1 /****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2013 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/ethtool.h>
21 #include <linux/topology.h>
22 #include <linux/gfp.h>
23 #include <linux/aer.h>
24 #include <linux/interrupt.h>
25 #include "net_driver.h"
32 #include "workarounds.h"
34 /**************************************************************************
38 **************************************************************************
41 /* Loopback mode names (see LOOPBACK_MODE()) */
42 const unsigned int efx_loopback_mode_max
= LOOPBACK_MAX
;
43 const char *const efx_loopback_mode_names
[] = {
44 [LOOPBACK_NONE
] = "NONE",
45 [LOOPBACK_DATA
] = "DATAPATH",
46 [LOOPBACK_GMAC
] = "GMAC",
47 [LOOPBACK_XGMII
] = "XGMII",
48 [LOOPBACK_XGXS
] = "XGXS",
49 [LOOPBACK_XAUI
] = "XAUI",
50 [LOOPBACK_GMII
] = "GMII",
51 [LOOPBACK_SGMII
] = "SGMII",
52 [LOOPBACK_XGBR
] = "XGBR",
53 [LOOPBACK_XFI
] = "XFI",
54 [LOOPBACK_XAUI_FAR
] = "XAUI_FAR",
55 [LOOPBACK_GMII_FAR
] = "GMII_FAR",
56 [LOOPBACK_SGMII_FAR
] = "SGMII_FAR",
57 [LOOPBACK_XFI_FAR
] = "XFI_FAR",
58 [LOOPBACK_GPHY
] = "GPHY",
59 [LOOPBACK_PHYXS
] = "PHYXS",
60 [LOOPBACK_PCS
] = "PCS",
61 [LOOPBACK_PMAPMD
] = "PMA/PMD",
62 [LOOPBACK_XPORT
] = "XPORT",
63 [LOOPBACK_XGMII_WS
] = "XGMII_WS",
64 [LOOPBACK_XAUI_WS
] = "XAUI_WS",
65 [LOOPBACK_XAUI_WS_FAR
] = "XAUI_WS_FAR",
66 [LOOPBACK_XAUI_WS_NEAR
] = "XAUI_WS_NEAR",
67 [LOOPBACK_GMII_WS
] = "GMII_WS",
68 [LOOPBACK_XFI_WS
] = "XFI_WS",
69 [LOOPBACK_XFI_WS_FAR
] = "XFI_WS_FAR",
70 [LOOPBACK_PHYXS_WS
] = "PHYXS_WS",
73 const unsigned int efx_reset_type_max
= RESET_TYPE_MAX
;
74 const char *const efx_reset_type_names
[] = {
75 [RESET_TYPE_INVISIBLE
] = "INVISIBLE",
76 [RESET_TYPE_ALL
] = "ALL",
77 [RESET_TYPE_RECOVER_OR_ALL
] = "RECOVER_OR_ALL",
78 [RESET_TYPE_WORLD
] = "WORLD",
79 [RESET_TYPE_RECOVER_OR_DISABLE
] = "RECOVER_OR_DISABLE",
80 [RESET_TYPE_MC_BIST
] = "MC_BIST",
81 [RESET_TYPE_DISABLE
] = "DISABLE",
82 [RESET_TYPE_TX_WATCHDOG
] = "TX_WATCHDOG",
83 [RESET_TYPE_INT_ERROR
] = "INT_ERROR",
84 [RESET_TYPE_RX_RECOVERY
] = "RX_RECOVERY",
85 [RESET_TYPE_DMA_ERROR
] = "DMA_ERROR",
86 [RESET_TYPE_TX_SKIP
] = "TX_SKIP",
87 [RESET_TYPE_MC_FAILURE
] = "MC_FAILURE",
88 [RESET_TYPE_MCDI_TIMEOUT
] = "MCDI_TIMEOUT (FLR)",
91 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
92 * queued onto this work queue. This is not a per-nic work queue, because
93 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
95 static struct workqueue_struct
*reset_workqueue
;
97 /* How often and how many times to poll for a reset while waiting for a
98 * BIST that another function started to complete.
100 #define BIST_WAIT_DELAY_MS 100
101 #define BIST_WAIT_DELAY_COUNT 100
103 /**************************************************************************
105 * Configurable values
107 *************************************************************************/
110 * Use separate channels for TX and RX events
112 * Set this to 1 to use separate channels for TX and RX. It allows us
113 * to control interrupt affinity separately for TX and RX.
115 * This is only used in MSI-X interrupt mode
117 static bool separate_tx_channels
;
118 module_param(separate_tx_channels
, bool, 0444);
119 MODULE_PARM_DESC(separate_tx_channels
,
120 "Use separate channels for TX and RX");
122 /* This is the weight assigned to each of the (per-channel) virtual
125 static int napi_weight
= 64;
127 /* This is the time (in jiffies) between invocations of the hardware
129 * On Falcon-based NICs, this will:
130 * - Check the on-board hardware monitor;
131 * - Poll the link state and reconfigure the hardware as necessary.
132 * On Siena-based NICs for power systems with EEH support, this will give EEH a
135 static unsigned int efx_monitor_interval
= 1 * HZ
;
137 /* Initial interrupt moderation settings. They can be modified after
138 * module load with ethtool.
140 * The default for RX should strike a balance between increasing the
141 * round-trip latency and reducing overhead.
143 static unsigned int rx_irq_mod_usec
= 60;
145 /* Initial interrupt moderation settings. They can be modified after
146 * module load with ethtool.
148 * This default is chosen to ensure that a 10G link does not go idle
149 * while a TX queue is stopped after it has become full. A queue is
150 * restarted when it drops below half full. The time this takes (assuming
151 * worst case 3 descriptors per packet and 1024 descriptors) is
152 * 512 / 3 * 1.2 = 205 usec.
154 static unsigned int tx_irq_mod_usec
= 150;
156 /* This is the first interrupt mode to try out of:
161 static unsigned int interrupt_mode
;
163 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
164 * i.e. the number of CPUs among which we may distribute simultaneous
165 * interrupt handling.
167 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
168 * The default (0) means to assign an interrupt to each core.
170 static unsigned int rss_cpus
;
171 module_param(rss_cpus
, uint
, 0444);
172 MODULE_PARM_DESC(rss_cpus
, "Number of CPUs to use for Receive-Side Scaling");
174 static bool phy_flash_cfg
;
175 module_param(phy_flash_cfg
, bool, 0644);
176 MODULE_PARM_DESC(phy_flash_cfg
, "Set PHYs into reflash mode initially");
178 static unsigned irq_adapt_low_thresh
= 8000;
179 module_param(irq_adapt_low_thresh
, uint
, 0644);
180 MODULE_PARM_DESC(irq_adapt_low_thresh
,
181 "Threshold score for reducing IRQ moderation");
183 static unsigned irq_adapt_high_thresh
= 16000;
184 module_param(irq_adapt_high_thresh
, uint
, 0644);
185 MODULE_PARM_DESC(irq_adapt_high_thresh
,
186 "Threshold score for increasing IRQ moderation");
188 static unsigned debug
= (NETIF_MSG_DRV
| NETIF_MSG_PROBE
|
189 NETIF_MSG_LINK
| NETIF_MSG_IFDOWN
|
190 NETIF_MSG_IFUP
| NETIF_MSG_RX_ERR
|
191 NETIF_MSG_TX_ERR
| NETIF_MSG_HW
);
192 module_param(debug
, uint
, 0);
193 MODULE_PARM_DESC(debug
, "Bitmapped debugging message enable value");
195 /**************************************************************************
197 * Utility functions and prototypes
199 *************************************************************************/
201 static int efx_soft_enable_interrupts(struct efx_nic
*efx
);
202 static void efx_soft_disable_interrupts(struct efx_nic
*efx
);
203 static void efx_remove_channel(struct efx_channel
*channel
);
204 static void efx_remove_channels(struct efx_nic
*efx
);
205 static const struct efx_channel_type efx_default_channel_type
;
206 static void efx_remove_port(struct efx_nic
*efx
);
207 static void efx_init_napi_channel(struct efx_channel
*channel
);
208 static void efx_fini_napi(struct efx_nic
*efx
);
209 static void efx_fini_napi_channel(struct efx_channel
*channel
);
210 static void efx_fini_struct(struct efx_nic
*efx
);
211 static void efx_start_all(struct efx_nic
*efx
);
212 static void efx_stop_all(struct efx_nic
*efx
);
214 #define EFX_ASSERT_RESET_SERIALISED(efx) \
216 if ((efx->state == STATE_READY) || \
217 (efx->state == STATE_RECOVERY) || \
218 (efx->state == STATE_DISABLED)) \
222 static int efx_check_disabled(struct efx_nic
*efx
)
224 if (efx
->state
== STATE_DISABLED
|| efx
->state
== STATE_RECOVERY
) {
225 netif_err(efx
, drv
, efx
->net_dev
,
226 "device is disabled due to earlier errors\n");
232 /**************************************************************************
234 * Event queue processing
236 *************************************************************************/
238 /* Process channel's event queue
240 * This function is responsible for processing the event queue of a
241 * single channel. The caller must guarantee that this function will
242 * never be concurrently called more than once on the same channel,
243 * though different channels may be being processed concurrently.
245 static int efx_process_channel(struct efx_channel
*channel
, int budget
)
249 if (unlikely(!channel
->enabled
))
252 spent
= efx_nic_process_eventq(channel
, budget
);
253 if (spent
&& efx_channel_has_rx_queue(channel
)) {
254 struct efx_rx_queue
*rx_queue
=
255 efx_channel_get_rx_queue(channel
);
257 efx_rx_flush_packet(channel
);
258 efx_fast_push_rx_descriptors(rx_queue
, true);
266 * NAPI guarantees serialisation of polls of the same device, which
267 * provides the guarantee required by efx_process_channel().
269 static int efx_poll(struct napi_struct
*napi
, int budget
)
271 struct efx_channel
*channel
=
272 container_of(napi
, struct efx_channel
, napi_str
);
273 struct efx_nic
*efx
= channel
->efx
;
276 if (!efx_channel_lock_napi(channel
))
279 netif_vdbg(efx
, intr
, efx
->net_dev
,
280 "channel %d NAPI poll executing on CPU %d\n",
281 channel
->channel
, raw_smp_processor_id());
283 spent
= efx_process_channel(channel
, budget
);
285 if (spent
< budget
) {
286 if (efx_channel_has_rx_queue(channel
) &&
287 efx
->irq_rx_adaptive
&&
288 unlikely(++channel
->irq_count
== 1000)) {
289 if (unlikely(channel
->irq_mod_score
<
290 irq_adapt_low_thresh
)) {
291 if (channel
->irq_moderation
> 1) {
292 channel
->irq_moderation
-= 1;
293 efx
->type
->push_irq_moderation(channel
);
295 } else if (unlikely(channel
->irq_mod_score
>
296 irq_adapt_high_thresh
)) {
297 if (channel
->irq_moderation
<
298 efx
->irq_rx_moderation
) {
299 channel
->irq_moderation
+= 1;
300 efx
->type
->push_irq_moderation(channel
);
303 channel
->irq_count
= 0;
304 channel
->irq_mod_score
= 0;
307 efx_filter_rfs_expire(channel
);
309 /* There is no race here; although napi_disable() will
310 * only wait for napi_complete(), this isn't a problem
311 * since efx_nic_eventq_read_ack() will have no effect if
312 * interrupts have already been disabled.
315 efx_nic_eventq_read_ack(channel
);
318 efx_channel_unlock_napi(channel
);
322 /* Create event queue
323 * Event queue memory allocations are done only once. If the channel
324 * is reset, the memory buffer will be reused; this guards against
325 * errors during channel reset and also simplifies interrupt handling.
327 static int efx_probe_eventq(struct efx_channel
*channel
)
329 struct efx_nic
*efx
= channel
->efx
;
330 unsigned long entries
;
332 netif_dbg(efx
, probe
, efx
->net_dev
,
333 "chan %d create event queue\n", channel
->channel
);
335 /* Build an event queue with room for one event per tx and rx buffer,
336 * plus some extra for link state events and MCDI completions. */
337 entries
= roundup_pow_of_two(efx
->rxq_entries
+ efx
->txq_entries
+ 128);
338 EFX_BUG_ON_PARANOID(entries
> EFX_MAX_EVQ_SIZE
);
339 channel
->eventq_mask
= max(entries
, EFX_MIN_EVQ_SIZE
) - 1;
341 return efx_nic_probe_eventq(channel
);
344 /* Prepare channel's event queue */
345 static int efx_init_eventq(struct efx_channel
*channel
)
347 struct efx_nic
*efx
= channel
->efx
;
350 EFX_WARN_ON_PARANOID(channel
->eventq_init
);
352 netif_dbg(efx
, drv
, efx
->net_dev
,
353 "chan %d init event queue\n", channel
->channel
);
355 rc
= efx_nic_init_eventq(channel
);
357 efx
->type
->push_irq_moderation(channel
);
358 channel
->eventq_read_ptr
= 0;
359 channel
->eventq_init
= true;
364 /* Enable event queue processing and NAPI */
365 void efx_start_eventq(struct efx_channel
*channel
)
367 netif_dbg(channel
->efx
, ifup
, channel
->efx
->net_dev
,
368 "chan %d start event queue\n", channel
->channel
);
370 /* Make sure the NAPI handler sees the enabled flag set */
371 channel
->enabled
= true;
374 efx_channel_enable(channel
);
375 napi_enable(&channel
->napi_str
);
376 efx_nic_eventq_read_ack(channel
);
379 /* Disable event queue processing and NAPI */
380 void efx_stop_eventq(struct efx_channel
*channel
)
382 if (!channel
->enabled
)
385 napi_disable(&channel
->napi_str
);
386 while (!efx_channel_disable(channel
))
387 usleep_range(1000, 20000);
388 channel
->enabled
= false;
391 static void efx_fini_eventq(struct efx_channel
*channel
)
393 if (!channel
->eventq_init
)
396 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
397 "chan %d fini event queue\n", channel
->channel
);
399 efx_nic_fini_eventq(channel
);
400 channel
->eventq_init
= false;
403 static void efx_remove_eventq(struct efx_channel
*channel
)
405 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
406 "chan %d remove event queue\n", channel
->channel
);
408 efx_nic_remove_eventq(channel
);
411 /**************************************************************************
415 *************************************************************************/
417 /* Allocate and initialise a channel structure. */
418 static struct efx_channel
*
419 efx_alloc_channel(struct efx_nic
*efx
, int i
, struct efx_channel
*old_channel
)
421 struct efx_channel
*channel
;
422 struct efx_rx_queue
*rx_queue
;
423 struct efx_tx_queue
*tx_queue
;
426 channel
= kzalloc(sizeof(*channel
), GFP_KERNEL
);
431 channel
->channel
= i
;
432 channel
->type
= &efx_default_channel_type
;
434 for (j
= 0; j
< EFX_TXQ_TYPES
; j
++) {
435 tx_queue
= &channel
->tx_queue
[j
];
437 tx_queue
->queue
= i
* EFX_TXQ_TYPES
+ j
;
438 tx_queue
->channel
= channel
;
441 rx_queue
= &channel
->rx_queue
;
443 setup_timer(&rx_queue
->slow_fill
, efx_rx_slow_fill
,
444 (unsigned long)rx_queue
);
449 /* Allocate and initialise a channel structure, copying parameters
450 * (but not resources) from an old channel structure.
452 static struct efx_channel
*
453 efx_copy_channel(const struct efx_channel
*old_channel
)
455 struct efx_channel
*channel
;
456 struct efx_rx_queue
*rx_queue
;
457 struct efx_tx_queue
*tx_queue
;
460 channel
= kmalloc(sizeof(*channel
), GFP_KERNEL
);
464 *channel
= *old_channel
;
466 channel
->napi_dev
= NULL
;
467 memset(&channel
->eventq
, 0, sizeof(channel
->eventq
));
469 for (j
= 0; j
< EFX_TXQ_TYPES
; j
++) {
470 tx_queue
= &channel
->tx_queue
[j
];
471 if (tx_queue
->channel
)
472 tx_queue
->channel
= channel
;
473 tx_queue
->buffer
= NULL
;
474 memset(&tx_queue
->txd
, 0, sizeof(tx_queue
->txd
));
477 rx_queue
= &channel
->rx_queue
;
478 rx_queue
->buffer
= NULL
;
479 memset(&rx_queue
->rxd
, 0, sizeof(rx_queue
->rxd
));
480 setup_timer(&rx_queue
->slow_fill
, efx_rx_slow_fill
,
481 (unsigned long)rx_queue
);
486 static int efx_probe_channel(struct efx_channel
*channel
)
488 struct efx_tx_queue
*tx_queue
;
489 struct efx_rx_queue
*rx_queue
;
492 netif_dbg(channel
->efx
, probe
, channel
->efx
->net_dev
,
493 "creating channel %d\n", channel
->channel
);
495 rc
= channel
->type
->pre_probe(channel
);
499 rc
= efx_probe_eventq(channel
);
503 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
504 rc
= efx_probe_tx_queue(tx_queue
);
509 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
510 rc
= efx_probe_rx_queue(rx_queue
);
518 efx_remove_channel(channel
);
523 efx_get_channel_name(struct efx_channel
*channel
, char *buf
, size_t len
)
525 struct efx_nic
*efx
= channel
->efx
;
529 number
= channel
->channel
;
530 if (efx
->tx_channel_offset
== 0) {
532 } else if (channel
->channel
< efx
->tx_channel_offset
) {
536 number
-= efx
->tx_channel_offset
;
538 snprintf(buf
, len
, "%s%s-%d", efx
->name
, type
, number
);
541 static void efx_set_channel_names(struct efx_nic
*efx
)
543 struct efx_channel
*channel
;
545 efx_for_each_channel(channel
, efx
)
546 channel
->type
->get_name(channel
,
547 efx
->msi_context
[channel
->channel
].name
,
548 sizeof(efx
->msi_context
[0].name
));
551 static int efx_probe_channels(struct efx_nic
*efx
)
553 struct efx_channel
*channel
;
556 /* Restart special buffer allocation */
557 efx
->next_buffer_table
= 0;
559 /* Probe channels in reverse, so that any 'extra' channels
560 * use the start of the buffer table. This allows the traffic
561 * channels to be resized without moving them or wasting the
562 * entries before them.
564 efx_for_each_channel_rev(channel
, efx
) {
565 rc
= efx_probe_channel(channel
);
567 netif_err(efx
, probe
, efx
->net_dev
,
568 "failed to create channel %d\n",
573 efx_set_channel_names(efx
);
578 efx_remove_channels(efx
);
582 /* Channels are shutdown and reinitialised whilst the NIC is running
583 * to propagate configuration changes (mtu, checksum offload), or
584 * to clear hardware error conditions
586 static void efx_start_datapath(struct efx_nic
*efx
)
588 bool old_rx_scatter
= efx
->rx_scatter
;
589 struct efx_tx_queue
*tx_queue
;
590 struct efx_rx_queue
*rx_queue
;
591 struct efx_channel
*channel
;
594 /* Calculate the rx buffer allocation parameters required to
595 * support the current MTU, including padding for header
596 * alignment and overruns.
598 efx
->rx_dma_len
= (efx
->rx_prefix_size
+
599 EFX_MAX_FRAME_LEN(efx
->net_dev
->mtu
) +
600 efx
->type
->rx_buffer_padding
);
601 rx_buf_len
= (sizeof(struct efx_rx_page_state
) +
602 efx
->rx_ip_align
+ efx
->rx_dma_len
);
603 if (rx_buf_len
<= PAGE_SIZE
) {
604 efx
->rx_scatter
= efx
->type
->always_rx_scatter
;
605 efx
->rx_buffer_order
= 0;
606 } else if (efx
->type
->can_rx_scatter
) {
607 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE
% L1_CACHE_BYTES
);
608 BUILD_BUG_ON(sizeof(struct efx_rx_page_state
) +
609 2 * ALIGN(NET_IP_ALIGN
+ EFX_RX_USR_BUF_SIZE
,
610 EFX_RX_BUF_ALIGNMENT
) >
612 efx
->rx_scatter
= true;
613 efx
->rx_dma_len
= EFX_RX_USR_BUF_SIZE
;
614 efx
->rx_buffer_order
= 0;
616 efx
->rx_scatter
= false;
617 efx
->rx_buffer_order
= get_order(rx_buf_len
);
620 efx_rx_config_page_split(efx
);
621 if (efx
->rx_buffer_order
)
622 netif_dbg(efx
, drv
, efx
->net_dev
,
623 "RX buf len=%u; page order=%u batch=%u\n",
624 efx
->rx_dma_len
, efx
->rx_buffer_order
,
625 efx
->rx_pages_per_batch
);
627 netif_dbg(efx
, drv
, efx
->net_dev
,
628 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
629 efx
->rx_dma_len
, efx
->rx_page_buf_step
,
630 efx
->rx_bufs_per_page
, efx
->rx_pages_per_batch
);
632 /* RX filters may also have scatter-enabled flags */
633 if (efx
->rx_scatter
!= old_rx_scatter
)
634 efx
->type
->filter_update_rx_scatter(efx
);
636 /* We must keep at least one descriptor in a TX ring empty.
637 * We could avoid this when the queue size does not exactly
638 * match the hardware ring size, but it's not that important.
639 * Therefore we stop the queue when one more skb might fill
640 * the ring completely. We wake it when half way back to
643 efx
->txq_stop_thresh
= efx
->txq_entries
- efx_tx_max_skb_descs(efx
);
644 efx
->txq_wake_thresh
= efx
->txq_stop_thresh
/ 2;
646 /* Initialise the channels */
647 efx_for_each_channel(channel
, efx
) {
648 efx_for_each_channel_tx_queue(tx_queue
, channel
) {
649 efx_init_tx_queue(tx_queue
);
650 atomic_inc(&efx
->active_queues
);
653 efx_for_each_channel_rx_queue(rx_queue
, channel
) {
654 efx_init_rx_queue(rx_queue
);
655 atomic_inc(&efx
->active_queues
);
656 efx_stop_eventq(channel
);
657 efx_fast_push_rx_descriptors(rx_queue
, false);
658 efx_start_eventq(channel
);
661 WARN_ON(channel
->rx_pkt_n_frags
);
664 efx_ptp_start_datapath(efx
);
666 if (netif_device_present(efx
->net_dev
))
667 netif_tx_wake_all_queues(efx
->net_dev
);
670 static void efx_stop_datapath(struct efx_nic
*efx
)
672 struct efx_channel
*channel
;
673 struct efx_tx_queue
*tx_queue
;
674 struct efx_rx_queue
*rx_queue
;
677 EFX_ASSERT_RESET_SERIALISED(efx
);
678 BUG_ON(efx
->port_enabled
);
680 efx_ptp_stop_datapath(efx
);
683 efx_for_each_channel(channel
, efx
) {
684 efx_for_each_channel_rx_queue(rx_queue
, channel
)
685 rx_queue
->refill_enabled
= false;
688 efx_for_each_channel(channel
, efx
) {
689 /* RX packet processing is pipelined, so wait for the
690 * NAPI handler to complete. At least event queue 0
691 * might be kept active by non-data events, so don't
692 * use napi_synchronize() but actually disable NAPI
695 if (efx_channel_has_rx_queue(channel
)) {
696 efx_stop_eventq(channel
);
697 efx_start_eventq(channel
);
701 rc
= efx
->type
->fini_dmaq(efx
);
702 if (rc
&& EFX_WORKAROUND_7803(efx
)) {
703 /* Schedule a reset to recover from the flush failure. The
704 * descriptor caches reference memory we're about to free,
705 * but falcon_reconfigure_mac_wrapper() won't reconnect
706 * the MACs because of the pending reset.
708 netif_err(efx
, drv
, efx
->net_dev
,
709 "Resetting to recover from flush failure\n");
710 efx_schedule_reset(efx
, RESET_TYPE_ALL
);
712 netif_err(efx
, drv
, efx
->net_dev
, "failed to flush queues\n");
714 netif_dbg(efx
, drv
, efx
->net_dev
,
715 "successfully flushed all queues\n");
718 efx_for_each_channel(channel
, efx
) {
719 efx_for_each_channel_rx_queue(rx_queue
, channel
)
720 efx_fini_rx_queue(rx_queue
);
721 efx_for_each_possible_channel_tx_queue(tx_queue
, channel
)
722 efx_fini_tx_queue(tx_queue
);
726 static void efx_remove_channel(struct efx_channel
*channel
)
728 struct efx_tx_queue
*tx_queue
;
729 struct efx_rx_queue
*rx_queue
;
731 netif_dbg(channel
->efx
, drv
, channel
->efx
->net_dev
,
732 "destroy chan %d\n", channel
->channel
);
734 efx_for_each_channel_rx_queue(rx_queue
, channel
)
735 efx_remove_rx_queue(rx_queue
);
736 efx_for_each_possible_channel_tx_queue(tx_queue
, channel
)
737 efx_remove_tx_queue(tx_queue
);
738 efx_remove_eventq(channel
);
739 channel
->type
->post_remove(channel
);
742 static void efx_remove_channels(struct efx_nic
*efx
)
744 struct efx_channel
*channel
;
746 efx_for_each_channel(channel
, efx
)
747 efx_remove_channel(channel
);
751 efx_realloc_channels(struct efx_nic
*efx
, u32 rxq_entries
, u32 txq_entries
)
753 struct efx_channel
*other_channel
[EFX_MAX_CHANNELS
], *channel
;
754 u32 old_rxq_entries
, old_txq_entries
;
755 unsigned i
, next_buffer_table
= 0;
758 rc
= efx_check_disabled(efx
);
762 /* Not all channels should be reallocated. We must avoid
763 * reallocating their buffer table entries.
765 efx_for_each_channel(channel
, efx
) {
766 struct efx_rx_queue
*rx_queue
;
767 struct efx_tx_queue
*tx_queue
;
769 if (channel
->type
->copy
)
771 next_buffer_table
= max(next_buffer_table
,
772 channel
->eventq
.index
+
773 channel
->eventq
.entries
);
774 efx_for_each_channel_rx_queue(rx_queue
, channel
)
775 next_buffer_table
= max(next_buffer_table
,
776 rx_queue
->rxd
.index
+
777 rx_queue
->rxd
.entries
);
778 efx_for_each_channel_tx_queue(tx_queue
, channel
)
779 next_buffer_table
= max(next_buffer_table
,
780 tx_queue
->txd
.index
+
781 tx_queue
->txd
.entries
);
784 efx_device_detach_sync(efx
);
786 efx_soft_disable_interrupts(efx
);
788 /* Clone channels (where possible) */
789 memset(other_channel
, 0, sizeof(other_channel
));
790 for (i
= 0; i
< efx
->n_channels
; i
++) {
791 channel
= efx
->channel
[i
];
792 if (channel
->type
->copy
)
793 channel
= channel
->type
->copy(channel
);
798 other_channel
[i
] = channel
;
801 /* Swap entry counts and channel pointers */
802 old_rxq_entries
= efx
->rxq_entries
;
803 old_txq_entries
= efx
->txq_entries
;
804 efx
->rxq_entries
= rxq_entries
;
805 efx
->txq_entries
= txq_entries
;
806 for (i
= 0; i
< efx
->n_channels
; i
++) {
807 channel
= efx
->channel
[i
];
808 efx
->channel
[i
] = other_channel
[i
];
809 other_channel
[i
] = channel
;
812 /* Restart buffer table allocation */
813 efx
->next_buffer_table
= next_buffer_table
;
815 for (i
= 0; i
< efx
->n_channels
; i
++) {
816 channel
= efx
->channel
[i
];
817 if (!channel
->type
->copy
)
819 rc
= efx_probe_channel(channel
);
822 efx_init_napi_channel(efx
->channel
[i
]);
826 /* Destroy unused channel structures */
827 for (i
= 0; i
< efx
->n_channels
; i
++) {
828 channel
= other_channel
[i
];
829 if (channel
&& channel
->type
->copy
) {
830 efx_fini_napi_channel(channel
);
831 efx_remove_channel(channel
);
836 rc2
= efx_soft_enable_interrupts(efx
);
839 netif_err(efx
, drv
, efx
->net_dev
,
840 "unable to restart interrupts on channel reallocation\n");
841 efx_schedule_reset(efx
, RESET_TYPE_DISABLE
);
844 netif_device_attach(efx
->net_dev
);
850 efx
->rxq_entries
= old_rxq_entries
;
851 efx
->txq_entries
= old_txq_entries
;
852 for (i
= 0; i
< efx
->n_channels
; i
++) {
853 channel
= efx
->channel
[i
];
854 efx
->channel
[i
] = other_channel
[i
];
855 other_channel
[i
] = channel
;
860 void efx_schedule_slow_fill(struct efx_rx_queue
*rx_queue
)
862 mod_timer(&rx_queue
->slow_fill
, jiffies
+ msecs_to_jiffies(100));
865 static const struct efx_channel_type efx_default_channel_type
= {
866 .pre_probe
= efx_channel_dummy_op_int
,
867 .post_remove
= efx_channel_dummy_op_void
,
868 .get_name
= efx_get_channel_name
,
869 .copy
= efx_copy_channel
,
870 .keep_eventq
= false,
873 int efx_channel_dummy_op_int(struct efx_channel
*channel
)
878 void efx_channel_dummy_op_void(struct efx_channel
*channel
)
882 /**************************************************************************
886 **************************************************************************/
888 /* This ensures that the kernel is kept informed (via
889 * netif_carrier_on/off) of the link status, and also maintains the
890 * link status's stop on the port's TX queue.
892 void efx_link_status_changed(struct efx_nic
*efx
)
894 struct efx_link_state
*link_state
= &efx
->link_state
;
896 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
897 * that no events are triggered between unregister_netdev() and the
898 * driver unloading. A more general condition is that NETDEV_CHANGE
899 * can only be generated between NETDEV_UP and NETDEV_DOWN */
900 if (!netif_running(efx
->net_dev
))
903 if (link_state
->up
!= netif_carrier_ok(efx
->net_dev
)) {
904 efx
->n_link_state_changes
++;
907 netif_carrier_on(efx
->net_dev
);
909 netif_carrier_off(efx
->net_dev
);
912 /* Status message for kernel log */
914 netif_info(efx
, link
, efx
->net_dev
,
915 "link up at %uMbps %s-duplex (MTU %d)\n",
916 link_state
->speed
, link_state
->fd
? "full" : "half",
919 netif_info(efx
, link
, efx
->net_dev
, "link down\n");
922 void efx_link_set_advertising(struct efx_nic
*efx
, u32 advertising
)
924 efx
->link_advertising
= advertising
;
926 if (advertising
& ADVERTISED_Pause
)
927 efx
->wanted_fc
|= (EFX_FC_TX
| EFX_FC_RX
);
929 efx
->wanted_fc
&= ~(EFX_FC_TX
| EFX_FC_RX
);
930 if (advertising
& ADVERTISED_Asym_Pause
)
931 efx
->wanted_fc
^= EFX_FC_TX
;
935 void efx_link_set_wanted_fc(struct efx_nic
*efx
, u8 wanted_fc
)
937 efx
->wanted_fc
= wanted_fc
;
938 if (efx
->link_advertising
) {
939 if (wanted_fc
& EFX_FC_RX
)
940 efx
->link_advertising
|= (ADVERTISED_Pause
|
941 ADVERTISED_Asym_Pause
);
943 efx
->link_advertising
&= ~(ADVERTISED_Pause
|
944 ADVERTISED_Asym_Pause
);
945 if (wanted_fc
& EFX_FC_TX
)
946 efx
->link_advertising
^= ADVERTISED_Asym_Pause
;
950 static void efx_fini_port(struct efx_nic
*efx
);
952 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
953 * the MAC appropriately. All other PHY configuration changes are pushed
954 * through phy_op->set_settings(), and pushed asynchronously to the MAC
955 * through efx_monitor().
957 * Callers must hold the mac_lock
959 int __efx_reconfigure_port(struct efx_nic
*efx
)
961 enum efx_phy_mode phy_mode
;
964 WARN_ON(!mutex_is_locked(&efx
->mac_lock
));
966 /* Disable PHY transmit in mac level loopbacks */
967 phy_mode
= efx
->phy_mode
;
968 if (LOOPBACK_INTERNAL(efx
))
969 efx
->phy_mode
|= PHY_MODE_TX_DISABLED
;
971 efx
->phy_mode
&= ~PHY_MODE_TX_DISABLED
;
973 rc
= efx
->type
->reconfigure_port(efx
);
976 efx
->phy_mode
= phy_mode
;
981 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
983 int efx_reconfigure_port(struct efx_nic
*efx
)
987 EFX_ASSERT_RESET_SERIALISED(efx
);
989 mutex_lock(&efx
->mac_lock
);
990 rc
= __efx_reconfigure_port(efx
);
991 mutex_unlock(&efx
->mac_lock
);
996 /* Asynchronous work item for changing MAC promiscuity and multicast
997 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
999 static void efx_mac_work(struct work_struct
*data
)
1001 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, mac_work
);
1003 mutex_lock(&efx
->mac_lock
);
1004 if (efx
->port_enabled
)
1005 efx
->type
->reconfigure_mac(efx
);
1006 mutex_unlock(&efx
->mac_lock
);
1009 static int efx_probe_port(struct efx_nic
*efx
)
1013 netif_dbg(efx
, probe
, efx
->net_dev
, "create port\n");
1016 efx
->phy_mode
= PHY_MODE_SPECIAL
;
1018 /* Connect up MAC/PHY operations table */
1019 rc
= efx
->type
->probe_port(efx
);
1023 /* Initialise MAC address to permanent address */
1024 ether_addr_copy(efx
->net_dev
->dev_addr
, efx
->net_dev
->perm_addr
);
1029 static int efx_init_port(struct efx_nic
*efx
)
1033 netif_dbg(efx
, drv
, efx
->net_dev
, "init port\n");
1035 mutex_lock(&efx
->mac_lock
);
1037 rc
= efx
->phy_op
->init(efx
);
1041 efx
->port_initialized
= true;
1043 /* Reconfigure the MAC before creating dma queues (required for
1044 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1045 efx
->type
->reconfigure_mac(efx
);
1047 /* Ensure the PHY advertises the correct flow control settings */
1048 rc
= efx
->phy_op
->reconfigure(efx
);
1049 if (rc
&& rc
!= -EPERM
)
1052 mutex_unlock(&efx
->mac_lock
);
1056 efx
->phy_op
->fini(efx
);
1058 mutex_unlock(&efx
->mac_lock
);
1062 static void efx_start_port(struct efx_nic
*efx
)
1064 netif_dbg(efx
, ifup
, efx
->net_dev
, "start port\n");
1065 BUG_ON(efx
->port_enabled
);
1067 mutex_lock(&efx
->mac_lock
);
1068 efx
->port_enabled
= true;
1070 /* Ensure MAC ingress/egress is enabled */
1071 efx
->type
->reconfigure_mac(efx
);
1073 mutex_unlock(&efx
->mac_lock
);
1076 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1077 * and the async self-test, wait for them to finish and prevent them
1078 * being scheduled again. This doesn't cover online resets, which
1079 * should only be cancelled when removing the device.
1081 static void efx_stop_port(struct efx_nic
*efx
)
1083 netif_dbg(efx
, ifdown
, efx
->net_dev
, "stop port\n");
1085 EFX_ASSERT_RESET_SERIALISED(efx
);
1087 mutex_lock(&efx
->mac_lock
);
1088 efx
->port_enabled
= false;
1089 mutex_unlock(&efx
->mac_lock
);
1091 /* Serialise against efx_set_multicast_list() */
1092 netif_addr_lock_bh(efx
->net_dev
);
1093 netif_addr_unlock_bh(efx
->net_dev
);
1095 cancel_delayed_work_sync(&efx
->monitor_work
);
1096 efx_selftest_async_cancel(efx
);
1097 cancel_work_sync(&efx
->mac_work
);
1100 static void efx_fini_port(struct efx_nic
*efx
)
1102 netif_dbg(efx
, drv
, efx
->net_dev
, "shut down port\n");
1104 if (!efx
->port_initialized
)
1107 efx
->phy_op
->fini(efx
);
1108 efx
->port_initialized
= false;
1110 efx
->link_state
.up
= false;
1111 efx_link_status_changed(efx
);
1114 static void efx_remove_port(struct efx_nic
*efx
)
1116 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying port\n");
1118 efx
->type
->remove_port(efx
);
1121 /**************************************************************************
1125 **************************************************************************/
1127 static LIST_HEAD(efx_primary_list
);
1128 static LIST_HEAD(efx_unassociated_list
);
1130 static bool efx_same_controller(struct efx_nic
*left
, struct efx_nic
*right
)
1132 return left
->type
== right
->type
&&
1133 left
->vpd_sn
&& right
->vpd_sn
&&
1134 !strcmp(left
->vpd_sn
, right
->vpd_sn
);
1137 static void efx_associate(struct efx_nic
*efx
)
1139 struct efx_nic
*other
, *next
;
1141 if (efx
->primary
== efx
) {
1142 /* Adding primary function; look for secondaries */
1144 netif_dbg(efx
, probe
, efx
->net_dev
, "adding to primary list\n");
1145 list_add_tail(&efx
->node
, &efx_primary_list
);
1147 list_for_each_entry_safe(other
, next
, &efx_unassociated_list
,
1149 if (efx_same_controller(efx
, other
)) {
1150 list_del(&other
->node
);
1151 netif_dbg(other
, probe
, other
->net_dev
,
1152 "moving to secondary list of %s %s\n",
1153 pci_name(efx
->pci_dev
),
1154 efx
->net_dev
->name
);
1155 list_add_tail(&other
->node
,
1156 &efx
->secondary_list
);
1157 other
->primary
= efx
;
1161 /* Adding secondary function; look for primary */
1163 list_for_each_entry(other
, &efx_primary_list
, node
) {
1164 if (efx_same_controller(efx
, other
)) {
1165 netif_dbg(efx
, probe
, efx
->net_dev
,
1166 "adding to secondary list of %s %s\n",
1167 pci_name(other
->pci_dev
),
1168 other
->net_dev
->name
);
1169 list_add_tail(&efx
->node
,
1170 &other
->secondary_list
);
1171 efx
->primary
= other
;
1176 netif_dbg(efx
, probe
, efx
->net_dev
,
1177 "adding to unassociated list\n");
1178 list_add_tail(&efx
->node
, &efx_unassociated_list
);
1182 static void efx_dissociate(struct efx_nic
*efx
)
1184 struct efx_nic
*other
, *next
;
1186 list_del(&efx
->node
);
1187 efx
->primary
= NULL
;
1189 list_for_each_entry_safe(other
, next
, &efx
->secondary_list
, node
) {
1190 list_del(&other
->node
);
1191 netif_dbg(other
, probe
, other
->net_dev
,
1192 "moving to unassociated list\n");
1193 list_add_tail(&other
->node
, &efx_unassociated_list
);
1194 other
->primary
= NULL
;
1198 /* This configures the PCI device to enable I/O and DMA. */
1199 static int efx_init_io(struct efx_nic
*efx
)
1201 struct pci_dev
*pci_dev
= efx
->pci_dev
;
1202 dma_addr_t dma_mask
= efx
->type
->max_dma_mask
;
1203 unsigned int mem_map_size
= efx
->type
->mem_map_size(efx
);
1206 netif_dbg(efx
, probe
, efx
->net_dev
, "initialising I/O\n");
1208 bar
= efx
->type
->mem_bar
;
1210 rc
= pci_enable_device(pci_dev
);
1212 netif_err(efx
, probe
, efx
->net_dev
,
1213 "failed to enable PCI device\n");
1217 pci_set_master(pci_dev
);
1219 /* Set the PCI DMA mask. Try all possibilities from our
1220 * genuine mask down to 32 bits, because some architectures
1221 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1222 * masks event though they reject 46 bit masks.
1224 while (dma_mask
> 0x7fffffffUL
) {
1225 if (dma_supported(&pci_dev
->dev
, dma_mask
)) {
1226 rc
= dma_set_mask_and_coherent(&pci_dev
->dev
, dma_mask
);
1233 netif_err(efx
, probe
, efx
->net_dev
,
1234 "could not find a suitable DMA mask\n");
1237 netif_dbg(efx
, probe
, efx
->net_dev
,
1238 "using DMA mask %llx\n", (unsigned long long) dma_mask
);
1240 efx
->membase_phys
= pci_resource_start(efx
->pci_dev
, bar
);
1241 rc
= pci_request_region(pci_dev
, bar
, "sfc");
1243 netif_err(efx
, probe
, efx
->net_dev
,
1244 "request for memory BAR failed\n");
1248 efx
->membase
= ioremap_nocache(efx
->membase_phys
, mem_map_size
);
1249 if (!efx
->membase
) {
1250 netif_err(efx
, probe
, efx
->net_dev
,
1251 "could not map memory BAR at %llx+%x\n",
1252 (unsigned long long)efx
->membase_phys
, mem_map_size
);
1256 netif_dbg(efx
, probe
, efx
->net_dev
,
1257 "memory BAR at %llx+%x (virtual %p)\n",
1258 (unsigned long long)efx
->membase_phys
, mem_map_size
,
1264 pci_release_region(efx
->pci_dev
, bar
);
1266 efx
->membase_phys
= 0;
1268 pci_disable_device(efx
->pci_dev
);
1273 static void efx_fini_io(struct efx_nic
*efx
)
1277 netif_dbg(efx
, drv
, efx
->net_dev
, "shutting down I/O\n");
1280 iounmap(efx
->membase
);
1281 efx
->membase
= NULL
;
1284 if (efx
->membase_phys
) {
1285 bar
= efx
->type
->mem_bar
;
1286 pci_release_region(efx
->pci_dev
, bar
);
1287 efx
->membase_phys
= 0;
1290 pci_disable_device(efx
->pci_dev
);
1293 void efx_set_default_rx_indir_table(struct efx_nic
*efx
)
1297 for (i
= 0; i
< ARRAY_SIZE(efx
->rx_indir_table
); i
++)
1298 efx
->rx_indir_table
[i
] =
1299 ethtool_rxfh_indir_default(i
, efx
->rss_spread
);
1302 static unsigned int efx_wanted_parallelism(struct efx_nic
*efx
)
1304 cpumask_var_t thread_mask
;
1311 if (unlikely(!zalloc_cpumask_var(&thread_mask
, GFP_KERNEL
))) {
1312 netif_warn(efx
, probe
, efx
->net_dev
,
1313 "RSS disabled due to allocation failure\n");
1318 for_each_online_cpu(cpu
) {
1319 if (!cpumask_test_cpu(cpu
, thread_mask
)) {
1321 cpumask_or(thread_mask
, thread_mask
,
1322 topology_thread_cpumask(cpu
));
1326 free_cpumask_var(thread_mask
);
1329 /* If RSS is requested for the PF *and* VFs then we can't write RSS
1330 * table entries that are inaccessible to VFs
1332 #ifdef CONFIG_SFC_SRIOV
1333 if (efx
->type
->sriov_wanted
) {
1334 if (efx
->type
->sriov_wanted(efx
) && efx_vf_size(efx
) > 1 &&
1335 count
> efx_vf_size(efx
)) {
1336 netif_warn(efx
, probe
, efx
->net_dev
,
1337 "Reducing number of RSS channels from %u to %u for "
1338 "VF support. Increase vf-msix-limit to use more "
1339 "channels on the PF.\n",
1340 count
, efx_vf_size(efx
));
1341 count
= efx_vf_size(efx
);
1349 /* Probe the number and type of interrupts we are able to obtain, and
1350 * the resulting numbers of channels and RX queues.
1352 static int efx_probe_interrupts(struct efx_nic
*efx
)
1354 unsigned int extra_channels
= 0;
1358 for (i
= 0; i
< EFX_MAX_EXTRA_CHANNELS
; i
++)
1359 if (efx
->extra_channel_type
[i
])
1362 if (efx
->interrupt_mode
== EFX_INT_MODE_MSIX
) {
1363 struct msix_entry xentries
[EFX_MAX_CHANNELS
];
1364 unsigned int n_channels
;
1366 n_channels
= efx_wanted_parallelism(efx
);
1367 if (separate_tx_channels
)
1369 n_channels
+= extra_channels
;
1370 n_channels
= min(n_channels
, efx
->max_channels
);
1372 for (i
= 0; i
< n_channels
; i
++)
1373 xentries
[i
].entry
= i
;
1374 rc
= pci_enable_msix_range(efx
->pci_dev
,
1375 xentries
, 1, n_channels
);
1377 /* Fall back to single channel MSI */
1378 efx
->interrupt_mode
= EFX_INT_MODE_MSI
;
1379 netif_err(efx
, drv
, efx
->net_dev
,
1380 "could not enable MSI-X\n");
1381 } else if (rc
< n_channels
) {
1382 netif_err(efx
, drv
, efx
->net_dev
,
1383 "WARNING: Insufficient MSI-X vectors"
1384 " available (%d < %u).\n", rc
, n_channels
);
1385 netif_err(efx
, drv
, efx
->net_dev
,
1386 "WARNING: Performance may be reduced.\n");
1391 efx
->n_channels
= n_channels
;
1392 if (n_channels
> extra_channels
)
1393 n_channels
-= extra_channels
;
1394 if (separate_tx_channels
) {
1395 efx
->n_tx_channels
= max(n_channels
/ 2, 1U);
1396 efx
->n_rx_channels
= max(n_channels
-
1400 efx
->n_tx_channels
= n_channels
;
1401 efx
->n_rx_channels
= n_channels
;
1403 for (i
= 0; i
< efx
->n_channels
; i
++)
1404 efx_get_channel(efx
, i
)->irq
=
1409 /* Try single interrupt MSI */
1410 if (efx
->interrupt_mode
== EFX_INT_MODE_MSI
) {
1411 efx
->n_channels
= 1;
1412 efx
->n_rx_channels
= 1;
1413 efx
->n_tx_channels
= 1;
1414 rc
= pci_enable_msi(efx
->pci_dev
);
1416 efx_get_channel(efx
, 0)->irq
= efx
->pci_dev
->irq
;
1418 netif_err(efx
, drv
, efx
->net_dev
,
1419 "could not enable MSI\n");
1420 efx
->interrupt_mode
= EFX_INT_MODE_LEGACY
;
1424 /* Assume legacy interrupts */
1425 if (efx
->interrupt_mode
== EFX_INT_MODE_LEGACY
) {
1426 efx
->n_channels
= 1 + (separate_tx_channels
? 1 : 0);
1427 efx
->n_rx_channels
= 1;
1428 efx
->n_tx_channels
= 1;
1429 efx
->legacy_irq
= efx
->pci_dev
->irq
;
1432 /* Assign extra channels if possible */
1433 j
= efx
->n_channels
;
1434 for (i
= 0; i
< EFX_MAX_EXTRA_CHANNELS
; i
++) {
1435 if (!efx
->extra_channel_type
[i
])
1437 if (efx
->interrupt_mode
!= EFX_INT_MODE_MSIX
||
1438 efx
->n_channels
<= extra_channels
) {
1439 efx
->extra_channel_type
[i
]->handle_no_channel(efx
);
1442 efx_get_channel(efx
, j
)->type
=
1443 efx
->extra_channel_type
[i
];
1447 /* RSS might be usable on VFs even if it is disabled on the PF */
1448 #ifdef CONFIG_SFC_SRIOV
1449 if (efx
->type
->sriov_wanted
) {
1450 efx
->rss_spread
= ((efx
->n_rx_channels
> 1 ||
1451 !efx
->type
->sriov_wanted(efx
)) ?
1452 efx
->n_rx_channels
: efx_vf_size(efx
));
1456 efx
->rss_spread
= efx
->n_rx_channels
;
1460 static int efx_soft_enable_interrupts(struct efx_nic
*efx
)
1462 struct efx_channel
*channel
, *end_channel
;
1465 BUG_ON(efx
->state
== STATE_DISABLED
);
1467 efx
->irq_soft_enabled
= true;
1470 efx_for_each_channel(channel
, efx
) {
1471 if (!channel
->type
->keep_eventq
) {
1472 rc
= efx_init_eventq(channel
);
1476 efx_start_eventq(channel
);
1479 efx_mcdi_mode_event(efx
);
1483 end_channel
= channel
;
1484 efx_for_each_channel(channel
, efx
) {
1485 if (channel
== end_channel
)
1487 efx_stop_eventq(channel
);
1488 if (!channel
->type
->keep_eventq
)
1489 efx_fini_eventq(channel
);
1495 static void efx_soft_disable_interrupts(struct efx_nic
*efx
)
1497 struct efx_channel
*channel
;
1499 if (efx
->state
== STATE_DISABLED
)
1502 efx_mcdi_mode_poll(efx
);
1504 efx
->irq_soft_enabled
= false;
1507 if (efx
->legacy_irq
)
1508 synchronize_irq(efx
->legacy_irq
);
1510 efx_for_each_channel(channel
, efx
) {
1512 synchronize_irq(channel
->irq
);
1514 efx_stop_eventq(channel
);
1515 if (!channel
->type
->keep_eventq
)
1516 efx_fini_eventq(channel
);
1519 /* Flush the asynchronous MCDI request queue */
1520 efx_mcdi_flush_async(efx
);
1523 static int efx_enable_interrupts(struct efx_nic
*efx
)
1525 struct efx_channel
*channel
, *end_channel
;
1528 BUG_ON(efx
->state
== STATE_DISABLED
);
1530 if (efx
->eeh_disabled_legacy_irq
) {
1531 enable_irq(efx
->legacy_irq
);
1532 efx
->eeh_disabled_legacy_irq
= false;
1535 efx
->type
->irq_enable_master(efx
);
1537 efx_for_each_channel(channel
, efx
) {
1538 if (channel
->type
->keep_eventq
) {
1539 rc
= efx_init_eventq(channel
);
1545 rc
= efx_soft_enable_interrupts(efx
);
1552 end_channel
= channel
;
1553 efx_for_each_channel(channel
, efx
) {
1554 if (channel
== end_channel
)
1556 if (channel
->type
->keep_eventq
)
1557 efx_fini_eventq(channel
);
1560 efx
->type
->irq_disable_non_ev(efx
);
1565 static void efx_disable_interrupts(struct efx_nic
*efx
)
1567 struct efx_channel
*channel
;
1569 efx_soft_disable_interrupts(efx
);
1571 efx_for_each_channel(channel
, efx
) {
1572 if (channel
->type
->keep_eventq
)
1573 efx_fini_eventq(channel
);
1576 efx
->type
->irq_disable_non_ev(efx
);
1579 static void efx_remove_interrupts(struct efx_nic
*efx
)
1581 struct efx_channel
*channel
;
1583 /* Remove MSI/MSI-X interrupts */
1584 efx_for_each_channel(channel
, efx
)
1586 pci_disable_msi(efx
->pci_dev
);
1587 pci_disable_msix(efx
->pci_dev
);
1589 /* Remove legacy interrupt */
1590 efx
->legacy_irq
= 0;
1593 static void efx_set_channels(struct efx_nic
*efx
)
1595 struct efx_channel
*channel
;
1596 struct efx_tx_queue
*tx_queue
;
1598 efx
->tx_channel_offset
=
1599 separate_tx_channels
? efx
->n_channels
- efx
->n_tx_channels
: 0;
1601 /* We need to mark which channels really have RX and TX
1602 * queues, and adjust the TX queue numbers if we have separate
1603 * RX-only and TX-only channels.
1605 efx_for_each_channel(channel
, efx
) {
1606 if (channel
->channel
< efx
->n_rx_channels
)
1607 channel
->rx_queue
.core_index
= channel
->channel
;
1609 channel
->rx_queue
.core_index
= -1;
1611 efx_for_each_channel_tx_queue(tx_queue
, channel
)
1612 tx_queue
->queue
-= (efx
->tx_channel_offset
*
1617 static int efx_probe_nic(struct efx_nic
*efx
)
1621 netif_dbg(efx
, probe
, efx
->net_dev
, "creating NIC\n");
1623 /* Carry out hardware-type specific initialisation */
1624 rc
= efx
->type
->probe(efx
);
1628 /* Determine the number of channels and queues by trying to hook
1629 * in MSI-X interrupts. */
1630 rc
= efx_probe_interrupts(efx
);
1634 efx_set_channels(efx
);
1636 rc
= efx
->type
->dimension_resources(efx
);
1640 if (efx
->n_channels
> 1)
1641 netdev_rss_key_fill(&efx
->rx_hash_key
,
1642 sizeof(efx
->rx_hash_key
));
1643 efx_set_default_rx_indir_table(efx
);
1645 netif_set_real_num_tx_queues(efx
->net_dev
, efx
->n_tx_channels
);
1646 netif_set_real_num_rx_queues(efx
->net_dev
, efx
->n_rx_channels
);
1648 /* Initialise the interrupt moderation settings */
1649 efx_init_irq_moderation(efx
, tx_irq_mod_usec
, rx_irq_mod_usec
, true,
1655 efx_remove_interrupts(efx
);
1657 efx
->type
->remove(efx
);
1661 static void efx_remove_nic(struct efx_nic
*efx
)
1663 netif_dbg(efx
, drv
, efx
->net_dev
, "destroying NIC\n");
1665 efx_remove_interrupts(efx
);
1666 efx
->type
->remove(efx
);
1669 static int efx_probe_filters(struct efx_nic
*efx
)
1673 spin_lock_init(&efx
->filter_lock
);
1675 rc
= efx
->type
->filter_table_probe(efx
);
1679 #ifdef CONFIG_RFS_ACCEL
1680 if (efx
->type
->offload_features
& NETIF_F_NTUPLE
) {
1681 efx
->rps_flow_id
= kcalloc(efx
->type
->max_rx_ip_filters
,
1682 sizeof(*efx
->rps_flow_id
),
1684 if (!efx
->rps_flow_id
) {
1685 efx
->type
->filter_table_remove(efx
);
1694 static void efx_remove_filters(struct efx_nic
*efx
)
1696 #ifdef CONFIG_RFS_ACCEL
1697 kfree(efx
->rps_flow_id
);
1699 efx
->type
->filter_table_remove(efx
);
1702 static void efx_restore_filters(struct efx_nic
*efx
)
1704 efx
->type
->filter_table_restore(efx
);
1707 /**************************************************************************
1709 * NIC startup/shutdown
1711 *************************************************************************/
1713 static int efx_probe_all(struct efx_nic
*efx
)
1717 rc
= efx_probe_nic(efx
);
1719 netif_err(efx
, probe
, efx
->net_dev
, "failed to create NIC\n");
1723 rc
= efx_probe_port(efx
);
1725 netif_err(efx
, probe
, efx
->net_dev
, "failed to create port\n");
1729 BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE
< EFX_RXQ_MIN_ENT
);
1730 if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE
< EFX_TXQ_MIN_ENT(efx
))) {
1734 efx
->rxq_entries
= efx
->txq_entries
= EFX_DEFAULT_DMAQ_SIZE
;
1736 #ifdef CONFIG_SFC_SRIOV
1737 rc
= efx
->type
->vswitching_probe(efx
);
1738 if (rc
) /* not fatal; the PF will still work fine */
1739 netif_warn(efx
, probe
, efx
->net_dev
,
1740 "failed to setup vswitching rc=%d;"
1741 " VFs may not function\n", rc
);
1744 rc
= efx_probe_filters(efx
);
1746 netif_err(efx
, probe
, efx
->net_dev
,
1747 "failed to create filter tables\n");
1751 rc
= efx_probe_channels(efx
);
1758 efx_remove_filters(efx
);
1760 #ifdef CONFIG_SFC_SRIOV
1761 efx
->type
->vswitching_remove(efx
);
1764 efx_remove_port(efx
);
1766 efx_remove_nic(efx
);
1771 /* If the interface is supposed to be running but is not, start
1772 * the hardware and software data path, regular activity for the port
1773 * (MAC statistics, link polling, etc.) and schedule the port to be
1774 * reconfigured. Interrupts must already be enabled. This function
1775 * is safe to call multiple times, so long as the NIC is not disabled.
1776 * Requires the RTNL lock.
1778 static void efx_start_all(struct efx_nic
*efx
)
1780 EFX_ASSERT_RESET_SERIALISED(efx
);
1781 BUG_ON(efx
->state
== STATE_DISABLED
);
1783 /* Check that it is appropriate to restart the interface. All
1784 * of these flags are safe to read under just the rtnl lock */
1785 if (efx
->port_enabled
|| !netif_running(efx
->net_dev
) ||
1789 efx_start_port(efx
);
1790 efx_start_datapath(efx
);
1792 /* Start the hardware monitor if there is one */
1793 if (efx
->type
->monitor
!= NULL
)
1794 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1795 efx_monitor_interval
);
1797 /* If link state detection is normally event-driven, we have
1798 * to poll now because we could have missed a change
1800 if (efx_nic_rev(efx
) >= EFX_REV_SIENA_A0
) {
1801 mutex_lock(&efx
->mac_lock
);
1802 if (efx
->phy_op
->poll(efx
))
1803 efx_link_status_changed(efx
);
1804 mutex_unlock(&efx
->mac_lock
);
1807 efx
->type
->start_stats(efx
);
1808 efx
->type
->pull_stats(efx
);
1809 spin_lock_bh(&efx
->stats_lock
);
1810 efx
->type
->update_stats(efx
, NULL
, NULL
);
1811 spin_unlock_bh(&efx
->stats_lock
);
1814 /* Quiesce the hardware and software data path, and regular activity
1815 * for the port without bringing the link down. Safe to call multiple
1816 * times with the NIC in almost any state, but interrupts should be
1817 * enabled. Requires the RTNL lock.
1819 static void efx_stop_all(struct efx_nic
*efx
)
1821 EFX_ASSERT_RESET_SERIALISED(efx
);
1823 /* port_enabled can be read safely under the rtnl lock */
1824 if (!efx
->port_enabled
)
1827 /* update stats before we go down so we can accurately count
1830 efx
->type
->pull_stats(efx
);
1831 spin_lock_bh(&efx
->stats_lock
);
1832 efx
->type
->update_stats(efx
, NULL
, NULL
);
1833 spin_unlock_bh(&efx
->stats_lock
);
1834 efx
->type
->stop_stats(efx
);
1837 /* Stop the kernel transmit interface. This is only valid if
1838 * the device is stopped or detached; otherwise the watchdog
1839 * may fire immediately.
1841 WARN_ON(netif_running(efx
->net_dev
) &&
1842 netif_device_present(efx
->net_dev
));
1843 netif_tx_disable(efx
->net_dev
);
1845 efx_stop_datapath(efx
);
1848 static void efx_remove_all(struct efx_nic
*efx
)
1850 efx_remove_channels(efx
);
1851 efx_remove_filters(efx
);
1852 #ifdef CONFIG_SFC_SRIOV
1853 efx
->type
->vswitching_remove(efx
);
1855 efx_remove_port(efx
);
1856 efx_remove_nic(efx
);
1859 /**************************************************************************
1861 * Interrupt moderation
1863 **************************************************************************/
1865 static unsigned int irq_mod_ticks(unsigned int usecs
, unsigned int quantum_ns
)
1869 if (usecs
* 1000 < quantum_ns
)
1870 return 1; /* never round down to 0 */
1871 return usecs
* 1000 / quantum_ns
;
1874 /* Set interrupt moderation parameters */
1875 int efx_init_irq_moderation(struct efx_nic
*efx
, unsigned int tx_usecs
,
1876 unsigned int rx_usecs
, bool rx_adaptive
,
1877 bool rx_may_override_tx
)
1879 struct efx_channel
*channel
;
1880 unsigned int irq_mod_max
= DIV_ROUND_UP(efx
->type
->timer_period_max
*
1881 efx
->timer_quantum_ns
,
1883 unsigned int tx_ticks
;
1884 unsigned int rx_ticks
;
1886 EFX_ASSERT_RESET_SERIALISED(efx
);
1888 if (tx_usecs
> irq_mod_max
|| rx_usecs
> irq_mod_max
)
1891 tx_ticks
= irq_mod_ticks(tx_usecs
, efx
->timer_quantum_ns
);
1892 rx_ticks
= irq_mod_ticks(rx_usecs
, efx
->timer_quantum_ns
);
1894 if (tx_ticks
!= rx_ticks
&& efx
->tx_channel_offset
== 0 &&
1895 !rx_may_override_tx
) {
1896 netif_err(efx
, drv
, efx
->net_dev
, "Channels are shared. "
1897 "RX and TX IRQ moderation must be equal\n");
1901 efx
->irq_rx_adaptive
= rx_adaptive
;
1902 efx
->irq_rx_moderation
= rx_ticks
;
1903 efx_for_each_channel(channel
, efx
) {
1904 if (efx_channel_has_rx_queue(channel
))
1905 channel
->irq_moderation
= rx_ticks
;
1906 else if (efx_channel_has_tx_queues(channel
))
1907 channel
->irq_moderation
= tx_ticks
;
1913 void efx_get_irq_moderation(struct efx_nic
*efx
, unsigned int *tx_usecs
,
1914 unsigned int *rx_usecs
, bool *rx_adaptive
)
1916 /* We must round up when converting ticks to microseconds
1917 * because we round down when converting the other way.
1920 *rx_adaptive
= efx
->irq_rx_adaptive
;
1921 *rx_usecs
= DIV_ROUND_UP(efx
->irq_rx_moderation
*
1922 efx
->timer_quantum_ns
,
1925 /* If channels are shared between RX and TX, so is IRQ
1926 * moderation. Otherwise, IRQ moderation is the same for all
1927 * TX channels and is not adaptive.
1929 if (efx
->tx_channel_offset
== 0)
1930 *tx_usecs
= *rx_usecs
;
1932 *tx_usecs
= DIV_ROUND_UP(
1933 efx
->channel
[efx
->tx_channel_offset
]->irq_moderation
*
1934 efx
->timer_quantum_ns
,
1938 /**************************************************************************
1942 **************************************************************************/
1944 /* Run periodically off the general workqueue */
1945 static void efx_monitor(struct work_struct
*data
)
1947 struct efx_nic
*efx
= container_of(data
, struct efx_nic
,
1950 netif_vdbg(efx
, timer
, efx
->net_dev
,
1951 "hardware monitor executing on CPU %d\n",
1952 raw_smp_processor_id());
1953 BUG_ON(efx
->type
->monitor
== NULL
);
1955 /* If the mac_lock is already held then it is likely a port
1956 * reconfiguration is already in place, which will likely do
1957 * most of the work of monitor() anyway. */
1958 if (mutex_trylock(&efx
->mac_lock
)) {
1959 if (efx
->port_enabled
)
1960 efx
->type
->monitor(efx
);
1961 mutex_unlock(&efx
->mac_lock
);
1964 queue_delayed_work(efx
->workqueue
, &efx
->monitor_work
,
1965 efx_monitor_interval
);
1968 /**************************************************************************
1972 *************************************************************************/
1975 * Context: process, rtnl_lock() held.
1977 static int efx_ioctl(struct net_device
*net_dev
, struct ifreq
*ifr
, int cmd
)
1979 struct efx_nic
*efx
= netdev_priv(net_dev
);
1980 struct mii_ioctl_data
*data
= if_mii(ifr
);
1982 if (cmd
== SIOCSHWTSTAMP
)
1983 return efx_ptp_set_ts_config(efx
, ifr
);
1984 if (cmd
== SIOCGHWTSTAMP
)
1985 return efx_ptp_get_ts_config(efx
, ifr
);
1987 /* Convert phy_id from older PRTAD/DEVAD format */
1988 if ((cmd
== SIOCGMIIREG
|| cmd
== SIOCSMIIREG
) &&
1989 (data
->phy_id
& 0xfc00) == 0x0400)
1990 data
->phy_id
^= MDIO_PHY_ID_C45
| 0x0400;
1992 return mdio_mii_ioctl(&efx
->mdio
, data
, cmd
);
1995 /**************************************************************************
1999 **************************************************************************/
2001 static void efx_init_napi_channel(struct efx_channel
*channel
)
2003 struct efx_nic
*efx
= channel
->efx
;
2005 channel
->napi_dev
= efx
->net_dev
;
2006 netif_napi_add(channel
->napi_dev
, &channel
->napi_str
,
2007 efx_poll
, napi_weight
);
2008 napi_hash_add(&channel
->napi_str
);
2009 efx_channel_init_lock(channel
);
2012 static void efx_init_napi(struct efx_nic
*efx
)
2014 struct efx_channel
*channel
;
2016 efx_for_each_channel(channel
, efx
)
2017 efx_init_napi_channel(channel
);
2020 static void efx_fini_napi_channel(struct efx_channel
*channel
)
2022 if (channel
->napi_dev
) {
2023 netif_napi_del(&channel
->napi_str
);
2024 napi_hash_del(&channel
->napi_str
);
2026 channel
->napi_dev
= NULL
;
2029 static void efx_fini_napi(struct efx_nic
*efx
)
2031 struct efx_channel
*channel
;
2033 efx_for_each_channel(channel
, efx
)
2034 efx_fini_napi_channel(channel
);
2037 /**************************************************************************
2039 * Kernel netpoll interface
2041 *************************************************************************/
2043 #ifdef CONFIG_NET_POLL_CONTROLLER
2045 /* Although in the common case interrupts will be disabled, this is not
2046 * guaranteed. However, all our work happens inside the NAPI callback,
2047 * so no locking is required.
2049 static void efx_netpoll(struct net_device
*net_dev
)
2051 struct efx_nic
*efx
= netdev_priv(net_dev
);
2052 struct efx_channel
*channel
;
2054 efx_for_each_channel(channel
, efx
)
2055 efx_schedule_channel(channel
);
2060 #ifdef CONFIG_NET_RX_BUSY_POLL
2061 static int efx_busy_poll(struct napi_struct
*napi
)
2063 struct efx_channel
*channel
=
2064 container_of(napi
, struct efx_channel
, napi_str
);
2065 struct efx_nic
*efx
= channel
->efx
;
2067 int old_rx_packets
, rx_packets
;
2069 if (!netif_running(efx
->net_dev
))
2070 return LL_FLUSH_FAILED
;
2072 if (!efx_channel_lock_poll(channel
))
2073 return LL_FLUSH_BUSY
;
2075 old_rx_packets
= channel
->rx_queue
.rx_packets
;
2076 efx_process_channel(channel
, budget
);
2078 rx_packets
= channel
->rx_queue
.rx_packets
- old_rx_packets
;
2080 /* There is no race condition with NAPI here.
2081 * NAPI will automatically be rescheduled if it yielded during busy
2082 * polling, because it was not able to take the lock and thus returned
2085 efx_channel_unlock_poll(channel
);
2091 /**************************************************************************
2093 * Kernel net device interface
2095 *************************************************************************/
2097 /* Context: process, rtnl_lock() held. */
2098 static int efx_net_open(struct net_device
*net_dev
)
2100 struct efx_nic
*efx
= netdev_priv(net_dev
);
2103 netif_dbg(efx
, ifup
, efx
->net_dev
, "opening device on CPU %d\n",
2104 raw_smp_processor_id());
2106 rc
= efx_check_disabled(efx
);
2109 if (efx
->phy_mode
& PHY_MODE_SPECIAL
)
2111 if (efx_mcdi_poll_reboot(efx
) && efx_reset(efx
, RESET_TYPE_ALL
))
2114 /* Notify the kernel of the link state polled during driver load,
2115 * before the monitor starts running */
2116 efx_link_status_changed(efx
);
2119 efx_selftest_async_start(efx
);
2123 /* Context: process, rtnl_lock() held.
2124 * Note that the kernel will ignore our return code; this method
2125 * should really be a void.
2127 static int efx_net_stop(struct net_device
*net_dev
)
2129 struct efx_nic
*efx
= netdev_priv(net_dev
);
2131 netif_dbg(efx
, ifdown
, efx
->net_dev
, "closing on CPU %d\n",
2132 raw_smp_processor_id());
2134 /* Stop the device and flush all the channels */
2140 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2141 static struct rtnl_link_stats64
*efx_net_stats(struct net_device
*net_dev
,
2142 struct rtnl_link_stats64
*stats
)
2144 struct efx_nic
*efx
= netdev_priv(net_dev
);
2146 spin_lock_bh(&efx
->stats_lock
);
2147 efx
->type
->update_stats(efx
, NULL
, stats
);
2148 spin_unlock_bh(&efx
->stats_lock
);
2153 /* Context: netif_tx_lock held, BHs disabled. */
2154 static void efx_watchdog(struct net_device
*net_dev
)
2156 struct efx_nic
*efx
= netdev_priv(net_dev
);
2158 netif_err(efx
, tx_err
, efx
->net_dev
,
2159 "TX stuck with port_enabled=%d: resetting channels\n",
2162 efx_schedule_reset(efx
, RESET_TYPE_TX_WATCHDOG
);
2166 /* Context: process, rtnl_lock() held. */
2167 static int efx_change_mtu(struct net_device
*net_dev
, int new_mtu
)
2169 struct efx_nic
*efx
= netdev_priv(net_dev
);
2172 rc
= efx_check_disabled(efx
);
2175 if (new_mtu
> EFX_MAX_MTU
)
2178 netif_dbg(efx
, drv
, efx
->net_dev
, "changing MTU to %d\n", new_mtu
);
2180 efx_device_detach_sync(efx
);
2183 mutex_lock(&efx
->mac_lock
);
2184 net_dev
->mtu
= new_mtu
;
2185 efx
->type
->reconfigure_mac(efx
);
2186 mutex_unlock(&efx
->mac_lock
);
2189 netif_device_attach(efx
->net_dev
);
2193 static int efx_set_mac_address(struct net_device
*net_dev
, void *data
)
2195 struct efx_nic
*efx
= netdev_priv(net_dev
);
2196 struct sockaddr
*addr
= data
;
2197 u8
*new_addr
= addr
->sa_data
;
2199 if (!is_valid_ether_addr(new_addr
)) {
2200 netif_err(efx
, drv
, efx
->net_dev
,
2201 "invalid ethernet MAC address requested: %pM\n",
2203 return -EADDRNOTAVAIL
;
2206 ether_addr_copy(net_dev
->dev_addr
, new_addr
);
2207 if (efx
->type
->sriov_mac_address_changed
)
2208 efx
->type
->sriov_mac_address_changed(efx
);
2210 /* Reconfigure the MAC */
2211 mutex_lock(&efx
->mac_lock
);
2212 efx
->type
->reconfigure_mac(efx
);
2213 mutex_unlock(&efx
->mac_lock
);
2218 /* Context: netif_addr_lock held, BHs disabled. */
2219 static void efx_set_rx_mode(struct net_device
*net_dev
)
2221 struct efx_nic
*efx
= netdev_priv(net_dev
);
2223 if (efx
->port_enabled
)
2224 queue_work(efx
->workqueue
, &efx
->mac_work
);
2225 /* Otherwise efx_start_port() will do this */
2228 static int efx_set_features(struct net_device
*net_dev
, netdev_features_t data
)
2230 struct efx_nic
*efx
= netdev_priv(net_dev
);
2232 /* If disabling RX n-tuple filtering, clear existing filters */
2233 if (net_dev
->features
& ~data
& NETIF_F_NTUPLE
)
2234 return efx
->type
->filter_clear_rx(efx
, EFX_FILTER_PRI_MANUAL
);
2239 static const struct net_device_ops efx_netdev_ops
= {
2240 .ndo_open
= efx_net_open
,
2241 .ndo_stop
= efx_net_stop
,
2242 .ndo_get_stats64
= efx_net_stats
,
2243 .ndo_tx_timeout
= efx_watchdog
,
2244 .ndo_start_xmit
= efx_hard_start_xmit
,
2245 .ndo_validate_addr
= eth_validate_addr
,
2246 .ndo_do_ioctl
= efx_ioctl
,
2247 .ndo_change_mtu
= efx_change_mtu
,
2248 .ndo_set_mac_address
= efx_set_mac_address
,
2249 .ndo_set_rx_mode
= efx_set_rx_mode
,
2250 .ndo_set_features
= efx_set_features
,
2251 #ifdef CONFIG_SFC_SRIOV
2252 .ndo_set_vf_mac
= efx_sriov_set_vf_mac
,
2253 .ndo_set_vf_vlan
= efx_sriov_set_vf_vlan
,
2254 .ndo_set_vf_spoofchk
= efx_sriov_set_vf_spoofchk
,
2255 .ndo_get_vf_config
= efx_sriov_get_vf_config
,
2257 #ifdef CONFIG_NET_POLL_CONTROLLER
2258 .ndo_poll_controller
= efx_netpoll
,
2260 .ndo_setup_tc
= efx_setup_tc
,
2261 #ifdef CONFIG_NET_RX_BUSY_POLL
2262 .ndo_busy_poll
= efx_busy_poll
,
2264 #ifdef CONFIG_RFS_ACCEL
2265 .ndo_rx_flow_steer
= efx_filter_rfs
,
2269 static void efx_update_name(struct efx_nic
*efx
)
2271 strcpy(efx
->name
, efx
->net_dev
->name
);
2272 efx_mtd_rename(efx
);
2273 efx_set_channel_names(efx
);
2276 static int efx_netdev_event(struct notifier_block
*this,
2277 unsigned long event
, void *ptr
)
2279 struct net_device
*net_dev
= netdev_notifier_info_to_dev(ptr
);
2281 if ((net_dev
->netdev_ops
== &efx_netdev_ops
) &&
2282 event
== NETDEV_CHANGENAME
)
2283 efx_update_name(netdev_priv(net_dev
));
2288 static struct notifier_block efx_netdev_notifier
= {
2289 .notifier_call
= efx_netdev_event
,
2293 show_phy_type(struct device
*dev
, struct device_attribute
*attr
, char *buf
)
2295 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
2296 return sprintf(buf
, "%d\n", efx
->phy_type
);
2298 static DEVICE_ATTR(phy_type
, 0444, show_phy_type
, NULL
);
2300 static int efx_register_netdev(struct efx_nic
*efx
)
2302 struct net_device
*net_dev
= efx
->net_dev
;
2303 struct efx_channel
*channel
;
2306 net_dev
->watchdog_timeo
= 5 * HZ
;
2307 net_dev
->irq
= efx
->pci_dev
->irq
;
2308 net_dev
->netdev_ops
= &efx_netdev_ops
;
2309 if (efx_nic_rev(efx
) >= EFX_REV_HUNT_A0
)
2310 net_dev
->priv_flags
|= IFF_UNICAST_FLT
;
2311 net_dev
->ethtool_ops
= &efx_ethtool_ops
;
2312 net_dev
->gso_max_segs
= EFX_TSO_MAX_SEGS
;
2316 /* Enable resets to be scheduled and check whether any were
2317 * already requested. If so, the NIC is probably hosed so we
2320 efx
->state
= STATE_READY
;
2321 smp_mb(); /* ensure we change state before checking reset_pending */
2322 if (efx
->reset_pending
) {
2323 netif_err(efx
, probe
, efx
->net_dev
,
2324 "aborting probe due to scheduled reset\n");
2329 rc
= dev_alloc_name(net_dev
, net_dev
->name
);
2332 efx_update_name(efx
);
2334 /* Always start with carrier off; PHY events will detect the link */
2335 netif_carrier_off(net_dev
);
2337 rc
= register_netdevice(net_dev
);
2341 efx_for_each_channel(channel
, efx
) {
2342 struct efx_tx_queue
*tx_queue
;
2343 efx_for_each_channel_tx_queue(tx_queue
, channel
)
2344 efx_init_tx_queue_core_txq(tx_queue
);
2351 rc
= device_create_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
2353 netif_err(efx
, drv
, efx
->net_dev
,
2354 "failed to init net dev attributes\n");
2355 goto fail_registered
;
2362 efx_dissociate(efx
);
2363 unregister_netdevice(net_dev
);
2365 efx
->state
= STATE_UNINIT
;
2367 netif_err(efx
, drv
, efx
->net_dev
, "could not register net dev\n");
2371 static void efx_unregister_netdev(struct efx_nic
*efx
)
2376 BUG_ON(netdev_priv(efx
->net_dev
) != efx
);
2378 strlcpy(efx
->name
, pci_name(efx
->pci_dev
), sizeof(efx
->name
));
2379 device_remove_file(&efx
->pci_dev
->dev
, &dev_attr_phy_type
);
2382 unregister_netdevice(efx
->net_dev
);
2383 efx
->state
= STATE_UNINIT
;
2387 /**************************************************************************
2389 * Device reset and suspend
2391 **************************************************************************/
2393 /* Tears down the entire software state and most of the hardware state
2395 void efx_reset_down(struct efx_nic
*efx
, enum reset_type method
)
2397 EFX_ASSERT_RESET_SERIALISED(efx
);
2399 if (method
== RESET_TYPE_MCDI_TIMEOUT
)
2400 efx
->type
->prepare_flr(efx
);
2403 efx_disable_interrupts(efx
);
2405 mutex_lock(&efx
->mac_lock
);
2406 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
)
2407 efx
->phy_op
->fini(efx
);
2408 efx
->type
->fini(efx
);
2411 /* This function will always ensure that the locks acquired in
2412 * efx_reset_down() are released. A failure return code indicates
2413 * that we were unable to reinitialise the hardware, and the
2414 * driver should be disabled. If ok is false, then the rx and tx
2415 * engines are not restarted, pending a RESET_DISABLE. */
2416 int efx_reset_up(struct efx_nic
*efx
, enum reset_type method
, bool ok
)
2420 EFX_ASSERT_RESET_SERIALISED(efx
);
2422 if (method
== RESET_TYPE_MCDI_TIMEOUT
)
2423 efx
->type
->finish_flr(efx
);
2425 /* Ensure that SRAM is initialised even if we're disabling the device */
2426 rc
= efx
->type
->init(efx
);
2428 netif_err(efx
, drv
, efx
->net_dev
, "failed to initialise NIC\n");
2435 if (efx
->port_initialized
&& method
!= RESET_TYPE_INVISIBLE
) {
2436 rc
= efx
->phy_op
->init(efx
);
2439 rc
= efx
->phy_op
->reconfigure(efx
);
2440 if (rc
&& rc
!= -EPERM
)
2441 netif_err(efx
, drv
, efx
->net_dev
,
2442 "could not restore PHY settings\n");
2445 rc
= efx_enable_interrupts(efx
);
2449 #ifdef CONFIG_SFC_SRIOV
2450 rc
= efx
->type
->vswitching_restore(efx
);
2451 if (rc
) /* not fatal; the PF will still work fine */
2452 netif_warn(efx
, probe
, efx
->net_dev
,
2453 "failed to restore vswitching rc=%d;"
2454 " VFs may not function\n", rc
);
2457 efx_restore_filters(efx
);
2458 if (efx
->type
->sriov_reset
)
2459 efx
->type
->sriov_reset(efx
);
2461 mutex_unlock(&efx
->mac_lock
);
2468 efx
->port_initialized
= false;
2470 mutex_unlock(&efx
->mac_lock
);
2475 /* Reset the NIC using the specified method. Note that the reset may
2476 * fail, in which case the card will be left in an unusable state.
2478 * Caller must hold the rtnl_lock.
2480 int efx_reset(struct efx_nic
*efx
, enum reset_type method
)
2485 netif_info(efx
, drv
, efx
->net_dev
, "resetting (%s)\n",
2486 RESET_TYPE(method
));
2488 efx_device_detach_sync(efx
);
2489 efx_reset_down(efx
, method
);
2491 rc
= efx
->type
->reset(efx
, method
);
2493 netif_err(efx
, drv
, efx
->net_dev
, "failed to reset hardware\n");
2497 /* Clear flags for the scopes we covered. We assume the NIC and
2498 * driver are now quiescent so that there is no race here.
2500 if (method
< RESET_TYPE_MAX_METHOD
)
2501 efx
->reset_pending
&= -(1 << (method
+ 1));
2502 else /* it doesn't fit into the well-ordered scope hierarchy */
2503 __clear_bit(method
, &efx
->reset_pending
);
2505 /* Reinitialise bus-mastering, which may have been turned off before
2506 * the reset was scheduled. This is still appropriate, even in the
2507 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2508 * can respond to requests. */
2509 pci_set_master(efx
->pci_dev
);
2512 /* Leave device stopped if necessary */
2514 method
== RESET_TYPE_DISABLE
||
2515 method
== RESET_TYPE_RECOVER_OR_DISABLE
;
2516 rc2
= efx_reset_up(efx
, method
, !disabled
);
2524 dev_close(efx
->net_dev
);
2525 netif_err(efx
, drv
, efx
->net_dev
, "has been disabled\n");
2526 efx
->state
= STATE_DISABLED
;
2528 netif_dbg(efx
, drv
, efx
->net_dev
, "reset complete\n");
2529 netif_device_attach(efx
->net_dev
);
2534 /* Try recovery mechanisms.
2535 * For now only EEH is supported.
2536 * Returns 0 if the recovery mechanisms are unsuccessful.
2537 * Returns a non-zero value otherwise.
2539 int efx_try_recovery(struct efx_nic
*efx
)
2542 /* A PCI error can occur and not be seen by EEH because nothing
2543 * happens on the PCI bus. In this case the driver may fail and
2544 * schedule a 'recover or reset', leading to this recovery handler.
2545 * Manually call the eeh failure check function.
2547 struct eeh_dev
*eehdev
= pci_dev_to_eeh_dev(efx
->pci_dev
);
2548 if (eeh_dev_check_failure(eehdev
)) {
2549 /* The EEH mechanisms will handle the error and reset the
2550 * device if necessary.
2558 static void efx_wait_for_bist_end(struct efx_nic
*efx
)
2562 for (i
= 0; i
< BIST_WAIT_DELAY_COUNT
; ++i
) {
2563 if (efx_mcdi_poll_reboot(efx
))
2565 msleep(BIST_WAIT_DELAY_MS
);
2568 netif_err(efx
, drv
, efx
->net_dev
, "Warning: No MC reboot after BIST mode\n");
2570 /* Either way unset the BIST flag. If we found no reboot we probably
2571 * won't recover, but we should try.
2573 efx
->mc_bist_for_other_fn
= false;
2576 /* The worker thread exists so that code that cannot sleep can
2577 * schedule a reset for later.
2579 static void efx_reset_work(struct work_struct
*data
)
2581 struct efx_nic
*efx
= container_of(data
, struct efx_nic
, reset_work
);
2582 unsigned long pending
;
2583 enum reset_type method
;
2585 pending
= ACCESS_ONCE(efx
->reset_pending
);
2586 method
= fls(pending
) - 1;
2588 if (method
== RESET_TYPE_MC_BIST
)
2589 efx_wait_for_bist_end(efx
);
2591 if ((method
== RESET_TYPE_RECOVER_OR_DISABLE
||
2592 method
== RESET_TYPE_RECOVER_OR_ALL
) &&
2593 efx_try_recovery(efx
))
2601 /* We checked the state in efx_schedule_reset() but it may
2602 * have changed by now. Now that we have the RTNL lock,
2603 * it cannot change again.
2605 if (efx
->state
== STATE_READY
)
2606 (void)efx_reset(efx
, method
);
2611 void efx_schedule_reset(struct efx_nic
*efx
, enum reset_type type
)
2613 enum reset_type method
;
2615 if (efx
->state
== STATE_RECOVERY
) {
2616 netif_dbg(efx
, drv
, efx
->net_dev
,
2617 "recovering: skip scheduling %s reset\n",
2623 case RESET_TYPE_INVISIBLE
:
2624 case RESET_TYPE_ALL
:
2625 case RESET_TYPE_RECOVER_OR_ALL
:
2626 case RESET_TYPE_WORLD
:
2627 case RESET_TYPE_DISABLE
:
2628 case RESET_TYPE_RECOVER_OR_DISABLE
:
2629 case RESET_TYPE_MC_BIST
:
2630 case RESET_TYPE_MCDI_TIMEOUT
:
2632 netif_dbg(efx
, drv
, efx
->net_dev
, "scheduling %s reset\n",
2633 RESET_TYPE(method
));
2636 method
= efx
->type
->map_reset_reason(type
);
2637 netif_dbg(efx
, drv
, efx
->net_dev
,
2638 "scheduling %s reset for %s\n",
2639 RESET_TYPE(method
), RESET_TYPE(type
));
2643 set_bit(method
, &efx
->reset_pending
);
2644 smp_mb(); /* ensure we change reset_pending before checking state */
2646 /* If we're not READY then just leave the flags set as the cue
2647 * to abort probing or reschedule the reset later.
2649 if (ACCESS_ONCE(efx
->state
) != STATE_READY
)
2652 /* efx_process_channel() will no longer read events once a
2653 * reset is scheduled. So switch back to poll'd MCDI completions. */
2654 efx_mcdi_mode_poll(efx
);
2656 queue_work(reset_workqueue
, &efx
->reset_work
);
2659 /**************************************************************************
2661 * List of NICs we support
2663 **************************************************************************/
2665 /* PCI device ID table */
2666 static const struct pci_device_id efx_pci_table
[] = {
2667 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
,
2668 PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0
),
2669 .driver_data
= (unsigned long) &falcon_a1_nic_type
},
2670 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
,
2671 PCI_DEVICE_ID_SOLARFLARE_SFC4000B
),
2672 .driver_data
= (unsigned long) &falcon_b0_nic_type
},
2673 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0803), /* SFC9020 */
2674 .driver_data
= (unsigned long) &siena_a0_nic_type
},
2675 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0813), /* SFL9021 */
2676 .driver_data
= (unsigned long) &siena_a0_nic_type
},
2677 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0903), /* SFC9120 PF */
2678 .driver_data
= (unsigned long) &efx_hunt_a0_nic_type
},
2679 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE
, 0x0923), /* SFC9140 PF */
2680 .driver_data
= (unsigned long) &efx_hunt_a0_nic_type
},
2681 {0} /* end of list */
2684 /**************************************************************************
2686 * Dummy PHY/MAC operations
2688 * Can be used for some unimplemented operations
2689 * Needed so all function pointers are valid and do not have to be tested
2692 **************************************************************************/
2693 int efx_port_dummy_op_int(struct efx_nic
*efx
)
2697 void efx_port_dummy_op_void(struct efx_nic
*efx
) {}
2699 static bool efx_port_dummy_op_poll(struct efx_nic
*efx
)
2704 static const struct efx_phy_operations efx_dummy_phy_operations
= {
2705 .init
= efx_port_dummy_op_int
,
2706 .reconfigure
= efx_port_dummy_op_int
,
2707 .poll
= efx_port_dummy_op_poll
,
2708 .fini
= efx_port_dummy_op_void
,
2711 /**************************************************************************
2715 **************************************************************************/
2717 /* This zeroes out and then fills in the invariants in a struct
2718 * efx_nic (including all sub-structures).
2720 static int efx_init_struct(struct efx_nic
*efx
,
2721 struct pci_dev
*pci_dev
, struct net_device
*net_dev
)
2725 /* Initialise common structures */
2726 INIT_LIST_HEAD(&efx
->node
);
2727 INIT_LIST_HEAD(&efx
->secondary_list
);
2728 spin_lock_init(&efx
->biu_lock
);
2729 #ifdef CONFIG_SFC_MTD
2730 INIT_LIST_HEAD(&efx
->mtd_list
);
2732 INIT_WORK(&efx
->reset_work
, efx_reset_work
);
2733 INIT_DELAYED_WORK(&efx
->monitor_work
, efx_monitor
);
2734 INIT_DELAYED_WORK(&efx
->selftest_work
, efx_selftest_async_work
);
2735 efx
->pci_dev
= pci_dev
;
2736 efx
->msg_enable
= debug
;
2737 efx
->state
= STATE_UNINIT
;
2738 strlcpy(efx
->name
, pci_name(pci_dev
), sizeof(efx
->name
));
2740 efx
->net_dev
= net_dev
;
2741 efx
->rx_prefix_size
= efx
->type
->rx_prefix_size
;
2743 NET_IP_ALIGN
? (efx
->rx_prefix_size
+ NET_IP_ALIGN
) % 4 : 0;
2744 efx
->rx_packet_hash_offset
=
2745 efx
->type
->rx_hash_offset
- efx
->type
->rx_prefix_size
;
2746 efx
->rx_packet_ts_offset
=
2747 efx
->type
->rx_ts_offset
- efx
->type
->rx_prefix_size
;
2748 spin_lock_init(&efx
->stats_lock
);
2749 mutex_init(&efx
->mac_lock
);
2750 efx
->phy_op
= &efx_dummy_phy_operations
;
2751 efx
->mdio
.dev
= net_dev
;
2752 INIT_WORK(&efx
->mac_work
, efx_mac_work
);
2753 init_waitqueue_head(&efx
->flush_wq
);
2755 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++) {
2756 efx
->channel
[i
] = efx_alloc_channel(efx
, i
, NULL
);
2757 if (!efx
->channel
[i
])
2759 efx
->msi_context
[i
].efx
= efx
;
2760 efx
->msi_context
[i
].index
= i
;
2763 /* Higher numbered interrupt modes are less capable! */
2764 efx
->interrupt_mode
= max(efx
->type
->max_interrupt_mode
,
2767 /* Would be good to use the net_dev name, but we're too early */
2768 snprintf(efx
->workqueue_name
, sizeof(efx
->workqueue_name
), "sfc%s",
2770 efx
->workqueue
= create_singlethread_workqueue(efx
->workqueue_name
);
2771 if (!efx
->workqueue
)
2777 efx_fini_struct(efx
);
2781 static void efx_fini_struct(struct efx_nic
*efx
)
2785 for (i
= 0; i
< EFX_MAX_CHANNELS
; i
++)
2786 kfree(efx
->channel
[i
]);
2790 if (efx
->workqueue
) {
2791 destroy_workqueue(efx
->workqueue
);
2792 efx
->workqueue
= NULL
;
2796 void efx_update_sw_stats(struct efx_nic
*efx
, u64
*stats
)
2798 u64 n_rx_nodesc_trunc
= 0;
2799 struct efx_channel
*channel
;
2801 efx_for_each_channel(channel
, efx
)
2802 n_rx_nodesc_trunc
+= channel
->n_rx_nodesc_trunc
;
2803 stats
[GENERIC_STAT_rx_nodesc_trunc
] = n_rx_nodesc_trunc
;
2804 stats
[GENERIC_STAT_rx_noskb_drops
] = atomic_read(&efx
->n_rx_noskb_drops
);
2807 /**************************************************************************
2811 **************************************************************************/
2813 /* Main body of final NIC shutdown code
2814 * This is called only at module unload (or hotplug removal).
2816 static void efx_pci_remove_main(struct efx_nic
*efx
)
2818 /* Flush reset_work. It can no longer be scheduled since we
2821 BUG_ON(efx
->state
== STATE_READY
);
2822 cancel_work_sync(&efx
->reset_work
);
2824 efx_disable_interrupts(efx
);
2825 efx_nic_fini_interrupt(efx
);
2827 efx
->type
->fini(efx
);
2829 efx_remove_all(efx
);
2832 /* Final NIC shutdown
2833 * This is called only at module unload (or hotplug removal).
2835 static void efx_pci_remove(struct pci_dev
*pci_dev
)
2837 struct efx_nic
*efx
;
2839 efx
= pci_get_drvdata(pci_dev
);
2843 /* Mark the NIC as fini, then stop the interface */
2845 efx_dissociate(efx
);
2846 dev_close(efx
->net_dev
);
2847 efx_disable_interrupts(efx
);
2850 if (efx
->type
->sriov_fini
)
2851 efx
->type
->sriov_fini(efx
);
2853 efx_unregister_netdev(efx
);
2855 efx_mtd_remove(efx
);
2857 efx_pci_remove_main(efx
);
2860 netif_dbg(efx
, drv
, efx
->net_dev
, "shutdown successful\n");
2862 efx_fini_struct(efx
);
2863 free_netdev(efx
->net_dev
);
2865 pci_disable_pcie_error_reporting(pci_dev
);
2868 /* NIC VPD information
2869 * Called during probe to display the part number of the
2870 * installed NIC. VPD is potentially very large but this should
2871 * always appear within the first 512 bytes.
2873 #define SFC_VPD_LEN 512
2874 static void efx_probe_vpd_strings(struct efx_nic
*efx
)
2876 struct pci_dev
*dev
= efx
->pci_dev
;
2877 char vpd_data
[SFC_VPD_LEN
];
2879 int ro_start
, ro_size
, i
, j
;
2881 /* Get the vpd data from the device */
2882 vpd_size
= pci_read_vpd(dev
, 0, sizeof(vpd_data
), vpd_data
);
2883 if (vpd_size
<= 0) {
2884 netif_err(efx
, drv
, efx
->net_dev
, "Unable to read VPD\n");
2888 /* Get the Read only section */
2889 ro_start
= pci_vpd_find_tag(vpd_data
, 0, vpd_size
, PCI_VPD_LRDT_RO_DATA
);
2891 netif_err(efx
, drv
, efx
->net_dev
, "VPD Read-only not found\n");
2895 ro_size
= pci_vpd_lrdt_size(&vpd_data
[ro_start
]);
2897 i
= ro_start
+ PCI_VPD_LRDT_TAG_SIZE
;
2898 if (i
+ j
> vpd_size
)
2901 /* Get the Part number */
2902 i
= pci_vpd_find_info_keyword(vpd_data
, i
, j
, "PN");
2904 netif_err(efx
, drv
, efx
->net_dev
, "Part number not found\n");
2908 j
= pci_vpd_info_field_size(&vpd_data
[i
]);
2909 i
+= PCI_VPD_INFO_FLD_HDR_SIZE
;
2910 if (i
+ j
> vpd_size
) {
2911 netif_err(efx
, drv
, efx
->net_dev
, "Incomplete part number\n");
2915 netif_info(efx
, drv
, efx
->net_dev
,
2916 "Part Number : %.*s\n", j
, &vpd_data
[i
]);
2918 i
= ro_start
+ PCI_VPD_LRDT_TAG_SIZE
;
2920 i
= pci_vpd_find_info_keyword(vpd_data
, i
, j
, "SN");
2922 netif_err(efx
, drv
, efx
->net_dev
, "Serial number not found\n");
2926 j
= pci_vpd_info_field_size(&vpd_data
[i
]);
2927 i
+= PCI_VPD_INFO_FLD_HDR_SIZE
;
2928 if (i
+ j
> vpd_size
) {
2929 netif_err(efx
, drv
, efx
->net_dev
, "Incomplete serial number\n");
2933 efx
->vpd_sn
= kmalloc(j
+ 1, GFP_KERNEL
);
2937 snprintf(efx
->vpd_sn
, j
+ 1, "%s", &vpd_data
[i
]);
2941 /* Main body of NIC initialisation
2942 * This is called at module load (or hotplug insertion, theoretically).
2944 static int efx_pci_probe_main(struct efx_nic
*efx
)
2948 /* Do start-of-day initialisation */
2949 rc
= efx_probe_all(efx
);
2955 rc
= efx
->type
->init(efx
);
2957 netif_err(efx
, probe
, efx
->net_dev
,
2958 "failed to initialise NIC\n");
2962 rc
= efx_init_port(efx
);
2964 netif_err(efx
, probe
, efx
->net_dev
,
2965 "failed to initialise port\n");
2969 rc
= efx_nic_init_interrupt(efx
);
2972 rc
= efx_enable_interrupts(efx
);
2979 efx_nic_fini_interrupt(efx
);
2983 efx
->type
->fini(efx
);
2986 efx_remove_all(efx
);
2991 /* NIC initialisation
2993 * This is called at module load (or hotplug insertion,
2994 * theoretically). It sets up PCI mappings, resets the NIC,
2995 * sets up and registers the network devices with the kernel and hooks
2996 * the interrupt service routine. It does not prepare the device for
2997 * transmission; this is left to the first time one of the network
2998 * interfaces is brought up (i.e. efx_net_open).
3000 static int efx_pci_probe(struct pci_dev
*pci_dev
,
3001 const struct pci_device_id
*entry
)
3003 struct net_device
*net_dev
;
3004 struct efx_nic
*efx
;
3007 /* Allocate and initialise a struct net_device and struct efx_nic */
3008 net_dev
= alloc_etherdev_mqs(sizeof(*efx
), EFX_MAX_CORE_TX_QUEUES
,
3012 efx
= netdev_priv(net_dev
);
3013 efx
->type
= (const struct efx_nic_type
*) entry
->driver_data
;
3014 net_dev
->features
|= (efx
->type
->offload_features
| NETIF_F_SG
|
3015 NETIF_F_HIGHDMA
| NETIF_F_TSO
|
3017 if (efx
->type
->offload_features
& NETIF_F_V6_CSUM
)
3018 net_dev
->features
|= NETIF_F_TSO6
;
3019 /* Mask for features that also apply to VLAN devices */
3020 net_dev
->vlan_features
|= (NETIF_F_ALL_CSUM
| NETIF_F_SG
|
3021 NETIF_F_HIGHDMA
| NETIF_F_ALL_TSO
|
3023 /* All offloads can be toggled */
3024 net_dev
->hw_features
= net_dev
->features
& ~NETIF_F_HIGHDMA
;
3025 pci_set_drvdata(pci_dev
, efx
);
3026 SET_NETDEV_DEV(net_dev
, &pci_dev
->dev
);
3027 rc
= efx_init_struct(efx
, pci_dev
, net_dev
);
3031 netif_info(efx
, probe
, efx
->net_dev
,
3032 "Solarflare NIC detected\n");
3034 efx_probe_vpd_strings(efx
);
3036 /* Set up basic I/O (BAR mappings etc) */
3037 rc
= efx_init_io(efx
);
3041 rc
= efx_pci_probe_main(efx
);
3045 rc
= efx_register_netdev(efx
);
3049 if (efx
->type
->sriov_init
) {
3050 rc
= efx
->type
->sriov_init(efx
);
3052 netif_err(efx
, probe
, efx
->net_dev
,
3053 "SR-IOV can't be enabled rc %d\n", rc
);
3056 netif_dbg(efx
, probe
, efx
->net_dev
, "initialisation successful\n");
3058 /* Try to create MTDs, but allow this to fail */
3060 rc
= efx_mtd_probe(efx
);
3063 netif_warn(efx
, probe
, efx
->net_dev
,
3064 "failed to create MTDs (%d)\n", rc
);
3066 rc
= pci_enable_pcie_error_reporting(pci_dev
);
3067 if (rc
&& rc
!= -EINVAL
)
3068 netif_warn(efx
, probe
, efx
->net_dev
,
3069 "pci_enable_pcie_error_reporting failed (%d)\n", rc
);
3074 efx_pci_remove_main(efx
);
3078 efx_fini_struct(efx
);
3081 netif_dbg(efx
, drv
, efx
->net_dev
, "initialisation failed. rc=%d\n", rc
);
3082 free_netdev(net_dev
);
3086 /* efx_pci_sriov_configure returns the actual number of Virtual Functions
3087 * enabled on success
3089 #ifdef CONFIG_SFC_SRIOV
3090 static int efx_pci_sriov_configure(struct pci_dev
*dev
, int num_vfs
)
3093 struct efx_nic
*efx
= pci_get_drvdata(dev
);
3095 if (efx
->type
->sriov_configure
) {
3096 rc
= efx
->type
->sriov_configure(efx
, num_vfs
);
3106 static int efx_pm_freeze(struct device
*dev
)
3108 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
3112 if (efx
->state
!= STATE_DISABLED
) {
3113 efx
->state
= STATE_UNINIT
;
3115 efx_device_detach_sync(efx
);
3118 efx_disable_interrupts(efx
);
3126 static int efx_pm_thaw(struct device
*dev
)
3129 struct efx_nic
*efx
= pci_get_drvdata(to_pci_dev(dev
));
3133 if (efx
->state
!= STATE_DISABLED
) {
3134 rc
= efx_enable_interrupts(efx
);
3138 mutex_lock(&efx
->mac_lock
);
3139 efx
->phy_op
->reconfigure(efx
);
3140 mutex_unlock(&efx
->mac_lock
);
3144 netif_device_attach(efx
->net_dev
);
3146 efx
->state
= STATE_READY
;
3148 efx
->type
->resume_wol(efx
);
3153 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3154 queue_work(reset_workqueue
, &efx
->reset_work
);
3164 static int efx_pm_poweroff(struct device
*dev
)
3166 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
3167 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
3169 efx
->type
->fini(efx
);
3171 efx
->reset_pending
= 0;
3173 pci_save_state(pci_dev
);
3174 return pci_set_power_state(pci_dev
, PCI_D3hot
);
3177 /* Used for both resume and restore */
3178 static int efx_pm_resume(struct device
*dev
)
3180 struct pci_dev
*pci_dev
= to_pci_dev(dev
);
3181 struct efx_nic
*efx
= pci_get_drvdata(pci_dev
);
3184 rc
= pci_set_power_state(pci_dev
, PCI_D0
);
3187 pci_restore_state(pci_dev
);
3188 rc
= pci_enable_device(pci_dev
);
3191 pci_set_master(efx
->pci_dev
);
3192 rc
= efx
->type
->reset(efx
, RESET_TYPE_ALL
);
3195 rc
= efx
->type
->init(efx
);
3198 rc
= efx_pm_thaw(dev
);
3202 static int efx_pm_suspend(struct device
*dev
)
3207 rc
= efx_pm_poweroff(dev
);
3213 static const struct dev_pm_ops efx_pm_ops
= {
3214 .suspend
= efx_pm_suspend
,
3215 .resume
= efx_pm_resume
,
3216 .freeze
= efx_pm_freeze
,
3217 .thaw
= efx_pm_thaw
,
3218 .poweroff
= efx_pm_poweroff
,
3219 .restore
= efx_pm_resume
,
3222 /* A PCI error affecting this device was detected.
3223 * At this point MMIO and DMA may be disabled.
3224 * Stop the software path and request a slot reset.
3226 static pci_ers_result_t
efx_io_error_detected(struct pci_dev
*pdev
,
3227 enum pci_channel_state state
)
3229 pci_ers_result_t status
= PCI_ERS_RESULT_RECOVERED
;
3230 struct efx_nic
*efx
= pci_get_drvdata(pdev
);
3232 if (state
== pci_channel_io_perm_failure
)
3233 return PCI_ERS_RESULT_DISCONNECT
;
3237 if (efx
->state
!= STATE_DISABLED
) {
3238 efx
->state
= STATE_RECOVERY
;
3239 efx
->reset_pending
= 0;
3241 efx_device_detach_sync(efx
);
3244 efx_disable_interrupts(efx
);
3246 status
= PCI_ERS_RESULT_NEED_RESET
;
3248 /* If the interface is disabled we don't want to do anything
3251 status
= PCI_ERS_RESULT_RECOVERED
;
3256 pci_disable_device(pdev
);
3261 /* Fake a successful reset, which will be performed later in efx_io_resume. */
3262 static pci_ers_result_t
efx_io_slot_reset(struct pci_dev
*pdev
)
3264 struct efx_nic
*efx
= pci_get_drvdata(pdev
);
3265 pci_ers_result_t status
= PCI_ERS_RESULT_RECOVERED
;
3268 if (pci_enable_device(pdev
)) {
3269 netif_err(efx
, hw
, efx
->net_dev
,
3270 "Cannot re-enable PCI device after reset.\n");
3271 status
= PCI_ERS_RESULT_DISCONNECT
;
3274 rc
= pci_cleanup_aer_uncorrect_error_status(pdev
);
3276 netif_err(efx
, hw
, efx
->net_dev
,
3277 "pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc
);
3278 /* Non-fatal error. Continue. */
3284 /* Perform the actual reset and resume I/O operations. */
3285 static void efx_io_resume(struct pci_dev
*pdev
)
3287 struct efx_nic
*efx
= pci_get_drvdata(pdev
);
3292 if (efx
->state
== STATE_DISABLED
)
3295 rc
= efx_reset(efx
, RESET_TYPE_ALL
);
3297 netif_err(efx
, hw
, efx
->net_dev
,
3298 "efx_reset failed after PCI error (%d)\n", rc
);
3300 efx
->state
= STATE_READY
;
3301 netif_dbg(efx
, hw
, efx
->net_dev
,
3302 "Done resetting and resuming IO after PCI error.\n");
3309 /* For simplicity and reliability, we always require a slot reset and try to
3310 * reset the hardware when a pci error affecting the device is detected.
3311 * We leave both the link_reset and mmio_enabled callback unimplemented:
3312 * with our request for slot reset the mmio_enabled callback will never be
3313 * called, and the link_reset callback is not used by AER or EEH mechanisms.
3315 static struct pci_error_handlers efx_err_handlers
= {
3316 .error_detected
= efx_io_error_detected
,
3317 .slot_reset
= efx_io_slot_reset
,
3318 .resume
= efx_io_resume
,
3321 static struct pci_driver efx_pci_driver
= {
3322 .name
= KBUILD_MODNAME
,
3323 .id_table
= efx_pci_table
,
3324 .probe
= efx_pci_probe
,
3325 .remove
= efx_pci_remove
,
3326 .driver
.pm
= &efx_pm_ops
,
3327 .err_handler
= &efx_err_handlers
,
3328 #ifdef CONFIG_SFC_SRIOV
3329 .sriov_configure
= efx_pci_sriov_configure
,
3333 /**************************************************************************
3335 * Kernel module interface
3337 *************************************************************************/
3339 module_param(interrupt_mode
, uint
, 0444);
3340 MODULE_PARM_DESC(interrupt_mode
,
3341 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3343 static int __init
efx_init_module(void)
3347 printk(KERN_INFO
"Solarflare NET driver v" EFX_DRIVER_VERSION
"\n");
3349 rc
= register_netdevice_notifier(&efx_netdev_notifier
);
3353 #ifdef CONFIG_SFC_SRIOV
3354 rc
= efx_init_sriov();
3359 reset_workqueue
= create_singlethread_workqueue("sfc_reset");
3360 if (!reset_workqueue
) {
3365 rc
= pci_register_driver(&efx_pci_driver
);
3372 destroy_workqueue(reset_workqueue
);
3374 #ifdef CONFIG_SFC_SRIOV
3378 unregister_netdevice_notifier(&efx_netdev_notifier
);
3383 static void __exit
efx_exit_module(void)
3385 printk(KERN_INFO
"Solarflare NET driver unloading\n");
3387 pci_unregister_driver(&efx_pci_driver
);
3388 destroy_workqueue(reset_workqueue
);
3389 #ifdef CONFIG_SFC_SRIOV
3392 unregister_netdevice_notifier(&efx_netdev_notifier
);
3396 module_init(efx_init_module
);
3397 module_exit(efx_exit_module
);
3399 MODULE_AUTHOR("Solarflare Communications and "
3400 "Michael Brown <mbrown@fensystems.co.uk>");
3401 MODULE_DESCRIPTION("Solarflare network driver");
3402 MODULE_LICENSE("GPL");
3403 MODULE_DEVICE_TABLE(pci
, efx_pci_table
);