|= (EFX_FC_TX
| EFX_FC_RX
);
813 efx
->wanted_fc
&= ~(EFX_FC_TX
| EFX_FC_RX
);
814 if (advertising
& ADVERTISED_Asym_Pause
)
815 efx
->wanted_fc
^= EFX_FC_TX
;
819 void efx_link_set_wanted_fc(struct efx_nic
*efx
, u8 wanted_fc
)
821 efx
->wanted_fc
= wanted_fc
;
822 if (efx
->link_advertising
) {
823 if (wanted_fc
& EFX_FC_RX
)
824 efx
->link_advertising
|= (ADVERTISED_Pause
|
825 ADVERTISED_Asym_Pause
);
827 efx
->link_advertising
&= ~(ADVERTISED_Pause
|
828 ADVERTISED_Asym_Pause
);
829 if (wanted_fc
& EFX_FC_TX
)
830 efx
->link_advertising
^= ADVERTISED_Asym_Pause
;
834 static void efx_fini_port(struct efx_nic
*efx
);
836 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
837 * the MAC appropriately. All other PHY configuration changes are pushed
838 * through phy_op->set_settings(), and pushed asynchronously to the MAC
839 * through efx_monitor().
841 * Callers must hold the mac_lock
843 int __efx_reconfigure_port(struct efx_nic
*efx
)
845 enum efx_phy_mode phy_mode
;
848 WARN_ON(!mutex_is_locked(&efx
->mac_lock
));
850 /* Serialise the promiscuous flag with efx_set_rx_mode. */
851 netif_addr_lock_bh(efx
->net_dev
);
852 netif_addr_unlock_bh(efx
->net_dev
);
854 /* Disable PHY transmit in mac level loopbacks */
855 phy_mode
= efx
->phy_mode
;
856 if (LOOPBACK_INTERNAL(efx
))
857 efx
->phy_mode
|= PHY_MODE_TX_DISABLED
;
859 efx
->phy_mode
&= ~PHY_MODE_TX_DISABLED
;
861 rc
= efx
->type
->reconfigure_port(efx
);
864 efx
->phy_mode
= phy_mode
;
869 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
871 int efx_reconfigure_port(struct efx_nic
*efx
)
875 EFX_ASSERT_RESET_SERIALISED(efx
);
877 mutex_lock(&efx
->mac_lock
);
878 rc
= __efx_reconfigure_port(efx
);
879 mutex_unlock(&efx
->mac_lock
);
884 /* Asynchronous work item for changing MAC promiscuity and multicast
885 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
887 static void efx_mac_work(struct work_struct
*data
)
889 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, mac_work
);
891 mutex_lock(&efx
->mac_lock
);
892 if (efx
->port_enabled
)
893 efx
->type
->reconfigure_mac(efx
);
894 mutex_unlock(&efx
->mac_lock
);
897 static int efx_probe_port(struct efx_nic
*efx
)
901 netif_dbg(efx
, probe
, efx
->net_dev
, "create port\n");
904 efx
->phy_mode
= PHY_MODE_SPECIAL
;
906 /* Connect up MAC/PHY operations table */
907 rc
= efx
->type
->probe_port(efx
);
911 /* Initialise MAC address to permanent address */
912 memcpy(efx
->net_dev
->dev_addr
, efx
->net_dev
->perm_addr
, ETH_ALEN
);
917 static int efx_init_port(struct efx_nic
*efx
)
921 netif_dbg(efx
, drv
, efx
->net_dev
, "init port\n");
923 mutex_lock(&efx
->mac_lock
);
925 rc
= efx
->phy_op
->init(efx
);
929 efx
->port_initialized
= true;
931 /* Reconfigure the MAC before creating dma queues (required for
932 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
933 efx
->type
->reconfigure_mac(efx
);
935 /* Ensure the PHY advertises the correct flow control settings */
936 rc
= efx
->phy_op
->reconfigure(efx
);
940 mutex_unlock(&efx
->mac_lock
);
944 efx
->phy_op
->fini(efx
);
946 mutex_unlock(&efx
->mac_lock
);
950 static void efx_start_port(struct efx_nic
*efx
)
952 netif_dbg(efx
, ifup
, efx
->net_dev
, "start port\n");
953 BUG_ON(efx
->port_enabled
);
955 mutex_lock(&efx
->mac_lock
);
956 efx
->port_enabled
= true;
958 /* efx_mac_work() might have been scheduled after efx_stop_port(),
959 * and then cancelled by efx_flush_all() */
960 efx
->type
->reconfigure_mac(efx
);
962 mutex_unlock(&efx
->mac_lock
);
965 /* Prevent efx_mac_work() and efx_monitor() from working */
966 static void efx_stop_port(struct efx_nic
*efx
)
968 netif_dbg(efx
, ifdown
, efx
->net_dev
, "stop port\n");
970 mutex_lock(&efx
->mac_lock
);
971 efx
->port_enabled
= false;
972 mutex_unlock(&efx
->mac_lock
);
974 /* Serialise against efx_set_multicast_list() */
975 netif_addr_lock_bh(efx
->net_dev
);
976 netif_addr_unlock_bh(efx
->net_dev
);
979 static void efx_fini_port(struct efx_nic
*efx
)
981 netif_dbg(efx
, drv
, efx
->net_dev
, "shut down port\n");
983 if (!efx
->port_initialized
)
986 efx
->phy_op
->fini(efx
);
987 efx
->port_initialized
= false;
989 efx
->link_state
.up
= false;
990 efx_link_status_changed(efx
);
993 static void efx_remove_port(struct efx_nic
*efx
)
995 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying port\n");
997 efx
->type
->remove_port(efx
);
1000 /**************************************************************************
1004 **************************************************************************/
1006 /* This configures the PCI device to enable I/O and DMA. */
1007 static int efx_init_io(struct efx_nic
*efx
)
1009 struct pci_dev
*pci_dev
= efx
->pci_dev
;
1010 dma_addr_t dma_mask
= efx
->type
->max_dma_mask
;
1013 netif_dbg(efx
, probe
, efx
->net_dev
, "initialising I/O\n");
1015 rc
= pci_enable_device(pci_dev
);
1017 netif_err(efx
, probe
, efx
->net_dev
,
1018 "failed to enable PCI device\n");
1022 pci_set_master(pci_dev
);
1024 /* Set the PCI DMA mask. Try all possibilities from our
1025 * genuine mask down to 32 bits, because some architectures
1026 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1027 * masks event though they reject 46 bit masks.
1029 while (dma_mask
> 0x7fffffffUL
) {
1030 if (pci_dma_supported(pci_dev
, dma_mask
)) {
1031 rc
= pci_set_dma_mask(pci_dev
, dma_mask
);
1038 netif_err(efx
, probe
, efx
->net_dev
,
1039 "could not find a suitable DMA mask\n");
1042 netif_dbg(efx
, probe
, efx
->net_dev
,
1043 "using DMA mask %llx\n", (unsigned long long) dma_mask
);
1044 rc
= pci_set_consistent_dma_mask(pci_dev
, dma_mask
);
1046 /* pci_set_consistent_dma_mask() is not *allowed* to
1047 * fail with a mask that pci_set_dma_mask() accepted,
1048 * but just in case...
1050 netif_err(efx
, probe
, efx
->net_dev
,
1051 "failed to set consistent DMA mask\n");
1055 efx
->membase_phys
= pci_resource_start(efx
->pci_dev
, EFX_MEM_BAR
);
1056 rc
= pci_request_region(pci_dev
, EFX_MEM_BAR
, "sfc");
1058 netif_err(efx
, probe
, efx
->net_dev
,
1059 "request for memory BAR failed\n");
1063 efx
->membase
= ioremap_nocache(efx
->membase_phys
,
1064 efx
->type
->mem_map_size
);
1065 if (!efx
->membase
) {
1066 netif_err(efx
, probe
, efx
->net_dev
,
1067 "could not map memory BAR at %llx+%x\n",
1068 (unsigned long long)efx
->membase_phys
,
1069 efx
->type
->mem_map_size
);
1073 netif_dbg(efx
, probe
, efx
->net_dev
,
1074 "memory BAR at %llx+%x (virtual %p)\n",
1075 (unsigned long long)efx
->membase_phys
,
1076 efx
->type
->mem_map_size
, efx
->membase
);
1081 pci_release_region(efx
->pci_dev
, EFX_MEM_BAR
);
1083 efx
->membase_phys
= 0;
1085 pci_disable_device(efx
->pci_dev
);
1090 static void efx_fini_io(struct efx_nic
*efx
)
1092 netif_dbg(efx
, drv
, efx
->net_dev
, "shutting down I/O\n");
1095 iounmap(efx
->membase
);
1096 efx
->membase
= NULL
;
1099 if (efx
->membase_phys
) {
1100 pci_release_region(efx
->pci_dev
, EFX_MEM_BAR
);
1101 efx
->membase_phys
= 0;
1104 pci_disable_device(efx
->pci_dev
);
1107 static unsigned int efx_wanted_parallelism(void)
1109 cpumask_var_t thread_mask
;
1116 if (unlikely(!zalloc_cpumask_var(&thread_mask
, GFP_KERNEL
))) {
1118 "sfc: RSS disabled due to allocation failure\n");
1123 for_each_online_cpu(cpu
) {
1124 if (!cpumask_test_cpu(cpu
, thread_mask
)) {
1126 cpumask_or(thread_mask
, thread_mask
,
1127 topology_thread_cpumask(cpu
));
1131 free_cpumask_var(thread_mask
);
1136 efx_init_rx_cpu_rmap(struct efx_nic
*efx
, struct msix_entry
*xentries
)
1138 #ifdef CONFIG_RFS_ACCEL
1142 efx
->net_dev
->rx_cpu_rmap
= alloc_irq_cpu_rmap(efx
->n_rx_channels
);
1143 if (!efx
->net_dev
->rx_cpu_rmap
)
1145 for (i
= 0; i
< efx
->n_rx_channels
; i
++) {
1146 rc
= irq_cpu_rmap_add(efx
->net_dev
->rx_cpu_rmap
,
1147 xentries
[i
].vector
);
1149 free_irq_cpu_rmap(efx
->net_dev
->rx_cpu_rmap
);
1150 efx
->net_dev
->rx_cpu_rmap
= NULL
;
1158 /* Probe the number and type of interrupts we are able to obtain, and
1159 * the resulting numbers of channels and RX queues.
1161 static int efx_probe_interrupts(struct efx_nic
*efx
)
1163 unsigned int max_channels
=
1164 min(efx
->type
->phys_addr_channels
, EFX_MAX_CHANNELS
);
1168 if (efx
->interrupt_mode
== EFX_INT_MODE_MSIX
) {
1169 struct msix_entry xentries
[EFX_MAX_CHANNELS
];
1170 unsigned int n_channels
;
1172 n_channels
= efx_wanted_parallelism();
1173 if (separate_tx_channels
)
1175 n_channels
= min(n_channels
, max_channels
);
1177 for (i
= 0; i
< n_channels
; i
++)
1178 xentries
[i
].entry
= i
;
1179 rc
= pci_enable_msix(efx
->pci_dev
, xentries
, n_channels
);
1181 netif_err(efx
, drv
, efx
->net_dev
,
1182 "WARNING: Insufficient MSI-X vectors"
1183 " available (%d < %u).\n", rc
, n_channels
);
1184 netif_err(efx
, drv
, efx
->net_dev
,
1185 "WARNING: Performance may be reduced.\n");
1186 EFX_BUG_ON_PARANOID(rc
>= n_channels
);
1188 rc
= pci_enable_msix(efx
->pci_dev
, xentries
,
1193 efx
->n_channels
= n_channels
;
1194 if (separate_tx_channels
) {
1195 efx
->n_tx_channels
=
1196 max(efx
->n_channels
/ 2, 1U);
1197 efx
->n_rx_channels
=
1198 max(efx
->n_channels
-
1199 efx
->n_tx_channels
, 1U);
1201 efx
->n_tx_channels
= efx
->n_channels
;
1202 efx
->n_rx_channels
= efx
->n_channels
;
1204 rc
= efx_init_rx_cpu_rmap(efx
, xentries
);
1206 pci_disable_msix(efx
->pci_dev
);
1209 for (i
= 0; i
< n_channels
; i
++)
1210 efx_get_channel(efx
, i
)->irq
=
1213 /* Fall back to single channel MSI */
1214 efx
->interrupt_mode
= EFX_INT_MODE_MSI
;
1215 netif_err(efx
, drv
, efx
->net_dev
,
1216 "could not enable MSI-X\n");
1220 /* Try single interrupt MSI */
1221 if (efx
->interrupt_mode
== EFX_INT_MODE_MSI
) {
1222 efx
->n_channels
= 1;
1223 efx
->n_rx_channels
= 1;
1224 efx
->n_tx_channels
= 1;
1225 rc
= pci_enable_msi(efx
->pci_dev
);
1227 efx_get_channel(efx
, 0)->irq
= efx
->pci_dev
->irq
;
1229 netif_err(efx
, drv
, efx
->net_dev
,
1230 "could not enable MSI\n");
1231 efx
->interrupt_mode
= EFX_INT_MODE_LEGACY
;
1235 /* Assume legacy interrupts */
1236 if (efx
->interrupt_mode
== EFX_INT_MODE_LEGACY
) {
1237 efx
->n_channels
= 1 + (separate_tx_channels
? 1 : 0);
1238 efx
->n_rx_channels
= 1;
1239 efx
->n_tx_channels
= 1;
1240 efx
->legacy_irq
= efx
->pci_dev
->irq
;
1246 /* Enable interrupts, then probe and start the event queues */
1247 static void efx_start_interrupts(struct efx_nic
*efx
)
1249 struct efx_channel
*channel
;
1251 if (efx
->legacy_irq
)
1252 efx
->legacy_irq_enabled
= true;
1253 efx_nic_enable_interrupts(efx
);
1255 efx_for_each_channel(channel
, efx
) {
1256 efx_init_eventq(channel
);
1257 efx_start_eventq(channel
);
1260 efx_mcdi_mode_event(efx
);
1263 static void efx_stop_interrupts(struct efx_nic
*efx
)
1265 struct efx_channel
*channel
;
1267 efx_mcdi_mode_poll(efx
);
1269 efx_nic_disable_interrupts(efx
);
1270 if (efx
->legacy_irq
) {
1271 synchronize_irq(efx
->legacy_irq
);
1272 efx
->legacy_irq_enabled
= false;
1275 efx_for_each_channel(channel
, efx
) {
1277 synchronize_irq(channel
->irq
);
1279 efx_stop_eventq(channel
);
1280 efx_fini_eventq(channel
);
1284 static void efx_remove_interrupts(struct efx_nic
*efx
)
1286 struct efx_channel
*channel
;
1288 /* Remove MSI/MSI-X interrupts */
1289 efx_for_each_channel(channel
, efx
)
1291 pci_disable_msi(efx
->pci_dev
);
1292 pci_disable_msix(efx
->pci_dev
);
1294 /* Remove legacy interrupt */
1295 efx
->legacy_irq
= 0;
1298 static void efx_set_channels(struct efx_nic
*efx
)
1300 struct efx_channel
*channel
;
1301 struct efx_tx_queue
*tx_queue
;
1303 efx
->tx_channel_offset
=
1304 separate_tx_channels
? efx
->n_channels
- efx
->n_tx_channels
: 0;
1306 /* We need to adjust the TX queue numbers if we have separate
1307 * RX-only and TX-only channels.
1309 efx_for_each_channel(channel
, efx
) {
1310 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1311 tx_queue
->queue
-= (efx
->tx_channel_offset
*
1316 static int efx_probe_nic(struct efx_nic
*efx
)
1321 netif_dbg(efx
, probe
, efx
->net_dev
, "creating NIC\n");
1323 /* Carry out hardware-type specific initialisation */
1324 rc
= efx
->type
->probe(efx
);
1328 /* Determine the number of channels and queues by trying to hook
1329 * in MSI-X interrupts. */
1330 rc
= efx_probe_interrupts(efx
);
1334 if (efx
->n_channels
> 1)
1335 get_random_bytes(&efx
->rx_hash_key
, sizeof(efx
->rx_hash_key
));
1336 for (i
= 0; i
< ARRAY_SIZE(efx
->rx_indir_table
); i
++)
1337 efx
->rx_indir_table
[i
] =
1338 ethtool_rxfh_indir_default(i
, efx
->n_rx_channels
);
1340 efx_set_channels(efx
);
1341 netif_set_real_num_tx_queues(efx
->net_dev
, efx
->n_tx_channels
);
1342 netif_set_real_num_rx_queues(efx
->net_dev
, efx
->n_rx_channels
);
1344 /* Initialise the interrupt moderation settings */
1345 efx_init_irq_moderation(efx
, tx_irq_mod_usec
, rx_irq_mod_usec
, true,
1351 efx
->type
->remove(efx
);
1355 static void efx_remove_nic(struct efx_nic
*efx
)
1357 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying NIC\n");
1359 efx_remove_interrupts(efx
);
1360 efx
->type
->remove(efx
);
1363 /**************************************************************************
1365 * NIC startup/shutdown
1367 *************************************************************************/
1369 static int efx_probe_all(struct efx_nic
*efx
)
1373 rc
= efx_probe_nic(efx
);
1375 netif_err(efx
, probe
, efx
->net_dev
, "failed to create NIC\n");
1379 rc
= efx_probe_port(efx
);
1381 netif_err(efx
, probe
, efx
->net_dev
, "failed to create port\n");
1385 efx
->rxq_entries
= efx
->txq_entries
= EFX_DEFAULT_DMAQ_SIZE
;
1386 rc
= efx_probe_channels(efx
);
1390 rc
= efx_probe_filters(efx
);
1392 netif_err(efx
, probe
, efx
->net_dev
,
1393 "failed to create filter tables\n");
1400 efx_remove_channels(efx
);
1402 efx_remove_port(efx
);
1404 efx_remove_nic(efx
);
1409 /* Called after previous invocation(s) of efx_stop_all, restarts the port,
1410 * kernel transmit queues and NAPI processing, and ensures that the port is
1411 * scheduled to be reconfigured. This function is safe to call multiple
1412 * times when the NIC is in any state.
1414 static void efx_start_all(struct efx_nic
*efx
)
1416 EFX_ASSERT_RESET_SERIALISED(efx
);
1418 /* Check that it is appropriate to restart the interface. All
1419 * of these flags are safe to read under just the rtnl lock */
1420 if (efx
->port_enabled
)
1422 if ((efx
->state
!= STATE_RUNNING
) && (efx
->state
!= STATE_INIT
))
1424 if (!netif_running(efx
->net_dev
))
1427 efx_start_port(efx
);
1428 efx_start_datapath(efx
);
1430 /* Start the hardware monitor if there is one. Otherwise (we're link
1431 * event driven), we have to poll the PHY because after an event queue
1432 * flush, we could have a missed a link state change */
1433 if (efx
->type
->monitor
!= NULL
) {
1434 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1435 efx_monitor_interval
);
1437 mutex_lock(&efx
->mac_lock
);
1438 if (efx
->phy_op
->poll(efx
))
1439 efx_link_status_changed(efx
);
1440 mutex_unlock(&efx
->mac_lock
);
1443 efx
->type
->start_stats(efx
);
1446 /* Flush all delayed work. Should only be called when no more delayed work
1447 * will be scheduled. This doesn't flush pending online resets (efx_reset),
1448 * since we're holding the rtnl_lock at this point. */
1449 static void efx_flush_all(struct efx_nic
*efx
)
1451 /* Make sure the hardware monitor is stopped */
1452 cancel_delayed_work_sync(&efx
->monitor_work
);
1453 /* Stop scheduled port reconfigurations */
1454 cancel_work_sync(&efx
->mac_work
);
1457 /* Quiesce hardware and software without bringing the link down.
1458 * Safe to call multiple times, when the nic and interface is in any
1459 * state. The caller is guaranteed to subsequently be in a position
1460 * to modify any hardware and software state they see fit without
1462 static void efx_stop_all(struct efx_nic
*efx
)
1464 EFX_ASSERT_RESET_SERIALISED(efx
);
1466 /* port_enabled can be read safely under the rtnl lock */
1467 if (!efx
->port_enabled
)
1470 efx
->type
->stop_stats(efx
);
1473 /* Flush efx_mac_work(), refill_workqueue, monitor_work */
1476 /* Stop the kernel transmit interface late, so the watchdog
1477 * timer isn't ticking over the flush */
1478 netif_tx_disable(efx
->net_dev
);
1480 efx_stop_datapath(efx
);
1483 static void efx_remove_all(struct efx_nic
*efx
)
1485 efx_remove_filters(efx
);
1486 efx_remove_channels(efx
);
1487 efx_remove_port(efx
);
1488 efx_remove_nic(efx
);
1491 /**************************************************************************
1493 * Interrupt moderation
1495 **************************************************************************/
1497 static unsigned int irq_mod_ticks(unsigned int usecs
, unsigned int quantum_ns
)
1501 if (usecs
* 1000 < quantum_ns
)
1502 return 1; /* never round down to 0 */
1503 return usecs
* 1000 / quantum_ns
;
1506 /* Set interrupt moderation parameters */
1507 int efx_init_irq_moderation(struct efx_nic
*efx
, unsigned int tx_usecs
,
1508 unsigned int rx_usecs
, bool rx_adaptive
,
1509 bool rx_may_override_tx
)
1511 struct efx_channel
*channel
;
1512 unsigned int irq_mod_max
= DIV_ROUND_UP(efx
->type
->timer_period_max
*
1513 efx
->timer_quantum_ns
,
1515 unsigned int tx_ticks
;
1516 unsigned int rx_ticks
;
1518 EFX_ASSERT_RESET_SERIALISED(efx
);
1520 if (tx_usecs
> irq_mod_max
|| rx_usecs
> irq_mod_max
)
1523 tx_ticks
= irq_mod_ticks(tx_usecs
, efx
->timer_quantum_ns
);
1524 rx_ticks
= irq_mod_ticks(rx_usecs
, efx
->timer_quantum_ns
);
1526 if (tx_ticks
!= rx_ticks
&& efx
->tx_channel_offset
== 0 &&
1527 !rx_may_override_tx
) {
1528 netif_err(efx
, drv
, efx
->net_dev
, "Channels are shared. "
1529 "RX and TX IRQ moderation must be equal\n");
1533 efx
->irq_rx_adaptive
= rx_adaptive
;
1534 efx
->irq_rx_moderation
= rx_ticks
;
1535 efx_for_each_channel(channel
, efx
) {
1536 if (efx_channel_has_rx_queue(channel
))
1537 channel
->irq_moderation
= rx_ticks
;
1538 else if (efx_channel_has_tx_queues(channel
))
1539 channel
->irq_moderation
= tx_ticks
;
1545 void efx_get_irq_moderation(struct efx_nic
*efx
, unsigned int *tx_usecs
,
1546 unsigned int *rx_usecs
, bool *rx_adaptive
)
1548 /* We must round up when converting ticks to microseconds
1549 * because we round down when converting the other way.
1552 *rx_adaptive
= efx
->irq_rx_adaptive
;
1553 *rx_usecs
= DIV_ROUND_UP(efx
->irq_rx_moderation
*
1554 efx
->timer_quantum_ns
,
1557 /* If channels are shared between RX and TX, so is IRQ
1558 * moderation. Otherwise, IRQ moderation is the same for all
1559 * TX channels and is not adaptive.
1561 if (efx
->tx_channel_offset
== 0)
1562 *tx_usecs
= *rx_usecs
;
1564 *tx_usecs
= DIV_ROUND_UP(
1565 efx
->channel
[efx
->tx_channel_offset
]->irq_moderation
*
1566 efx
->timer_quantum_ns
,
1570 /**************************************************************************
1574 **************************************************************************/
1576 /* Run periodically off the general workqueue */
1577 static void efx_monitor(struct work_struct
*data
)
1579 struct efx_nic
*efx
= container_of(data
, struct efx_nic
,
1582 netif_vdbg(efx
, timer
, efx
->net_dev
,
1583 "hardware monitor executing on CPU %d\n",
1584 raw_smp_processor_id());
1585 BUG_ON(efx
->type
->monitor
== NULL
);
1587 /* If the mac_lock is already held then it is likely a port
1588 * reconfiguration is already in place, which will likely do
1589 * most of the work of monitor() anyway. */
1590 if (mutex_trylock(&efx
->mac_lock
)) {
1591 if (efx
->port_enabled
)
1592 efx
->type
->monitor(efx
);
1593 mutex_unlock(&efx
->mac_lock
);
1596 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1597 efx_monitor_interval
);
1600 /**************************************************************************
1604 *************************************************************************/
1607 * Context: process, rtnl_lock() held.
1609 static int efx_ioctl(struct net_device
*net_dev
, struct ifreq
*ifr
, int cmd
)
1611 struct efx_nic
*efx
= netdev_priv(net_dev
);
1612 struct mii_ioctl_data
*data
= if_mii(ifr
);
1614 EFX_ASSERT_RESET_SERIALISED(efx
);
1616 /* Convert phy_id from older PRTAD/DEVAD format */
1617 if ((cmd
== SIOCGMIIREG
|| cmd
== SIOCSMIIREG
) &&
1618 (data
->phy_id
& 0xfc00) == 0x0400)
1619 data
->phy_id
^= MDIO_PHY_ID_C45
| 0x0400;
1621 return mdio_mii_ioctl(&efx
->mdio
, data
, cmd
);
1624 /**************************************************************************
1628 **************************************************************************/
1630 static void efx_init_napi(struct efx_nic
*efx
)
1632 struct efx_channel
*channel
;
1634 efx_for_each_channel(channel
, efx
) {
1635 channel
->napi_dev
= efx
->net_dev
;
1636 netif_napi_add(channel
->napi_dev
, &channel
->napi_str
,
1637 efx_poll
, napi_weight
);
1641 static void efx_fini_napi_channel(struct efx_channel
*channel
)
1643 if (channel
->napi_dev
)
1644 netif_napi_del(&channel
->napi_str
);
1645 channel
->napi_dev
= NULL
;
1648 static void efx_fini_napi(struct efx_nic
*efx
)
1650 struct efx_channel
*channel
;
1652 efx_for_each_channel(channel
, efx
)
1653 efx_fini_napi_channel(channel
);
1656 /**************************************************************************
1658 * Kernel netpoll interface
1660 *************************************************************************/
1662 #ifdef CONFIG_NET_POLL_CONTROLLER
1664 /* Although in the common case interrupts will be disabled, this is not
1665 * guaranteed. However, all our work happens inside the NAPI callback,
1666 * so no locking is required.
1668 static void efx_netpoll(struct net_device
*net_dev
)
1670 struct efx_nic
*efx
= netdev_priv(net_dev
);
1671 struct efx_channel
*channel
;
1673 efx_for_each_channel(channel
, efx
)
1674 efx_schedule_channel(channel
);
1679 /**************************************************************************
1681 * Kernel net device interface
1683 *************************************************************************/
1685 /* Context: process, rtnl_lock() held. */
1686 static int efx_net_open(struct net_device
*net_dev
)
1688 struct efx_nic
*efx
= netdev_priv(net_dev
);
1689 EFX_ASSERT_RESET_SERIALISED(efx
);
1691 netif_dbg(efx
, ifup
, efx
->net_dev
, "opening device on CPU %d\n",
1692 raw_smp_processor_id());
1694 if (efx
->state
== STATE_DISABLED
)
1696 if (efx
->phy_mode
& PHY_MODE_SPECIAL
)
1698 if (efx_mcdi_poll_reboot(efx
) && efx_reset(efx
, RESET_TYPE_ALL
))
1701 /* Notify the kernel of the link state polled during driver load,
1702 * before the monitor starts running */
1703 efx_link_status_changed(efx
);
1709 /* Context: process, rtnl_lock() held.
1710 * Note that the kernel will ignore our return code; this method
1711 * should really be a void.
1713 static int efx_net_stop(struct net_device
*net_dev
)
1715 struct efx_nic
*efx
= netdev_priv(net_dev
);
1717 netif_dbg(efx
, ifdown
, efx
->net_dev
, "closing on CPU %d\n",
1718 raw_smp_processor_id());
1720 if (efx
->state
!= STATE_DISABLED
) {
1721 /* Stop the device and flush all the channels */
1728 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
1729 static struct rtnl_link_stats64
*efx_net_stats(struct net_device
*net_dev
,
1730 struct rtnl_link_stats64
*stats
)
1732 struct efx_nic
*efx
= netdev_priv(net_dev
);
1733 struct efx_mac_stats
*mac_stats
= &efx
->mac_stats
;
1735 spin_lock_bh(&efx
->stats_lock
);
1737 efx
->type
->update_stats(efx
);
1739 stats
->rx_packets
= mac_stats
->rx_packets
;
1740 stats
->tx_packets
= mac_stats
->tx_packets
;
1741 stats
->rx_bytes
= mac_stats
->rx_bytes
;
1742 stats
->tx_bytes
= mac_stats
->tx_bytes
;
1743 stats
->rx_dropped
= efx
->n_rx_nodesc_drop_cnt
;
1744 stats
->multicast
= mac_stats
->rx_multicast
;
1745 stats
->collisions
= mac_stats
->tx_collision
;
1746 stats
->rx_length_errors
= (mac_stats
->rx_gtjumbo
+
1747 mac_stats
->rx_length_error
);
1748 stats
->rx_crc_errors
= mac_stats
->rx_bad
;
1749 stats
->rx_frame_errors
= mac_stats
->rx_align_error
;
1750 stats
->rx_fifo_errors
= mac_stats
->rx_overflow
;
1751 stats
->rx_missed_errors
= mac_stats
->rx_missed
;
1752 stats
->tx_window_errors
= mac_stats
->tx_late_collision
;
1754 stats
->rx_errors
= (stats
->rx_length_errors
+
1755 stats
->rx_crc_errors
+
1756 stats
->rx_frame_errors
+
1757 mac_stats
->rx_symbol_error
);
1758 stats
->tx_errors
= (stats
->tx_window_errors
+
1761 spin_unlock_bh(&efx
->stats_lock
);
1766 /* Context: netif_tx_lock held, BHs disabled. */
1767 static void efx_watchdog(struct net_device
*net_dev
)
1769 struct efx_nic
*efx
= netdev_priv(net_dev
);
1771 netif_err(efx
, tx_err
, efx
->net_dev
,
1772 "TX stuck with port_enabled=%d: resetting channels\n",
1775 efx_schedule_reset(efx
, RESET_TYPE_TX_WATCHDOG
);
1779 /* Context: process, rtnl_lock() held. */
1780 static int efx_change_mtu(struct net_device
*net_dev
, int new_mtu
)
1782 struct efx_nic
*efx
= netdev_priv(net_dev
);
1784 EFX_ASSERT_RESET_SERIALISED(efx
);
1786 if (new_mtu
> EFX_MAX_MTU
)
1791 netif_dbg(efx
, drv
, efx
->net_dev
, "changing MTU to %d\n", new_mtu
);
1793 mutex_lock(&efx
->mac_lock
);
1794 /* Reconfigure the MAC before enabling the dma queues so that
1795 * the RX buffers don't overflow */
1796 net_dev
->mtu
= new_mtu
;
1797 efx
->type
->reconfigure_mac(efx
);
1798 mutex_unlock(&efx
->mac_lock
);
1804 static int efx_set_mac_address(struct net_device
*net_dev
, void *data
)
1806 struct efx_nic
*efx
= netdev_priv(net_dev
);
1807 struct sockaddr
*addr
= data
;
1808 char *new_addr
= addr
->sa_data
;
1810 EFX_ASSERT_RESET_SERIALISED(efx
);
1812 if (!is_valid_ether_addr(new_addr
)) {
1813 netif_err(efx
, drv
, efx
->net_dev
,
1814 "invalid ethernet MAC address requested: %pM\n",
1819 memcpy(net_dev
->dev_addr
, new_addr
, net_dev
->addr_len
);
1821 /* Reconfigure the MAC */
1822 mutex_lock(&efx
->mac_lock
);
1823 efx
->type
->reconfigure_mac(efx
);
1824 mutex_unlock(&efx
->mac_lock
);
1829 /* Context: netif_addr_lock held, BHs disabled. */
1830 static void efx_set_rx_mode(struct net_device
*net_dev
)
1832 struct efx_nic
*efx
= netdev_priv(net_dev
);
1833 struct netdev_hw_addr
*ha
;
1834 union efx_multicast_hash
*mc_hash
= &efx
->multicast_hash
;
1838 efx
->promiscuous
= !!(net_dev
->flags
& IFF_PROMISC
);
1840 /* Build multicast hash table */
1841 if (efx
->promiscuous
|| (net_dev
->flags
& IFF_ALLMULTI
)) {
1842 memset(mc_hash
, 0xff, sizeof(*mc_hash
));
1844 memset(mc_hash
, 0x00, sizeof(*mc_hash
));
1845 netdev_for_each_mc_addr(ha
, net_dev
) {
1846 crc
= ether_crc_le(ETH_ALEN
, ha
->addr
);
1847 bit
= crc
& (EFX_MCAST_HASH_ENTRIES
- 1);
1848 set_bit_le(bit
, mc_hash
->byte
);
1851 /* Broadcast packets go through the multicast hash filter.
1852 * ether_crc_le() of the broadcast address is 0xbe2612ff
1853 * so we always add bit 0xff to the mask.
1855 set_bit_le(0xff, mc_hash
->byte
);
1858 if (efx
->port_enabled
)
1859 queue_work(efx
->workqueue
, &efx
->mac_work
);
1860 /* Otherwise efx_start_port() will do this */
1863 static int efx_set_features(struct net_device
*net_dev
, netdev_features_t data
)
1865 struct efx_nic
*efx
= netdev_priv(net_dev
);
1867 /* If disabling RX n-tuple filtering, clear existing filters */
1868 if (net_dev
->features
& ~data
& NETIF_F_NTUPLE
)
1869 efx_filter_clear_rx(efx
, EFX_FILTER_PRI_MANUAL
);
1874 static const struct net_device_ops efx_netdev_ops
= {
1875 .ndo_open
= efx_net_open
,
1876 .ndo_stop
= efx_net_stop
,
1877 .ndo_get_stats64
= efx_net_stats
,
1878 .ndo_tx_timeout
= efx_watchdog
,
1879 .ndo_start_xmit
= efx_hard_start_xmit
,
1880 .ndo_validate_addr
= eth_validate_addr
,
1881 .ndo_do_ioctl
= efx_ioctl
,
1882 .ndo_change_mtu
= efx_change_mtu
,
1883 .ndo_set_mac_address
= efx_set_mac_address
,
1884 .ndo_set_rx_mode
= efx_set_rx_mode
,
1885 .ndo_set_features
= efx_set_features
,
1886 #ifdef CONFIG_NET_POLL_CONTROLLER
1887 .ndo_poll_controller
= efx_netpoll
,
1889 .ndo_setup_tc
= efx_setup_tc
,
1890 #ifdef CONFIG_RFS_ACCEL
1891 .ndo_rx_flow_steer
= efx_filter_rfs
,
1895 static void efx_update_name(struct efx_nic
*efx
)
1897 strcpy(efx
->name
, efx
->net_dev
->name
);
1898 efx_mtd_rename(efx
);
1899 efx_set_channel_names(efx
);
1902 static int efx_netdev_event(struct notifier_block
*this,
1903 unsigned long event
, void *ptr
)
1905 struct net_device
*net_dev
= ptr
;
1907 if (net_dev
->netdev_ops
== &efx_netdev_ops
&&
1908 event
== NETDEV_CHANGENAME
)
1909 efx_update_name(netdev_priv(net_dev
));
1914 static struct notifier_block efx_netdev_notifier
= {
1915 .notifier_call
= efx_netdev_event
,
1919 show_phy_type(struct device
*dev
, struct device_attribute
*attr
, char *buf
)
1921 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
1922 return sprintf(buf
, "%d\n", efx
->phy_type
);
1924 static DEVICE_ATTR(phy_type
, 0644, show_phy_type
, NULL
);
1926 static int efx_register_netdev(struct efx_nic
*efx
)
1928 struct net_device
*net_dev
= efx
->net_dev
;
1929 struct efx_channel
*channel
;
1932 net_dev
->watchdog_timeo
= 5 * HZ
;
1933 net_dev
->irq
= efx
->pci_dev
->irq
;
1934 net_dev
->netdev_ops
= &efx_netdev_ops
;
1935 SET_ETHTOOL_OPS(net_dev
, &efx_ethtool_ops
);
1939 rc
= dev_alloc_name(net_dev
, net_dev
->name
);
1942 efx_update_name(efx
);
1944 rc
= register_netdevice(net_dev
);
1948 efx_for_each_channel(channel
, efx
) {
1949 struct efx_tx_queue
*tx_queue
;
1950 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1951 efx_init_tx_queue_core_txq(tx_queue
);
1954 /* Always start with carrier off; PHY events will detect the link */
1955 netif_carrier_off(net_dev
);
1959 rc
= device_create_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1961 netif_err(efx
, drv
, efx
->net_dev
,
1962 "failed to init net dev attributes\n");
1963 goto fail_registered
;
1970 netif_err(efx
, drv
, efx
->net_dev
, "could not register net dev\n");
1974 unregister_netdev(net_dev
);
1978 static void efx_unregister_netdev(struct efx_nic
*efx
)
1980 struct efx_channel
*channel
;
1981 struct efx_tx_queue
*tx_queue
;
1986 BUG_ON(netdev_priv(efx
->net_dev
) != efx
);
1988 /* Free up any skbs still remaining. This has to happen before
1989 * we try to unregister the netdev as running their destructors
1990 * may be needed to get the device ref. count to 0. */
1991 efx_for_each_channel(channel
, efx
) {
1992 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1993 efx_release_tx_buffers(tx_queue
);
1996 strlcpy(efx
->name
, pci_name(efx
->pci_dev
), sizeof(efx
->name
));
1997 device_remove_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
1998 unregister_netdev(efx
->net_dev
);
2001 /**************************************************************************
2003 * Device reset and suspend
2005 **************************************************************************/
2007 /* Tears down the entire software state and most of the hardware state
2009 void efx_reset_down(struct efx_nic
*efx
, enum reset_type method
)
2011 EFX_ASSERT_RESET_SERIALISED(efx
);
2014 mutex_lock(&efx
->mac_lock
);
2016 efx_stop_interrupts(efx
);
2017 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
)
2018 efx
->phy_op
->fini(efx
);
2019 efx
->type
->fini(efx
);
2022 /* This function will always ensure that the locks acquired in
2023 * efx_reset_down() are released. A failure return code indicates
2024 * that we were unable to reinitialise the hardware, and the
2025 * driver should be disabled. If ok is false, then the rx and tx
2026 * engines are not restarted, pending a RESET_DISABLE. */
2027 int efx_reset_up(struct efx_nic
*efx
, enum reset_type method
, bool ok
)
2031 EFX_ASSERT_RESET_SERIALISED(efx
);
2033 rc
= efx
->type
->init(efx
);
2035 netif_err(efx
, drv
, efx
->net_dev
, "failed to initialise NIC\n");
2042 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
) {
2043 rc
= efx
->phy_op
->init(efx
);
2046 if (efx
->phy_op
->reconfigure(efx
))
2047 netif_err(efx
, drv
, efx
->net_dev
,
2048 "could not restore PHY settings\n");
2051 efx
->type
->reconfigure_mac(efx
);
2053 efx_start_interrupts(efx
);
2054 efx_restore_filters(efx
);
2056 mutex_unlock(&efx
->mac_lock
);
2063 efx
->port_initialized
= false;
2065 mutex_unlock(&efx
->mac_lock
);
2070 /* Reset the NIC using the specified method. Note that the reset may
2071 * fail, in which case the card will be left in an unusable state.
2073 * Caller must hold the rtnl_lock.
2075 int efx_reset(struct efx_nic
*efx
, enum reset_type method
)
2080 netif_info(efx
, drv
, efx
->net_dev
, "resetting (%s)\n",
2081 RESET_TYPE(method
));
2083 netif_device_detach(efx
->net_dev
);
2084 efx_reset_down(efx
, method
);
2086 rc
= efx
->type
->reset(efx
, method
);
2088 netif_err(efx
, drv
, efx
->net_dev
, "failed to reset hardware\n");
2092 /* Clear flags for the scopes we covered. We assume the NIC and
2093 * driver are now quiescent so that there is no race here.
2095 efx
->reset_pending
&= -(1 << (method
+ 1));
2097 /* Reinitialise bus-mastering, which may have been turned off before
2098 * the reset was scheduled. This is still appropriate, even in the
2099 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2100 * can respond to requests. */
2101 pci_set_master(efx
->pci_dev
);
2104 /* Leave device stopped if necessary */
2105 disabled
= rc
|| method
== RESET_TYPE_DISABLE
;
2106 rc2
= efx_reset_up(efx
, method
, !disabled
);
2114 dev_close(efx
->net_dev
);
2115 netif_err(efx
, drv
, efx
->net_dev
, "has been disabled\n");
2116 efx
->state
= STATE_DISABLED
;
2118 netif_dbg(efx
, drv
, efx
->net_dev
, "reset complete\n");
2119 netif_device_attach(efx
->net_dev
);
2124 /* The worker thread exists so that code that cannot sleep can
2125 * schedule a reset for later.
2127 static void efx_reset_work(struct work_struct
*data
)
2129 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, reset_work
);
2130 unsigned long pending
= ACCESS_ONCE(efx
->reset_pending
);
2135 /* If we're not RUNNING then don't reset. Leave the reset_pending
2136 * flags set so that efx_pci_probe_main will be retried */
2137 if (efx
->state
!= STATE_RUNNING
) {
2138 netif_info(efx
, drv
, efx
->net_dev
,
2139 "scheduled reset quenched. NIC not RUNNING\n");
2144 (void)efx_reset(efx
, fls(pending
) - 1);
2148 void efx_schedule_reset(struct efx_nic
*efx
, enum reset_type type
)
2150 enum reset_type method
;
2153 case RESET_TYPE_INVISIBLE
:
2154 case RESET_TYPE_ALL
:
2155 case RESET_TYPE_WORLD
:
2156 case RESET_TYPE_DISABLE
:
2158 netif_dbg(efx
, drv
, efx
->net_dev
, "scheduling %s reset\n",
2159 RESET_TYPE(method
));
2162 method
= efx
->type
->map_reset_reason(type
);
2163 netif_dbg(efx
, drv
, efx
->net_dev
,
2164 "scheduling %s reset for %s\n",
2165 RESET_TYPE(method
), RESET_TYPE(type
));
2169 set_bit(method
, &efx
->reset_pending
);
2171 /* efx_process_channel() will no longer read events once a
2172 * reset is scheduled. So switch back to poll'd MCDI completions. */
2173 efx_mcdi_mode_poll(efx
);
2175 queue_work(reset_workqueue
, &efx
->reset_work
);
2178 /**************************************************************************
2180 * List of NICs we support
2182 **************************************************************************/
2184 /* PCI device ID table */
2185 static DEFINE_PCI_DEVICE_TABLE(efx_pci_table
) = {
2186 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
,
2187 PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0
),
2188 .driver_data
= (unsigned long) &falcon_a1_nic_type
},
2189 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
,
2190 PCI_DEVICE_ID_SOLARFLARE_SFC4000B
),
2191 .driver_data
= (unsigned long) &falcon_b0_nic_type
},
2192 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0803), /* SFC9020 */
2193 .driver_data
= (unsigned long) &siena_a0_nic_type
},
2194 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0813), /* SFL9021 */
2195 .driver_data
= (unsigned long) &siena_a0_nic_type
},
2196 {0} /* end of list */
2199 /**************************************************************************
2201 * Dummy PHY/MAC operations
2203 * Can be used for some unimplemented operations
2204 * Needed so all function pointers are valid and do not have to be tested
2207 **************************************************************************/
2208 int efx_port_dummy_op_int(struct efx_nic
*efx
)
2212 void efx_port_dummy_op_void(struct efx_nic
*efx
) {}
2214 static bool efx_port_dummy_op_poll(struct efx_nic
*efx
)
2219 static const struct efx_phy_operations efx_dummy_phy_operations
= {
2220 .init
= efx_port_dummy_op_int
,
2221 .reconfigure
= efx_port_dummy_op_int
,
2222 .poll
= efx_port_dummy_op_poll
,
2223 .fini
= efx_port_dummy_op_void
,
2226 /**************************************************************************
2230 **************************************************************************/
2232 /* This zeroes out and then fills in the invariants in a struct
2233 * efx_nic (including all sub-structures).
2235 static int efx_init_struct(struct efx_nic
*efx
, const struct efx_nic_type
*type
,
2236 struct pci_dev
*pci_dev
, struct net_device
*net_dev
)
2240 /* Initialise common structures */
2241 memset(efx
, 0, sizeof(*efx
));
2242 spin_lock_init(&efx
->biu_lock
);
2243 #ifdef CONFIG_SFC_MTD
2244 INIT_LIST_HEAD(&efx
->mtd_list
);
2246 INIT_WORK(&efx
->reset_work
, efx_reset_work
);
2247 INIT_DELAYED_WORK(&efx
->monitor_work
, efx_monitor
);
2248 efx
->pci_dev
= pci_dev
;
2249 efx
->msg_enable
= debug
;
2250 efx
->state
= STATE_INIT
;
2251 strlcpy(efx
->name
, pci_name(pci_dev
), sizeof(efx
->name
));
2253 efx
->net_dev
= net_dev
;
2254 spin_lock_init(&efx
->stats_lock
);
2255 mutex_init(&efx
->mac_lock
);
2256 efx
->phy_op
= &efx_dummy_phy_operations
;
2257 efx
->mdio
.dev
= net_dev
;
2258 INIT_WORK(&efx
->mac_work
, efx_mac_work
);
2259 init_waitqueue_head(&efx
->flush_wq
);
2261 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++) {
2262 efx
->channel
[i
] = efx_alloc_channel(efx
, i
, NULL
);
2263 if (!efx
->channel
[i
])
2269 EFX_BUG_ON_PARANOID(efx
->type
->phys_addr_channels
> EFX_MAX_CHANNELS
);
2271 /* Higher numbered interrupt modes are less capable! */
2272 efx
->interrupt_mode
= max(efx
->type
->max_interrupt_mode
,
2275 /* Would be good to use the net_dev name, but we're too early */
2276 snprintf(efx
->workqueue_name
, sizeof(efx
->workqueue_name
), "sfc%s",
2278 efx
->workqueue
= create_singlethread_workqueue(efx
->workqueue_name
);
2279 if (!efx
->workqueue
)
2285 efx_fini_struct(efx
);
2289 static void efx_fini_struct(struct efx_nic
*efx
)
2293 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++)
2294 kfree(efx
->channel
[i
]);
2296 if (efx
->workqueue
) {
2297 destroy_workqueue(efx
->workqueue
);
2298 efx
->workqueue
= NULL
;
2302 /**************************************************************************
2306 **************************************************************************/
2308 /* Main body of final NIC shutdown code
2309 * This is called only at module unload (or hotplug removal).
2311 static void efx_pci_remove_main(struct efx_nic
*efx
)
2313 #ifdef CONFIG_RFS_ACCEL
2314 free_irq_cpu_rmap(efx
->net_dev
->rx_cpu_rmap
);
2315 efx
->net_dev
->rx_cpu_rmap
= NULL
;
2317 efx_stop_interrupts(efx
);
2318 efx_nic_fini_interrupt(efx
);
2320 efx
->type
->fini(efx
);
2322 efx_remove_all(efx
);
2325 /* Final NIC shutdown
2326 * This is called only at module unload (or hotplug removal).
2328 static void efx_pci_remove(struct pci_dev
*pci_dev
)
2330 struct efx_nic
*efx
;
2332 efx
= pci_get_drvdata(pci_dev
);
2336 /* Mark the NIC as fini, then stop the interface */
2338 efx
->state
= STATE_FINI
;
2339 dev_close(efx
->net_dev
);
2341 /* Allow any queued efx_resets() to complete */
2344 efx_stop_interrupts(efx
);
2345 efx_unregister_netdev(efx
);
2347 efx_mtd_remove(efx
);
2349 /* Wait for any scheduled resets to complete. No more will be
2350 * scheduled from this point because efx_stop_all() has been
2351 * called, we are no longer registered with driverlink, and
2352 * the net_device's have been removed. */
2353 cancel_work_sync(&efx
->reset_work
);
2355 efx_pci_remove_main(efx
);
2358 netif_dbg(efx
, drv
, efx
->net_dev
, "shutdown successful\n");
2360 pci_set_drvdata(pci_dev
, NULL
);
2361 efx_fini_struct(efx
);
2362 free_netdev(efx
->net_dev
);
2365 /* Main body of NIC initialisation
2366 * This is called at module load (or hotplug insertion, theoretically).
2368 static int efx_pci_probe_main(struct efx_nic
*efx
)
2372 /* Do start-of-day initialisation */
2373 rc
= efx_probe_all(efx
);
2379 rc
= efx
->type
->init(efx
);
2381 netif_err(efx
, probe
, efx
->net_dev
,
2382 "failed to initialise NIC\n");
2386 rc
= efx_init_port(efx
);
2388 netif_err(efx
, probe
, efx
->net_dev
,
2389 "failed to initialise port\n");
2393 rc
= efx_nic_init_interrupt(efx
);
2396 efx_start_interrupts(efx
);
2403 efx
->type
->fini(efx
);
2406 efx_remove_all(efx
);
2411 /* NIC initialisation
2413 * This is called at module load (or hotplug insertion,
2414 * theoretically). It sets up PCI mappings, resets the NIC,
2415 * sets up and registers the network devices with the kernel and hooks
2416 * the interrupt service routine. It does not prepare the device for
2417 * transmission; this is left to the first time one of the network
2418 * interfaces is brought up (i.e. efx_net_open).
2420 static int __devinit
efx_pci_probe(struct pci_dev
*pci_dev
,
2421 const struct pci_device_id
*entry
)
2423 const struct efx_nic_type
*type
= (const struct efx_nic_type
*) entry
->driver_data
;
2424 struct net_device
*net_dev
;
2425 struct efx_nic
*efx
;
2428 /* Allocate and initialise a struct net_device and struct efx_nic */
2429 net_dev
= alloc_etherdev_mqs(sizeof(*efx
), EFX_MAX_CORE_TX_QUEUES
,
2433 net_dev
->features
|= (type
->offload_features
| NETIF_F_SG
|
2434 NETIF_F_HIGHDMA
| NETIF_F_TSO
|
2436 if (type
->offload_features
& NETIF_F_V6_CSUM
)
2437 net_dev
->features
|= NETIF_F_TSO6
;
2438 /* Mask for features that also apply to VLAN devices */
2439 net_dev
->vlan_features
|= (NETIF_F_ALL_CSUM
| NETIF_F_SG
|
2440 NETIF_F_HIGHDMA
| NETIF_F_ALL_TSO
|
2442 /* All offloads can be toggled */
2443 net_dev
->hw_features
= net_dev
->features
& ~NETIF_F_HIGHDMA
;
2444 efx
= netdev_priv(net_dev
);
2445 pci_set_drvdata(pci_dev
, efx
);
2446 SET_NETDEV_DEV(net_dev
, &pci_dev
->dev
);
2447 rc
= efx_init_struct(efx
, type
, pci_dev
, net_dev
);
2451 netif_info(efx
, probe
, efx
->net_dev
,
2452 "Solarflare NIC detected\n");
2454 /* Set up basic I/O (BAR mappings etc) */
2455 rc
= efx_init_io(efx
);
2459 rc
= efx_pci_probe_main(efx
);
2461 /* Serialise against efx_reset(). No more resets will be
2462 * scheduled since efx_stop_all() has been called, and we have
2463 * not and never have been registered.
2465 cancel_work_sync(&efx
->reset_work
);
2470 /* If there was a scheduled reset during probe, the NIC is
2471 * probably hosed anyway.
2473 if (efx
->reset_pending
) {
2478 /* Switch to the running state before we expose the device to the OS,
2479 * so that dev_open()|efx_start_all() will actually start the device */
2480 efx
->state
= STATE_RUNNING
;
2482 rc
= efx_register_netdev(efx
);
2486 netif_dbg(efx
, probe
, efx
->net_dev
, "initialisation successful\n");
2488 /* Try to create MTDs, but allow this to fail */
2490 rc
= efx_mtd_probe(efx
);
2493 netif_warn(efx
, probe
, efx
->net_dev
,
2494 "failed to create MTDs (%d)\n", rc
);
2499 efx_pci_remove_main(efx
);
2503 efx_fini_struct(efx
);
2506 netif_dbg(efx
, drv
, efx
->net_dev
, "initialisation failed. rc=%d\n", rc
);
2507 free_netdev(net_dev
);
2511 static int efx_pm_freeze(struct device
*dev
)
2513 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2515 efx
->state
= STATE_FINI
;
2517 netif_device_detach(efx
->net_dev
);
2520 efx_stop_interrupts(efx
);
2525 static int efx_pm_thaw(struct device
*dev
)
2527 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2529 efx
->state
= STATE_INIT
;
2531 efx_start_interrupts(efx
);
2533 mutex_lock(&efx
->mac_lock
);
2534 efx
->phy_op
->reconfigure(efx
);
2535 mutex_unlock(&efx
->mac_lock
);
2539 netif_device_attach(efx
->net_dev
);
2541 efx
->state
= STATE_RUNNING
;
2543 efx
->type
->resume_wol(efx
);
2545 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
2546 queue_work(reset_workqueue
, &efx
->reset_work
);
2551 static int efx_pm_poweroff(struct device
*dev
)
2553 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
2554 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
2556 efx
->type
->fini(efx
);
2558 efx
->reset_pending
= 0;
2560 pci_save_state(pci_dev
);
2561 return pci_set_power_state(pci_dev
, PCI_D3hot
);
2564 /* Used for both resume and restore */
2565 static int efx_pm_resume(struct device
*dev
)
2567 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
2568 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
2571 rc
= pci_set_power_state(pci_dev
, PCI_D0
);
2574 pci_restore_state(pci_dev
);
2575 rc
= pci_enable_device(pci_dev
);
2578 pci_set_master(efx
->pci_dev
);
2579 rc
= efx
->type
->reset(efx
, RESET_TYPE_ALL
);
2582 rc
= efx
->type
->init(efx
);
2589 static int efx_pm_suspend(struct device
*dev
)
2594 rc
= efx_pm_poweroff(dev
);
2600 static const struct dev_pm_ops efx_pm_ops
= {
2601 .suspend
= efx_pm_suspend
,
2602 .resume
= efx_pm_resume
,
2603 .freeze
= efx_pm_freeze
,
2604 .thaw
= efx_pm_thaw
,
2605 .poweroff
= efx_pm_poweroff
,
2606 .restore
= efx_pm_resume
,
2609 static struct pci_driver efx_pci_driver
= {
2610 .name
= KBUILD_MODNAME
,
2611 .id_table
= efx_pci_table
,
2612 .probe
= efx_pci_probe
,
2613 .remove
= efx_pci_remove
,
2614 .driver
.pm
= &efx_pm_ops
,
2617 /**************************************************************************
2619 * Kernel module interface
2621 *************************************************************************/
2623 module_param(interrupt_mode
, uint
, 0444);
2624 MODULE_PARM_DESC(interrupt_mode
,
2625 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
2627 static int __init
efx_init_module(void)
2631 printk(KERN_INFO
"Solarflare NET driver v" EFX_DRIVER_VERSION
"\n");
2633 rc
= register_netdevice_notifier(&efx_netdev_notifier
);
2637 reset_workqueue
= create_singlethread_workqueue("sfc_reset");
2638 if (!reset_workqueue
) {
2643 rc
= pci_register_driver(&efx_pci_driver
);
2650 destroy_workqueue(reset_workqueue
);
2652 unregister_netdevice_notifier(&efx_netdev_notifier
);
2657 static void __exit
efx_exit_module(void)
2659 printk(KERN_INFO
"Solarflare NET driver unloading\n");
2661 pci_unregister_driver(&efx_pci_driver
);
2662 destroy_workqueue(reset_workqueue
);
2663 unregister_netdevice_notifier(&efx_netdev_notifier
);
2667 module_init(efx_init_module
);
2668 module_exit(efx_exit_module
);
2670 MODULE_AUTHOR("Solarflare Communications and "
2671 "Michael Brown <mbrown@fensystems.co.uk>");
2672 MODULE_DESCRIPTION("Solarflare Communications network driver");
2673 MODULE_LICENSE("GPL");
2674 MODULE_DEVICE_TABLE(pci
, efx_pci_table
);