Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2005-2006 Fen Systems Ltd.
5 * Copyright 2005-2013 Solarflare Communications Inc.
6 */
7
8#include <linux/filter.h>
9#include <linux/module.h>
10#include <linux/pci.h>
11#include <linux/netdevice.h>
12#include <linux/etherdevice.h>
13#include <linux/delay.h>
14#include <linux/notifier.h>
15#include <linux/ip.h>
16#include <linux/tcp.h>
17#include <linux/in.h>
18#include <linux/ethtool.h>
19#include <linux/topology.h>
20#include <linux/gfp.h>
21#include <linux/interrupt.h>
22#include "net_driver.h"
23#include <net/gre.h>
24#include <net/udp_tunnel.h>
25#include <net/netdev_queues.h>
26#include "efx.h"
27#include "efx_common.h"
28#include "efx_channels.h"
29#include "ef100.h"
30#include "rx_common.h"
31#include "tx_common.h"
32#include "nic.h"
33#include "io.h"
34#include "selftest.h"
35#include "sriov.h"
36#include "efx_devlink.h"
37
38#include "mcdi_port_common.h"
39#include "mcdi_pcol.h"
40#include "workarounds.h"
41
42/**************************************************************************
43 *
44 * Configurable values
45 *
46 *************************************************************************/
47
48module_param_named(interrupt_mode, efx_interrupt_mode, uint, 0444);
49MODULE_PARM_DESC(interrupt_mode,
50 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
51
52module_param(rss_cpus, uint, 0444);
53MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
54
55/*
56 * Use separate channels for TX and RX events
57 *
58 * Set this to 1 to use separate channels for TX and RX. It allows us
59 * to control interrupt affinity separately for TX and RX.
60 *
61 * This is only used in MSI-X interrupt mode
62 */
63bool efx_separate_tx_channels;
64module_param(efx_separate_tx_channels, bool, 0444);
65MODULE_PARM_DESC(efx_separate_tx_channels,
66 "Use separate channels for TX and RX");
67
68/* Initial interrupt moderation settings. They can be modified after
69 * module load with ethtool.
70 *
71 * The default for RX should strike a balance between increasing the
72 * round-trip latency and reducing overhead.
73 */
74static unsigned int rx_irq_mod_usec = 60;
75
76/* Initial interrupt moderation settings. They can be modified after
77 * module load with ethtool.
78 *
79 * This default is chosen to ensure that a 10G link does not go idle
80 * while a TX queue is stopped after it has become full. A queue is
81 * restarted when it drops below half full. The time this takes (assuming
82 * worst case 3 descriptors per packet and 1024 descriptors) is
83 * 512 / 3 * 1.2 = 205 usec.
84 */
85static unsigned int tx_irq_mod_usec = 150;
86
87static bool phy_flash_cfg;
88module_param(phy_flash_cfg, bool, 0644);
89MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
90
91static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
92 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
93 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
94 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
95module_param(debug, uint, 0);
96MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
97
98/**************************************************************************
99 *
100 * Utility functions and prototypes
101 *
102 *************************************************************************/
103
104static void efx_remove_port(struct efx_nic *efx);
105static int efx_xdp_setup_prog(struct efx_nic *efx, struct bpf_prog *prog);
106static int efx_xdp(struct net_device *dev, struct netdev_bpf *xdp);
107static int efx_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **xdpfs,
108 u32 flags);
109
110/**************************************************************************
111 *
112 * Port handling
113 *
114 **************************************************************************/
115
116static void efx_fini_port(struct efx_nic *efx);
117
118static int efx_probe_port(struct efx_nic *efx)
119{
120 int rc;
121
122 netif_dbg(efx, probe, efx->net_dev, "create port\n");
123
124 if (phy_flash_cfg)
125 efx->phy_mode = PHY_MODE_SPECIAL;
126
127 /* Connect up MAC/PHY operations table */
128 rc = efx->type->probe_port(efx);
129 if (rc)
130 return rc;
131
132 /* Initialise MAC address to permanent address */
133 eth_hw_addr_set(efx->net_dev, efx->net_dev->perm_addr);
134
135 return 0;
136}
137
138static int efx_init_port(struct efx_nic *efx)
139{
140 int rc;
141
142 netif_dbg(efx, drv, efx->net_dev, "init port\n");
143
144 mutex_lock(&efx->mac_lock);
145
146 efx->port_initialized = true;
147
148 /* Ensure the PHY advertises the correct flow control settings */
149 rc = efx_mcdi_port_reconfigure(efx);
150 if (rc && rc != -EPERM)
151 goto fail;
152
153 mutex_unlock(&efx->mac_lock);
154 return 0;
155
156fail:
157 mutex_unlock(&efx->mac_lock);
158 return rc;
159}
160
161static void efx_fini_port(struct efx_nic *efx)
162{
163 netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
164
165 if (!efx->port_initialized)
166 return;
167
168 efx->port_initialized = false;
169
170 efx->link_state.up = false;
171 efx_link_status_changed(efx);
172}
173
174static void efx_remove_port(struct efx_nic *efx)
175{
176 netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
177
178 efx->type->remove_port(efx);
179}
180
181/**************************************************************************
182 *
183 * NIC handling
184 *
185 **************************************************************************/
186
187static LIST_HEAD(efx_primary_list);
188static LIST_HEAD(efx_unassociated_list);
189
190static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
191{
192 return left->type == right->type &&
193 left->vpd_sn && right->vpd_sn &&
194 !strcmp(left->vpd_sn, right->vpd_sn);
195}
196
197static void efx_associate(struct efx_nic *efx)
198{
199 struct efx_nic *other, *next;
200
201 if (efx->primary == efx) {
202 /* Adding primary function; look for secondaries */
203
204 netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
205 list_add_tail(&efx->node, &efx_primary_list);
206
207 list_for_each_entry_safe(other, next, &efx_unassociated_list,
208 node) {
209 if (efx_same_controller(efx, other)) {
210 list_del(&other->node);
211 netif_dbg(other, probe, other->net_dev,
212 "moving to secondary list of %s %s\n",
213 pci_name(efx->pci_dev),
214 efx->net_dev->name);
215 list_add_tail(&other->node,
216 &efx->secondary_list);
217 other->primary = efx;
218 }
219 }
220 } else {
221 /* Adding secondary function; look for primary */
222
223 list_for_each_entry(other, &efx_primary_list, node) {
224 if (efx_same_controller(efx, other)) {
225 netif_dbg(efx, probe, efx->net_dev,
226 "adding to secondary list of %s %s\n",
227 pci_name(other->pci_dev),
228 other->net_dev->name);
229 list_add_tail(&efx->node,
230 &other->secondary_list);
231 efx->primary = other;
232 return;
233 }
234 }
235
236 netif_dbg(efx, probe, efx->net_dev,
237 "adding to unassociated list\n");
238 list_add_tail(&efx->node, &efx_unassociated_list);
239 }
240}
241
242static void efx_dissociate(struct efx_nic *efx)
243{
244 struct efx_nic *other, *next;
245
246 list_del(&efx->node);
247 efx->primary = NULL;
248
249 list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
250 list_del(&other->node);
251 netif_dbg(other, probe, other->net_dev,
252 "moving to unassociated list\n");
253 list_add_tail(&other->node, &efx_unassociated_list);
254 other->primary = NULL;
255 }
256}
257
258static int efx_probe_nic(struct efx_nic *efx)
259{
260 int rc;
261
262 netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
263
264 /* Carry out hardware-type specific initialisation */
265 rc = efx->type->probe(efx);
266 if (rc)
267 return rc;
268
269 do {
270 if (!efx->max_channels || !efx->max_tx_channels) {
271 netif_err(efx, drv, efx->net_dev,
272 "Insufficient resources to allocate"
273 " any channels\n");
274 rc = -ENOSPC;
275 goto fail1;
276 }
277
278 /* Determine the number of channels and queues by trying
279 * to hook in MSI-X interrupts.
280 */
281 rc = efx_probe_interrupts(efx);
282 if (rc)
283 goto fail1;
284
285 rc = efx_set_channels(efx);
286 if (rc)
287 goto fail1;
288
289 /* dimension_resources can fail with EAGAIN */
290 rc = efx->type->dimension_resources(efx);
291 if (rc != 0 && rc != -EAGAIN)
292 goto fail2;
293
294 if (rc == -EAGAIN)
295 /* try again with new max_channels */
296 efx_remove_interrupts(efx);
297
298 } while (rc == -EAGAIN);
299
300 if (efx->n_channels > 1)
301 netdev_rss_key_fill(efx->rss_context.rx_hash_key,
302 sizeof(efx->rss_context.rx_hash_key));
303 efx_set_default_rx_indir_table(efx, efx->rss_context.rx_indir_table);
304
305 /* Initialise the interrupt moderation settings */
306 efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
307 efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
308 true);
309
310 return 0;
311
312fail2:
313 efx_remove_interrupts(efx);
314fail1:
315 efx->type->remove(efx);
316 return rc;
317}
318
319static void efx_remove_nic(struct efx_nic *efx)
320{
321 netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
322
323 efx_remove_interrupts(efx);
324 efx->type->remove(efx);
325}
326
327/**************************************************************************
328 *
329 * NIC startup/shutdown
330 *
331 *************************************************************************/
332
333static int efx_probe_all(struct efx_nic *efx)
334{
335 int rc;
336
337 rc = efx_probe_nic(efx);
338 if (rc) {
339 netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
340 goto fail1;
341 }
342
343 rc = efx_probe_port(efx);
344 if (rc) {
345 netif_err(efx, probe, efx->net_dev, "failed to create port\n");
346 goto fail2;
347 }
348
349 BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
350 if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
351 rc = -EINVAL;
352 goto fail3;
353 }
354
355#ifdef CONFIG_SFC_SRIOV
356 rc = efx->type->vswitching_probe(efx);
357 if (rc) /* not fatal; the PF will still work fine */
358 netif_warn(efx, probe, efx->net_dev,
359 "failed to setup vswitching rc=%d;"
360 " VFs may not function\n", rc);
361#endif
362
363 rc = efx_probe_filters(efx);
364 if (rc) {
365 netif_err(efx, probe, efx->net_dev,
366 "failed to create filter tables\n");
367 goto fail4;
368 }
369
370 rc = efx_probe_channels(efx);
371 if (rc)
372 goto fail5;
373
374 efx->state = STATE_NET_DOWN;
375
376 return 0;
377
378 fail5:
379 efx_remove_filters(efx);
380 fail4:
381#ifdef CONFIG_SFC_SRIOV
382 efx->type->vswitching_remove(efx);
383#endif
384 fail3:
385 efx_remove_port(efx);
386 fail2:
387 efx_remove_nic(efx);
388 fail1:
389 return rc;
390}
391
392static void efx_remove_all(struct efx_nic *efx)
393{
394 rtnl_lock();
395 efx_xdp_setup_prog(efx, NULL);
396 rtnl_unlock();
397
398 efx_remove_channels(efx);
399 efx_remove_filters(efx);
400#ifdef CONFIG_SFC_SRIOV
401 efx->type->vswitching_remove(efx);
402#endif
403 efx_remove_port(efx);
404 efx_remove_nic(efx);
405}
406
407/**************************************************************************
408 *
409 * Interrupt moderation
410 *
411 **************************************************************************/
412unsigned int efx_usecs_to_ticks(struct efx_nic *efx, unsigned int usecs)
413{
414 if (usecs == 0)
415 return 0;
416 if (usecs * 1000 < efx->timer_quantum_ns)
417 return 1; /* never round down to 0 */
418 return usecs * 1000 / efx->timer_quantum_ns;
419}
420
421/* Set interrupt moderation parameters */
422int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
423 unsigned int rx_usecs, bool rx_adaptive,
424 bool rx_may_override_tx)
425{
426 struct efx_channel *channel;
427 unsigned int timer_max_us;
428
429 EFX_ASSERT_RESET_SERIALISED(efx);
430
431 timer_max_us = efx->timer_max_ns / 1000;
432
433 if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
434 return -EINVAL;
435
436 if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
437 !rx_may_override_tx) {
438 netif_err(efx, drv, efx->net_dev, "Channels are shared. "
439 "RX and TX IRQ moderation must be equal\n");
440 return -EINVAL;
441 }
442
443 efx->irq_rx_adaptive = rx_adaptive;
444 efx->irq_rx_moderation_us = rx_usecs;
445 efx_for_each_channel(channel, efx) {
446 if (efx_channel_has_rx_queue(channel))
447 channel->irq_moderation_us = rx_usecs;
448 else if (efx_channel_has_tx_queues(channel))
449 channel->irq_moderation_us = tx_usecs;
450 else if (efx_channel_is_xdp_tx(channel))
451 channel->irq_moderation_us = tx_usecs;
452 }
453
454 return 0;
455}
456
457void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
458 unsigned int *rx_usecs, bool *rx_adaptive)
459{
460 *rx_adaptive = efx->irq_rx_adaptive;
461 *rx_usecs = efx->irq_rx_moderation_us;
462
463 /* If channels are shared between RX and TX, so is IRQ
464 * moderation. Otherwise, IRQ moderation is the same for all
465 * TX channels and is not adaptive.
466 */
467 if (efx->tx_channel_offset == 0) {
468 *tx_usecs = *rx_usecs;
469 } else {
470 struct efx_channel *tx_channel;
471
472 tx_channel = efx->channel[efx->tx_channel_offset];
473 *tx_usecs = tx_channel->irq_moderation_us;
474 }
475}
476
477/**************************************************************************
478 *
479 * ioctls
480 *
481 *************************************************************************/
482
483/* Net device ioctl
484 * Context: process, rtnl_lock() held.
485 */
486static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
487{
488 struct efx_nic *efx = efx_netdev_priv(net_dev);
489 struct mii_ioctl_data *data = if_mii(ifr);
490
491 /* Convert phy_id from older PRTAD/DEVAD format */
492 if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
493 (data->phy_id & 0xfc00) == 0x0400)
494 data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
495
496 return mdio_mii_ioctl(&efx->mdio, data, cmd);
497}
498
499/**************************************************************************
500 *
501 * Kernel net device interface
502 *
503 *************************************************************************/
504
505/* Context: process, rtnl_lock() held. */
506int efx_net_open(struct net_device *net_dev)
507{
508 struct efx_nic *efx = efx_netdev_priv(net_dev);
509 int rc;
510
511 netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
512 raw_smp_processor_id());
513
514 rc = efx_check_disabled(efx);
515 if (rc)
516 return rc;
517 if (efx->phy_mode & PHY_MODE_SPECIAL)
518 return -EBUSY;
519 if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
520 return -EIO;
521
522 /* Notify the kernel of the link state polled during driver load,
523 * before the monitor starts running */
524 efx_link_status_changed(efx);
525
526 efx_start_all(efx);
527 if (efx->state == STATE_DISABLED || efx->reset_pending)
528 netif_device_detach(efx->net_dev);
529 else
530 efx->state = STATE_NET_UP;
531
532 return 0;
533}
534
535/* Context: process, rtnl_lock() held.
536 * Note that the kernel will ignore our return code; this method
537 * should really be a void.
538 */
539int efx_net_stop(struct net_device *net_dev)
540{
541 struct efx_nic *efx = efx_netdev_priv(net_dev);
542
543 netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
544 raw_smp_processor_id());
545
546 /* Stop the device and flush all the channels */
547 efx_stop_all(efx);
548
549 return 0;
550}
551
552static int efx_vlan_rx_add_vid(struct net_device *net_dev, __be16 proto, u16 vid)
553{
554 struct efx_nic *efx = efx_netdev_priv(net_dev);
555
556 if (efx->type->vlan_rx_add_vid)
557 return efx->type->vlan_rx_add_vid(efx, proto, vid);
558 else
559 return -EOPNOTSUPP;
560}
561
562static int efx_vlan_rx_kill_vid(struct net_device *net_dev, __be16 proto, u16 vid)
563{
564 struct efx_nic *efx = efx_netdev_priv(net_dev);
565
566 if (efx->type->vlan_rx_kill_vid)
567 return efx->type->vlan_rx_kill_vid(efx, proto, vid);
568 else
569 return -EOPNOTSUPP;
570}
571
572static int efx_hwtstamp_set(struct net_device *net_dev,
573 struct kernel_hwtstamp_config *config,
574 struct netlink_ext_ack *extack)
575{
576 struct efx_nic *efx = efx_netdev_priv(net_dev);
577
578 return efx_ptp_set_ts_config(efx, config, extack);
579}
580
581static int efx_hwtstamp_get(struct net_device *net_dev,
582 struct kernel_hwtstamp_config *config)
583{
584 struct efx_nic *efx = efx_netdev_priv(net_dev);
585
586 return efx_ptp_get_ts_config(efx, config);
587}
588
589static const struct net_device_ops efx_netdev_ops = {
590 .ndo_open = efx_net_open,
591 .ndo_stop = efx_net_stop,
592 .ndo_get_stats64 = efx_net_stats,
593 .ndo_tx_timeout = efx_watchdog,
594 .ndo_start_xmit = efx_hard_start_xmit,
595 .ndo_validate_addr = eth_validate_addr,
596 .ndo_eth_ioctl = efx_ioctl,
597 .ndo_change_mtu = efx_change_mtu,
598 .ndo_set_mac_address = efx_set_mac_address,
599 .ndo_set_rx_mode = efx_set_rx_mode,
600 .ndo_set_features = efx_set_features,
601 .ndo_features_check = efx_features_check,
602 .ndo_vlan_rx_add_vid = efx_vlan_rx_add_vid,
603 .ndo_vlan_rx_kill_vid = efx_vlan_rx_kill_vid,
604 .ndo_hwtstamp_set = efx_hwtstamp_set,
605 .ndo_hwtstamp_get = efx_hwtstamp_get,
606#ifdef CONFIG_SFC_SRIOV
607 .ndo_set_vf_mac = efx_sriov_set_vf_mac,
608 .ndo_set_vf_vlan = efx_sriov_set_vf_vlan,
609 .ndo_set_vf_spoofchk = efx_sriov_set_vf_spoofchk,
610 .ndo_get_vf_config = efx_sriov_get_vf_config,
611 .ndo_set_vf_link_state = efx_sriov_set_vf_link_state,
612#endif
613 .ndo_get_phys_port_id = efx_get_phys_port_id,
614 .ndo_get_phys_port_name = efx_get_phys_port_name,
615#ifdef CONFIG_RFS_ACCEL
616 .ndo_rx_flow_steer = efx_filter_rfs,
617#endif
618 .ndo_xdp_xmit = efx_xdp_xmit,
619 .ndo_bpf = efx_xdp
620};
621
622static void efx_get_queue_stats_rx(struct net_device *net_dev, int idx,
623 struct netdev_queue_stats_rx *stats)
624{
625 struct efx_nic *efx = efx_netdev_priv(net_dev);
626 struct efx_rx_queue *rx_queue;
627 struct efx_channel *channel;
628
629 channel = efx_get_channel(efx, idx);
630 rx_queue = efx_channel_get_rx_queue(channel);
631 /* Count only packets since last time datapath was started */
632 stats->packets = rx_queue->rx_packets - rx_queue->old_rx_packets;
633 stats->bytes = rx_queue->rx_bytes - rx_queue->old_rx_bytes;
634 stats->hw_drops = efx_get_queue_stat_rx_hw_drops(channel) -
635 channel->old_n_rx_hw_drops;
636 stats->hw_drop_overruns = channel->n_rx_nodesc_trunc -
637 channel->old_n_rx_hw_drop_overruns;
638}
639
640static void efx_get_queue_stats_tx(struct net_device *net_dev, int idx,
641 struct netdev_queue_stats_tx *stats)
642{
643 struct efx_nic *efx = efx_netdev_priv(net_dev);
644 struct efx_tx_queue *tx_queue;
645 struct efx_channel *channel;
646
647 channel = efx_get_tx_channel(efx, idx);
648 stats->packets = 0;
649 stats->bytes = 0;
650 stats->hw_gso_packets = 0;
651 stats->hw_gso_wire_packets = 0;
652 efx_for_each_channel_tx_queue(tx_queue, channel) {
653 stats->packets += tx_queue->complete_packets -
654 tx_queue->old_complete_packets;
655 stats->bytes += tx_queue->complete_bytes -
656 tx_queue->old_complete_bytes;
657 /* Note that, unlike stats->packets and stats->bytes,
658 * these count TXes enqueued, rather than completed,
659 * which may not be what users expect.
660 */
661 stats->hw_gso_packets += tx_queue->tso_bursts -
662 tx_queue->old_tso_bursts;
663 stats->hw_gso_wire_packets += tx_queue->tso_packets -
664 tx_queue->old_tso_packets;
665 }
666}
667
668static void efx_get_base_stats(struct net_device *net_dev,
669 struct netdev_queue_stats_rx *rx,
670 struct netdev_queue_stats_tx *tx)
671{
672 struct efx_nic *efx = efx_netdev_priv(net_dev);
673 struct efx_tx_queue *tx_queue;
674 struct efx_rx_queue *rx_queue;
675 struct efx_channel *channel;
676
677 rx->packets = 0;
678 rx->bytes = 0;
679 rx->hw_drops = 0;
680 rx->hw_drop_overruns = 0;
681 tx->packets = 0;
682 tx->bytes = 0;
683 tx->hw_gso_packets = 0;
684 tx->hw_gso_wire_packets = 0;
685
686 /* Count all packets on non-core queues, and packets before last
687 * datapath start on core queues.
688 */
689 efx_for_each_channel(channel, efx) {
690 rx_queue = efx_channel_get_rx_queue(channel);
691 if (channel->channel >= net_dev->real_num_rx_queues) {
692 rx->packets += rx_queue->rx_packets;
693 rx->bytes += rx_queue->rx_bytes;
694 rx->hw_drops += efx_get_queue_stat_rx_hw_drops(channel);
695 rx->hw_drop_overruns += channel->n_rx_nodesc_trunc;
696 } else {
697 rx->packets += rx_queue->old_rx_packets;
698 rx->bytes += rx_queue->old_rx_bytes;
699 rx->hw_drops += channel->old_n_rx_hw_drops;
700 rx->hw_drop_overruns += channel->old_n_rx_hw_drop_overruns;
701 }
702 efx_for_each_channel_tx_queue(tx_queue, channel) {
703 if (channel->channel < efx->tx_channel_offset ||
704 channel->channel >= efx->tx_channel_offset +
705 net_dev->real_num_tx_queues) {
706 tx->packets += tx_queue->complete_packets;
707 tx->bytes += tx_queue->complete_bytes;
708 tx->hw_gso_packets += tx_queue->tso_bursts;
709 tx->hw_gso_wire_packets += tx_queue->tso_packets;
710 } else {
711 tx->packets += tx_queue->old_complete_packets;
712 tx->bytes += tx_queue->old_complete_bytes;
713 tx->hw_gso_packets += tx_queue->old_tso_bursts;
714 tx->hw_gso_wire_packets += tx_queue->old_tso_packets;
715 }
716 /* Include XDP TX in device-wide stats */
717 tx->packets += tx_queue->complete_xdp_packets;
718 tx->bytes += tx_queue->complete_xdp_bytes;
719 }
720 }
721}
722
723static const struct netdev_stat_ops efx_stat_ops = {
724 .get_queue_stats_rx = efx_get_queue_stats_rx,
725 .get_queue_stats_tx = efx_get_queue_stats_tx,
726 .get_base_stats = efx_get_base_stats,
727};
728
729static int efx_xdp_setup_prog(struct efx_nic *efx, struct bpf_prog *prog)
730{
731 struct bpf_prog *old_prog;
732
733 if (efx->xdp_rxq_info_failed) {
734 netif_err(efx, drv, efx->net_dev,
735 "Unable to bind XDP program due to previous failure of rxq_info\n");
736 return -EINVAL;
737 }
738
739 if (prog && efx->net_dev->mtu > efx_xdp_max_mtu(efx)) {
740 netif_err(efx, drv, efx->net_dev,
741 "Unable to configure XDP with MTU of %d (max: %d)\n",
742 efx->net_dev->mtu, efx_xdp_max_mtu(efx));
743 return -EINVAL;
744 }
745
746 old_prog = rtnl_dereference(efx->xdp_prog);
747 rcu_assign_pointer(efx->xdp_prog, prog);
748 /* Release the reference that was originally passed by the caller. */
749 if (old_prog)
750 bpf_prog_put(old_prog);
751
752 return 0;
753}
754
755/* Context: process, rtnl_lock() held. */
756static int efx_xdp(struct net_device *dev, struct netdev_bpf *xdp)
757{
758 struct efx_nic *efx = efx_netdev_priv(dev);
759
760 switch (xdp->command) {
761 case XDP_SETUP_PROG:
762 return efx_xdp_setup_prog(efx, xdp->prog);
763 default:
764 return -EINVAL;
765 }
766}
767
768static int efx_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **xdpfs,
769 u32 flags)
770{
771 struct efx_nic *efx = efx_netdev_priv(dev);
772
773 if (!netif_running(dev))
774 return -EINVAL;
775
776 return efx_xdp_tx_buffers(efx, n, xdpfs, flags & XDP_XMIT_FLUSH);
777}
778
779static void efx_update_name(struct efx_nic *efx)
780{
781 strcpy(efx->name, efx->net_dev->name);
782 efx_mtd_rename(efx);
783 efx_set_channel_names(efx);
784}
785
786static int efx_netdev_event(struct notifier_block *this,
787 unsigned long event, void *ptr)
788{
789 struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
790
791 if ((net_dev->netdev_ops == &efx_netdev_ops) &&
792 event == NETDEV_CHANGENAME)
793 efx_update_name(efx_netdev_priv(net_dev));
794
795 return NOTIFY_DONE;
796}
797
798static struct notifier_block efx_netdev_notifier = {
799 .notifier_call = efx_netdev_event,
800};
801
802static ssize_t phy_type_show(struct device *dev,
803 struct device_attribute *attr, char *buf)
804{
805 struct efx_nic *efx = dev_get_drvdata(dev);
806 return sprintf(buf, "%d\n", efx->phy_type);
807}
808static DEVICE_ATTR_RO(phy_type);
809
810static int efx_register_netdev(struct efx_nic *efx)
811{
812 struct net_device *net_dev = efx->net_dev;
813 struct efx_channel *channel;
814 int rc;
815
816 net_dev->watchdog_timeo = 5 * HZ;
817 net_dev->irq = efx->pci_dev->irq;
818 net_dev->netdev_ops = &efx_netdev_ops;
819 net_dev->stat_ops = &efx_stat_ops;
820 if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
821 net_dev->priv_flags |= IFF_UNICAST_FLT;
822 net_dev->ethtool_ops = &efx_ethtool_ops;
823 netif_set_tso_max_segs(net_dev, EFX_TSO_MAX_SEGS);
824 net_dev->min_mtu = EFX_MIN_MTU;
825 net_dev->max_mtu = EFX_MAX_MTU;
826
827 rtnl_lock();
828
829 /* Enable resets to be scheduled and check whether any were
830 * already requested. If so, the NIC is probably hosed so we
831 * abort.
832 */
833 if (efx->reset_pending) {
834 pci_err(efx->pci_dev, "aborting probe due to scheduled reset\n");
835 rc = -EIO;
836 goto fail_locked;
837 }
838
839 rc = dev_alloc_name(net_dev, net_dev->name);
840 if (rc < 0)
841 goto fail_locked;
842 efx_update_name(efx);
843
844 /* Always start with carrier off; PHY events will detect the link */
845 netif_carrier_off(net_dev);
846
847 rc = register_netdevice(net_dev);
848 if (rc)
849 goto fail_locked;
850
851 efx_for_each_channel(channel, efx) {
852 struct efx_tx_queue *tx_queue;
853 efx_for_each_channel_tx_queue(tx_queue, channel)
854 efx_init_tx_queue_core_txq(tx_queue);
855 }
856
857 efx_associate(efx);
858
859 efx->state = STATE_NET_DOWN;
860
861 rtnl_unlock();
862
863 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
864 if (rc) {
865 netif_err(efx, drv, efx->net_dev,
866 "failed to init net dev attributes\n");
867 goto fail_registered;
868 }
869
870 efx_init_mcdi_logging(efx);
871
872 return 0;
873
874fail_registered:
875 rtnl_lock();
876 efx_dissociate(efx);
877 unregister_netdevice(net_dev);
878fail_locked:
879 efx->state = STATE_UNINIT;
880 rtnl_unlock();
881 netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
882 return rc;
883}
884
885static void efx_unregister_netdev(struct efx_nic *efx)
886{
887 if (!efx->net_dev)
888 return;
889
890 if (WARN_ON(efx_netdev_priv(efx->net_dev) != efx))
891 return;
892
893 if (efx_dev_registered(efx)) {
894 strscpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
895 efx_fini_mcdi_logging(efx);
896 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
897 unregister_netdev(efx->net_dev);
898 }
899}
900
901/**************************************************************************
902 *
903 * List of NICs we support
904 *
905 **************************************************************************/
906
907/* PCI device ID table */
908static const struct pci_device_id efx_pci_table[] = {
909 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903), /* SFC9120 PF */
910 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
911 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1903), /* SFC9120 VF */
912 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
913 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0923), /* SFC9140 PF */
914 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
915 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1923), /* SFC9140 VF */
916 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
917 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0a03), /* SFC9220 PF */
918 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
919 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1a03), /* SFC9220 VF */
920 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
921 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0b03), /* SFC9250 PF */
922 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
923 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x1b03), /* SFC9250 VF */
924 .driver_data = (unsigned long) &efx_hunt_a0_vf_nic_type},
925 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0c03), /* X4 PF (FF/LL) */
926 .driver_data = (unsigned long)&efx_x4_nic_type},
927 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x2c03), /* X4 PF (FF only) */
928 .driver_data = (unsigned long)&efx_x4_nic_type},
929 {0} /* end of list */
930};
931
932/**************************************************************************
933 *
934 * Data housekeeping
935 *
936 **************************************************************************/
937
938void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
939{
940 u64 n_rx_nodesc_trunc = 0;
941 struct efx_channel *channel;
942
943 efx_for_each_channel(channel, efx)
944 n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
945 stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
946 stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
947}
948
949/**************************************************************************
950 *
951 * PCI interface
952 *
953 **************************************************************************/
954
955/* Main body of final NIC shutdown code
956 * This is called only at module unload (or hotplug removal).
957 */
958static void efx_pci_remove_main(struct efx_nic *efx)
959{
960 /* Flush reset_work. It can no longer be scheduled since we
961 * are not READY.
962 */
963 WARN_ON(efx_net_active(efx->state));
964 efx_flush_reset_workqueue(efx);
965
966 efx_disable_interrupts(efx);
967 efx_clear_interrupt_affinity(efx);
968 efx_nic_fini_interrupt(efx);
969 efx_fini_port(efx);
970 efx->type->fini(efx);
971 efx_fini_napi(efx);
972 efx_remove_all(efx);
973}
974
975/* Final NIC shutdown
976 * This is called only at module unload (or hotplug removal). A PF can call
977 * this on its VFs to ensure they are unbound first.
978 */
979static void efx_pci_remove(struct pci_dev *pci_dev)
980{
981 struct efx_probe_data *probe_data;
982 struct efx_nic *efx;
983
984 efx = pci_get_drvdata(pci_dev);
985 if (!efx)
986 return;
987
988 /* Mark the NIC as fini, then stop the interface */
989 rtnl_lock();
990 efx_dissociate(efx);
991 dev_close(efx->net_dev);
992 efx_disable_interrupts(efx);
993 efx->state = STATE_UNINIT;
994 rtnl_unlock();
995
996 if (efx->type->sriov_fini)
997 efx->type->sriov_fini(efx);
998
999 efx_fini_devlink_lock(efx);
1000 efx_unregister_netdev(efx);
1001
1002 efx_mtd_remove(efx);
1003
1004 efx_pci_remove_main(efx);
1005
1006 efx_fini_io(efx);
1007 pci_dbg(efx->pci_dev, "shutdown successful\n");
1008
1009 efx_fini_devlink_and_unlock(efx);
1010 efx_fini_struct(efx);
1011 free_netdev(efx->net_dev);
1012 probe_data = container_of(efx, struct efx_probe_data, efx);
1013 kfree(probe_data);
1014};
1015
1016/* NIC VPD information
1017 * Called during probe to display the part number of the
1018 * installed NIC.
1019 */
1020static void efx_probe_vpd_strings(struct efx_nic *efx)
1021{
1022 struct pci_dev *dev = efx->pci_dev;
1023 unsigned int vpd_size, kw_len;
1024 u8 *vpd_data;
1025 int start;
1026
1027 vpd_data = pci_vpd_alloc(dev, &vpd_size);
1028 if (IS_ERR(vpd_data)) {
1029 pci_warn(dev, "Unable to read VPD\n");
1030 return;
1031 }
1032
1033 start = pci_vpd_find_ro_info_keyword(vpd_data, vpd_size,
1034 PCI_VPD_RO_KEYWORD_PARTNO, &kw_len);
1035 if (start < 0)
1036 pci_err(dev, "Part number not found or incomplete\n");
1037 else
1038 pci_info(dev, "Part Number : %.*s\n", kw_len, vpd_data + start);
1039
1040 start = pci_vpd_find_ro_info_keyword(vpd_data, vpd_size,
1041 PCI_VPD_RO_KEYWORD_SERIALNO, &kw_len);
1042 if (start < 0)
1043 pci_err(dev, "Serial number not found or incomplete\n");
1044 else
1045 efx->vpd_sn = kmemdup_nul(vpd_data + start, kw_len, GFP_KERNEL);
1046
1047 kfree(vpd_data);
1048}
1049
1050
1051/* Main body of NIC initialisation
1052 * This is called at module load (or hotplug insertion, theoretically).
1053 */
1054static int efx_pci_probe_main(struct efx_nic *efx)
1055{
1056 int rc;
1057
1058 /* Do start-of-day initialisation */
1059 rc = efx_probe_all(efx);
1060 if (rc)
1061 goto fail1;
1062
1063 efx_init_napi(efx);
1064
1065 down_write(&efx->filter_sem);
1066 rc = efx->type->init(efx);
1067 up_write(&efx->filter_sem);
1068 if (rc) {
1069 pci_err(efx->pci_dev, "failed to initialise NIC\n");
1070 goto fail3;
1071 }
1072
1073 rc = efx_init_port(efx);
1074 if (rc) {
1075 netif_err(efx, probe, efx->net_dev,
1076 "failed to initialise port\n");
1077 goto fail4;
1078 }
1079
1080 rc = efx_nic_init_interrupt(efx);
1081 if (rc)
1082 goto fail5;
1083
1084 efx_set_interrupt_affinity(efx);
1085 rc = efx_enable_interrupts(efx);
1086 if (rc)
1087 goto fail6;
1088
1089 return 0;
1090
1091 fail6:
1092 efx_clear_interrupt_affinity(efx);
1093 efx_nic_fini_interrupt(efx);
1094 fail5:
1095 efx_fini_port(efx);
1096 fail4:
1097 efx->type->fini(efx);
1098 fail3:
1099 efx_fini_napi(efx);
1100 efx_remove_all(efx);
1101 fail1:
1102 return rc;
1103}
1104
1105static int efx_pci_probe_post_io(struct efx_nic *efx)
1106{
1107 struct net_device *net_dev = efx->net_dev;
1108 int rc = efx_pci_probe_main(efx);
1109
1110 if (rc)
1111 return rc;
1112
1113 if (efx->type->sriov_init) {
1114 rc = efx->type->sriov_init(efx);
1115 if (rc)
1116 pci_err(efx->pci_dev, "SR-IOV can't be enabled rc %d\n",
1117 rc);
1118 }
1119
1120 /* Determine netdevice features */
1121 net_dev->features |= efx->type->offload_features;
1122
1123 /* Add TSO features */
1124 if (efx->type->tso_versions && efx->type->tso_versions(efx))
1125 net_dev->features |= NETIF_F_TSO | NETIF_F_TSO6;
1126
1127 /* Mask for features that also apply to VLAN devices */
1128 net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
1129 NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
1130 NETIF_F_RXCSUM);
1131
1132 /* Determine user configurable features */
1133 net_dev->hw_features |= net_dev->features & ~efx->fixed_features;
1134
1135 /* Disable receiving frames with bad FCS, by default. */
1136 net_dev->features &= ~NETIF_F_RXALL;
1137
1138 /* Disable VLAN filtering by default. It may be enforced if
1139 * the feature is fixed (i.e. VLAN filters are required to
1140 * receive VLAN tagged packets due to vPort restrictions).
1141 */
1142 net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
1143 net_dev->features |= efx->fixed_features;
1144
1145 net_dev->xdp_features = NETDEV_XDP_ACT_BASIC |
1146 NETDEV_XDP_ACT_REDIRECT |
1147 NETDEV_XDP_ACT_NDO_XMIT;
1148
1149 /* devlink creation, registration and lock */
1150 rc = efx_probe_devlink_and_lock(efx);
1151 if (rc)
1152 pci_err(efx->pci_dev, "devlink registration failed");
1153
1154 rc = efx_register_netdev(efx);
1155 efx_probe_devlink_unlock(efx);
1156 if (!rc)
1157 return 0;
1158
1159 efx_pci_remove_main(efx);
1160 return rc;
1161}
1162
1163/* NIC initialisation
1164 *
1165 * This is called at module load (or hotplug insertion,
1166 * theoretically). It sets up PCI mappings, resets the NIC,
1167 * sets up and registers the network devices with the kernel and hooks
1168 * the interrupt service routine. It does not prepare the device for
1169 * transmission; this is left to the first time one of the network
1170 * interfaces is brought up (i.e. efx_net_open).
1171 */
1172static int efx_pci_probe(struct pci_dev *pci_dev,
1173 const struct pci_device_id *entry)
1174{
1175 struct efx_probe_data *probe_data, **probe_ptr;
1176 struct net_device *net_dev;
1177 struct efx_nic *efx;
1178 int rc;
1179
1180 /* Allocate probe data and struct efx_nic */
1181 probe_data = kzalloc(sizeof(*probe_data), GFP_KERNEL);
1182 if (!probe_data)
1183 return -ENOMEM;
1184 probe_data->pci_dev = pci_dev;
1185 efx = &probe_data->efx;
1186
1187 /* Allocate and initialise a struct net_device */
1188 net_dev = alloc_etherdev_mq(sizeof(probe_data), EFX_MAX_CORE_TX_QUEUES);
1189 if (!net_dev) {
1190 rc = -ENOMEM;
1191 goto fail0;
1192 }
1193 probe_ptr = netdev_priv(net_dev);
1194 *probe_ptr = probe_data;
1195 efx->net_dev = net_dev;
1196 efx->type = (const struct efx_nic_type *) entry->driver_data;
1197 efx->fixed_features |= NETIF_F_HIGHDMA;
1198
1199 pci_set_drvdata(pci_dev, efx);
1200 SET_NETDEV_DEV(net_dev, &pci_dev->dev);
1201 rc = efx_init_struct(efx, pci_dev);
1202 if (rc)
1203 goto fail1;
1204 efx->mdio.dev = net_dev;
1205
1206 pci_info(pci_dev, "Solarflare NIC detected\n");
1207
1208 if (!efx->type->is_vf)
1209 efx_probe_vpd_strings(efx);
1210
1211 /* Set up basic I/O (BAR mappings etc) */
1212 rc = efx_init_io(efx, efx->type->mem_bar(efx), efx->type->max_dma_mask,
1213 efx->type->mem_map_size(efx));
1214 if (rc)
1215 goto fail2;
1216
1217 rc = efx_pci_probe_post_io(efx);
1218 if (rc) {
1219 /* On failure, retry once immediately.
1220 * If we aborted probe due to a scheduled reset, dismiss it.
1221 */
1222 efx->reset_pending = 0;
1223 rc = efx_pci_probe_post_io(efx);
1224 if (rc) {
1225 /* On another failure, retry once more
1226 * after a 50-305ms delay.
1227 */
1228 unsigned char r;
1229
1230 get_random_bytes(&r, 1);
1231 msleep((unsigned int)r + 50);
1232 efx->reset_pending = 0;
1233 rc = efx_pci_probe_post_io(efx);
1234 }
1235 }
1236 if (rc)
1237 goto fail3;
1238
1239 netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
1240
1241 /* Try to create MTDs, but allow this to fail */
1242 rtnl_lock();
1243 rc = efx_mtd_probe(efx);
1244 rtnl_unlock();
1245 if (rc && rc != -EPERM)
1246 netif_warn(efx, probe, efx->net_dev,
1247 "failed to create MTDs (%d)\n", rc);
1248
1249 if (efx->type->udp_tnl_push_ports)
1250 efx->type->udp_tnl_push_ports(efx);
1251
1252 return 0;
1253
1254 fail3:
1255 efx_fini_io(efx);
1256 fail2:
1257 efx_fini_struct(efx);
1258 fail1:
1259 WARN_ON(rc > 0);
1260 netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
1261 free_netdev(net_dev);
1262 fail0:
1263 kfree(probe_data);
1264 return rc;
1265}
1266
1267/* efx_pci_sriov_configure returns the actual number of Virtual Functions
1268 * enabled on success
1269 */
1270#ifdef CONFIG_SFC_SRIOV
1271static int efx_pci_sriov_configure(struct pci_dev *dev, int num_vfs)
1272{
1273 int rc;
1274 struct efx_nic *efx = pci_get_drvdata(dev);
1275
1276 if (efx->type->sriov_configure) {
1277 rc = efx->type->sriov_configure(efx, num_vfs);
1278 if (rc)
1279 return rc;
1280 else
1281 return num_vfs;
1282 } else
1283 return -EOPNOTSUPP;
1284}
1285#endif
1286
1287static int efx_pm_freeze(struct device *dev)
1288{
1289 struct efx_nic *efx = dev_get_drvdata(dev);
1290
1291 rtnl_lock();
1292
1293 if (efx_net_active(efx->state)) {
1294 efx_device_detach_sync(efx);
1295
1296 efx_stop_all(efx);
1297 efx_disable_interrupts(efx);
1298
1299 efx->state = efx_freeze(efx->state);
1300 }
1301
1302 rtnl_unlock();
1303
1304 return 0;
1305}
1306
1307static void efx_pci_shutdown(struct pci_dev *pci_dev)
1308{
1309 struct efx_nic *efx = pci_get_drvdata(pci_dev);
1310
1311 if (!efx)
1312 return;
1313
1314 efx_pm_freeze(&pci_dev->dev);
1315 pci_disable_device(pci_dev);
1316}
1317
1318static int efx_pm_thaw(struct device *dev)
1319{
1320 int rc;
1321 struct efx_nic *efx = dev_get_drvdata(dev);
1322
1323 rtnl_lock();
1324
1325 if (efx_frozen(efx->state)) {
1326 rc = efx_enable_interrupts(efx);
1327 if (rc)
1328 goto fail;
1329
1330 mutex_lock(&efx->mac_lock);
1331 efx_mcdi_port_reconfigure(efx);
1332 mutex_unlock(&efx->mac_lock);
1333
1334 efx_start_all(efx);
1335
1336 efx_device_attach_if_not_resetting(efx);
1337
1338 efx->state = efx_thaw(efx->state);
1339
1340 efx->type->resume_wol(efx);
1341 }
1342
1343 rtnl_unlock();
1344
1345 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
1346 efx_queue_reset_work(efx);
1347
1348 return 0;
1349
1350fail:
1351 rtnl_unlock();
1352
1353 return rc;
1354}
1355
1356static int efx_pm_poweroff(struct device *dev)
1357{
1358 struct pci_dev *pci_dev = to_pci_dev(dev);
1359 struct efx_nic *efx = pci_get_drvdata(pci_dev);
1360
1361 efx->type->fini(efx);
1362
1363 efx->reset_pending = 0;
1364
1365 pci_save_state(pci_dev);
1366 return pci_set_power_state(pci_dev, PCI_D3hot);
1367}
1368
1369/* Used for both resume and restore */
1370static int efx_pm_resume(struct device *dev)
1371{
1372 struct pci_dev *pci_dev = to_pci_dev(dev);
1373 struct efx_nic *efx = pci_get_drvdata(pci_dev);
1374 int rc;
1375
1376 rc = pci_set_power_state(pci_dev, PCI_D0);
1377 if (rc)
1378 return rc;
1379 pci_restore_state(pci_dev);
1380 rc = pci_enable_device(pci_dev);
1381 if (rc)
1382 return rc;
1383 pci_set_master(efx->pci_dev);
1384 rc = efx->type->reset(efx, RESET_TYPE_ALL);
1385 if (rc)
1386 return rc;
1387 down_write(&efx->filter_sem);
1388 rc = efx->type->init(efx);
1389 up_write(&efx->filter_sem);
1390 if (rc)
1391 return rc;
1392 rc = efx_pm_thaw(dev);
1393 return rc;
1394}
1395
1396static int efx_pm_suspend(struct device *dev)
1397{
1398 int rc;
1399
1400 efx_pm_freeze(dev);
1401 rc = efx_pm_poweroff(dev);
1402 if (rc)
1403 efx_pm_resume(dev);
1404 return rc;
1405}
1406
1407static const struct dev_pm_ops efx_pm_ops = {
1408 .suspend = efx_pm_suspend,
1409 .resume = efx_pm_resume,
1410 .freeze = efx_pm_freeze,
1411 .thaw = efx_pm_thaw,
1412 .poweroff = efx_pm_poweroff,
1413 .restore = efx_pm_resume,
1414};
1415
1416static struct pci_driver efx_pci_driver = {
1417 .name = KBUILD_MODNAME,
1418 .id_table = efx_pci_table,
1419 .probe = efx_pci_probe,
1420 .remove = efx_pci_remove,
1421 .driver.pm = &efx_pm_ops,
1422 .shutdown = efx_pci_shutdown,
1423 .err_handler = &efx_err_handlers,
1424#ifdef CONFIG_SFC_SRIOV
1425 .sriov_configure = efx_pci_sriov_configure,
1426#endif
1427};
1428
1429/**************************************************************************
1430 *
1431 * Kernel module interface
1432 *
1433 *************************************************************************/
1434
1435static int __init efx_init_module(void)
1436{
1437 int rc;
1438
1439 printk(KERN_INFO "Solarflare NET driver\n");
1440
1441 rc = register_netdevice_notifier(&efx_netdev_notifier);
1442 if (rc)
1443 goto err_notifier;
1444
1445 rc = efx_create_reset_workqueue();
1446 if (rc)
1447 goto err_reset;
1448
1449 rc = pci_register_driver(&efx_pci_driver);
1450 if (rc < 0)
1451 goto err_pci;
1452
1453 rc = pci_register_driver(&ef100_pci_driver);
1454 if (rc < 0)
1455 goto err_pci_ef100;
1456
1457 return 0;
1458
1459 err_pci_ef100:
1460 pci_unregister_driver(&efx_pci_driver);
1461 err_pci:
1462 efx_destroy_reset_workqueue();
1463 err_reset:
1464 unregister_netdevice_notifier(&efx_netdev_notifier);
1465 err_notifier:
1466 return rc;
1467}
1468
1469static void __exit efx_exit_module(void)
1470{
1471 printk(KERN_INFO "Solarflare NET driver unloading\n");
1472
1473 pci_unregister_driver(&ef100_pci_driver);
1474 pci_unregister_driver(&efx_pci_driver);
1475 efx_destroy_reset_workqueue();
1476 unregister_netdevice_notifier(&efx_netdev_notifier);
1477
1478}
1479
1480module_init(efx_init_module);
1481module_exit(efx_exit_module);
1482
1483MODULE_AUTHOR("Solarflare Communications and "
1484 "Michael Brown <mbrown@fensystems.co.uk>");
1485MODULE_DESCRIPTION("Solarflare network driver");
1486MODULE_LICENSE("GPL");
1487MODULE_DEVICE_TABLE(pci, efx_pci_table);
1/****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2013 Solarflare Communications Inc.
5 *
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.
9 */
10
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>
17#include <linux/ip.h>
18#include <linux/tcp.h>
19#include <linux/in.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"
26#include "efx.h"
27#include "nic.h"
28#include "selftest.h"
29
30#include "mcdi.h"
31#include "workarounds.h"
32
33/**************************************************************************
34 *
35 * Type name strings
36 *
37 **************************************************************************
38 */
39
40/* Loopback mode names (see LOOPBACK_MODE()) */
41const unsigned int efx_loopback_mode_max = LOOPBACK_MAX;
42const char *const efx_loopback_mode_names[] = {
43 [LOOPBACK_NONE] = "NONE",
44 [LOOPBACK_DATA] = "DATAPATH",
45 [LOOPBACK_GMAC] = "GMAC",
46 [LOOPBACK_XGMII] = "XGMII",
47 [LOOPBACK_XGXS] = "XGXS",
48 [LOOPBACK_XAUI] = "XAUI",
49 [LOOPBACK_GMII] = "GMII",
50 [LOOPBACK_SGMII] = "SGMII",
51 [LOOPBACK_XGBR] = "XGBR",
52 [LOOPBACK_XFI] = "XFI",
53 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
54 [LOOPBACK_GMII_FAR] = "GMII_FAR",
55 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
56 [LOOPBACK_XFI_FAR] = "XFI_FAR",
57 [LOOPBACK_GPHY] = "GPHY",
58 [LOOPBACK_PHYXS] = "PHYXS",
59 [LOOPBACK_PCS] = "PCS",
60 [LOOPBACK_PMAPMD] = "PMA/PMD",
61 [LOOPBACK_XPORT] = "XPORT",
62 [LOOPBACK_XGMII_WS] = "XGMII_WS",
63 [LOOPBACK_XAUI_WS] = "XAUI_WS",
64 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
65 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
66 [LOOPBACK_GMII_WS] = "GMII_WS",
67 [LOOPBACK_XFI_WS] = "XFI_WS",
68 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
69 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
70};
71
72const unsigned int efx_reset_type_max = RESET_TYPE_MAX;
73const char *const efx_reset_type_names[] = {
74 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
75 [RESET_TYPE_ALL] = "ALL",
76 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
77 [RESET_TYPE_WORLD] = "WORLD",
78 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
79 [RESET_TYPE_MC_BIST] = "MC_BIST",
80 [RESET_TYPE_DISABLE] = "DISABLE",
81 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
82 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
83 [RESET_TYPE_RX_RECOVERY] = "RX_RECOVERY",
84 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
85 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
86 [RESET_TYPE_MC_FAILURE] = "MC_FAILURE",
87 [RESET_TYPE_MCDI_TIMEOUT] = "MCDI_TIMEOUT (FLR)",
88};
89
90/* Reset workqueue. If any NIC has a hardware failure then a reset will be
91 * queued onto this work queue. This is not a per-nic work queue, because
92 * efx_reset_work() acquires the rtnl lock, so resets are naturally serialised.
93 */
94static struct workqueue_struct *reset_workqueue;
95
96/* How often and how many times to poll for a reset while waiting for a
97 * BIST that another function started to complete.
98 */
99#define BIST_WAIT_DELAY_MS 100
100#define BIST_WAIT_DELAY_COUNT 100
101
102/**************************************************************************
103 *
104 * Configurable values
105 *
106 *************************************************************************/
107
108/*
109 * Use separate channels for TX and RX events
110 *
111 * Set this to 1 to use separate channels for TX and RX. It allows us
112 * to control interrupt affinity separately for TX and RX.
113 *
114 * This is only used in MSI-X interrupt mode
115 */
116static bool separate_tx_channels;
117module_param(separate_tx_channels, bool, 0444);
118MODULE_PARM_DESC(separate_tx_channels,
119 "Use separate channels for TX and RX");
120
121/* This is the weight assigned to each of the (per-channel) virtual
122 * NAPI devices.
123 */
124static int napi_weight = 64;
125
126/* This is the time (in jiffies) between invocations of the hardware
127 * monitor.
128 * On Falcon-based NICs, this will:
129 * - Check the on-board hardware monitor;
130 * - Poll the link state and reconfigure the hardware as necessary.
131 * On Siena-based NICs for power systems with EEH support, this will give EEH a
132 * chance to start.
133 */
134static unsigned int efx_monitor_interval = 1 * HZ;
135
136/* Initial interrupt moderation settings. They can be modified after
137 * module load with ethtool.
138 *
139 * The default for RX should strike a balance between increasing the
140 * round-trip latency and reducing overhead.
141 */
142static unsigned int rx_irq_mod_usec = 60;
143
144/* Initial interrupt moderation settings. They can be modified after
145 * module load with ethtool.
146 *
147 * This default is chosen to ensure that a 10G link does not go idle
148 * while a TX queue is stopped after it has become full. A queue is
149 * restarted when it drops below half full. The time this takes (assuming
150 * worst case 3 descriptors per packet and 1024 descriptors) is
151 * 512 / 3 * 1.2 = 205 usec.
152 */
153static unsigned int tx_irq_mod_usec = 150;
154
155/* This is the first interrupt mode to try out of:
156 * 0 => MSI-X
157 * 1 => MSI
158 * 2 => legacy
159 */
160static unsigned int interrupt_mode;
161
162/* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
163 * i.e. the number of CPUs among which we may distribute simultaneous
164 * interrupt handling.
165 *
166 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
167 * The default (0) means to assign an interrupt to each core.
168 */
169static unsigned int rss_cpus;
170module_param(rss_cpus, uint, 0444);
171MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
172
173static bool phy_flash_cfg;
174module_param(phy_flash_cfg, bool, 0644);
175MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
176
177static unsigned irq_adapt_low_thresh = 8000;
178module_param(irq_adapt_low_thresh, uint, 0644);
179MODULE_PARM_DESC(irq_adapt_low_thresh,
180 "Threshold score for reducing IRQ moderation");
181
182static unsigned irq_adapt_high_thresh = 16000;
183module_param(irq_adapt_high_thresh, uint, 0644);
184MODULE_PARM_DESC(irq_adapt_high_thresh,
185 "Threshold score for increasing IRQ moderation");
186
187static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
188 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
189 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
190 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
191module_param(debug, uint, 0);
192MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
193
194/**************************************************************************
195 *
196 * Utility functions and prototypes
197 *
198 *************************************************************************/
199
200static int efx_soft_enable_interrupts(struct efx_nic *efx);
201static void efx_soft_disable_interrupts(struct efx_nic *efx);
202static void efx_remove_channel(struct efx_channel *channel);
203static void efx_remove_channels(struct efx_nic *efx);
204static const struct efx_channel_type efx_default_channel_type;
205static void efx_remove_port(struct efx_nic *efx);
206static void efx_init_napi_channel(struct efx_channel *channel);
207static void efx_fini_napi(struct efx_nic *efx);
208static void efx_fini_napi_channel(struct efx_channel *channel);
209static void efx_fini_struct(struct efx_nic *efx);
210static void efx_start_all(struct efx_nic *efx);
211static void efx_stop_all(struct efx_nic *efx);
212
213#define EFX_ASSERT_RESET_SERIALISED(efx) \
214 do { \
215 if ((efx->state == STATE_READY) || \
216 (efx->state == STATE_RECOVERY) || \
217 (efx->state == STATE_DISABLED)) \
218 ASSERT_RTNL(); \
219 } while (0)
220
221static int efx_check_disabled(struct efx_nic *efx)
222{
223 if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
224 netif_err(efx, drv, efx->net_dev,
225 "device is disabled due to earlier errors\n");
226 return -EIO;
227 }
228 return 0;
229}
230
231/**************************************************************************
232 *
233 * Event queue processing
234 *
235 *************************************************************************/
236
237/* Process channel's event queue
238 *
239 * This function is responsible for processing the event queue of a
240 * single channel. The caller must guarantee that this function will
241 * never be concurrently called more than once on the same channel,
242 * though different channels may be being processed concurrently.
243 */
244static int efx_process_channel(struct efx_channel *channel, int budget)
245{
246 int spent;
247
248 if (unlikely(!channel->enabled))
249 return 0;
250
251 spent = efx_nic_process_eventq(channel, budget);
252 if (spent && efx_channel_has_rx_queue(channel)) {
253 struct efx_rx_queue *rx_queue =
254 efx_channel_get_rx_queue(channel);
255
256 efx_rx_flush_packet(channel);
257 efx_fast_push_rx_descriptors(rx_queue, true);
258 }
259
260 return spent;
261}
262
263/* NAPI poll handler
264 *
265 * NAPI guarantees serialisation of polls of the same device, which
266 * provides the guarantee required by efx_process_channel().
267 */
268static int efx_poll(struct napi_struct *napi, int budget)
269{
270 struct efx_channel *channel =
271 container_of(napi, struct efx_channel, napi_str);
272 struct efx_nic *efx = channel->efx;
273 int spent;
274
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());
278
279 spent = efx_process_channel(channel, budget);
280
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);
290 }
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);
297 }
298 }
299 channel->irq_count = 0;
300 channel->irq_mod_score = 0;
301 }
302
303 efx_filter_rfs_expire(channel);
304
305 /* There is no race here; although napi_disable() will
306 * only wait for napi_complete(), this isn't a problem
307 * since efx_nic_eventq_read_ack() will have no effect if
308 * interrupts have already been disabled.
309 */
310 napi_complete(napi);
311 efx_nic_eventq_read_ack(channel);
312 }
313
314 return spent;
315}
316
317/* Create event queue
318 * Event queue memory allocations are done only once. If the channel
319 * is reset, the memory buffer will be reused; this guards against
320 * errors during channel reset and also simplifies interrupt handling.
321 */
322static int efx_probe_eventq(struct efx_channel *channel)
323{
324 struct efx_nic *efx = channel->efx;
325 unsigned long entries;
326
327 netif_dbg(efx, probe, efx->net_dev,
328 "chan %d create event queue\n", channel->channel);
329
330 /* Build an event queue with room for one event per tx and rx buffer,
331 * plus some extra for link state events and MCDI completions. */
332 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
333 EFX_BUG_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
334 channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
335
336 return efx_nic_probe_eventq(channel);
337}
338
339/* Prepare channel's event queue */
340static int efx_init_eventq(struct efx_channel *channel)
341{
342 struct efx_nic *efx = channel->efx;
343 int rc;
344
345 EFX_WARN_ON_PARANOID(channel->eventq_init);
346
347 netif_dbg(efx, drv, efx->net_dev,
348 "chan %d init event queue\n", channel->channel);
349
350 rc = efx_nic_init_eventq(channel);
351 if (rc == 0) {
352 efx->type->push_irq_moderation(channel);
353 channel->eventq_read_ptr = 0;
354 channel->eventq_init = true;
355 }
356 return rc;
357}
358
359/* Enable event queue processing and NAPI */
360static void efx_start_eventq(struct efx_channel *channel)
361{
362 netif_dbg(channel->efx, ifup, channel->efx->net_dev,
363 "chan %d start event queue\n", channel->channel);
364
365 /* Make sure the NAPI handler sees the enabled flag set */
366 channel->enabled = true;
367 smp_wmb();
368
369 napi_enable(&channel->napi_str);
370 efx_nic_eventq_read_ack(channel);
371}
372
373/* Disable event queue processing and NAPI */
374static void efx_stop_eventq(struct efx_channel *channel)
375{
376 if (!channel->enabled)
377 return;
378
379 napi_disable(&channel->napi_str);
380 channel->enabled = false;
381}
382
383static void efx_fini_eventq(struct efx_channel *channel)
384{
385 if (!channel->eventq_init)
386 return;
387
388 netif_dbg(channel->efx, drv, channel->efx->net_dev,
389 "chan %d fini event queue\n", channel->channel);
390
391 efx_nic_fini_eventq(channel);
392 channel->eventq_init = false;
393}
394
395static void efx_remove_eventq(struct efx_channel *channel)
396{
397 netif_dbg(channel->efx, drv, channel->efx->net_dev,
398 "chan %d remove event queue\n", channel->channel);
399
400 efx_nic_remove_eventq(channel);
401}
402
403/**************************************************************************
404 *
405 * Channel handling
406 *
407 *************************************************************************/
408
409/* Allocate and initialise a channel structure. */
410static struct efx_channel *
411efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
412{
413 struct efx_channel *channel;
414 struct efx_rx_queue *rx_queue;
415 struct efx_tx_queue *tx_queue;
416 int j;
417
418 channel = kzalloc(sizeof(*channel), GFP_KERNEL);
419 if (!channel)
420 return NULL;
421
422 channel->efx = efx;
423 channel->channel = i;
424 channel->type = &efx_default_channel_type;
425
426 for (j = 0; j < EFX_TXQ_TYPES; j++) {
427 tx_queue = &channel->tx_queue[j];
428 tx_queue->efx = efx;
429 tx_queue->queue = i * EFX_TXQ_TYPES + j;
430 tx_queue->channel = channel;
431 }
432
433 rx_queue = &channel->rx_queue;
434 rx_queue->efx = efx;
435 setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
436 (unsigned long)rx_queue);
437
438 return channel;
439}
440
441/* Allocate and initialise a channel structure, copying parameters
442 * (but not resources) from an old channel structure.
443 */
444static struct efx_channel *
445efx_copy_channel(const struct efx_channel *old_channel)
446{
447 struct efx_channel *channel;
448 struct efx_rx_queue *rx_queue;
449 struct efx_tx_queue *tx_queue;
450 int j;
451
452 channel = kmalloc(sizeof(*channel), GFP_KERNEL);
453 if (!channel)
454 return NULL;
455
456 *channel = *old_channel;
457
458 channel->napi_dev = NULL;
459 memset(&channel->eventq, 0, sizeof(channel->eventq));
460
461 for (j = 0; j < EFX_TXQ_TYPES; j++) {
462 tx_queue = &channel->tx_queue[j];
463 if (tx_queue->channel)
464 tx_queue->channel = channel;
465 tx_queue->buffer = NULL;
466 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
467 }
468
469 rx_queue = &channel->rx_queue;
470 rx_queue->buffer = NULL;
471 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
472 setup_timer(&rx_queue->slow_fill, efx_rx_slow_fill,
473 (unsigned long)rx_queue);
474
475 return channel;
476}
477
478static int efx_probe_channel(struct efx_channel *channel)
479{
480 struct efx_tx_queue *tx_queue;
481 struct efx_rx_queue *rx_queue;
482 int rc;
483
484 netif_dbg(channel->efx, probe, channel->efx->net_dev,
485 "creating channel %d\n", channel->channel);
486
487 rc = channel->type->pre_probe(channel);
488 if (rc)
489 goto fail;
490
491 rc = efx_probe_eventq(channel);
492 if (rc)
493 goto fail;
494
495 efx_for_each_channel_tx_queue(tx_queue, channel) {
496 rc = efx_probe_tx_queue(tx_queue);
497 if (rc)
498 goto fail;
499 }
500
501 efx_for_each_channel_rx_queue(rx_queue, channel) {
502 rc = efx_probe_rx_queue(rx_queue);
503 if (rc)
504 goto fail;
505 }
506
507 return 0;
508
509fail:
510 efx_remove_channel(channel);
511 return rc;
512}
513
514static void
515efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
516{
517 struct efx_nic *efx = channel->efx;
518 const char *type;
519 int number;
520
521 number = channel->channel;
522 if (efx->tx_channel_offset == 0) {
523 type = "";
524 } else if (channel->channel < efx->tx_channel_offset) {
525 type = "-rx";
526 } else {
527 type = "-tx";
528 number -= efx->tx_channel_offset;
529 }
530 snprintf(buf, len, "%s%s-%d", efx->name, type, number);
531}
532
533static void efx_set_channel_names(struct efx_nic *efx)
534{
535 struct efx_channel *channel;
536
537 efx_for_each_channel(channel, efx)
538 channel->type->get_name(channel,
539 efx->msi_context[channel->channel].name,
540 sizeof(efx->msi_context[0].name));
541}
542
543static int efx_probe_channels(struct efx_nic *efx)
544{
545 struct efx_channel *channel;
546 int rc;
547
548 /* Restart special buffer allocation */
549 efx->next_buffer_table = 0;
550
551 /* Probe channels in reverse, so that any 'extra' channels
552 * use the start of the buffer table. This allows the traffic
553 * channels to be resized without moving them or wasting the
554 * entries before them.
555 */
556 efx_for_each_channel_rev(channel, efx) {
557 rc = efx_probe_channel(channel);
558 if (rc) {
559 netif_err(efx, probe, efx->net_dev,
560 "failed to create channel %d\n",
561 channel->channel);
562 goto fail;
563 }
564 }
565 efx_set_channel_names(efx);
566
567 return 0;
568
569fail:
570 efx_remove_channels(efx);
571 return rc;
572}
573
574/* Channels are shutdown and reinitialised whilst the NIC is running
575 * to propagate configuration changes (mtu, checksum offload), or
576 * to clear hardware error conditions
577 */
578static void efx_start_datapath(struct efx_nic *efx)
579{
580 bool old_rx_scatter = efx->rx_scatter;
581 struct efx_tx_queue *tx_queue;
582 struct efx_rx_queue *rx_queue;
583 struct efx_channel *channel;
584 size_t rx_buf_len;
585
586 /* Calculate the rx buffer allocation parameters required to
587 * support the current MTU, including padding for header
588 * alignment and overruns.
589 */
590 efx->rx_dma_len = (efx->rx_prefix_size +
591 EFX_MAX_FRAME_LEN(efx->net_dev->mtu) +
592 efx->type->rx_buffer_padding);
593 rx_buf_len = (sizeof(struct efx_rx_page_state) +
594 efx->rx_ip_align + efx->rx_dma_len);
595 if (rx_buf_len <= PAGE_SIZE) {
596 efx->rx_scatter = efx->type->always_rx_scatter;
597 efx->rx_buffer_order = 0;
598 } else if (efx->type->can_rx_scatter) {
599 BUILD_BUG_ON(EFX_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
600 BUILD_BUG_ON(sizeof(struct efx_rx_page_state) +
601 2 * ALIGN(NET_IP_ALIGN + EFX_RX_USR_BUF_SIZE,
602 EFX_RX_BUF_ALIGNMENT) >
603 PAGE_SIZE);
604 efx->rx_scatter = true;
605 efx->rx_dma_len = EFX_RX_USR_BUF_SIZE;
606 efx->rx_buffer_order = 0;
607 } else {
608 efx->rx_scatter = false;
609 efx->rx_buffer_order = get_order(rx_buf_len);
610 }
611
612 efx_rx_config_page_split(efx);
613 if (efx->rx_buffer_order)
614 netif_dbg(efx, drv, efx->net_dev,
615 "RX buf len=%u; page order=%u batch=%u\n",
616 efx->rx_dma_len, efx->rx_buffer_order,
617 efx->rx_pages_per_batch);
618 else
619 netif_dbg(efx, drv, efx->net_dev,
620 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
621 efx->rx_dma_len, efx->rx_page_buf_step,
622 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
623
624 /* RX filters may also have scatter-enabled flags */
625 if (efx->rx_scatter != old_rx_scatter)
626 efx->type->filter_update_rx_scatter(efx);
627
628 /* We must keep at least one descriptor in a TX ring empty.
629 * We could avoid this when the queue size does not exactly
630 * match the hardware ring size, but it's not that important.
631 * Therefore we stop the queue when one more skb might fill
632 * the ring completely. We wake it when half way back to
633 * empty.
634 */
635 efx->txq_stop_thresh = efx->txq_entries - efx_tx_max_skb_descs(efx);
636 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
637
638 /* Initialise the channels */
639 efx_for_each_channel(channel, efx) {
640 efx_for_each_channel_tx_queue(tx_queue, channel) {
641 efx_init_tx_queue(tx_queue);
642 atomic_inc(&efx->active_queues);
643 }
644
645 efx_for_each_channel_rx_queue(rx_queue, channel) {
646 efx_init_rx_queue(rx_queue);
647 atomic_inc(&efx->active_queues);
648 efx_stop_eventq(channel);
649 efx_fast_push_rx_descriptors(rx_queue, false);
650 efx_start_eventq(channel);
651 }
652
653 WARN_ON(channel->rx_pkt_n_frags);
654 }
655
656 efx_ptp_start_datapath(efx);
657
658 if (netif_device_present(efx->net_dev))
659 netif_tx_wake_all_queues(efx->net_dev);
660}
661
662static void efx_stop_datapath(struct efx_nic *efx)
663{
664 struct efx_channel *channel;
665 struct efx_tx_queue *tx_queue;
666 struct efx_rx_queue *rx_queue;
667 int rc;
668
669 EFX_ASSERT_RESET_SERIALISED(efx);
670 BUG_ON(efx->port_enabled);
671
672 efx_ptp_stop_datapath(efx);
673
674 /* Stop RX refill */
675 efx_for_each_channel(channel, efx) {
676 efx_for_each_channel_rx_queue(rx_queue, channel)
677 rx_queue->refill_enabled = false;
678 }
679
680 efx_for_each_channel(channel, efx) {
681 /* RX packet processing is pipelined, so wait for the
682 * NAPI handler to complete. At least event queue 0
683 * might be kept active by non-data events, so don't
684 * use napi_synchronize() but actually disable NAPI
685 * temporarily.
686 */
687 if (efx_channel_has_rx_queue(channel)) {
688 efx_stop_eventq(channel);
689 efx_start_eventq(channel);
690 }
691 }
692
693 rc = efx->type->fini_dmaq(efx);
694 if (rc && EFX_WORKAROUND_7803(efx)) {
695 /* Schedule a reset to recover from the flush failure. The
696 * descriptor caches reference memory we're about to free,
697 * but falcon_reconfigure_mac_wrapper() won't reconnect
698 * the MACs because of the pending reset.
699 */
700 netif_err(efx, drv, efx->net_dev,
701 "Resetting to recover from flush failure\n");
702 efx_schedule_reset(efx, RESET_TYPE_ALL);
703 } else if (rc) {
704 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
705 } else {
706 netif_dbg(efx, drv, efx->net_dev,
707 "successfully flushed all queues\n");
708 }
709
710 efx_for_each_channel(channel, efx) {
711 efx_for_each_channel_rx_queue(rx_queue, channel)
712 efx_fini_rx_queue(rx_queue);
713 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
714 efx_fini_tx_queue(tx_queue);
715 }
716}
717
718static void efx_remove_channel(struct efx_channel *channel)
719{
720 struct efx_tx_queue *tx_queue;
721 struct efx_rx_queue *rx_queue;
722
723 netif_dbg(channel->efx, drv, channel->efx->net_dev,
724 "destroy chan %d\n", channel->channel);
725
726 efx_for_each_channel_rx_queue(rx_queue, channel)
727 efx_remove_rx_queue(rx_queue);
728 efx_for_each_possible_channel_tx_queue(tx_queue, channel)
729 efx_remove_tx_queue(tx_queue);
730 efx_remove_eventq(channel);
731 channel->type->post_remove(channel);
732}
733
734static void efx_remove_channels(struct efx_nic *efx)
735{
736 struct efx_channel *channel;
737
738 efx_for_each_channel(channel, efx)
739 efx_remove_channel(channel);
740}
741
742int
743efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
744{
745 struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
746 u32 old_rxq_entries, old_txq_entries;
747 unsigned i, next_buffer_table = 0;
748 int rc, rc2;
749
750 rc = efx_check_disabled(efx);
751 if (rc)
752 return rc;
753
754 /* Not all channels should be reallocated. We must avoid
755 * reallocating their buffer table entries.
756 */
757 efx_for_each_channel(channel, efx) {
758 struct efx_rx_queue *rx_queue;
759 struct efx_tx_queue *tx_queue;
760
761 if (channel->type->copy)
762 continue;
763 next_buffer_table = max(next_buffer_table,
764 channel->eventq.index +
765 channel->eventq.entries);
766 efx_for_each_channel_rx_queue(rx_queue, channel)
767 next_buffer_table = max(next_buffer_table,
768 rx_queue->rxd.index +
769 rx_queue->rxd.entries);
770 efx_for_each_channel_tx_queue(tx_queue, channel)
771 next_buffer_table = max(next_buffer_table,
772 tx_queue->txd.index +
773 tx_queue->txd.entries);
774 }
775
776 efx_device_detach_sync(efx);
777 efx_stop_all(efx);
778 efx_soft_disable_interrupts(efx);
779
780 /* Clone channels (where possible) */
781 memset(other_channel, 0, sizeof(other_channel));
782 for (i = 0; i < efx->n_channels; i++) {
783 channel = efx->channel[i];
784 if (channel->type->copy)
785 channel = channel->type->copy(channel);
786 if (!channel) {
787 rc = -ENOMEM;
788 goto out;
789 }
790 other_channel[i] = channel;
791 }
792
793 /* Swap entry counts and channel pointers */
794 old_rxq_entries = efx->rxq_entries;
795 old_txq_entries = efx->txq_entries;
796 efx->rxq_entries = rxq_entries;
797 efx->txq_entries = txq_entries;
798 for (i = 0; i < efx->n_channels; i++) {
799 channel = efx->channel[i];
800 efx->channel[i] = other_channel[i];
801 other_channel[i] = channel;
802 }
803
804 /* Restart buffer table allocation */
805 efx->next_buffer_table = next_buffer_table;
806
807 for (i = 0; i < efx->n_channels; i++) {
808 channel = efx->channel[i];
809 if (!channel->type->copy)
810 continue;
811 rc = efx_probe_channel(channel);
812 if (rc)
813 goto rollback;
814 efx_init_napi_channel(efx->channel[i]);
815 }
816
817out:
818 /* Destroy unused channel structures */
819 for (i = 0; i < efx->n_channels; i++) {
820 channel = other_channel[i];
821 if (channel && channel->type->copy) {
822 efx_fini_napi_channel(channel);
823 efx_remove_channel(channel);
824 kfree(channel);
825 }
826 }
827
828 rc2 = efx_soft_enable_interrupts(efx);
829 if (rc2) {
830 rc = rc ? rc : rc2;
831 netif_err(efx, drv, efx->net_dev,
832 "unable to restart interrupts on channel reallocation\n");
833 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
834 } else {
835 efx_start_all(efx);
836 netif_device_attach(efx->net_dev);
837 }
838 return rc;
839
840rollback:
841 /* Swap back */
842 efx->rxq_entries = old_rxq_entries;
843 efx->txq_entries = old_txq_entries;
844 for (i = 0; i < efx->n_channels; i++) {
845 channel = efx->channel[i];
846 efx->channel[i] = other_channel[i];
847 other_channel[i] = channel;
848 }
849 goto out;
850}
851
852void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
853{
854 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
855}
856
857static const struct efx_channel_type efx_default_channel_type = {
858 .pre_probe = efx_channel_dummy_op_int,
859 .post_remove = efx_channel_dummy_op_void,
860 .get_name = efx_get_channel_name,
861 .copy = efx_copy_channel,
862 .keep_eventq = false,
863};
864
865int efx_channel_dummy_op_int(struct efx_channel *channel)
866{
867 return 0;
868}
869
870void efx_channel_dummy_op_void(struct efx_channel *channel)
871{
872}
873
874/**************************************************************************
875 *
876 * Port handling
877 *
878 **************************************************************************/
879
880/* This ensures that the kernel is kept informed (via
881 * netif_carrier_on/off) of the link status, and also maintains the
882 * link status's stop on the port's TX queue.
883 */
884void efx_link_status_changed(struct efx_nic *efx)
885{
886 struct efx_link_state *link_state = &efx->link_state;
887
888 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
889 * that no events are triggered between unregister_netdev() and the
890 * driver unloading. A more general condition is that NETDEV_CHANGE
891 * can only be generated between NETDEV_UP and NETDEV_DOWN */
892 if (!netif_running(efx->net_dev))
893 return;
894
895 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
896 efx->n_link_state_changes++;
897
898 if (link_state->up)
899 netif_carrier_on(efx->net_dev);
900 else
901 netif_carrier_off(efx->net_dev);
902 }
903
904 /* Status message for kernel log */
905 if (link_state->up)
906 netif_info(efx, link, efx->net_dev,
907 "link up at %uMbps %s-duplex (MTU %d)\n",
908 link_state->speed, link_state->fd ? "full" : "half",
909 efx->net_dev->mtu);
910 else
911 netif_info(efx, link, efx->net_dev, "link down\n");
912}
913
914void efx_link_set_advertising(struct efx_nic *efx, u32 advertising)
915{
916 efx->link_advertising = advertising;
917 if (advertising) {
918 if (advertising & ADVERTISED_Pause)
919 efx->wanted_fc |= (EFX_FC_TX | EFX_FC_RX);
920 else
921 efx->wanted_fc &= ~(EFX_FC_TX | EFX_FC_RX);
922 if (advertising & ADVERTISED_Asym_Pause)
923 efx->wanted_fc ^= EFX_FC_TX;
924 }
925}
926
927void efx_link_set_wanted_fc(struct efx_nic *efx, u8 wanted_fc)
928{
929 efx->wanted_fc = wanted_fc;
930 if (efx->link_advertising) {
931 if (wanted_fc & EFX_FC_RX)
932 efx->link_advertising |= (ADVERTISED_Pause |
933 ADVERTISED_Asym_Pause);
934 else
935 efx->link_advertising &= ~(ADVERTISED_Pause |
936 ADVERTISED_Asym_Pause);
937 if (wanted_fc & EFX_FC_TX)
938 efx->link_advertising ^= ADVERTISED_Asym_Pause;
939 }
940}
941
942static void efx_fini_port(struct efx_nic *efx);
943
944/* Push loopback/power/transmit disable settings to the PHY, and reconfigure
945 * the MAC appropriately. All other PHY configuration changes are pushed
946 * through phy_op->set_settings(), and pushed asynchronously to the MAC
947 * through efx_monitor().
948 *
949 * Callers must hold the mac_lock
950 */
951int __efx_reconfigure_port(struct efx_nic *efx)
952{
953 enum efx_phy_mode phy_mode;
954 int rc;
955
956 WARN_ON(!mutex_is_locked(&efx->mac_lock));
957
958 /* Disable PHY transmit in mac level loopbacks */
959 phy_mode = efx->phy_mode;
960 if (LOOPBACK_INTERNAL(efx))
961 efx->phy_mode |= PHY_MODE_TX_DISABLED;
962 else
963 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
964
965 rc = efx->type->reconfigure_port(efx);
966
967 if (rc)
968 efx->phy_mode = phy_mode;
969
970 return rc;
971}
972
973/* Reinitialise the MAC to pick up new PHY settings, even if the port is
974 * disabled. */
975int efx_reconfigure_port(struct efx_nic *efx)
976{
977 int rc;
978
979 EFX_ASSERT_RESET_SERIALISED(efx);
980
981 mutex_lock(&efx->mac_lock);
982 rc = __efx_reconfigure_port(efx);
983 mutex_unlock(&efx->mac_lock);
984
985 return rc;
986}
987
988/* Asynchronous work item for changing MAC promiscuity and multicast
989 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
990 * MAC directly. */
991static void efx_mac_work(struct work_struct *data)
992{
993 struct efx_nic *efx = container_of(data, struct efx_nic, mac_work);
994
995 mutex_lock(&efx->mac_lock);
996 if (efx->port_enabled)
997 efx->type->reconfigure_mac(efx);
998 mutex_unlock(&efx->mac_lock);
999}
1000
1001static int efx_probe_port(struct efx_nic *efx)
1002{
1003 int rc;
1004
1005 netif_dbg(efx, probe, efx->net_dev, "create port\n");
1006
1007 if (phy_flash_cfg)
1008 efx->phy_mode = PHY_MODE_SPECIAL;
1009
1010 /* Connect up MAC/PHY operations table */
1011 rc = efx->type->probe_port(efx);
1012 if (rc)
1013 return rc;
1014
1015 /* Initialise MAC address to permanent address */
1016 ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1017
1018 return 0;
1019}
1020
1021static int efx_init_port(struct efx_nic *efx)
1022{
1023 int rc;
1024
1025 netif_dbg(efx, drv, efx->net_dev, "init port\n");
1026
1027 mutex_lock(&efx->mac_lock);
1028
1029 rc = efx->phy_op->init(efx);
1030 if (rc)
1031 goto fail1;
1032
1033 efx->port_initialized = true;
1034
1035 /* Reconfigure the MAC before creating dma queues (required for
1036 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1037 efx->type->reconfigure_mac(efx);
1038
1039 /* Ensure the PHY advertises the correct flow control settings */
1040 rc = efx->phy_op->reconfigure(efx);
1041 if (rc)
1042 goto fail2;
1043
1044 mutex_unlock(&efx->mac_lock);
1045 return 0;
1046
1047fail2:
1048 efx->phy_op->fini(efx);
1049fail1:
1050 mutex_unlock(&efx->mac_lock);
1051 return rc;
1052}
1053
1054static void efx_start_port(struct efx_nic *efx)
1055{
1056 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1057 BUG_ON(efx->port_enabled);
1058
1059 mutex_lock(&efx->mac_lock);
1060 efx->port_enabled = true;
1061
1062 /* Ensure MAC ingress/egress is enabled */
1063 efx->type->reconfigure_mac(efx);
1064
1065 mutex_unlock(&efx->mac_lock);
1066}
1067
1068/* Cancel work for MAC reconfiguration, periodic hardware monitoring
1069 * and the async self-test, wait for them to finish and prevent them
1070 * being scheduled again. This doesn't cover online resets, which
1071 * should only be cancelled when removing the device.
1072 */
1073static void efx_stop_port(struct efx_nic *efx)
1074{
1075 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1076
1077 EFX_ASSERT_RESET_SERIALISED(efx);
1078
1079 mutex_lock(&efx->mac_lock);
1080 efx->port_enabled = false;
1081 mutex_unlock(&efx->mac_lock);
1082
1083 /* Serialise against efx_set_multicast_list() */
1084 netif_addr_lock_bh(efx->net_dev);
1085 netif_addr_unlock_bh(efx->net_dev);
1086
1087 cancel_delayed_work_sync(&efx->monitor_work);
1088 efx_selftest_async_cancel(efx);
1089 cancel_work_sync(&efx->mac_work);
1090}
1091
1092static void efx_fini_port(struct efx_nic *efx)
1093{
1094 netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1095
1096 if (!efx->port_initialized)
1097 return;
1098
1099 efx->phy_op->fini(efx);
1100 efx->port_initialized = false;
1101
1102 efx->link_state.up = false;
1103 efx_link_status_changed(efx);
1104}
1105
1106static void efx_remove_port(struct efx_nic *efx)
1107{
1108 netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1109
1110 efx->type->remove_port(efx);
1111}
1112
1113/**************************************************************************
1114 *
1115 * NIC handling
1116 *
1117 **************************************************************************/
1118
1119static LIST_HEAD(efx_primary_list);
1120static LIST_HEAD(efx_unassociated_list);
1121
1122static bool efx_same_controller(struct efx_nic *left, struct efx_nic *right)
1123{
1124 return left->type == right->type &&
1125 left->vpd_sn && right->vpd_sn &&
1126 !strcmp(left->vpd_sn, right->vpd_sn);
1127}
1128
1129static void efx_associate(struct efx_nic *efx)
1130{
1131 struct efx_nic *other, *next;
1132
1133 if (efx->primary == efx) {
1134 /* Adding primary function; look for secondaries */
1135
1136 netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1137 list_add_tail(&efx->node, &efx_primary_list);
1138
1139 list_for_each_entry_safe(other, next, &efx_unassociated_list,
1140 node) {
1141 if (efx_same_controller(efx, other)) {
1142 list_del(&other->node);
1143 netif_dbg(other, probe, other->net_dev,
1144 "moving to secondary list of %s %s\n",
1145 pci_name(efx->pci_dev),
1146 efx->net_dev->name);
1147 list_add_tail(&other->node,
1148 &efx->secondary_list);
1149 other->primary = efx;
1150 }
1151 }
1152 } else {
1153 /* Adding secondary function; look for primary */
1154
1155 list_for_each_entry(other, &efx_primary_list, node) {
1156 if (efx_same_controller(efx, other)) {
1157 netif_dbg(efx, probe, efx->net_dev,
1158 "adding to secondary list of %s %s\n",
1159 pci_name(other->pci_dev),
1160 other->net_dev->name);
1161 list_add_tail(&efx->node,
1162 &other->secondary_list);
1163 efx->primary = other;
1164 return;
1165 }
1166 }
1167
1168 netif_dbg(efx, probe, efx->net_dev,
1169 "adding to unassociated list\n");
1170 list_add_tail(&efx->node, &efx_unassociated_list);
1171 }
1172}
1173
1174static void efx_dissociate(struct efx_nic *efx)
1175{
1176 struct efx_nic *other, *next;
1177
1178 list_del(&efx->node);
1179 efx->primary = NULL;
1180
1181 list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1182 list_del(&other->node);
1183 netif_dbg(other, probe, other->net_dev,
1184 "moving to unassociated list\n");
1185 list_add_tail(&other->node, &efx_unassociated_list);
1186 other->primary = NULL;
1187 }
1188}
1189
1190/* This configures the PCI device to enable I/O and DMA. */
1191static int efx_init_io(struct efx_nic *efx)
1192{
1193 struct pci_dev *pci_dev = efx->pci_dev;
1194 dma_addr_t dma_mask = efx->type->max_dma_mask;
1195 unsigned int mem_map_size = efx->type->mem_map_size(efx);
1196 int rc;
1197
1198 netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1199
1200 rc = pci_enable_device(pci_dev);
1201 if (rc) {
1202 netif_err(efx, probe, efx->net_dev,
1203 "failed to enable PCI device\n");
1204 goto fail1;
1205 }
1206
1207 pci_set_master(pci_dev);
1208
1209 /* Set the PCI DMA mask. Try all possibilities from our
1210 * genuine mask down to 32 bits, because some architectures
1211 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1212 * masks event though they reject 46 bit masks.
1213 */
1214 while (dma_mask > 0x7fffffffUL) {
1215 if (dma_supported(&pci_dev->dev, dma_mask)) {
1216 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1217 if (rc == 0)
1218 break;
1219 }
1220 dma_mask >>= 1;
1221 }
1222 if (rc) {
1223 netif_err(efx, probe, efx->net_dev,
1224 "could not find a suitable DMA mask\n");
1225 goto fail2;
1226 }
1227 netif_dbg(efx, probe, efx->net_dev,
1228 "using DMA mask %llx\n", (unsigned long long) dma_mask);
1229
1230 efx->membase_phys = pci_resource_start(efx->pci_dev, EFX_MEM_BAR);
1231 rc = pci_request_region(pci_dev, EFX_MEM_BAR, "sfc");
1232 if (rc) {
1233 netif_err(efx, probe, efx->net_dev,
1234 "request for memory BAR failed\n");
1235 rc = -EIO;
1236 goto fail3;
1237 }
1238 efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1239 if (!efx->membase) {
1240 netif_err(efx, probe, efx->net_dev,
1241 "could not map memory BAR at %llx+%x\n",
1242 (unsigned long long)efx->membase_phys, mem_map_size);
1243 rc = -ENOMEM;
1244 goto fail4;
1245 }
1246 netif_dbg(efx, probe, efx->net_dev,
1247 "memory BAR at %llx+%x (virtual %p)\n",
1248 (unsigned long long)efx->membase_phys, mem_map_size,
1249 efx->membase);
1250
1251 return 0;
1252
1253 fail4:
1254 pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1255 fail3:
1256 efx->membase_phys = 0;
1257 fail2:
1258 pci_disable_device(efx->pci_dev);
1259 fail1:
1260 return rc;
1261}
1262
1263static void efx_fini_io(struct efx_nic *efx)
1264{
1265 netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1266
1267 if (efx->membase) {
1268 iounmap(efx->membase);
1269 efx->membase = NULL;
1270 }
1271
1272 if (efx->membase_phys) {
1273 pci_release_region(efx->pci_dev, EFX_MEM_BAR);
1274 efx->membase_phys = 0;
1275 }
1276
1277 pci_disable_device(efx->pci_dev);
1278}
1279
1280static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
1281{
1282 cpumask_var_t thread_mask;
1283 unsigned int count;
1284 int cpu;
1285
1286 if (rss_cpus) {
1287 count = rss_cpus;
1288 } else {
1289 if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1290 netif_warn(efx, probe, efx->net_dev,
1291 "RSS disabled due to allocation failure\n");
1292 return 1;
1293 }
1294
1295 count = 0;
1296 for_each_online_cpu(cpu) {
1297 if (!cpumask_test_cpu(cpu, thread_mask)) {
1298 ++count;
1299 cpumask_or(thread_mask, thread_mask,
1300 topology_thread_cpumask(cpu));
1301 }
1302 }
1303
1304 free_cpumask_var(thread_mask);
1305 }
1306
1307 /* If RSS is requested for the PF *and* VFs then we can't write RSS
1308 * table entries that are inaccessible to VFs
1309 */
1310 if (efx_sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
1311 count > efx_vf_size(efx)) {
1312 netif_warn(efx, probe, efx->net_dev,
1313 "Reducing number of RSS channels from %u to %u for "
1314 "VF support. Increase vf-msix-limit to use more "
1315 "channels on the PF.\n",
1316 count, efx_vf_size(efx));
1317 count = efx_vf_size(efx);
1318 }
1319
1320 return count;
1321}
1322
1323/* Probe the number and type of interrupts we are able to obtain, and
1324 * the resulting numbers of channels and RX queues.
1325 */
1326static int efx_probe_interrupts(struct efx_nic *efx)
1327{
1328 unsigned int extra_channels = 0;
1329 unsigned int i, j;
1330 int rc;
1331
1332 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
1333 if (efx->extra_channel_type[i])
1334 ++extra_channels;
1335
1336 if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1337 struct msix_entry xentries[EFX_MAX_CHANNELS];
1338 unsigned int n_channels;
1339
1340 n_channels = efx_wanted_parallelism(efx);
1341 if (separate_tx_channels)
1342 n_channels *= 2;
1343 n_channels += extra_channels;
1344 n_channels = min(n_channels, efx->max_channels);
1345
1346 for (i = 0; i < n_channels; i++)
1347 xentries[i].entry = i;
1348 rc = pci_enable_msix_range(efx->pci_dev,
1349 xentries, 1, n_channels);
1350 if (rc < 0) {
1351 /* Fall back to single channel MSI */
1352 efx->interrupt_mode = EFX_INT_MODE_MSI;
1353 netif_err(efx, drv, efx->net_dev,
1354 "could not enable MSI-X\n");
1355 } else if (rc < n_channels) {
1356 netif_err(efx, drv, efx->net_dev,
1357 "WARNING: Insufficient MSI-X vectors"
1358 " available (%d < %u).\n", rc, n_channels);
1359 netif_err(efx, drv, efx->net_dev,
1360 "WARNING: Performance may be reduced.\n");
1361 n_channels = rc;
1362 }
1363
1364 if (rc > 0) {
1365 efx->n_channels = n_channels;
1366 if (n_channels > extra_channels)
1367 n_channels -= extra_channels;
1368 if (separate_tx_channels) {
1369 efx->n_tx_channels = max(n_channels / 2, 1U);
1370 efx->n_rx_channels = max(n_channels -
1371 efx->n_tx_channels,
1372 1U);
1373 } else {
1374 efx->n_tx_channels = n_channels;
1375 efx->n_rx_channels = n_channels;
1376 }
1377 for (i = 0; i < efx->n_channels; i++)
1378 efx_get_channel(efx, i)->irq =
1379 xentries[i].vector;
1380 }
1381 }
1382
1383 /* Try single interrupt MSI */
1384 if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
1385 efx->n_channels = 1;
1386 efx->n_rx_channels = 1;
1387 efx->n_tx_channels = 1;
1388 rc = pci_enable_msi(efx->pci_dev);
1389 if (rc == 0) {
1390 efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1391 } else {
1392 netif_err(efx, drv, efx->net_dev,
1393 "could not enable MSI\n");
1394 efx->interrupt_mode = EFX_INT_MODE_LEGACY;
1395 }
1396 }
1397
1398 /* Assume legacy interrupts */
1399 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
1400 efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
1401 efx->n_rx_channels = 1;
1402 efx->n_tx_channels = 1;
1403 efx->legacy_irq = efx->pci_dev->irq;
1404 }
1405
1406 /* Assign extra channels if possible */
1407 j = efx->n_channels;
1408 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
1409 if (!efx->extra_channel_type[i])
1410 continue;
1411 if (efx->interrupt_mode != EFX_INT_MODE_MSIX ||
1412 efx->n_channels <= extra_channels) {
1413 efx->extra_channel_type[i]->handle_no_channel(efx);
1414 } else {
1415 --j;
1416 efx_get_channel(efx, j)->type =
1417 efx->extra_channel_type[i];
1418 }
1419 }
1420
1421 /* RSS might be usable on VFs even if it is disabled on the PF */
1422 efx->rss_spread = ((efx->n_rx_channels > 1 || !efx_sriov_wanted(efx)) ?
1423 efx->n_rx_channels : efx_vf_size(efx));
1424
1425 return 0;
1426}
1427
1428static int efx_soft_enable_interrupts(struct efx_nic *efx)
1429{
1430 struct efx_channel *channel, *end_channel;
1431 int rc;
1432
1433 BUG_ON(efx->state == STATE_DISABLED);
1434
1435 efx->irq_soft_enabled = true;
1436 smp_wmb();
1437
1438 efx_for_each_channel(channel, efx) {
1439 if (!channel->type->keep_eventq) {
1440 rc = efx_init_eventq(channel);
1441 if (rc)
1442 goto fail;
1443 }
1444 efx_start_eventq(channel);
1445 }
1446
1447 efx_mcdi_mode_event(efx);
1448
1449 return 0;
1450fail:
1451 end_channel = channel;
1452 efx_for_each_channel(channel, efx) {
1453 if (channel == end_channel)
1454 break;
1455 efx_stop_eventq(channel);
1456 if (!channel->type->keep_eventq)
1457 efx_fini_eventq(channel);
1458 }
1459
1460 return rc;
1461}
1462
1463static void efx_soft_disable_interrupts(struct efx_nic *efx)
1464{
1465 struct efx_channel *channel;
1466
1467 if (efx->state == STATE_DISABLED)
1468 return;
1469
1470 efx_mcdi_mode_poll(efx);
1471
1472 efx->irq_soft_enabled = false;
1473 smp_wmb();
1474
1475 if (efx->legacy_irq)
1476 synchronize_irq(efx->legacy_irq);
1477
1478 efx_for_each_channel(channel, efx) {
1479 if (channel->irq)
1480 synchronize_irq(channel->irq);
1481
1482 efx_stop_eventq(channel);
1483 if (!channel->type->keep_eventq)
1484 efx_fini_eventq(channel);
1485 }
1486
1487 /* Flush the asynchronous MCDI request queue */
1488 efx_mcdi_flush_async(efx);
1489}
1490
1491static int efx_enable_interrupts(struct efx_nic *efx)
1492{
1493 struct efx_channel *channel, *end_channel;
1494 int rc;
1495
1496 BUG_ON(efx->state == STATE_DISABLED);
1497
1498 if (efx->eeh_disabled_legacy_irq) {
1499 enable_irq(efx->legacy_irq);
1500 efx->eeh_disabled_legacy_irq = false;
1501 }
1502
1503 efx->type->irq_enable_master(efx);
1504
1505 efx_for_each_channel(channel, efx) {
1506 if (channel->type->keep_eventq) {
1507 rc = efx_init_eventq(channel);
1508 if (rc)
1509 goto fail;
1510 }
1511 }
1512
1513 rc = efx_soft_enable_interrupts(efx);
1514 if (rc)
1515 goto fail;
1516
1517 return 0;
1518
1519fail:
1520 end_channel = channel;
1521 efx_for_each_channel(channel, efx) {
1522 if (channel == end_channel)
1523 break;
1524 if (channel->type->keep_eventq)
1525 efx_fini_eventq(channel);
1526 }
1527
1528 efx->type->irq_disable_non_ev(efx);
1529
1530 return rc;
1531}
1532
1533static void efx_disable_interrupts(struct efx_nic *efx)
1534{
1535 struct efx_channel *channel;
1536
1537 efx_soft_disable_interrupts(efx);
1538
1539 efx_for_each_channel(channel, efx) {
1540 if (channel->type->keep_eventq)
1541 efx_fini_eventq(channel);
1542 }
1543
1544 efx->type->irq_disable_non_ev(efx);
1545}
1546
1547static void efx_remove_interrupts(struct efx_nic *efx)
1548{
1549 struct efx_channel *channel;
1550
1551 /* Remove MSI/MSI-X interrupts */
1552 efx_for_each_channel(channel, efx)
1553 channel->irq = 0;
1554 pci_disable_msi(efx->pci_dev);
1555 pci_disable_msix(efx->pci_dev);
1556
1557 /* Remove legacy interrupt */
1558 efx->legacy_irq = 0;
1559}
1560
1561static void efx_set_channels(struct efx_nic *efx)
1562{
1563 struct efx_channel *channel;
1564 struct efx_tx_queue *tx_queue;
1565
1566 efx->tx_channel_offset =
1567 separate_tx_channels ? efx->n_channels - efx->n_tx_channels : 0;
1568
1569 /* We need to mark which channels really have RX and TX
1570 * queues, and adjust the TX queue numbers if we have separate
1571 * RX-only and TX-only channels.
1572 */
1573 efx_for_each_channel(channel, efx) {
1574 if (channel->channel < efx->n_rx_channels)
1575 channel->rx_queue.core_index = channel->channel;
1576 else
1577 channel->rx_queue.core_index = -1;
1578
1579 efx_for_each_channel_tx_queue(tx_queue, channel)
1580 tx_queue->queue -= (efx->tx_channel_offset *
1581 EFX_TXQ_TYPES);
1582 }
1583}
1584
1585static int efx_probe_nic(struct efx_nic *efx)
1586{
1587 size_t i;
1588 int rc;
1589
1590 netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1591
1592 /* Carry out hardware-type specific initialisation */
1593 rc = efx->type->probe(efx);
1594 if (rc)
1595 return rc;
1596
1597 /* Determine the number of channels and queues by trying to hook
1598 * in MSI-X interrupts. */
1599 rc = efx_probe_interrupts(efx);
1600 if (rc)
1601 goto fail1;
1602
1603 efx_set_channels(efx);
1604
1605 rc = efx->type->dimension_resources(efx);
1606 if (rc)
1607 goto fail2;
1608
1609 if (efx->n_channels > 1)
1610 get_random_bytes(&efx->rx_hash_key, sizeof(efx->rx_hash_key));
1611 for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1612 efx->rx_indir_table[i] =
1613 ethtool_rxfh_indir_default(i, efx->rss_spread);
1614
1615 netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1616 netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1617
1618 /* Initialise the interrupt moderation settings */
1619 efx_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1620 true);
1621
1622 return 0;
1623
1624fail2:
1625 efx_remove_interrupts(efx);
1626fail1:
1627 efx->type->remove(efx);
1628 return rc;
1629}
1630
1631static void efx_remove_nic(struct efx_nic *efx)
1632{
1633 netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1634
1635 efx_remove_interrupts(efx);
1636 efx->type->remove(efx);
1637}
1638
1639static int efx_probe_filters(struct efx_nic *efx)
1640{
1641 int rc;
1642
1643 spin_lock_init(&efx->filter_lock);
1644
1645 rc = efx->type->filter_table_probe(efx);
1646 if (rc)
1647 return rc;
1648
1649#ifdef CONFIG_RFS_ACCEL
1650 if (efx->type->offload_features & NETIF_F_NTUPLE) {
1651 efx->rps_flow_id = kcalloc(efx->type->max_rx_ip_filters,
1652 sizeof(*efx->rps_flow_id),
1653 GFP_KERNEL);
1654 if (!efx->rps_flow_id) {
1655 efx->type->filter_table_remove(efx);
1656 return -ENOMEM;
1657 }
1658 }
1659#endif
1660
1661 return 0;
1662}
1663
1664static void efx_remove_filters(struct efx_nic *efx)
1665{
1666#ifdef CONFIG_RFS_ACCEL
1667 kfree(efx->rps_flow_id);
1668#endif
1669 efx->type->filter_table_remove(efx);
1670}
1671
1672static void efx_restore_filters(struct efx_nic *efx)
1673{
1674 efx->type->filter_table_restore(efx);
1675}
1676
1677/**************************************************************************
1678 *
1679 * NIC startup/shutdown
1680 *
1681 *************************************************************************/
1682
1683static int efx_probe_all(struct efx_nic *efx)
1684{
1685 int rc;
1686
1687 rc = efx_probe_nic(efx);
1688 if (rc) {
1689 netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1690 goto fail1;
1691 }
1692
1693 rc = efx_probe_port(efx);
1694 if (rc) {
1695 netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1696 goto fail2;
1697 }
1698
1699 BUILD_BUG_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_RXQ_MIN_ENT);
1700 if (WARN_ON(EFX_DEFAULT_DMAQ_SIZE < EFX_TXQ_MIN_ENT(efx))) {
1701 rc = -EINVAL;
1702 goto fail3;
1703 }
1704 efx->rxq_entries = efx->txq_entries = EFX_DEFAULT_DMAQ_SIZE;
1705
1706 rc = efx_probe_filters(efx);
1707 if (rc) {
1708 netif_err(efx, probe, efx->net_dev,
1709 "failed to create filter tables\n");
1710 goto fail3;
1711 }
1712
1713 rc = efx_probe_channels(efx);
1714 if (rc)
1715 goto fail4;
1716
1717 return 0;
1718
1719 fail4:
1720 efx_remove_filters(efx);
1721 fail3:
1722 efx_remove_port(efx);
1723 fail2:
1724 efx_remove_nic(efx);
1725 fail1:
1726 return rc;
1727}
1728
1729/* If the interface is supposed to be running but is not, start
1730 * the hardware and software data path, regular activity for the port
1731 * (MAC statistics, link polling, etc.) and schedule the port to be
1732 * reconfigured. Interrupts must already be enabled. This function
1733 * is safe to call multiple times, so long as the NIC is not disabled.
1734 * Requires the RTNL lock.
1735 */
1736static void efx_start_all(struct efx_nic *efx)
1737{
1738 EFX_ASSERT_RESET_SERIALISED(efx);
1739 BUG_ON(efx->state == STATE_DISABLED);
1740
1741 /* Check that it is appropriate to restart the interface. All
1742 * of these flags are safe to read under just the rtnl lock */
1743 if (efx->port_enabled || !netif_running(efx->net_dev) ||
1744 efx->reset_pending)
1745 return;
1746
1747 efx_start_port(efx);
1748 efx_start_datapath(efx);
1749
1750 /* Start the hardware monitor if there is one */
1751 if (efx->type->monitor != NULL)
1752 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1753 efx_monitor_interval);
1754
1755 /* If link state detection is normally event-driven, we have
1756 * to poll now because we could have missed a change
1757 */
1758 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1759 mutex_lock(&efx->mac_lock);
1760 if (efx->phy_op->poll(efx))
1761 efx_link_status_changed(efx);
1762 mutex_unlock(&efx->mac_lock);
1763 }
1764
1765 efx->type->start_stats(efx);
1766 efx->type->pull_stats(efx);
1767 spin_lock_bh(&efx->stats_lock);
1768 efx->type->update_stats(efx, NULL, NULL);
1769 spin_unlock_bh(&efx->stats_lock);
1770}
1771
1772/* Quiesce the hardware and software data path, and regular activity
1773 * for the port without bringing the link down. Safe to call multiple
1774 * times with the NIC in almost any state, but interrupts should be
1775 * enabled. Requires the RTNL lock.
1776 */
1777static void efx_stop_all(struct efx_nic *efx)
1778{
1779 EFX_ASSERT_RESET_SERIALISED(efx);
1780
1781 /* port_enabled can be read safely under the rtnl lock */
1782 if (!efx->port_enabled)
1783 return;
1784
1785 /* update stats before we go down so we can accurately count
1786 * rx_nodesc_drops
1787 */
1788 efx->type->pull_stats(efx);
1789 spin_lock_bh(&efx->stats_lock);
1790 efx->type->update_stats(efx, NULL, NULL);
1791 spin_unlock_bh(&efx->stats_lock);
1792 efx->type->stop_stats(efx);
1793 efx_stop_port(efx);
1794
1795 /* Stop the kernel transmit interface. This is only valid if
1796 * the device is stopped or detached; otherwise the watchdog
1797 * may fire immediately.
1798 */
1799 WARN_ON(netif_running(efx->net_dev) &&
1800 netif_device_present(efx->net_dev));
1801 netif_tx_disable(efx->net_dev);
1802
1803 efx_stop_datapath(efx);
1804}
1805
1806static void efx_remove_all(struct efx_nic *efx)
1807{
1808 efx_remove_channels(efx);
1809 efx_remove_filters(efx);
1810 efx_remove_port(efx);
1811 efx_remove_nic(efx);
1812}
1813
1814/**************************************************************************
1815 *
1816 * Interrupt moderation
1817 *
1818 **************************************************************************/
1819
1820static unsigned int irq_mod_ticks(unsigned int usecs, unsigned int quantum_ns)
1821{
1822 if (usecs == 0)
1823 return 0;
1824 if (usecs * 1000 < quantum_ns)
1825 return 1; /* never round down to 0 */
1826 return usecs * 1000 / quantum_ns;
1827}
1828
1829/* Set interrupt moderation parameters */
1830int efx_init_irq_moderation(struct efx_nic *efx, unsigned int tx_usecs,
1831 unsigned int rx_usecs, bool rx_adaptive,
1832 bool rx_may_override_tx)
1833{
1834 struct efx_channel *channel;
1835 unsigned int irq_mod_max = DIV_ROUND_UP(efx->type->timer_period_max *
1836 efx->timer_quantum_ns,
1837 1000);
1838 unsigned int tx_ticks;
1839 unsigned int rx_ticks;
1840
1841 EFX_ASSERT_RESET_SERIALISED(efx);
1842
1843 if (tx_usecs > irq_mod_max || rx_usecs > irq_mod_max)
1844 return -EINVAL;
1845
1846 tx_ticks = irq_mod_ticks(tx_usecs, efx->timer_quantum_ns);
1847 rx_ticks = irq_mod_ticks(rx_usecs, efx->timer_quantum_ns);
1848
1849 if (tx_ticks != rx_ticks && efx->tx_channel_offset == 0 &&
1850 !rx_may_override_tx) {
1851 netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1852 "RX and TX IRQ moderation must be equal\n");
1853 return -EINVAL;
1854 }
1855
1856 efx->irq_rx_adaptive = rx_adaptive;
1857 efx->irq_rx_moderation = rx_ticks;
1858 efx_for_each_channel(channel, efx) {
1859 if (efx_channel_has_rx_queue(channel))
1860 channel->irq_moderation = rx_ticks;
1861 else if (efx_channel_has_tx_queues(channel))
1862 channel->irq_moderation = tx_ticks;
1863 }
1864
1865 return 0;
1866}
1867
1868void efx_get_irq_moderation(struct efx_nic *efx, unsigned int *tx_usecs,
1869 unsigned int *rx_usecs, bool *rx_adaptive)
1870{
1871 /* We must round up when converting ticks to microseconds
1872 * because we round down when converting the other way.
1873 */
1874
1875 *rx_adaptive = efx->irq_rx_adaptive;
1876 *rx_usecs = DIV_ROUND_UP(efx->irq_rx_moderation *
1877 efx->timer_quantum_ns,
1878 1000);
1879
1880 /* If channels are shared between RX and TX, so is IRQ
1881 * moderation. Otherwise, IRQ moderation is the same for all
1882 * TX channels and is not adaptive.
1883 */
1884 if (efx->tx_channel_offset == 0)
1885 *tx_usecs = *rx_usecs;
1886 else
1887 *tx_usecs = DIV_ROUND_UP(
1888 efx->channel[efx->tx_channel_offset]->irq_moderation *
1889 efx->timer_quantum_ns,
1890 1000);
1891}
1892
1893/**************************************************************************
1894 *
1895 * Hardware monitor
1896 *
1897 **************************************************************************/
1898
1899/* Run periodically off the general workqueue */
1900static void efx_monitor(struct work_struct *data)
1901{
1902 struct efx_nic *efx = container_of(data, struct efx_nic,
1903 monitor_work.work);
1904
1905 netif_vdbg(efx, timer, efx->net_dev,
1906 "hardware monitor executing on CPU %d\n",
1907 raw_smp_processor_id());
1908 BUG_ON(efx->type->monitor == NULL);
1909
1910 /* If the mac_lock is already held then it is likely a port
1911 * reconfiguration is already in place, which will likely do
1912 * most of the work of monitor() anyway. */
1913 if (mutex_trylock(&efx->mac_lock)) {
1914 if (efx->port_enabled)
1915 efx->type->monitor(efx);
1916 mutex_unlock(&efx->mac_lock);
1917 }
1918
1919 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1920 efx_monitor_interval);
1921}
1922
1923/**************************************************************************
1924 *
1925 * ioctls
1926 *
1927 *************************************************************************/
1928
1929/* Net device ioctl
1930 * Context: process, rtnl_lock() held.
1931 */
1932static int efx_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
1933{
1934 struct efx_nic *efx = netdev_priv(net_dev);
1935 struct mii_ioctl_data *data = if_mii(ifr);
1936
1937 if (cmd == SIOCSHWTSTAMP)
1938 return efx_ptp_set_ts_config(efx, ifr);
1939 if (cmd == SIOCGHWTSTAMP)
1940 return efx_ptp_get_ts_config(efx, ifr);
1941
1942 /* Convert phy_id from older PRTAD/DEVAD format */
1943 if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
1944 (data->phy_id & 0xfc00) == 0x0400)
1945 data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
1946
1947 return mdio_mii_ioctl(&efx->mdio, data, cmd);
1948}
1949
1950/**************************************************************************
1951 *
1952 * NAPI interface
1953 *
1954 **************************************************************************/
1955
1956static void efx_init_napi_channel(struct efx_channel *channel)
1957{
1958 struct efx_nic *efx = channel->efx;
1959
1960 channel->napi_dev = efx->net_dev;
1961 netif_napi_add(channel->napi_dev, &channel->napi_str,
1962 efx_poll, napi_weight);
1963}
1964
1965static void efx_init_napi(struct efx_nic *efx)
1966{
1967 struct efx_channel *channel;
1968
1969 efx_for_each_channel(channel, efx)
1970 efx_init_napi_channel(channel);
1971}
1972
1973static void efx_fini_napi_channel(struct efx_channel *channel)
1974{
1975 if (channel->napi_dev)
1976 netif_napi_del(&channel->napi_str);
1977 channel->napi_dev = NULL;
1978}
1979
1980static void efx_fini_napi(struct efx_nic *efx)
1981{
1982 struct efx_channel *channel;
1983
1984 efx_for_each_channel(channel, efx)
1985 efx_fini_napi_channel(channel);
1986}
1987
1988/**************************************************************************
1989 *
1990 * Kernel netpoll interface
1991 *
1992 *************************************************************************/
1993
1994#ifdef CONFIG_NET_POLL_CONTROLLER
1995
1996/* Although in the common case interrupts will be disabled, this is not
1997 * guaranteed. However, all our work happens inside the NAPI callback,
1998 * so no locking is required.
1999 */
2000static void efx_netpoll(struct net_device *net_dev)
2001{
2002 struct efx_nic *efx = netdev_priv(net_dev);
2003 struct efx_channel *channel;
2004
2005 efx_for_each_channel(channel, efx)
2006 efx_schedule_channel(channel);
2007}
2008
2009#endif
2010
2011/**************************************************************************
2012 *
2013 * Kernel net device interface
2014 *
2015 *************************************************************************/
2016
2017/* Context: process, rtnl_lock() held. */
2018static int efx_net_open(struct net_device *net_dev)
2019{
2020 struct efx_nic *efx = netdev_priv(net_dev);
2021 int rc;
2022
2023 netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2024 raw_smp_processor_id());
2025
2026 rc = efx_check_disabled(efx);
2027 if (rc)
2028 return rc;
2029 if (efx->phy_mode & PHY_MODE_SPECIAL)
2030 return -EBUSY;
2031 if (efx_mcdi_poll_reboot(efx) && efx_reset(efx, RESET_TYPE_ALL))
2032 return -EIO;
2033
2034 /* Notify the kernel of the link state polled during driver load,
2035 * before the monitor starts running */
2036 efx_link_status_changed(efx);
2037
2038 efx_start_all(efx);
2039 efx_selftest_async_start(efx);
2040 return 0;
2041}
2042
2043/* Context: process, rtnl_lock() held.
2044 * Note that the kernel will ignore our return code; this method
2045 * should really be a void.
2046 */
2047static int efx_net_stop(struct net_device *net_dev)
2048{
2049 struct efx_nic *efx = netdev_priv(net_dev);
2050
2051 netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2052 raw_smp_processor_id());
2053
2054 /* Stop the device and flush all the channels */
2055 efx_stop_all(efx);
2056
2057 return 0;
2058}
2059
2060/* Context: process, dev_base_lock or RTNL held, non-blocking. */
2061static struct rtnl_link_stats64 *efx_net_stats(struct net_device *net_dev,
2062 struct rtnl_link_stats64 *stats)
2063{
2064 struct efx_nic *efx = netdev_priv(net_dev);
2065
2066 spin_lock_bh(&efx->stats_lock);
2067 efx->type->update_stats(efx, NULL, stats);
2068 spin_unlock_bh(&efx->stats_lock);
2069
2070 return stats;
2071}
2072
2073/* Context: netif_tx_lock held, BHs disabled. */
2074static void efx_watchdog(struct net_device *net_dev)
2075{
2076 struct efx_nic *efx = netdev_priv(net_dev);
2077
2078 netif_err(efx, tx_err, efx->net_dev,
2079 "TX stuck with port_enabled=%d: resetting channels\n",
2080 efx->port_enabled);
2081
2082 efx_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2083}
2084
2085
2086/* Context: process, rtnl_lock() held. */
2087static int efx_change_mtu(struct net_device *net_dev, int new_mtu)
2088{
2089 struct efx_nic *efx = netdev_priv(net_dev);
2090 int rc;
2091
2092 rc = efx_check_disabled(efx);
2093 if (rc)
2094 return rc;
2095 if (new_mtu > EFX_MAX_MTU)
2096 return -EINVAL;
2097
2098 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2099
2100 efx_device_detach_sync(efx);
2101 efx_stop_all(efx);
2102
2103 mutex_lock(&efx->mac_lock);
2104 net_dev->mtu = new_mtu;
2105 efx->type->reconfigure_mac(efx);
2106 mutex_unlock(&efx->mac_lock);
2107
2108 efx_start_all(efx);
2109 netif_device_attach(efx->net_dev);
2110 return 0;
2111}
2112
2113static int efx_set_mac_address(struct net_device *net_dev, void *data)
2114{
2115 struct efx_nic *efx = netdev_priv(net_dev);
2116 struct sockaddr *addr = data;
2117 u8 *new_addr = addr->sa_data;
2118
2119 if (!is_valid_ether_addr(new_addr)) {
2120 netif_err(efx, drv, efx->net_dev,
2121 "invalid ethernet MAC address requested: %pM\n",
2122 new_addr);
2123 return -EADDRNOTAVAIL;
2124 }
2125
2126 ether_addr_copy(net_dev->dev_addr, new_addr);
2127 efx_sriov_mac_address_changed(efx);
2128
2129 /* Reconfigure the MAC */
2130 mutex_lock(&efx->mac_lock);
2131 efx->type->reconfigure_mac(efx);
2132 mutex_unlock(&efx->mac_lock);
2133
2134 return 0;
2135}
2136
2137/* Context: netif_addr_lock held, BHs disabled. */
2138static void efx_set_rx_mode(struct net_device *net_dev)
2139{
2140 struct efx_nic *efx = netdev_priv(net_dev);
2141
2142 if (efx->port_enabled)
2143 queue_work(efx->workqueue, &efx->mac_work);
2144 /* Otherwise efx_start_port() will do this */
2145}
2146
2147static int efx_set_features(struct net_device *net_dev, netdev_features_t data)
2148{
2149 struct efx_nic *efx = netdev_priv(net_dev);
2150
2151 /* If disabling RX n-tuple filtering, clear existing filters */
2152 if (net_dev->features & ~data & NETIF_F_NTUPLE)
2153 return efx->type->filter_clear_rx(efx, EFX_FILTER_PRI_MANUAL);
2154
2155 return 0;
2156}
2157
2158static const struct net_device_ops efx_farch_netdev_ops = {
2159 .ndo_open = efx_net_open,
2160 .ndo_stop = efx_net_stop,
2161 .ndo_get_stats64 = efx_net_stats,
2162 .ndo_tx_timeout = efx_watchdog,
2163 .ndo_start_xmit = efx_hard_start_xmit,
2164 .ndo_validate_addr = eth_validate_addr,
2165 .ndo_do_ioctl = efx_ioctl,
2166 .ndo_change_mtu = efx_change_mtu,
2167 .ndo_set_mac_address = efx_set_mac_address,
2168 .ndo_set_rx_mode = efx_set_rx_mode,
2169 .ndo_set_features = efx_set_features,
2170#ifdef CONFIG_SFC_SRIOV
2171 .ndo_set_vf_mac = efx_sriov_set_vf_mac,
2172 .ndo_set_vf_vlan = efx_sriov_set_vf_vlan,
2173 .ndo_set_vf_spoofchk = efx_sriov_set_vf_spoofchk,
2174 .ndo_get_vf_config = efx_sriov_get_vf_config,
2175#endif
2176#ifdef CONFIG_NET_POLL_CONTROLLER
2177 .ndo_poll_controller = efx_netpoll,
2178#endif
2179 .ndo_setup_tc = efx_setup_tc,
2180#ifdef CONFIG_RFS_ACCEL
2181 .ndo_rx_flow_steer = efx_filter_rfs,
2182#endif
2183};
2184
2185static const struct net_device_ops efx_ef10_netdev_ops = {
2186 .ndo_open = efx_net_open,
2187 .ndo_stop = efx_net_stop,
2188 .ndo_get_stats64 = efx_net_stats,
2189 .ndo_tx_timeout = efx_watchdog,
2190 .ndo_start_xmit = efx_hard_start_xmit,
2191 .ndo_validate_addr = eth_validate_addr,
2192 .ndo_do_ioctl = efx_ioctl,
2193 .ndo_change_mtu = efx_change_mtu,
2194 .ndo_set_mac_address = efx_set_mac_address,
2195 .ndo_set_rx_mode = efx_set_rx_mode,
2196 .ndo_set_features = efx_set_features,
2197#ifdef CONFIG_NET_POLL_CONTROLLER
2198 .ndo_poll_controller = efx_netpoll,
2199#endif
2200#ifdef CONFIG_RFS_ACCEL
2201 .ndo_rx_flow_steer = efx_filter_rfs,
2202#endif
2203};
2204
2205static void efx_update_name(struct efx_nic *efx)
2206{
2207 strcpy(efx->name, efx->net_dev->name);
2208 efx_mtd_rename(efx);
2209 efx_set_channel_names(efx);
2210}
2211
2212static int efx_netdev_event(struct notifier_block *this,
2213 unsigned long event, void *ptr)
2214{
2215 struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2216
2217 if ((net_dev->netdev_ops == &efx_farch_netdev_ops ||
2218 net_dev->netdev_ops == &efx_ef10_netdev_ops) &&
2219 event == NETDEV_CHANGENAME)
2220 efx_update_name(netdev_priv(net_dev));
2221
2222 return NOTIFY_DONE;
2223}
2224
2225static struct notifier_block efx_netdev_notifier = {
2226 .notifier_call = efx_netdev_event,
2227};
2228
2229static ssize_t
2230show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2231{
2232 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2233 return sprintf(buf, "%d\n", efx->phy_type);
2234}
2235static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2236
2237static int efx_register_netdev(struct efx_nic *efx)
2238{
2239 struct net_device *net_dev = efx->net_dev;
2240 struct efx_channel *channel;
2241 int rc;
2242
2243 net_dev->watchdog_timeo = 5 * HZ;
2244 net_dev->irq = efx->pci_dev->irq;
2245 if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0) {
2246 net_dev->netdev_ops = &efx_ef10_netdev_ops;
2247 net_dev->priv_flags |= IFF_UNICAST_FLT;
2248 } else {
2249 net_dev->netdev_ops = &efx_farch_netdev_ops;
2250 }
2251 SET_ETHTOOL_OPS(net_dev, &efx_ethtool_ops);
2252 net_dev->gso_max_segs = EFX_TSO_MAX_SEGS;
2253
2254 rtnl_lock();
2255
2256 /* Enable resets to be scheduled and check whether any were
2257 * already requested. If so, the NIC is probably hosed so we
2258 * abort.
2259 */
2260 efx->state = STATE_READY;
2261 smp_mb(); /* ensure we change state before checking reset_pending */
2262 if (efx->reset_pending) {
2263 netif_err(efx, probe, efx->net_dev,
2264 "aborting probe due to scheduled reset\n");
2265 rc = -EIO;
2266 goto fail_locked;
2267 }
2268
2269 rc = dev_alloc_name(net_dev, net_dev->name);
2270 if (rc < 0)
2271 goto fail_locked;
2272 efx_update_name(efx);
2273
2274 /* Always start with carrier off; PHY events will detect the link */
2275 netif_carrier_off(net_dev);
2276
2277 rc = register_netdevice(net_dev);
2278 if (rc)
2279 goto fail_locked;
2280
2281 efx_for_each_channel(channel, efx) {
2282 struct efx_tx_queue *tx_queue;
2283 efx_for_each_channel_tx_queue(tx_queue, channel)
2284 efx_init_tx_queue_core_txq(tx_queue);
2285 }
2286
2287 efx_associate(efx);
2288
2289 rtnl_unlock();
2290
2291 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2292 if (rc) {
2293 netif_err(efx, drv, efx->net_dev,
2294 "failed to init net dev attributes\n");
2295 goto fail_registered;
2296 }
2297
2298 return 0;
2299
2300fail_registered:
2301 rtnl_lock();
2302 efx_dissociate(efx);
2303 unregister_netdevice(net_dev);
2304fail_locked:
2305 efx->state = STATE_UNINIT;
2306 rtnl_unlock();
2307 netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2308 return rc;
2309}
2310
2311static void efx_unregister_netdev(struct efx_nic *efx)
2312{
2313 if (!efx->net_dev)
2314 return;
2315
2316 BUG_ON(netdev_priv(efx->net_dev) != efx);
2317
2318 strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2319 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2320
2321 rtnl_lock();
2322 unregister_netdevice(efx->net_dev);
2323 efx->state = STATE_UNINIT;
2324 rtnl_unlock();
2325}
2326
2327/**************************************************************************
2328 *
2329 * Device reset and suspend
2330 *
2331 **************************************************************************/
2332
2333/* Tears down the entire software state and most of the hardware state
2334 * before reset. */
2335void efx_reset_down(struct efx_nic *efx, enum reset_type method)
2336{
2337 EFX_ASSERT_RESET_SERIALISED(efx);
2338
2339 if (method == RESET_TYPE_MCDI_TIMEOUT)
2340 efx->type->prepare_flr(efx);
2341
2342 efx_stop_all(efx);
2343 efx_disable_interrupts(efx);
2344
2345 mutex_lock(&efx->mac_lock);
2346 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE)
2347 efx->phy_op->fini(efx);
2348 efx->type->fini(efx);
2349}
2350
2351/* This function will always ensure that the locks acquired in
2352 * efx_reset_down() are released. A failure return code indicates
2353 * that we were unable to reinitialise the hardware, and the
2354 * driver should be disabled. If ok is false, then the rx and tx
2355 * engines are not restarted, pending a RESET_DISABLE. */
2356int efx_reset_up(struct efx_nic *efx, enum reset_type method, bool ok)
2357{
2358 int rc;
2359
2360 EFX_ASSERT_RESET_SERIALISED(efx);
2361
2362 if (method == RESET_TYPE_MCDI_TIMEOUT)
2363 efx->type->finish_flr(efx);
2364
2365 /* Ensure that SRAM is initialised even if we're disabling the device */
2366 rc = efx->type->init(efx);
2367 if (rc) {
2368 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2369 goto fail;
2370 }
2371
2372 if (!ok)
2373 goto fail;
2374
2375 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE) {
2376 rc = efx->phy_op->init(efx);
2377 if (rc)
2378 goto fail;
2379 if (efx->phy_op->reconfigure(efx))
2380 netif_err(efx, drv, efx->net_dev,
2381 "could not restore PHY settings\n");
2382 }
2383
2384 rc = efx_enable_interrupts(efx);
2385 if (rc)
2386 goto fail;
2387 efx_restore_filters(efx);
2388 efx_sriov_reset(efx);
2389
2390 mutex_unlock(&efx->mac_lock);
2391
2392 efx_start_all(efx);
2393
2394 return 0;
2395
2396fail:
2397 efx->port_initialized = false;
2398
2399 mutex_unlock(&efx->mac_lock);
2400
2401 return rc;
2402}
2403
2404/* Reset the NIC using the specified method. Note that the reset may
2405 * fail, in which case the card will be left in an unusable state.
2406 *
2407 * Caller must hold the rtnl_lock.
2408 */
2409int efx_reset(struct efx_nic *efx, enum reset_type method)
2410{
2411 int rc, rc2;
2412 bool disabled;
2413
2414 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2415 RESET_TYPE(method));
2416
2417 efx_device_detach_sync(efx);
2418 efx_reset_down(efx, method);
2419
2420 rc = efx->type->reset(efx, method);
2421 if (rc) {
2422 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2423 goto out;
2424 }
2425
2426 /* Clear flags for the scopes we covered. We assume the NIC and
2427 * driver are now quiescent so that there is no race here.
2428 */
2429 if (method < RESET_TYPE_MAX_METHOD)
2430 efx->reset_pending &= -(1 << (method + 1));
2431 else /* it doesn't fit into the well-ordered scope hierarchy */
2432 __clear_bit(method, &efx->reset_pending);
2433
2434 /* Reinitialise bus-mastering, which may have been turned off before
2435 * the reset was scheduled. This is still appropriate, even in the
2436 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2437 * can respond to requests. */
2438 pci_set_master(efx->pci_dev);
2439
2440out:
2441 /* Leave device stopped if necessary */
2442 disabled = rc ||
2443 method == RESET_TYPE_DISABLE ||
2444 method == RESET_TYPE_RECOVER_OR_DISABLE;
2445 rc2 = efx_reset_up(efx, method, !disabled);
2446 if (rc2) {
2447 disabled = true;
2448 if (!rc)
2449 rc = rc2;
2450 }
2451
2452 if (disabled) {
2453 dev_close(efx->net_dev);
2454 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2455 efx->state = STATE_DISABLED;
2456 } else {
2457 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2458 netif_device_attach(efx->net_dev);
2459 }
2460 return rc;
2461}
2462
2463/* Try recovery mechanisms.
2464 * For now only EEH is supported.
2465 * Returns 0 if the recovery mechanisms are unsuccessful.
2466 * Returns a non-zero value otherwise.
2467 */
2468int efx_try_recovery(struct efx_nic *efx)
2469{
2470#ifdef CONFIG_EEH
2471 /* A PCI error can occur and not be seen by EEH because nothing
2472 * happens on the PCI bus. In this case the driver may fail and
2473 * schedule a 'recover or reset', leading to this recovery handler.
2474 * Manually call the eeh failure check function.
2475 */
2476 struct eeh_dev *eehdev =
2477 of_node_to_eeh_dev(pci_device_to_OF_node(efx->pci_dev));
2478
2479 if (eeh_dev_check_failure(eehdev)) {
2480 /* The EEH mechanisms will handle the error and reset the
2481 * device if necessary.
2482 */
2483 return 1;
2484 }
2485#endif
2486 return 0;
2487}
2488
2489static void efx_wait_for_bist_end(struct efx_nic *efx)
2490{
2491 int i;
2492
2493 for (i = 0; i < BIST_WAIT_DELAY_COUNT; ++i) {
2494 if (efx_mcdi_poll_reboot(efx))
2495 goto out;
2496 msleep(BIST_WAIT_DELAY_MS);
2497 }
2498
2499 netif_err(efx, drv, efx->net_dev, "Warning: No MC reboot after BIST mode\n");
2500out:
2501 /* Either way unset the BIST flag. If we found no reboot we probably
2502 * won't recover, but we should try.
2503 */
2504 efx->mc_bist_for_other_fn = false;
2505}
2506
2507/* The worker thread exists so that code that cannot sleep can
2508 * schedule a reset for later.
2509 */
2510static void efx_reset_work(struct work_struct *data)
2511{
2512 struct efx_nic *efx = container_of(data, struct efx_nic, reset_work);
2513 unsigned long pending;
2514 enum reset_type method;
2515
2516 pending = ACCESS_ONCE(efx->reset_pending);
2517 method = fls(pending) - 1;
2518
2519 if (method == RESET_TYPE_MC_BIST)
2520 efx_wait_for_bist_end(efx);
2521
2522 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2523 method == RESET_TYPE_RECOVER_OR_ALL) &&
2524 efx_try_recovery(efx))
2525 return;
2526
2527 if (!pending)
2528 return;
2529
2530 rtnl_lock();
2531
2532 /* We checked the state in efx_schedule_reset() but it may
2533 * have changed by now. Now that we have the RTNL lock,
2534 * it cannot change again.
2535 */
2536 if (efx->state == STATE_READY)
2537 (void)efx_reset(efx, method);
2538
2539 rtnl_unlock();
2540}
2541
2542void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
2543{
2544 enum reset_type method;
2545
2546 if (efx->state == STATE_RECOVERY) {
2547 netif_dbg(efx, drv, efx->net_dev,
2548 "recovering: skip scheduling %s reset\n",
2549 RESET_TYPE(type));
2550 return;
2551 }
2552
2553 switch (type) {
2554 case RESET_TYPE_INVISIBLE:
2555 case RESET_TYPE_ALL:
2556 case RESET_TYPE_RECOVER_OR_ALL:
2557 case RESET_TYPE_WORLD:
2558 case RESET_TYPE_DISABLE:
2559 case RESET_TYPE_RECOVER_OR_DISABLE:
2560 case RESET_TYPE_MC_BIST:
2561 case RESET_TYPE_MCDI_TIMEOUT:
2562 method = type;
2563 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2564 RESET_TYPE(method));
2565 break;
2566 default:
2567 method = efx->type->map_reset_reason(type);
2568 netif_dbg(efx, drv, efx->net_dev,
2569 "scheduling %s reset for %s\n",
2570 RESET_TYPE(method), RESET_TYPE(type));
2571 break;
2572 }
2573
2574 set_bit(method, &efx->reset_pending);
2575 smp_mb(); /* ensure we change reset_pending before checking state */
2576
2577 /* If we're not READY then just leave the flags set as the cue
2578 * to abort probing or reschedule the reset later.
2579 */
2580 if (ACCESS_ONCE(efx->state) != STATE_READY)
2581 return;
2582
2583 /* efx_process_channel() will no longer read events once a
2584 * reset is scheduled. So switch back to poll'd MCDI completions. */
2585 efx_mcdi_mode_poll(efx);
2586
2587 queue_work(reset_workqueue, &efx->reset_work);
2588}
2589
2590/**************************************************************************
2591 *
2592 * List of NICs we support
2593 *
2594 **************************************************************************/
2595
2596/* PCI device ID table */
2597static DEFINE_PCI_DEVICE_TABLE(efx_pci_table) = {
2598 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2599 PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2600 .driver_data = (unsigned long) &falcon_a1_nic_type},
2601 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2602 PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2603 .driver_data = (unsigned long) &falcon_b0_nic_type},
2604 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0803), /* SFC9020 */
2605 .driver_data = (unsigned long) &siena_a0_nic_type},
2606 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0813), /* SFL9021 */
2607 .driver_data = (unsigned long) &siena_a0_nic_type},
2608 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE, 0x0903), /* SFC9120 PF */
2609 .driver_data = (unsigned long) &efx_hunt_a0_nic_type},
2610 {0} /* end of list */
2611};
2612
2613/**************************************************************************
2614 *
2615 * Dummy PHY/MAC operations
2616 *
2617 * Can be used for some unimplemented operations
2618 * Needed so all function pointers are valid and do not have to be tested
2619 * before use
2620 *
2621 **************************************************************************/
2622int efx_port_dummy_op_int(struct efx_nic *efx)
2623{
2624 return 0;
2625}
2626void efx_port_dummy_op_void(struct efx_nic *efx) {}
2627
2628static bool efx_port_dummy_op_poll(struct efx_nic *efx)
2629{
2630 return false;
2631}
2632
2633static const struct efx_phy_operations efx_dummy_phy_operations = {
2634 .init = efx_port_dummy_op_int,
2635 .reconfigure = efx_port_dummy_op_int,
2636 .poll = efx_port_dummy_op_poll,
2637 .fini = efx_port_dummy_op_void,
2638};
2639
2640/**************************************************************************
2641 *
2642 * Data housekeeping
2643 *
2644 **************************************************************************/
2645
2646/* This zeroes out and then fills in the invariants in a struct
2647 * efx_nic (including all sub-structures).
2648 */
2649static int efx_init_struct(struct efx_nic *efx,
2650 struct pci_dev *pci_dev, struct net_device *net_dev)
2651{
2652 int i;
2653
2654 /* Initialise common structures */
2655 INIT_LIST_HEAD(&efx->node);
2656 INIT_LIST_HEAD(&efx->secondary_list);
2657 spin_lock_init(&efx->biu_lock);
2658#ifdef CONFIG_SFC_MTD
2659 INIT_LIST_HEAD(&efx->mtd_list);
2660#endif
2661 INIT_WORK(&efx->reset_work, efx_reset_work);
2662 INIT_DELAYED_WORK(&efx->monitor_work, efx_monitor);
2663 INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);
2664 efx->pci_dev = pci_dev;
2665 efx->msg_enable = debug;
2666 efx->state = STATE_UNINIT;
2667 strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2668
2669 efx->net_dev = net_dev;
2670 efx->rx_prefix_size = efx->type->rx_prefix_size;
2671 efx->rx_ip_align =
2672 NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2673 efx->rx_packet_hash_offset =
2674 efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2675 efx->rx_packet_ts_offset =
2676 efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2677 spin_lock_init(&efx->stats_lock);
2678 mutex_init(&efx->mac_lock);
2679 efx->phy_op = &efx_dummy_phy_operations;
2680 efx->mdio.dev = net_dev;
2681 INIT_WORK(&efx->mac_work, efx_mac_work);
2682 init_waitqueue_head(&efx->flush_wq);
2683
2684 for (i = 0; i < EFX_MAX_CHANNELS; i++) {
2685 efx->channel[i] = efx_alloc_channel(efx, i, NULL);
2686 if (!efx->channel[i])
2687 goto fail;
2688 efx->msi_context[i].efx = efx;
2689 efx->msi_context[i].index = i;
2690 }
2691
2692 /* Higher numbered interrupt modes are less capable! */
2693 efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2694 interrupt_mode);
2695
2696 /* Would be good to use the net_dev name, but we're too early */
2697 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2698 pci_name(pci_dev));
2699 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2700 if (!efx->workqueue)
2701 goto fail;
2702
2703 return 0;
2704
2705fail:
2706 efx_fini_struct(efx);
2707 return -ENOMEM;
2708}
2709
2710static void efx_fini_struct(struct efx_nic *efx)
2711{
2712 int i;
2713
2714 for (i = 0; i < EFX_MAX_CHANNELS; i++)
2715 kfree(efx->channel[i]);
2716
2717 kfree(efx->vpd_sn);
2718
2719 if (efx->workqueue) {
2720 destroy_workqueue(efx->workqueue);
2721 efx->workqueue = NULL;
2722 }
2723}
2724
2725/**************************************************************************
2726 *
2727 * PCI interface
2728 *
2729 **************************************************************************/
2730
2731/* Main body of final NIC shutdown code
2732 * This is called only at module unload (or hotplug removal).
2733 */
2734static void efx_pci_remove_main(struct efx_nic *efx)
2735{
2736 /* Flush reset_work. It can no longer be scheduled since we
2737 * are not READY.
2738 */
2739 BUG_ON(efx->state == STATE_READY);
2740 cancel_work_sync(&efx->reset_work);
2741
2742 efx_disable_interrupts(efx);
2743 efx_nic_fini_interrupt(efx);
2744 efx_fini_port(efx);
2745 efx->type->fini(efx);
2746 efx_fini_napi(efx);
2747 efx_remove_all(efx);
2748}
2749
2750/* Final NIC shutdown
2751 * This is called only at module unload (or hotplug removal).
2752 */
2753static void efx_pci_remove(struct pci_dev *pci_dev)
2754{
2755 struct efx_nic *efx;
2756
2757 efx = pci_get_drvdata(pci_dev);
2758 if (!efx)
2759 return;
2760
2761 /* Mark the NIC as fini, then stop the interface */
2762 rtnl_lock();
2763 efx_dissociate(efx);
2764 dev_close(efx->net_dev);
2765 efx_disable_interrupts(efx);
2766 rtnl_unlock();
2767
2768 efx_sriov_fini(efx);
2769 efx_unregister_netdev(efx);
2770
2771 efx_mtd_remove(efx);
2772
2773 efx_pci_remove_main(efx);
2774
2775 efx_fini_io(efx);
2776 netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2777
2778 efx_fini_struct(efx);
2779 free_netdev(efx->net_dev);
2780
2781 pci_disable_pcie_error_reporting(pci_dev);
2782};
2783
2784/* NIC VPD information
2785 * Called during probe to display the part number of the
2786 * installed NIC. VPD is potentially very large but this should
2787 * always appear within the first 512 bytes.
2788 */
2789#define SFC_VPD_LEN 512
2790static void efx_probe_vpd_strings(struct efx_nic *efx)
2791{
2792 struct pci_dev *dev = efx->pci_dev;
2793 char vpd_data[SFC_VPD_LEN];
2794 ssize_t vpd_size;
2795 int ro_start, ro_size, i, j;
2796
2797 /* Get the vpd data from the device */
2798 vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2799 if (vpd_size <= 0) {
2800 netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2801 return;
2802 }
2803
2804 /* Get the Read only section */
2805 ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2806 if (ro_start < 0) {
2807 netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2808 return;
2809 }
2810
2811 ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
2812 j = ro_size;
2813 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2814 if (i + j > vpd_size)
2815 j = vpd_size - i;
2816
2817 /* Get the Part number */
2818 i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2819 if (i < 0) {
2820 netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2821 return;
2822 }
2823
2824 j = pci_vpd_info_field_size(&vpd_data[i]);
2825 i += PCI_VPD_INFO_FLD_HDR_SIZE;
2826 if (i + j > vpd_size) {
2827 netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2828 return;
2829 }
2830
2831 netif_info(efx, drv, efx->net_dev,
2832 "Part Number : %.*s\n", j, &vpd_data[i]);
2833
2834 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2835 j = ro_size;
2836 i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
2837 if (i < 0) {
2838 netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
2839 return;
2840 }
2841
2842 j = pci_vpd_info_field_size(&vpd_data[i]);
2843 i += PCI_VPD_INFO_FLD_HDR_SIZE;
2844 if (i + j > vpd_size) {
2845 netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
2846 return;
2847 }
2848
2849 efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
2850 if (!efx->vpd_sn)
2851 return;
2852
2853 snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
2854}
2855
2856
2857/* Main body of NIC initialisation
2858 * This is called at module load (or hotplug insertion, theoretically).
2859 */
2860static int efx_pci_probe_main(struct efx_nic *efx)
2861{
2862 int rc;
2863
2864 /* Do start-of-day initialisation */
2865 rc = efx_probe_all(efx);
2866 if (rc)
2867 goto fail1;
2868
2869 efx_init_napi(efx);
2870
2871 rc = efx->type->init(efx);
2872 if (rc) {
2873 netif_err(efx, probe, efx->net_dev,
2874 "failed to initialise NIC\n");
2875 goto fail3;
2876 }
2877
2878 rc = efx_init_port(efx);
2879 if (rc) {
2880 netif_err(efx, probe, efx->net_dev,
2881 "failed to initialise port\n");
2882 goto fail4;
2883 }
2884
2885 rc = efx_nic_init_interrupt(efx);
2886 if (rc)
2887 goto fail5;
2888 rc = efx_enable_interrupts(efx);
2889 if (rc)
2890 goto fail6;
2891
2892 return 0;
2893
2894 fail6:
2895 efx_nic_fini_interrupt(efx);
2896 fail5:
2897 efx_fini_port(efx);
2898 fail4:
2899 efx->type->fini(efx);
2900 fail3:
2901 efx_fini_napi(efx);
2902 efx_remove_all(efx);
2903 fail1:
2904 return rc;
2905}
2906
2907/* NIC initialisation
2908 *
2909 * This is called at module load (or hotplug insertion,
2910 * theoretically). It sets up PCI mappings, resets the NIC,
2911 * sets up and registers the network devices with the kernel and hooks
2912 * the interrupt service routine. It does not prepare the device for
2913 * transmission; this is left to the first time one of the network
2914 * interfaces is brought up (i.e. efx_net_open).
2915 */
2916static int efx_pci_probe(struct pci_dev *pci_dev,
2917 const struct pci_device_id *entry)
2918{
2919 struct net_device *net_dev;
2920 struct efx_nic *efx;
2921 int rc;
2922
2923 /* Allocate and initialise a struct net_device and struct efx_nic */
2924 net_dev = alloc_etherdev_mqs(sizeof(*efx), EFX_MAX_CORE_TX_QUEUES,
2925 EFX_MAX_RX_QUEUES);
2926 if (!net_dev)
2927 return -ENOMEM;
2928 efx = netdev_priv(net_dev);
2929 efx->type = (const struct efx_nic_type *) entry->driver_data;
2930 net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2931 NETIF_F_HIGHDMA | NETIF_F_TSO |
2932 NETIF_F_RXCSUM);
2933 if (efx->type->offload_features & NETIF_F_V6_CSUM)
2934 net_dev->features |= NETIF_F_TSO6;
2935 /* Mask for features that also apply to VLAN devices */
2936 net_dev->vlan_features |= (NETIF_F_ALL_CSUM | NETIF_F_SG |
2937 NETIF_F_HIGHDMA | NETIF_F_ALL_TSO |
2938 NETIF_F_RXCSUM);
2939 /* All offloads can be toggled */
2940 net_dev->hw_features = net_dev->features & ~NETIF_F_HIGHDMA;
2941 pci_set_drvdata(pci_dev, efx);
2942 SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2943 rc = efx_init_struct(efx, pci_dev, net_dev);
2944 if (rc)
2945 goto fail1;
2946
2947 netif_info(efx, probe, efx->net_dev,
2948 "Solarflare NIC detected\n");
2949
2950 efx_probe_vpd_strings(efx);
2951
2952 /* Set up basic I/O (BAR mappings etc) */
2953 rc = efx_init_io(efx);
2954 if (rc)
2955 goto fail2;
2956
2957 rc = efx_pci_probe_main(efx);
2958 if (rc)
2959 goto fail3;
2960
2961 rc = efx_register_netdev(efx);
2962 if (rc)
2963 goto fail4;
2964
2965 rc = efx_sriov_init(efx);
2966 if (rc)
2967 netif_err(efx, probe, efx->net_dev,
2968 "SR-IOV can't be enabled rc %d\n", rc);
2969
2970 netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
2971
2972 /* Try to create MTDs, but allow this to fail */
2973 rtnl_lock();
2974 rc = efx_mtd_probe(efx);
2975 rtnl_unlock();
2976 if (rc)
2977 netif_warn(efx, probe, efx->net_dev,
2978 "failed to create MTDs (%d)\n", rc);
2979
2980 rc = pci_enable_pcie_error_reporting(pci_dev);
2981 if (rc && rc != -EINVAL)
2982 netif_warn(efx, probe, efx->net_dev,
2983 "pci_enable_pcie_error_reporting failed (%d)\n", rc);
2984
2985 return 0;
2986
2987 fail4:
2988 efx_pci_remove_main(efx);
2989 fail3:
2990 efx_fini_io(efx);
2991 fail2:
2992 efx_fini_struct(efx);
2993 fail1:
2994 WARN_ON(rc > 0);
2995 netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
2996 free_netdev(net_dev);
2997 return rc;
2998}
2999
3000static int efx_pm_freeze(struct device *dev)
3001{
3002 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3003
3004 rtnl_lock();
3005
3006 if (efx->state != STATE_DISABLED) {
3007 efx->state = STATE_UNINIT;
3008
3009 efx_device_detach_sync(efx);
3010
3011 efx_stop_all(efx);
3012 efx_disable_interrupts(efx);
3013 }
3014
3015 rtnl_unlock();
3016
3017 return 0;
3018}
3019
3020static int efx_pm_thaw(struct device *dev)
3021{
3022 int rc;
3023 struct efx_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3024
3025 rtnl_lock();
3026
3027 if (efx->state != STATE_DISABLED) {
3028 rc = efx_enable_interrupts(efx);
3029 if (rc)
3030 goto fail;
3031
3032 mutex_lock(&efx->mac_lock);
3033 efx->phy_op->reconfigure(efx);
3034 mutex_unlock(&efx->mac_lock);
3035
3036 efx_start_all(efx);
3037
3038 netif_device_attach(efx->net_dev);
3039
3040 efx->state = STATE_READY;
3041
3042 efx->type->resume_wol(efx);
3043 }
3044
3045 rtnl_unlock();
3046
3047 /* Reschedule any quenched resets scheduled during efx_pm_freeze() */
3048 queue_work(reset_workqueue, &efx->reset_work);
3049
3050 return 0;
3051
3052fail:
3053 rtnl_unlock();
3054
3055 return rc;
3056}
3057
3058static int efx_pm_poweroff(struct device *dev)
3059{
3060 struct pci_dev *pci_dev = to_pci_dev(dev);
3061 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3062
3063 efx->type->fini(efx);
3064
3065 efx->reset_pending = 0;
3066
3067 pci_save_state(pci_dev);
3068 return pci_set_power_state(pci_dev, PCI_D3hot);
3069}
3070
3071/* Used for both resume and restore */
3072static int efx_pm_resume(struct device *dev)
3073{
3074 struct pci_dev *pci_dev = to_pci_dev(dev);
3075 struct efx_nic *efx = pci_get_drvdata(pci_dev);
3076 int rc;
3077
3078 rc = pci_set_power_state(pci_dev, PCI_D0);
3079 if (rc)
3080 return rc;
3081 pci_restore_state(pci_dev);
3082 rc = pci_enable_device(pci_dev);
3083 if (rc)
3084 return rc;
3085 pci_set_master(efx->pci_dev);
3086 rc = efx->type->reset(efx, RESET_TYPE_ALL);
3087 if (rc)
3088 return rc;
3089 rc = efx->type->init(efx);
3090 if (rc)
3091 return rc;
3092 rc = efx_pm_thaw(dev);
3093 return rc;
3094}
3095
3096static int efx_pm_suspend(struct device *dev)
3097{
3098 int rc;
3099
3100 efx_pm_freeze(dev);
3101 rc = efx_pm_poweroff(dev);
3102 if (rc)
3103 efx_pm_resume(dev);
3104 return rc;
3105}
3106
3107static const struct dev_pm_ops efx_pm_ops = {
3108 .suspend = efx_pm_suspend,
3109 .resume = efx_pm_resume,
3110 .freeze = efx_pm_freeze,
3111 .thaw = efx_pm_thaw,
3112 .poweroff = efx_pm_poweroff,
3113 .restore = efx_pm_resume,
3114};
3115
3116/* A PCI error affecting this device was detected.
3117 * At this point MMIO and DMA may be disabled.
3118 * Stop the software path and request a slot reset.
3119 */
3120static pci_ers_result_t efx_io_error_detected(struct pci_dev *pdev,
3121 enum pci_channel_state state)
3122{
3123 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3124 struct efx_nic *efx = pci_get_drvdata(pdev);
3125
3126 if (state == pci_channel_io_perm_failure)
3127 return PCI_ERS_RESULT_DISCONNECT;
3128
3129 rtnl_lock();
3130
3131 if (efx->state != STATE_DISABLED) {
3132 efx->state = STATE_RECOVERY;
3133 efx->reset_pending = 0;
3134
3135 efx_device_detach_sync(efx);
3136
3137 efx_stop_all(efx);
3138 efx_disable_interrupts(efx);
3139
3140 status = PCI_ERS_RESULT_NEED_RESET;
3141 } else {
3142 /* If the interface is disabled we don't want to do anything
3143 * with it.
3144 */
3145 status = PCI_ERS_RESULT_RECOVERED;
3146 }
3147
3148 rtnl_unlock();
3149
3150 pci_disable_device(pdev);
3151
3152 return status;
3153}
3154
3155/* Fake a successfull reset, which will be performed later in efx_io_resume. */
3156static pci_ers_result_t efx_io_slot_reset(struct pci_dev *pdev)
3157{
3158 struct efx_nic *efx = pci_get_drvdata(pdev);
3159 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3160 int rc;
3161
3162 if (pci_enable_device(pdev)) {
3163 netif_err(efx, hw, efx->net_dev,
3164 "Cannot re-enable PCI device after reset.\n");
3165 status = PCI_ERS_RESULT_DISCONNECT;
3166 }
3167
3168 rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3169 if (rc) {
3170 netif_err(efx, hw, efx->net_dev,
3171 "pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3172 /* Non-fatal error. Continue. */
3173 }
3174
3175 return status;
3176}
3177
3178/* Perform the actual reset and resume I/O operations. */
3179static void efx_io_resume(struct pci_dev *pdev)
3180{
3181 struct efx_nic *efx = pci_get_drvdata(pdev);
3182 int rc;
3183
3184 rtnl_lock();
3185
3186 if (efx->state == STATE_DISABLED)
3187 goto out;
3188
3189 rc = efx_reset(efx, RESET_TYPE_ALL);
3190 if (rc) {
3191 netif_err(efx, hw, efx->net_dev,
3192 "efx_reset failed after PCI error (%d)\n", rc);
3193 } else {
3194 efx->state = STATE_READY;
3195 netif_dbg(efx, hw, efx->net_dev,
3196 "Done resetting and resuming IO after PCI error.\n");
3197 }
3198
3199out:
3200 rtnl_unlock();
3201}
3202
3203/* For simplicity and reliability, we always require a slot reset and try to
3204 * reset the hardware when a pci error affecting the device is detected.
3205 * We leave both the link_reset and mmio_enabled callback unimplemented:
3206 * with our request for slot reset the mmio_enabled callback will never be
3207 * called, and the link_reset callback is not used by AER or EEH mechanisms.
3208 */
3209static struct pci_error_handlers efx_err_handlers = {
3210 .error_detected = efx_io_error_detected,
3211 .slot_reset = efx_io_slot_reset,
3212 .resume = efx_io_resume,
3213};
3214
3215static struct pci_driver efx_pci_driver = {
3216 .name = KBUILD_MODNAME,
3217 .id_table = efx_pci_table,
3218 .probe = efx_pci_probe,
3219 .remove = efx_pci_remove,
3220 .driver.pm = &efx_pm_ops,
3221 .err_handler = &efx_err_handlers,
3222};
3223
3224/**************************************************************************
3225 *
3226 * Kernel module interface
3227 *
3228 *************************************************************************/
3229
3230module_param(interrupt_mode, uint, 0444);
3231MODULE_PARM_DESC(interrupt_mode,
3232 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3233
3234static int __init efx_init_module(void)
3235{
3236 int rc;
3237
3238 printk(KERN_INFO "Solarflare NET driver v" EFX_DRIVER_VERSION "\n");
3239
3240 rc = register_netdevice_notifier(&efx_netdev_notifier);
3241 if (rc)
3242 goto err_notifier;
3243
3244 rc = efx_init_sriov();
3245 if (rc)
3246 goto err_sriov;
3247
3248 reset_workqueue = create_singlethread_workqueue("sfc_reset");
3249 if (!reset_workqueue) {
3250 rc = -ENOMEM;
3251 goto err_reset;
3252 }
3253
3254 rc = pci_register_driver(&efx_pci_driver);
3255 if (rc < 0)
3256 goto err_pci;
3257
3258 return 0;
3259
3260 err_pci:
3261 destroy_workqueue(reset_workqueue);
3262 err_reset:
3263 efx_fini_sriov();
3264 err_sriov:
3265 unregister_netdevice_notifier(&efx_netdev_notifier);
3266 err_notifier:
3267 return rc;
3268}
3269
3270static void __exit efx_exit_module(void)
3271{
3272 printk(KERN_INFO "Solarflare NET driver unloading\n");
3273
3274 pci_unregister_driver(&efx_pci_driver);
3275 destroy_workqueue(reset_workqueue);
3276 efx_fini_sriov();
3277 unregister_netdevice_notifier(&efx_netdev_notifier);
3278
3279}
3280
3281module_init(efx_init_module);
3282module_exit(efx_exit_module);
3283
3284MODULE_AUTHOR("Solarflare Communications and "
3285 "Michael Brown <mbrown@fensystems.co.uk>");
3286MODULE_DESCRIPTION("Solarflare network driver");
3287MODULE_LICENSE("GPL");
3288MODULE_DEVICE_TABLE(pci, efx_pci_table);