?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
x86/mm/pat: Don't report PAT on CPUs that don't support it
[linux/fpc-iii.git] / drivers / net / ethernet / sfc / falcon / efx.c
blob29614da91cbf919f91841d8e644ab4b246741ec7
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 "workarounds.h"
31
32 /**************************************************************************
33 *
34 * Type name strings
35 *
36 **************************************************************************
37 */
38
39 /* Loopback mode names (see LOOPBACK_MODE()) */
40 const unsigned int ef4_loopback_mode_max = LOOPBACK_MAX;
41 const char *const ef4_loopback_mode_names[] = {
42 [LOOPBACK_NONE] = "NONE",
43 [LOOPBACK_DATA] = "DATAPATH",
44 [LOOPBACK_GMAC] = "GMAC",
45 [LOOPBACK_XGMII] = "XGMII",
46 [LOOPBACK_XGXS] = "XGXS",
47 [LOOPBACK_XAUI] = "XAUI",
48 [LOOPBACK_GMII] = "GMII",
49 [LOOPBACK_SGMII] = "SGMII",
50 [LOOPBACK_XGBR] = "XGBR",
51 [LOOPBACK_XFI] = "XFI",
52 [LOOPBACK_XAUI_FAR] = "XAUI_FAR",
53 [LOOPBACK_GMII_FAR] = "GMII_FAR",
54 [LOOPBACK_SGMII_FAR] = "SGMII_FAR",
55 [LOOPBACK_XFI_FAR] = "XFI_FAR",
56 [LOOPBACK_GPHY] = "GPHY",
57 [LOOPBACK_PHYXS] = "PHYXS",
58 [LOOPBACK_PCS] = "PCS",
59 [LOOPBACK_PMAPMD] = "PMA/PMD",
60 [LOOPBACK_XPORT] = "XPORT",
61 [LOOPBACK_XGMII_WS] = "XGMII_WS",
62 [LOOPBACK_XAUI_WS] = "XAUI_WS",
63 [LOOPBACK_XAUI_WS_FAR] = "XAUI_WS_FAR",
64 [LOOPBACK_XAUI_WS_NEAR] = "XAUI_WS_NEAR",
65 [LOOPBACK_GMII_WS] = "GMII_WS",
66 [LOOPBACK_XFI_WS] = "XFI_WS",
67 [LOOPBACK_XFI_WS_FAR] = "XFI_WS_FAR",
68 [LOOPBACK_PHYXS_WS] = "PHYXS_WS",
69 };
70
71 const unsigned int ef4_reset_type_max = RESET_TYPE_MAX;
72 const char *const ef4_reset_type_names[] = {
73 [RESET_TYPE_INVISIBLE] = "INVISIBLE",
74 [RESET_TYPE_ALL] = "ALL",
75 [RESET_TYPE_RECOVER_OR_ALL] = "RECOVER_OR_ALL",
76 [RESET_TYPE_WORLD] = "WORLD",
77 [RESET_TYPE_RECOVER_OR_DISABLE] = "RECOVER_OR_DISABLE",
78 [RESET_TYPE_DATAPATH] = "DATAPATH",
79 [RESET_TYPE_DISABLE] = "DISABLE",
80 [RESET_TYPE_TX_WATCHDOG] = "TX_WATCHDOG",
81 [RESET_TYPE_INT_ERROR] = "INT_ERROR",
82 [RESET_TYPE_RX_RECOVERY] = "RX_RECOVERY",
83 [RESET_TYPE_DMA_ERROR] = "DMA_ERROR",
84 [RESET_TYPE_TX_SKIP] = "TX_SKIP",
85 };
86
87 /* Reset workqueue. If any NIC has a hardware failure then a reset will be
88 * queued onto this work queue. This is not a per-nic work queue, because
89 * ef4_reset_work() acquires the rtnl lock, so resets are naturally serialised.
90 */
91 static struct workqueue_struct *reset_workqueue;
92
93 /* How often and how many times to poll for a reset while waiting for a
94 * BIST that another function started to complete.
95 */
96 #define BIST_WAIT_DELAY_MS 100
97 #define BIST_WAIT_DELAY_COUNT 100
98
99 /**************************************************************************
100 *
101 * Configurable values
102 *
103 *************************************************************************/
104
105 /*
106 * Use separate channels for TX and RX events
107 *
108 * Set this to 1 to use separate channels for TX and RX. It allows us
109 * to control interrupt affinity separately for TX and RX.
110 *
111 * This is only used in MSI-X interrupt mode
112 */
113 bool ef4_separate_tx_channels;
114 module_param(ef4_separate_tx_channels, bool, 0444);
115 MODULE_PARM_DESC(ef4_separate_tx_channels,
116 "Use separate channels for TX and RX");
117
118 /* This is the weight assigned to each of the (per-channel) virtual
119 * NAPI devices.
120 */
121 static int napi_weight = 64;
122
123 /* This is the time (in jiffies) between invocations of the hardware
124 * monitor.
125 * On Falcon-based NICs, this will:
126 * - Check the on-board hardware monitor;
127 * - Poll the link state and reconfigure the hardware as necessary.
128 * On Siena-based NICs for power systems with EEH support, this will give EEH a
129 * chance to start.
130 */
131 static unsigned int ef4_monitor_interval = 1 * HZ;
132
133 /* Initial interrupt moderation settings. They can be modified after
134 * module load with ethtool.
135 *
136 * The default for RX should strike a balance between increasing the
137 * round-trip latency and reducing overhead.
138 */
139 static unsigned int rx_irq_mod_usec = 60;
140
141 /* Initial interrupt moderation settings. They can be modified after
142 * module load with ethtool.
143 *
144 * This default is chosen to ensure that a 10G link does not go idle
145 * while a TX queue is stopped after it has become full. A queue is
146 * restarted when it drops below half full. The time this takes (assuming
147 * worst case 3 descriptors per packet and 1024 descriptors) is
148 * 512 / 3 * 1.2 = 205 usec.
149 */
150 static unsigned int tx_irq_mod_usec = 150;
151
152 /* This is the first interrupt mode to try out of:
153 * 0 => MSI-X
154 * 1 => MSI
155 * 2 => legacy
156 */
157 static unsigned int interrupt_mode;
158
159 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
160 * i.e. the number of CPUs among which we may distribute simultaneous
161 * interrupt handling.
162 *
163 * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
164 * The default (0) means to assign an interrupt to each core.
165 */
166 static unsigned int rss_cpus;
167 module_param(rss_cpus, uint, 0444);
168 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
169
170 static bool phy_flash_cfg;
171 module_param(phy_flash_cfg, bool, 0644);
172 MODULE_PARM_DESC(phy_flash_cfg, "Set PHYs into reflash mode initially");
173
174 static unsigned irq_adapt_low_thresh = 8000;
175 module_param(irq_adapt_low_thresh, uint, 0644);
176 MODULE_PARM_DESC(irq_adapt_low_thresh,
177 "Threshold score for reducing IRQ moderation");
178
179 static unsigned irq_adapt_high_thresh = 16000;
180 module_param(irq_adapt_high_thresh, uint, 0644);
181 MODULE_PARM_DESC(irq_adapt_high_thresh,
182 "Threshold score for increasing IRQ moderation");
183
184 static unsigned debug = (NETIF_MSG_DRV | NETIF_MSG_PROBE |
185 NETIF_MSG_LINK | NETIF_MSG_IFDOWN |
186 NETIF_MSG_IFUP | NETIF_MSG_RX_ERR |
187 NETIF_MSG_TX_ERR | NETIF_MSG_HW);
188 module_param(debug, uint, 0);
189 MODULE_PARM_DESC(debug, "Bitmapped debugging message enable value");
190
191 /**************************************************************************
192 *
193 * Utility functions and prototypes
194 *
195 *************************************************************************/
196
197 static int ef4_soft_enable_interrupts(struct ef4_nic *efx);
198 static void ef4_soft_disable_interrupts(struct ef4_nic *efx);
199 static void ef4_remove_channel(struct ef4_channel *channel);
200 static void ef4_remove_channels(struct ef4_nic *efx);
201 static const struct ef4_channel_type ef4_default_channel_type;
202 static void ef4_remove_port(struct ef4_nic *efx);
203 static void ef4_init_napi_channel(struct ef4_channel *channel);
204 static void ef4_fini_napi(struct ef4_nic *efx);
205 static void ef4_fini_napi_channel(struct ef4_channel *channel);
206 static void ef4_fini_struct(struct ef4_nic *efx);
207 static void ef4_start_all(struct ef4_nic *efx);
208 static void ef4_stop_all(struct ef4_nic *efx);
209
210 #define EF4_ASSERT_RESET_SERIALISED(efx) \
211 do { \
212 if ((efx->state == STATE_READY) || \
213 (efx->state == STATE_RECOVERY) || \
214 (efx->state == STATE_DISABLED)) \
215 ASSERT_RTNL(); \
216 } while (0)
217
218 static int ef4_check_disabled(struct ef4_nic *efx)
219 {
220 if (efx->state == STATE_DISABLED || efx->state == STATE_RECOVERY) {
221 netif_err(efx, drv, efx->net_dev,
222 "device is disabled due to earlier errors\n");
223 return -EIO;
224 }
225 return 0;
226 }
227
228 /**************************************************************************
229 *
230 * Event queue processing
231 *
232 *************************************************************************/
233
234 /* Process channel's event queue
235 *
236 * This function is responsible for processing the event queue of a
237 * single channel. The caller must guarantee that this function will
238 * never be concurrently called more than once on the same channel,
239 * though different channels may be being processed concurrently.
240 */
241 static int ef4_process_channel(struct ef4_channel *channel, int budget)
242 {
243 struct ef4_tx_queue *tx_queue;
244 int spent;
245
246 if (unlikely(!channel->enabled))
247 return 0;
248
249 ef4_for_each_channel_tx_queue(tx_queue, channel) {
250 tx_queue->pkts_compl = 0;
251 tx_queue->bytes_compl = 0;
252 }
253
254 spent = ef4_nic_process_eventq(channel, budget);
255 if (spent && ef4_channel_has_rx_queue(channel)) {
256 struct ef4_rx_queue *rx_queue =
257 ef4_channel_get_rx_queue(channel);
258
259 ef4_rx_flush_packet(channel);
260 ef4_fast_push_rx_descriptors(rx_queue, true);
261 }
262
263 /* Update BQL */
264 ef4_for_each_channel_tx_queue(tx_queue, channel) {
265 if (tx_queue->bytes_compl) {
266 netdev_tx_completed_queue(tx_queue->core_txq,
267 tx_queue->pkts_compl, tx_queue->bytes_compl);
268 }
269 }
270
271 return spent;
272 }
273
274 /* NAPI poll handler
275 *
276 * NAPI guarantees serialisation of polls of the same device, which
277 * provides the guarantee required by ef4_process_channel().
278 */
279 static void ef4_update_irq_mod(struct ef4_nic *efx, struct ef4_channel *channel)
280 {
281 int step = efx->irq_mod_step_us;
282
283 if (channel->irq_mod_score < irq_adapt_low_thresh) {
284 if (channel->irq_moderation_us > step) {
285 channel->irq_moderation_us -= step;
286 efx->type->push_irq_moderation(channel);
287 }
288 } else if (channel->irq_mod_score > irq_adapt_high_thresh) {
289 if (channel->irq_moderation_us <
290 efx->irq_rx_moderation_us) {
291 channel->irq_moderation_us += step;
292 efx->type->push_irq_moderation(channel);
293 }
294 }
295
296 channel->irq_count = 0;
297 channel->irq_mod_score = 0;
298 }
299
300 static int ef4_poll(struct napi_struct *napi, int budget)
301 {
302 struct ef4_channel *channel =
303 container_of(napi, struct ef4_channel, napi_str);
304 struct ef4_nic *efx = channel->efx;
305 int spent;
306
307 netif_vdbg(efx, intr, efx->net_dev,
308 "channel %d NAPI poll executing on CPU %d\n",
309 channel->channel, raw_smp_processor_id());
310
311 spent = ef4_process_channel(channel, budget);
312
313 if (spent < budget) {
314 if (ef4_channel_has_rx_queue(channel) &&
315 efx->irq_rx_adaptive &&
316 unlikely(++channel->irq_count == 1000)) {
317 ef4_update_irq_mod(efx, channel);
318 }
319
320 ef4_filter_rfs_expire(channel);
321
322 /* There is no race here; although napi_disable() will
323 * only wait for napi_complete(), this isn't a problem
324 * since ef4_nic_eventq_read_ack() will have no effect if
325 * interrupts have already been disabled.
326 */
327 napi_complete_done(napi, spent);
328 ef4_nic_eventq_read_ack(channel);
329 }
330
331 return spent;
332 }
333
334 /* Create event queue
335 * Event queue memory allocations are done only once. If the channel
336 * is reset, the memory buffer will be reused; this guards against
337 * errors during channel reset and also simplifies interrupt handling.
338 */
339 static int ef4_probe_eventq(struct ef4_channel *channel)
340 {
341 struct ef4_nic *efx = channel->efx;
342 unsigned long entries;
343
344 netif_dbg(efx, probe, efx->net_dev,
345 "chan %d create event queue\n", channel->channel);
346
347 /* Build an event queue with room for one event per tx and rx buffer,
348 * plus some extra for link state events and MCDI completions. */
349 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
350 EF4_BUG_ON_PARANOID(entries > EF4_MAX_EVQ_SIZE);
351 channel->eventq_mask = max(entries, EF4_MIN_EVQ_SIZE) - 1;
352
353 return ef4_nic_probe_eventq(channel);
354 }
355
356 /* Prepare channel's event queue */
357 static int ef4_init_eventq(struct ef4_channel *channel)
358 {
359 struct ef4_nic *efx = channel->efx;
360 int rc;
361
362 EF4_WARN_ON_PARANOID(channel->eventq_init);
363
364 netif_dbg(efx, drv, efx->net_dev,
365 "chan %d init event queue\n", channel->channel);
366
367 rc = ef4_nic_init_eventq(channel);
368 if (rc == 0) {
369 efx->type->push_irq_moderation(channel);
370 channel->eventq_read_ptr = 0;
371 channel->eventq_init = true;
372 }
373 return rc;
374 }
375
376 /* Enable event queue processing and NAPI */
377 void ef4_start_eventq(struct ef4_channel *channel)
378 {
379 netif_dbg(channel->efx, ifup, channel->efx->net_dev,
380 "chan %d start event queue\n", channel->channel);
381
382 /* Make sure the NAPI handler sees the enabled flag set */
383 channel->enabled = true;
384 smp_wmb();
385
386 napi_enable(&channel->napi_str);
387 ef4_nic_eventq_read_ack(channel);
388 }
389
390 /* Disable event queue processing and NAPI */
391 void ef4_stop_eventq(struct ef4_channel *channel)
392 {
393 if (!channel->enabled)
394 return;
395
396 napi_disable(&channel->napi_str);
397 channel->enabled = false;
398 }
399
400 static void ef4_fini_eventq(struct ef4_channel *channel)
401 {
402 if (!channel->eventq_init)
403 return;
404
405 netif_dbg(channel->efx, drv, channel->efx->net_dev,
406 "chan %d fini event queue\n", channel->channel);
407
408 ef4_nic_fini_eventq(channel);
409 channel->eventq_init = false;
410 }
411
412 static void ef4_remove_eventq(struct ef4_channel *channel)
413 {
414 netif_dbg(channel->efx, drv, channel->efx->net_dev,
415 "chan %d remove event queue\n", channel->channel);
416
417 ef4_nic_remove_eventq(channel);
418 }
419
420 /**************************************************************************
421 *
422 * Channel handling
423 *
424 *************************************************************************/
425
426 /* Allocate and initialise a channel structure. */
427 static struct ef4_channel *
428 ef4_alloc_channel(struct ef4_nic *efx, int i, struct ef4_channel *old_channel)
429 {
430 struct ef4_channel *channel;
431 struct ef4_rx_queue *rx_queue;
432 struct ef4_tx_queue *tx_queue;
433 int j;
434
435 channel = kzalloc(sizeof(*channel), GFP_KERNEL);
436 if (!channel)
437 return NULL;
438
439 channel->efx = efx;
440 channel->channel = i;
441 channel->type = &ef4_default_channel_type;
442
443 for (j = 0; j < EF4_TXQ_TYPES; j++) {
444 tx_queue = &channel->tx_queue[j];
445 tx_queue->efx = efx;
446 tx_queue->queue = i * EF4_TXQ_TYPES + j;
447 tx_queue->channel = channel;
448 }
449
450 rx_queue = &channel->rx_queue;
451 rx_queue->efx = efx;
452 setup_timer(&rx_queue->slow_fill, ef4_rx_slow_fill,
453 (unsigned long)rx_queue);
454
455 return channel;
456 }
457
458 /* Allocate and initialise a channel structure, copying parameters
459 * (but not resources) from an old channel structure.
460 */
461 static struct ef4_channel *
462 ef4_copy_channel(const struct ef4_channel *old_channel)
463 {
464 struct ef4_channel *channel;
465 struct ef4_rx_queue *rx_queue;
466 struct ef4_tx_queue *tx_queue;
467 int j;
468
469 channel = kmalloc(sizeof(*channel), GFP_KERNEL);
470 if (!channel)
471 return NULL;
472
473 *channel = *old_channel;
474
475 channel->napi_dev = NULL;
476 INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
477 channel->napi_str.napi_id = 0;
478 channel->napi_str.state = 0;
479 memset(&channel->eventq, 0, sizeof(channel->eventq));
480
481 for (j = 0; j < EF4_TXQ_TYPES; j++) {
482 tx_queue = &channel->tx_queue[j];
483 if (tx_queue->channel)
484 tx_queue->channel = channel;
485 tx_queue->buffer = NULL;
486 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
487 }
488
489 rx_queue = &channel->rx_queue;
490 rx_queue->buffer = NULL;
491 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
492 setup_timer(&rx_queue->slow_fill, ef4_rx_slow_fill,
493 (unsigned long)rx_queue);
494
495 return channel;
496 }
497
498 static int ef4_probe_channel(struct ef4_channel *channel)
499 {
500 struct ef4_tx_queue *tx_queue;
501 struct ef4_rx_queue *rx_queue;
502 int rc;
503
504 netif_dbg(channel->efx, probe, channel->efx->net_dev,
505 "creating channel %d\n", channel->channel);
506
507 rc = channel->type->pre_probe(channel);
508 if (rc)
509 goto fail;
510
511 rc = ef4_probe_eventq(channel);
512 if (rc)
513 goto fail;
514
515 ef4_for_each_channel_tx_queue(tx_queue, channel) {
516 rc = ef4_probe_tx_queue(tx_queue);
517 if (rc)
518 goto fail;
519 }
520
521 ef4_for_each_channel_rx_queue(rx_queue, channel) {
522 rc = ef4_probe_rx_queue(rx_queue);
523 if (rc)
524 goto fail;
525 }
526
527 return 0;
528
529 fail:
530 ef4_remove_channel(channel);
531 return rc;
532 }
533
534 static void
535 ef4_get_channel_name(struct ef4_channel *channel, char *buf, size_t len)
536 {
537 struct ef4_nic *efx = channel->efx;
538 const char *type;
539 int number;
540
541 number = channel->channel;
542 if (efx->tx_channel_offset == 0) {
543 type = "";
544 } else if (channel->channel < efx->tx_channel_offset) {
545 type = "-rx";
546 } else {
547 type = "-tx";
548 number -= efx->tx_channel_offset;
549 }
550 snprintf(buf, len, "%s%s-%d", efx->name, type, number);
551 }
552
553 static void ef4_set_channel_names(struct ef4_nic *efx)
554 {
555 struct ef4_channel *channel;
556
557 ef4_for_each_channel(channel, efx)
558 channel->type->get_name(channel,
559 efx->msi_context[channel->channel].name,
560 sizeof(efx->msi_context[0].name));
561 }
562
563 static int ef4_probe_channels(struct ef4_nic *efx)
564 {
565 struct ef4_channel *channel;
566 int rc;
567
568 /* Restart special buffer allocation */
569 efx->next_buffer_table = 0;
570
571 /* Probe channels in reverse, so that any 'extra' channels
572 * use the start of the buffer table. This allows the traffic
573 * channels to be resized without moving them or wasting the
574 * entries before them.
575 */
576 ef4_for_each_channel_rev(channel, efx) {
577 rc = ef4_probe_channel(channel);
578 if (rc) {
579 netif_err(efx, probe, efx->net_dev,
580 "failed to create channel %d\n",
581 channel->channel);
582 goto fail;
583 }
584 }
585 ef4_set_channel_names(efx);
586
587 return 0;
588
589 fail:
590 ef4_remove_channels(efx);
591 return rc;
592 }
593
594 /* Channels are shutdown and reinitialised whilst the NIC is running
595 * to propagate configuration changes (mtu, checksum offload), or
596 * to clear hardware error conditions
597 */
598 static void ef4_start_datapath(struct ef4_nic *efx)
599 {
600 netdev_features_t old_features = efx->net_dev->features;
601 bool old_rx_scatter = efx->rx_scatter;
602 struct ef4_tx_queue *tx_queue;
603 struct ef4_rx_queue *rx_queue;
604 struct ef4_channel *channel;
605 size_t rx_buf_len;
606
607 /* Calculate the rx buffer allocation parameters required to
608 * support the current MTU, including padding for header
609 * alignment and overruns.
610 */
611 efx->rx_dma_len = (efx->rx_prefix_size +
612 EF4_MAX_FRAME_LEN(efx->net_dev->mtu) +
613 efx->type->rx_buffer_padding);
614 rx_buf_len = (sizeof(struct ef4_rx_page_state) +
615 efx->rx_ip_align + efx->rx_dma_len);
616 if (rx_buf_len <= PAGE_SIZE) {
617 efx->rx_scatter = efx->type->always_rx_scatter;
618 efx->rx_buffer_order = 0;
619 } else if (efx->type->can_rx_scatter) {
620 BUILD_BUG_ON(EF4_RX_USR_BUF_SIZE % L1_CACHE_BYTES);
621 BUILD_BUG_ON(sizeof(struct ef4_rx_page_state) +
622 2 * ALIGN(NET_IP_ALIGN + EF4_RX_USR_BUF_SIZE,
623 EF4_RX_BUF_ALIGNMENT) >
624 PAGE_SIZE);
625 efx->rx_scatter = true;
626 efx->rx_dma_len = EF4_RX_USR_BUF_SIZE;
627 efx->rx_buffer_order = 0;
628 } else {
629 efx->rx_scatter = false;
630 efx->rx_buffer_order = get_order(rx_buf_len);
631 }
632
633 ef4_rx_config_page_split(efx);
634 if (efx->rx_buffer_order)
635 netif_dbg(efx, drv, efx->net_dev,
636 "RX buf len=%u; page order=%u batch=%u\n",
637 efx->rx_dma_len, efx->rx_buffer_order,
638 efx->rx_pages_per_batch);
639 else
640 netif_dbg(efx, drv, efx->net_dev,
641 "RX buf len=%u step=%u bpp=%u; page batch=%u\n",
642 efx->rx_dma_len, efx->rx_page_buf_step,
643 efx->rx_bufs_per_page, efx->rx_pages_per_batch);
644
645 /* Restore previously fixed features in hw_features and remove
646 * features which are fixed now
647 */
648 efx->net_dev->hw_features |= efx->net_dev->features;
649 efx->net_dev->hw_features &= ~efx->fixed_features;
650 efx->net_dev->features |= efx->fixed_features;
651 if (efx->net_dev->features != old_features)
652 netdev_features_change(efx->net_dev);
653
654 /* RX filters may also have scatter-enabled flags */
655 if (efx->rx_scatter != old_rx_scatter)
656 efx->type->filter_update_rx_scatter(efx);
657
658 /* We must keep at least one descriptor in a TX ring empty.
659 * We could avoid this when the queue size does not exactly
660 * match the hardware ring size, but it's not that important.
661 * Therefore we stop the queue when one more skb might fill
662 * the ring completely. We wake it when half way back to
663 * empty.
664 */
665 efx->txq_stop_thresh = efx->txq_entries - ef4_tx_max_skb_descs(efx);
666 efx->txq_wake_thresh = efx->txq_stop_thresh / 2;
667
668 /* Initialise the channels */
669 ef4_for_each_channel(channel, efx) {
670 ef4_for_each_channel_tx_queue(tx_queue, channel) {
671 ef4_init_tx_queue(tx_queue);
672 atomic_inc(&efx->active_queues);
673 }
674
675 ef4_for_each_channel_rx_queue(rx_queue, channel) {
676 ef4_init_rx_queue(rx_queue);
677 atomic_inc(&efx->active_queues);
678 ef4_stop_eventq(channel);
679 ef4_fast_push_rx_descriptors(rx_queue, false);
680 ef4_start_eventq(channel);
681 }
682
683 WARN_ON(channel->rx_pkt_n_frags);
684 }
685
686 if (netif_device_present(efx->net_dev))
687 netif_tx_wake_all_queues(efx->net_dev);
688 }
689
690 static void ef4_stop_datapath(struct ef4_nic *efx)
691 {
692 struct ef4_channel *channel;
693 struct ef4_tx_queue *tx_queue;
694 struct ef4_rx_queue *rx_queue;
695 int rc;
696
697 EF4_ASSERT_RESET_SERIALISED(efx);
698 BUG_ON(efx->port_enabled);
699
700 /* Stop RX refill */
701 ef4_for_each_channel(channel, efx) {
702 ef4_for_each_channel_rx_queue(rx_queue, channel)
703 rx_queue->refill_enabled = false;
704 }
705
706 ef4_for_each_channel(channel, efx) {
707 /* RX packet processing is pipelined, so wait for the
708 * NAPI handler to complete. At least event queue 0
709 * might be kept active by non-data events, so don't
710 * use napi_synchronize() but actually disable NAPI
711 * temporarily.
712 */
713 if (ef4_channel_has_rx_queue(channel)) {
714 ef4_stop_eventq(channel);
715 ef4_start_eventq(channel);
716 }
717 }
718
719 rc = efx->type->fini_dmaq(efx);
720 if (rc && EF4_WORKAROUND_7803(efx)) {
721 /* Schedule a reset to recover from the flush failure. The
722 * descriptor caches reference memory we're about to free,
723 * but falcon_reconfigure_mac_wrapper() won't reconnect
724 * the MACs because of the pending reset.
725 */
726 netif_err(efx, drv, efx->net_dev,
727 "Resetting to recover from flush failure\n");
728 ef4_schedule_reset(efx, RESET_TYPE_ALL);
729 } else if (rc) {
730 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
731 } else {
732 netif_dbg(efx, drv, efx->net_dev,
733 "successfully flushed all queues\n");
734 }
735
736 ef4_for_each_channel(channel, efx) {
737 ef4_for_each_channel_rx_queue(rx_queue, channel)
738 ef4_fini_rx_queue(rx_queue);
739 ef4_for_each_possible_channel_tx_queue(tx_queue, channel)
740 ef4_fini_tx_queue(tx_queue);
741 }
742 }
743
744 static void ef4_remove_channel(struct ef4_channel *channel)
745 {
746 struct ef4_tx_queue *tx_queue;
747 struct ef4_rx_queue *rx_queue;
748
749 netif_dbg(channel->efx, drv, channel->efx->net_dev,
750 "destroy chan %d\n", channel->channel);
751
752 ef4_for_each_channel_rx_queue(rx_queue, channel)
753 ef4_remove_rx_queue(rx_queue);
754 ef4_for_each_possible_channel_tx_queue(tx_queue, channel)
755 ef4_remove_tx_queue(tx_queue);
756 ef4_remove_eventq(channel);
757 channel->type->post_remove(channel);
758 }
759
760 static void ef4_remove_channels(struct ef4_nic *efx)
761 {
762 struct ef4_channel *channel;
763
764 ef4_for_each_channel(channel, efx)
765 ef4_remove_channel(channel);
766 }
767
768 int
769 ef4_realloc_channels(struct ef4_nic *efx, u32 rxq_entries, u32 txq_entries)
770 {
771 struct ef4_channel *other_channel[EF4_MAX_CHANNELS], *channel;
772 u32 old_rxq_entries, old_txq_entries;
773 unsigned i, next_buffer_table = 0;
774 int rc, rc2;
775
776 rc = ef4_check_disabled(efx);
777 if (rc)
778 return rc;
779
780 /* Not all channels should be reallocated. We must avoid
781 * reallocating their buffer table entries.
782 */
783 ef4_for_each_channel(channel, efx) {
784 struct ef4_rx_queue *rx_queue;
785 struct ef4_tx_queue *tx_queue;
786
787 if (channel->type->copy)
788 continue;
789 next_buffer_table = max(next_buffer_table,
790 channel->eventq.index +
791 channel->eventq.entries);
792 ef4_for_each_channel_rx_queue(rx_queue, channel)
793 next_buffer_table = max(next_buffer_table,
794 rx_queue->rxd.index +
795 rx_queue->rxd.entries);
796 ef4_for_each_channel_tx_queue(tx_queue, channel)
797 next_buffer_table = max(next_buffer_table,
798 tx_queue->txd.index +
799 tx_queue->txd.entries);
800 }
801
802 ef4_device_detach_sync(efx);
803 ef4_stop_all(efx);
804 ef4_soft_disable_interrupts(efx);
805
806 /* Clone channels (where possible) */
807 memset(other_channel, 0, sizeof(other_channel));
808 for (i = 0; i < efx->n_channels; i++) {
809 channel = efx->channel[i];
810 if (channel->type->copy)
811 channel = channel->type->copy(channel);
812 if (!channel) {
813 rc = -ENOMEM;
814 goto out;
815 }
816 other_channel[i] = channel;
817 }
818
819 /* Swap entry counts and channel pointers */
820 old_rxq_entries = efx->rxq_entries;
821 old_txq_entries = efx->txq_entries;
822 efx->rxq_entries = rxq_entries;
823 efx->txq_entries = txq_entries;
824 for (i = 0; i < efx->n_channels; i++) {
825 channel = efx->channel[i];
826 efx->channel[i] = other_channel[i];
827 other_channel[i] = channel;
828 }
829
830 /* Restart buffer table allocation */
831 efx->next_buffer_table = next_buffer_table;
832
833 for (i = 0; i < efx->n_channels; i++) {
834 channel = efx->channel[i];
835 if (!channel->type->copy)
836 continue;
837 rc = ef4_probe_channel(channel);
838 if (rc)
839 goto rollback;
840 ef4_init_napi_channel(efx->channel[i]);
841 }
842
843 out:
844 /* Destroy unused channel structures */
845 for (i = 0; i < efx->n_channels; i++) {
846 channel = other_channel[i];
847 if (channel && channel->type->copy) {
848 ef4_fini_napi_channel(channel);
849 ef4_remove_channel(channel);
850 kfree(channel);
851 }
852 }
853
854 rc2 = ef4_soft_enable_interrupts(efx);
855 if (rc2) {
856 rc = rc ? rc : rc2;
857 netif_err(efx, drv, efx->net_dev,
858 "unable to restart interrupts on channel reallocation\n");
859 ef4_schedule_reset(efx, RESET_TYPE_DISABLE);
860 } else {
861 ef4_start_all(efx);
862 netif_device_attach(efx->net_dev);
863 }
864 return rc;
865
866 rollback:
867 /* Swap back */
868 efx->rxq_entries = old_rxq_entries;
869 efx->txq_entries = old_txq_entries;
870 for (i = 0; i < efx->n_channels; i++) {
871 channel = efx->channel[i];
872 efx->channel[i] = other_channel[i];
873 other_channel[i] = channel;
874 }
875 goto out;
876 }
877
878 void ef4_schedule_slow_fill(struct ef4_rx_queue *rx_queue)
879 {
880 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(100));
881 }
882
883 static const struct ef4_channel_type ef4_default_channel_type = {
884 .pre_probe = ef4_channel_dummy_op_int,
885 .post_remove = ef4_channel_dummy_op_void,
886 .get_name = ef4_get_channel_name,
887 .copy = ef4_copy_channel,
888 .keep_eventq = false,
889 };
890
891 int ef4_channel_dummy_op_int(struct ef4_channel *channel)
892 {
893 return 0;
894 }
895
896 void ef4_channel_dummy_op_void(struct ef4_channel *channel)
897 {
898 }
899
900 /**************************************************************************
901 *
902 * Port handling
903 *
904 **************************************************************************/
905
906 /* This ensures that the kernel is kept informed (via
907 * netif_carrier_on/off) of the link status, and also maintains the
908 * link status's stop on the port's TX queue.
909 */
910 void ef4_link_status_changed(struct ef4_nic *efx)
911 {
912 struct ef4_link_state *link_state = &efx->link_state;
913
914 /* SFC Bug 5356: A net_dev notifier is registered, so we must ensure
915 * that no events are triggered between unregister_netdev() and the
916 * driver unloading. A more general condition is that NETDEV_CHANGE
917 * can only be generated between NETDEV_UP and NETDEV_DOWN */
918 if (!netif_running(efx->net_dev))
919 return;
920
921 if (link_state->up != netif_carrier_ok(efx->net_dev)) {
922 efx->n_link_state_changes++;
923
924 if (link_state->up)
925 netif_carrier_on(efx->net_dev);
926 else
927 netif_carrier_off(efx->net_dev);
928 }
929
930 /* Status message for kernel log */
931 if (link_state->up)
932 netif_info(efx, link, efx->net_dev,
933 "link up at %uMbps %s-duplex (MTU %d)\n",
934 link_state->speed, link_state->fd ? "full" : "half",
935 efx->net_dev->mtu);
936 else
937 netif_info(efx, link, efx->net_dev, "link down\n");
938 }
939
940 void ef4_link_set_advertising(struct ef4_nic *efx, u32 advertising)
941 {
942 efx->link_advertising = advertising;
943 if (advertising) {
944 if (advertising & ADVERTISED_Pause)
945 efx->wanted_fc |= (EF4_FC_TX | EF4_FC_RX);
946 else
947 efx->wanted_fc &= ~(EF4_FC_TX | EF4_FC_RX);
948 if (advertising & ADVERTISED_Asym_Pause)
949 efx->wanted_fc ^= EF4_FC_TX;
950 }
951 }
952
953 void ef4_link_set_wanted_fc(struct ef4_nic *efx, u8 wanted_fc)
954 {
955 efx->wanted_fc = wanted_fc;
956 if (efx->link_advertising) {
957 if (wanted_fc & EF4_FC_RX)
958 efx->link_advertising |= (ADVERTISED_Pause |
959 ADVERTISED_Asym_Pause);
960 else
961 efx->link_advertising &= ~(ADVERTISED_Pause |
962 ADVERTISED_Asym_Pause);
963 if (wanted_fc & EF4_FC_TX)
964 efx->link_advertising ^= ADVERTISED_Asym_Pause;
965 }
966 }
967
968 static void ef4_fini_port(struct ef4_nic *efx);
969
970 /* We assume that efx->type->reconfigure_mac will always try to sync RX
971 * filters and therefore needs to read-lock the filter table against freeing
972 */
973 void ef4_mac_reconfigure(struct ef4_nic *efx)
974 {
975 down_read(&efx->filter_sem);
976 efx->type->reconfigure_mac(efx);
977 up_read(&efx->filter_sem);
978 }
979
980 /* Push loopback/power/transmit disable settings to the PHY, and reconfigure
981 * the MAC appropriately. All other PHY configuration changes are pushed
982 * through phy_op->set_link_ksettings(), and pushed asynchronously to the MAC
983 * through ef4_monitor().
984 *
985 * Callers must hold the mac_lock
986 */
987 int __ef4_reconfigure_port(struct ef4_nic *efx)
988 {
989 enum ef4_phy_mode phy_mode;
990 int rc;
991
992 WARN_ON(!mutex_is_locked(&efx->mac_lock));
993
994 /* Disable PHY transmit in mac level loopbacks */
995 phy_mode = efx->phy_mode;
996 if (LOOPBACK_INTERNAL(efx))
997 efx->phy_mode |= PHY_MODE_TX_DISABLED;
998 else
999 efx->phy_mode &= ~PHY_MODE_TX_DISABLED;
1000
1001 rc = efx->type->reconfigure_port(efx);
1002
1003 if (rc)
1004 efx->phy_mode = phy_mode;
1005
1006 return rc;
1007 }
1008
1009 /* Reinitialise the MAC to pick up new PHY settings, even if the port is
1010 * disabled. */
1011 int ef4_reconfigure_port(struct ef4_nic *efx)
1012 {
1013 int rc;
1014
1015 EF4_ASSERT_RESET_SERIALISED(efx);
1016
1017 mutex_lock(&efx->mac_lock);
1018 rc = __ef4_reconfigure_port(efx);
1019 mutex_unlock(&efx->mac_lock);
1020
1021 return rc;
1022 }
1023
1024 /* Asynchronous work item for changing MAC promiscuity and multicast
1025 * hash. Avoid a drain/rx_ingress enable by reconfiguring the current
1026 * MAC directly. */
1027 static void ef4_mac_work(struct work_struct *data)
1028 {
1029 struct ef4_nic *efx = container_of(data, struct ef4_nic, mac_work);
1030
1031 mutex_lock(&efx->mac_lock);
1032 if (efx->port_enabled)
1033 ef4_mac_reconfigure(efx);
1034 mutex_unlock(&efx->mac_lock);
1035 }
1036
1037 static int ef4_probe_port(struct ef4_nic *efx)
1038 {
1039 int rc;
1040
1041 netif_dbg(efx, probe, efx->net_dev, "create port\n");
1042
1043 if (phy_flash_cfg)
1044 efx->phy_mode = PHY_MODE_SPECIAL;
1045
1046 /* Connect up MAC/PHY operations table */
1047 rc = efx->type->probe_port(efx);
1048 if (rc)
1049 return rc;
1050
1051 /* Initialise MAC address to permanent address */
1052 ether_addr_copy(efx->net_dev->dev_addr, efx->net_dev->perm_addr);
1053
1054 return 0;
1055 }
1056
1057 static int ef4_init_port(struct ef4_nic *efx)
1058 {
1059 int rc;
1060
1061 netif_dbg(efx, drv, efx->net_dev, "init port\n");
1062
1063 mutex_lock(&efx->mac_lock);
1064
1065 rc = efx->phy_op->init(efx);
1066 if (rc)
1067 goto fail1;
1068
1069 efx->port_initialized = true;
1070
1071 /* Reconfigure the MAC before creating dma queues (required for
1072 * Falcon/A1 where RX_INGR_EN/TX_DRAIN_EN isn't supported) */
1073 ef4_mac_reconfigure(efx);
1074
1075 /* Ensure the PHY advertises the correct flow control settings */
1076 rc = efx->phy_op->reconfigure(efx);
1077 if (rc && rc != -EPERM)
1078 goto fail2;
1079
1080 mutex_unlock(&efx->mac_lock);
1081 return 0;
1082
1083 fail2:
1084 efx->phy_op->fini(efx);
1085 fail1:
1086 mutex_unlock(&efx->mac_lock);
1087 return rc;
1088 }
1089
1090 static void ef4_start_port(struct ef4_nic *efx)
1091 {
1092 netif_dbg(efx, ifup, efx->net_dev, "start port\n");
1093 BUG_ON(efx->port_enabled);
1094
1095 mutex_lock(&efx->mac_lock);
1096 efx->port_enabled = true;
1097
1098 /* Ensure MAC ingress/egress is enabled */
1099 ef4_mac_reconfigure(efx);
1100
1101 mutex_unlock(&efx->mac_lock);
1102 }
1103
1104 /* Cancel work for MAC reconfiguration, periodic hardware monitoring
1105 * and the async self-test, wait for them to finish and prevent them
1106 * being scheduled again. This doesn't cover online resets, which
1107 * should only be cancelled when removing the device.
1108 */
1109 static void ef4_stop_port(struct ef4_nic *efx)
1110 {
1111 netif_dbg(efx, ifdown, efx->net_dev, "stop port\n");
1112
1113 EF4_ASSERT_RESET_SERIALISED(efx);
1114
1115 mutex_lock(&efx->mac_lock);
1116 efx->port_enabled = false;
1117 mutex_unlock(&efx->mac_lock);
1118
1119 /* Serialise against ef4_set_multicast_list() */
1120 netif_addr_lock_bh(efx->net_dev);
1121 netif_addr_unlock_bh(efx->net_dev);
1122
1123 cancel_delayed_work_sync(&efx->monitor_work);
1124 ef4_selftest_async_cancel(efx);
1125 cancel_work_sync(&efx->mac_work);
1126 }
1127
1128 static void ef4_fini_port(struct ef4_nic *efx)
1129 {
1130 netif_dbg(efx, drv, efx->net_dev, "shut down port\n");
1131
1132 if (!efx->port_initialized)
1133 return;
1134
1135 efx->phy_op->fini(efx);
1136 efx->port_initialized = false;
1137
1138 efx->link_state.up = false;
1139 ef4_link_status_changed(efx);
1140 }
1141
1142 static void ef4_remove_port(struct ef4_nic *efx)
1143 {
1144 netif_dbg(efx, drv, efx->net_dev, "destroying port\n");
1145
1146 efx->type->remove_port(efx);
1147 }
1148
1149 /**************************************************************************
1150 *
1151 * NIC handling
1152 *
1153 **************************************************************************/
1154
1155 static LIST_HEAD(ef4_primary_list);
1156 static LIST_HEAD(ef4_unassociated_list);
1157
1158 static bool ef4_same_controller(struct ef4_nic *left, struct ef4_nic *right)
1159 {
1160 return left->type == right->type &&
1161 left->vpd_sn && right->vpd_sn &&
1162 !strcmp(left->vpd_sn, right->vpd_sn);
1163 }
1164
1165 static void ef4_associate(struct ef4_nic *efx)
1166 {
1167 struct ef4_nic *other, *next;
1168
1169 if (efx->primary == efx) {
1170 /* Adding primary function; look for secondaries */
1171
1172 netif_dbg(efx, probe, efx->net_dev, "adding to primary list\n");
1173 list_add_tail(&efx->node, &ef4_primary_list);
1174
1175 list_for_each_entry_safe(other, next, &ef4_unassociated_list,
1176 node) {
1177 if (ef4_same_controller(efx, other)) {
1178 list_del(&other->node);
1179 netif_dbg(other, probe, other->net_dev,
1180 "moving to secondary list of %s %s\n",
1181 pci_name(efx->pci_dev),
1182 efx->net_dev->name);
1183 list_add_tail(&other->node,
1184 &efx->secondary_list);
1185 other->primary = efx;
1186 }
1187 }
1188 } else {
1189 /* Adding secondary function; look for primary */
1190
1191 list_for_each_entry(other, &ef4_primary_list, node) {
1192 if (ef4_same_controller(efx, other)) {
1193 netif_dbg(efx, probe, efx->net_dev,
1194 "adding to secondary list of %s %s\n",
1195 pci_name(other->pci_dev),
1196 other->net_dev->name);
1197 list_add_tail(&efx->node,
1198 &other->secondary_list);
1199 efx->primary = other;
1200 return;
1201 }
1202 }
1203
1204 netif_dbg(efx, probe, efx->net_dev,
1205 "adding to unassociated list\n");
1206 list_add_tail(&efx->node, &ef4_unassociated_list);
1207 }
1208 }
1209
1210 static void ef4_dissociate(struct ef4_nic *efx)
1211 {
1212 struct ef4_nic *other, *next;
1213
1214 list_del(&efx->node);
1215 efx->primary = NULL;
1216
1217 list_for_each_entry_safe(other, next, &efx->secondary_list, node) {
1218 list_del(&other->node);
1219 netif_dbg(other, probe, other->net_dev,
1220 "moving to unassociated list\n");
1221 list_add_tail(&other->node, &ef4_unassociated_list);
1222 other->primary = NULL;
1223 }
1224 }
1225
1226 /* This configures the PCI device to enable I/O and DMA. */
1227 static int ef4_init_io(struct ef4_nic *efx)
1228 {
1229 struct pci_dev *pci_dev = efx->pci_dev;
1230 dma_addr_t dma_mask = efx->type->max_dma_mask;
1231 unsigned int mem_map_size = efx->type->mem_map_size(efx);
1232 int rc, bar;
1233
1234 netif_dbg(efx, probe, efx->net_dev, "initialising I/O\n");
1235
1236 bar = efx->type->mem_bar;
1237
1238 rc = pci_enable_device(pci_dev);
1239 if (rc) {
1240 netif_err(efx, probe, efx->net_dev,
1241 "failed to enable PCI device\n");
1242 goto fail1;
1243 }
1244
1245 pci_set_master(pci_dev);
1246
1247 /* Set the PCI DMA mask. Try all possibilities from our
1248 * genuine mask down to 32 bits, because some architectures
1249 * (e.g. x86_64 with iommu_sac_force set) will allow 40 bit
1250 * masks event though they reject 46 bit masks.
1251 */
1252 while (dma_mask > 0x7fffffffUL) {
1253 rc = dma_set_mask_and_coherent(&pci_dev->dev, dma_mask);
1254 if (rc == 0)
1255 break;
1256 dma_mask >>= 1;
1257 }
1258 if (rc) {
1259 netif_err(efx, probe, efx->net_dev,
1260 "could not find a suitable DMA mask\n");
1261 goto fail2;
1262 }
1263 netif_dbg(efx, probe, efx->net_dev,
1264 "using DMA mask %llx\n", (unsigned long long) dma_mask);
1265
1266 efx->membase_phys = pci_resource_start(efx->pci_dev, bar);
1267 rc = pci_request_region(pci_dev, bar, "sfc");
1268 if (rc) {
1269 netif_err(efx, probe, efx->net_dev,
1270 "request for memory BAR failed\n");
1271 rc = -EIO;
1272 goto fail3;
1273 }
1274 efx->membase = ioremap_nocache(efx->membase_phys, mem_map_size);
1275 if (!efx->membase) {
1276 netif_err(efx, probe, efx->net_dev,
1277 "could not map memory BAR at %llx+%x\n",
1278 (unsigned long long)efx->membase_phys, mem_map_size);
1279 rc = -ENOMEM;
1280 goto fail4;
1281 }
1282 netif_dbg(efx, probe, efx->net_dev,
1283 "memory BAR at %llx+%x (virtual %p)\n",
1284 (unsigned long long)efx->membase_phys, mem_map_size,
1285 efx->membase);
1286
1287 return 0;
1288
1289 fail4:
1290 pci_release_region(efx->pci_dev, bar);
1291 fail3:
1292 efx->membase_phys = 0;
1293 fail2:
1294 pci_disable_device(efx->pci_dev);
1295 fail1:
1296 return rc;
1297 }
1298
1299 static void ef4_fini_io(struct ef4_nic *efx)
1300 {
1301 int bar;
1302
1303 netif_dbg(efx, drv, efx->net_dev, "shutting down I/O\n");
1304
1305 if (efx->membase) {
1306 iounmap(efx->membase);
1307 efx->membase = NULL;
1308 }
1309
1310 if (efx->membase_phys) {
1311 bar = efx->type->mem_bar;
1312 pci_release_region(efx->pci_dev, bar);
1313 efx->membase_phys = 0;
1314 }
1315
1316 /* Don't disable bus-mastering if VFs are assigned */
1317 if (!pci_vfs_assigned(efx->pci_dev))
1318 pci_disable_device(efx->pci_dev);
1319 }
1320
1321 void ef4_set_default_rx_indir_table(struct ef4_nic *efx)
1322 {
1323 size_t i;
1324
1325 for (i = 0; i < ARRAY_SIZE(efx->rx_indir_table); i++)
1326 efx->rx_indir_table[i] =
1327 ethtool_rxfh_indir_default(i, efx->rss_spread);
1328 }
1329
1330 static unsigned int ef4_wanted_parallelism(struct ef4_nic *efx)
1331 {
1332 cpumask_var_t thread_mask;
1333 unsigned int count;
1334 int cpu;
1335
1336 if (rss_cpus) {
1337 count = rss_cpus;
1338 } else {
1339 if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
1340 netif_warn(efx, probe, efx->net_dev,
1341 "RSS disabled due to allocation failure\n");
1342 return 1;
1343 }
1344
1345 count = 0;
1346 for_each_online_cpu(cpu) {
1347 if (!cpumask_test_cpu(cpu, thread_mask)) {
1348 ++count;
1349 cpumask_or(thread_mask, thread_mask,
1350 topology_sibling_cpumask(cpu));
1351 }
1352 }
1353
1354 free_cpumask_var(thread_mask);
1355 }
1356
1357 if (count > EF4_MAX_RX_QUEUES) {
1358 netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
1359 "Reducing number of rx queues from %u to %u.\n",
1360 count, EF4_MAX_RX_QUEUES);
1361 count = EF4_MAX_RX_QUEUES;
1362 }
1363
1364 return count;
1365 }
1366
1367 /* Probe the number and type of interrupts we are able to obtain, and
1368 * the resulting numbers of channels and RX queues.
1369 */
1370 static int ef4_probe_interrupts(struct ef4_nic *efx)
1371 {
1372 unsigned int extra_channels = 0;
1373 unsigned int i, j;
1374 int rc;
1375
1376 for (i = 0; i < EF4_MAX_EXTRA_CHANNELS; i++)
1377 if (efx->extra_channel_type[i])
1378 ++extra_channels;
1379
1380 if (efx->interrupt_mode == EF4_INT_MODE_MSIX) {
1381 struct msix_entry xentries[EF4_MAX_CHANNELS];
1382 unsigned int n_channels;
1383
1384 n_channels = ef4_wanted_parallelism(efx);
1385 if (ef4_separate_tx_channels)
1386 n_channels *= 2;
1387 n_channels += extra_channels;
1388 n_channels = min(n_channels, efx->max_channels);
1389
1390 for (i = 0; i < n_channels; i++)
1391 xentries[i].entry = i;
1392 rc = pci_enable_msix_range(efx->pci_dev,
1393 xentries, 1, n_channels);
1394 if (rc < 0) {
1395 /* Fall back to single channel MSI */
1396 efx->interrupt_mode = EF4_INT_MODE_MSI;
1397 netif_err(efx, drv, efx->net_dev,
1398 "could not enable MSI-X\n");
1399 } else if (rc < n_channels) {
1400 netif_err(efx, drv, efx->net_dev,
1401 "WARNING: Insufficient MSI-X vectors"
1402 " available (%d < %u).\n", rc, n_channels);
1403 netif_err(efx, drv, efx->net_dev,
1404 "WARNING: Performance may be reduced.\n");
1405 n_channels = rc;
1406 }
1407
1408 if (rc > 0) {
1409 efx->n_channels = n_channels;
1410 if (n_channels > extra_channels)
1411 n_channels -= extra_channels;
1412 if (ef4_separate_tx_channels) {
1413 efx->n_tx_channels = min(max(n_channels / 2,
1414 1U),
1415 efx->max_tx_channels);
1416 efx->n_rx_channels = max(n_channels -
1417 efx->n_tx_channels,
1418 1U);
1419 } else {
1420 efx->n_tx_channels = min(n_channels,
1421 efx->max_tx_channels);
1422 efx->n_rx_channels = n_channels;
1423 }
1424 for (i = 0; i < efx->n_channels; i++)
1425 ef4_get_channel(efx, i)->irq =
1426 xentries[i].vector;
1427 }
1428 }
1429
1430 /* Try single interrupt MSI */
1431 if (efx->interrupt_mode == EF4_INT_MODE_MSI) {
1432 efx->n_channels = 1;
1433 efx->n_rx_channels = 1;
1434 efx->n_tx_channels = 1;
1435 rc = pci_enable_msi(efx->pci_dev);
1436 if (rc == 0) {
1437 ef4_get_channel(efx, 0)->irq = efx->pci_dev->irq;
1438 } else {
1439 netif_err(efx, drv, efx->net_dev,
1440 "could not enable MSI\n");
1441 efx->interrupt_mode = EF4_INT_MODE_LEGACY;
1442 }
1443 }
1444
1445 /* Assume legacy interrupts */
1446 if (efx->interrupt_mode == EF4_INT_MODE_LEGACY) {
1447 efx->n_channels = 1 + (ef4_separate_tx_channels ? 1 : 0);
1448 efx->n_rx_channels = 1;
1449 efx->n_tx_channels = 1;
1450 efx->legacy_irq = efx->pci_dev->irq;
1451 }
1452
1453 /* Assign extra channels if possible */
1454 j = efx->n_channels;
1455 for (i = 0; i < EF4_MAX_EXTRA_CHANNELS; i++) {
1456 if (!efx->extra_channel_type[i])
1457 continue;
1458 if (efx->interrupt_mode != EF4_INT_MODE_MSIX ||
1459 efx->n_channels <= extra_channels) {
1460 efx->extra_channel_type[i]->handle_no_channel(efx);
1461 } else {
1462 --j;
1463 ef4_get_channel(efx, j)->type =
1464 efx->extra_channel_type[i];
1465 }
1466 }
1467
1468 efx->rss_spread = efx->n_rx_channels;
1469
1470 return 0;
1471 }
1472
1473 static int ef4_soft_enable_interrupts(struct ef4_nic *efx)
1474 {
1475 struct ef4_channel *channel, *end_channel;
1476 int rc;
1477
1478 BUG_ON(efx->state == STATE_DISABLED);
1479
1480 efx->irq_soft_enabled = true;
1481 smp_wmb();
1482
1483 ef4_for_each_channel(channel, efx) {
1484 if (!channel->type->keep_eventq) {
1485 rc = ef4_init_eventq(channel);
1486 if (rc)
1487 goto fail;
1488 }
1489 ef4_start_eventq(channel);
1490 }
1491
1492 return 0;
1493 fail:
1494 end_channel = channel;
1495 ef4_for_each_channel(channel, efx) {
1496 if (channel == end_channel)
1497 break;
1498 ef4_stop_eventq(channel);
1499 if (!channel->type->keep_eventq)
1500 ef4_fini_eventq(channel);
1501 }
1502
1503 return rc;
1504 }
1505
1506 static void ef4_soft_disable_interrupts(struct ef4_nic *efx)
1507 {
1508 struct ef4_channel *channel;
1509
1510 if (efx->state == STATE_DISABLED)
1511 return;
1512
1513 efx->irq_soft_enabled = false;
1514 smp_wmb();
1515
1516 if (efx->legacy_irq)
1517 synchronize_irq(efx->legacy_irq);
1518
1519 ef4_for_each_channel(channel, efx) {
1520 if (channel->irq)
1521 synchronize_irq(channel->irq);
1522
1523 ef4_stop_eventq(channel);
1524 if (!channel->type->keep_eventq)
1525 ef4_fini_eventq(channel);
1526 }
1527 }
1528
1529 static int ef4_enable_interrupts(struct ef4_nic *efx)
1530 {
1531 struct ef4_channel *channel, *end_channel;
1532 int rc;
1533
1534 BUG_ON(efx->state == STATE_DISABLED);
1535
1536 if (efx->eeh_disabled_legacy_irq) {
1537 enable_irq(efx->legacy_irq);
1538 efx->eeh_disabled_legacy_irq = false;
1539 }
1540
1541 efx->type->irq_enable_master(efx);
1542
1543 ef4_for_each_channel(channel, efx) {
1544 if (channel->type->keep_eventq) {
1545 rc = ef4_init_eventq(channel);
1546 if (rc)
1547 goto fail;
1548 }
1549 }
1550
1551 rc = ef4_soft_enable_interrupts(efx);
1552 if (rc)
1553 goto fail;
1554
1555 return 0;
1556
1557 fail:
1558 end_channel = channel;
1559 ef4_for_each_channel(channel, efx) {
1560 if (channel == end_channel)
1561 break;
1562 if (channel->type->keep_eventq)
1563 ef4_fini_eventq(channel);
1564 }
1565
1566 efx->type->irq_disable_non_ev(efx);
1567
1568 return rc;
1569 }
1570
1571 static void ef4_disable_interrupts(struct ef4_nic *efx)
1572 {
1573 struct ef4_channel *channel;
1574
1575 ef4_soft_disable_interrupts(efx);
1576
1577 ef4_for_each_channel(channel, efx) {
1578 if (channel->type->keep_eventq)
1579 ef4_fini_eventq(channel);
1580 }
1581
1582 efx->type->irq_disable_non_ev(efx);
1583 }
1584
1585 static void ef4_remove_interrupts(struct ef4_nic *efx)
1586 {
1587 struct ef4_channel *channel;
1588
1589 /* Remove MSI/MSI-X interrupts */
1590 ef4_for_each_channel(channel, efx)
1591 channel->irq = 0;
1592 pci_disable_msi(efx->pci_dev);
1593 pci_disable_msix(efx->pci_dev);
1594
1595 /* Remove legacy interrupt */
1596 efx->legacy_irq = 0;
1597 }
1598
1599 static void ef4_set_channels(struct ef4_nic *efx)
1600 {
1601 struct ef4_channel *channel;
1602 struct ef4_tx_queue *tx_queue;
1603
1604 efx->tx_channel_offset =
1605 ef4_separate_tx_channels ?
1606 efx->n_channels - efx->n_tx_channels : 0;
1607
1608 /* We need to mark which channels really have RX and TX
1609 * queues, and adjust the TX queue numbers if we have separate
1610 * RX-only and TX-only channels.
1611 */
1612 ef4_for_each_channel(channel, efx) {
1613 if (channel->channel < efx->n_rx_channels)
1614 channel->rx_queue.core_index = channel->channel;
1615 else
1616 channel->rx_queue.core_index = -1;
1617
1618 ef4_for_each_channel_tx_queue(tx_queue, channel)
1619 tx_queue->queue -= (efx->tx_channel_offset *
1620 EF4_TXQ_TYPES);
1621 }
1622 }
1623
1624 static int ef4_probe_nic(struct ef4_nic *efx)
1625 {
1626 int rc;
1627
1628 netif_dbg(efx, probe, efx->net_dev, "creating NIC\n");
1629
1630 /* Carry out hardware-type specific initialisation */
1631 rc = efx->type->probe(efx);
1632 if (rc)
1633 return rc;
1634
1635 do {
1636 if (!efx->max_channels || !efx->max_tx_channels) {
1637 netif_err(efx, drv, efx->net_dev,
1638 "Insufficient resources to allocate"
1639 " any channels\n");
1640 rc = -ENOSPC;
1641 goto fail1;
1642 }
1643
1644 /* Determine the number of channels and queues by trying
1645 * to hook in MSI-X interrupts.
1646 */
1647 rc = ef4_probe_interrupts(efx);
1648 if (rc)
1649 goto fail1;
1650
1651 ef4_set_channels(efx);
1652
1653 /* dimension_resources can fail with EAGAIN */
1654 rc = efx->type->dimension_resources(efx);
1655 if (rc != 0 && rc != -EAGAIN)
1656 goto fail2;
1657
1658 if (rc == -EAGAIN)
1659 /* try again with new max_channels */
1660 ef4_remove_interrupts(efx);
1661
1662 } while (rc == -EAGAIN);
1663
1664 if (efx->n_channels > 1)
1665 netdev_rss_key_fill(&efx->rx_hash_key,
1666 sizeof(efx->rx_hash_key));
1667 ef4_set_default_rx_indir_table(efx);
1668
1669 netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
1670 netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
1671
1672 /* Initialise the interrupt moderation settings */
1673 efx->irq_mod_step_us = DIV_ROUND_UP(efx->timer_quantum_ns, 1000);
1674 ef4_init_irq_moderation(efx, tx_irq_mod_usec, rx_irq_mod_usec, true,
1675 true);
1676
1677 return 0;
1678
1679 fail2:
1680 ef4_remove_interrupts(efx);
1681 fail1:
1682 efx->type->remove(efx);
1683 return rc;
1684 }
1685
1686 static void ef4_remove_nic(struct ef4_nic *efx)
1687 {
1688 netif_dbg(efx, drv, efx->net_dev, "destroying NIC\n");
1689
1690 ef4_remove_interrupts(efx);
1691 efx->type->remove(efx);
1692 }
1693
1694 static int ef4_probe_filters(struct ef4_nic *efx)
1695 {
1696 int rc;
1697
1698 spin_lock_init(&efx->filter_lock);
1699 init_rwsem(&efx->filter_sem);
1700 mutex_lock(&efx->mac_lock);
1701 down_write(&efx->filter_sem);
1702 rc = efx->type->filter_table_probe(efx);
1703 if (rc)
1704 goto out_unlock;
1705
1706 #ifdef CONFIG_RFS_ACCEL
1707 if (efx->type->offload_features & NETIF_F_NTUPLE) {
1708 struct ef4_channel *channel;
1709 int i, success = 1;
1710
1711 ef4_for_each_channel(channel, efx) {
1712 channel->rps_flow_id =
1713 kcalloc(efx->type->max_rx_ip_filters,
1714 sizeof(*channel->rps_flow_id),
1715 GFP_KERNEL);
1716 if (!channel->rps_flow_id)
1717 success = 0;
1718 else
1719 for (i = 0;
1720 i < efx->type->max_rx_ip_filters;
1721 ++i)
1722 channel->rps_flow_id[i] =
1723 RPS_FLOW_ID_INVALID;
1724 }
1725
1726 if (!success) {
1727 ef4_for_each_channel(channel, efx)
1728 kfree(channel->rps_flow_id);
1729 efx->type->filter_table_remove(efx);
1730 rc = -ENOMEM;
1731 goto out_unlock;
1732 }
1733
1734 efx->rps_expire_index = efx->rps_expire_channel = 0;
1735 }
1736 #endif
1737 out_unlock:
1738 up_write(&efx->filter_sem);
1739 mutex_unlock(&efx->mac_lock);
1740 return rc;
1741 }
1742
1743 static void ef4_remove_filters(struct ef4_nic *efx)
1744 {
1745 #ifdef CONFIG_RFS_ACCEL
1746 struct ef4_channel *channel;
1747
1748 ef4_for_each_channel(channel, efx)
1749 kfree(channel->rps_flow_id);
1750 #endif
1751 down_write(&efx->filter_sem);
1752 efx->type->filter_table_remove(efx);
1753 up_write(&efx->filter_sem);
1754 }
1755
1756 static void ef4_restore_filters(struct ef4_nic *efx)
1757 {
1758 down_read(&efx->filter_sem);
1759 efx->type->filter_table_restore(efx);
1760 up_read(&efx->filter_sem);
1761 }
1762
1763 /**************************************************************************
1764 *
1765 * NIC startup/shutdown
1766 *
1767 *************************************************************************/
1768
1769 static int ef4_probe_all(struct ef4_nic *efx)
1770 {
1771 int rc;
1772
1773 rc = ef4_probe_nic(efx);
1774 if (rc) {
1775 netif_err(efx, probe, efx->net_dev, "failed to create NIC\n");
1776 goto fail1;
1777 }
1778
1779 rc = ef4_probe_port(efx);
1780 if (rc) {
1781 netif_err(efx, probe, efx->net_dev, "failed to create port\n");
1782 goto fail2;
1783 }
1784
1785 BUILD_BUG_ON(EF4_DEFAULT_DMAQ_SIZE < EF4_RXQ_MIN_ENT);
1786 if (WARN_ON(EF4_DEFAULT_DMAQ_SIZE < EF4_TXQ_MIN_ENT(efx))) {
1787 rc = -EINVAL;
1788 goto fail3;
1789 }
1790 efx->rxq_entries = efx->txq_entries = EF4_DEFAULT_DMAQ_SIZE;
1791
1792 rc = ef4_probe_filters(efx);
1793 if (rc) {
1794 netif_err(efx, probe, efx->net_dev,
1795 "failed to create filter tables\n");
1796 goto fail4;
1797 }
1798
1799 rc = ef4_probe_channels(efx);
1800 if (rc)
1801 goto fail5;
1802
1803 return 0;
1804
1805 fail5:
1806 ef4_remove_filters(efx);
1807 fail4:
1808 fail3:
1809 ef4_remove_port(efx);
1810 fail2:
1811 ef4_remove_nic(efx);
1812 fail1:
1813 return rc;
1814 }
1815
1816 /* If the interface is supposed to be running but is not, start
1817 * the hardware and software data path, regular activity for the port
1818 * (MAC statistics, link polling, etc.) and schedule the port to be
1819 * reconfigured. Interrupts must already be enabled. This function
1820 * is safe to call multiple times, so long as the NIC is not disabled.
1821 * Requires the RTNL lock.
1822 */
1823 static void ef4_start_all(struct ef4_nic *efx)
1824 {
1825 EF4_ASSERT_RESET_SERIALISED(efx);
1826 BUG_ON(efx->state == STATE_DISABLED);
1827
1828 /* Check that it is appropriate to restart the interface. All
1829 * of these flags are safe to read under just the rtnl lock */
1830 if (efx->port_enabled || !netif_running(efx->net_dev) ||
1831 efx->reset_pending)
1832 return;
1833
1834 ef4_start_port(efx);
1835 ef4_start_datapath(efx);
1836
1837 /* Start the hardware monitor if there is one */
1838 if (efx->type->monitor != NULL)
1839 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1840 ef4_monitor_interval);
1841
1842 efx->type->start_stats(efx);
1843 efx->type->pull_stats(efx);
1844 spin_lock_bh(&efx->stats_lock);
1845 efx->type->update_stats(efx, NULL, NULL);
1846 spin_unlock_bh(&efx->stats_lock);
1847 }
1848
1849 /* Quiesce the hardware and software data path, and regular activity
1850 * for the port without bringing the link down. Safe to call multiple
1851 * times with the NIC in almost any state, but interrupts should be
1852 * enabled. Requires the RTNL lock.
1853 */
1854 static void ef4_stop_all(struct ef4_nic *efx)
1855 {
1856 EF4_ASSERT_RESET_SERIALISED(efx);
1857
1858 /* port_enabled can be read safely under the rtnl lock */
1859 if (!efx->port_enabled)
1860 return;
1861
1862 /* update stats before we go down so we can accurately count
1863 * rx_nodesc_drops
1864 */
1865 efx->type->pull_stats(efx);
1866 spin_lock_bh(&efx->stats_lock);
1867 efx->type->update_stats(efx, NULL, NULL);
1868 spin_unlock_bh(&efx->stats_lock);
1869 efx->type->stop_stats(efx);
1870 ef4_stop_port(efx);
1871
1872 /* Stop the kernel transmit interface. This is only valid if
1873 * the device is stopped or detached; otherwise the watchdog
1874 * may fire immediately.
1875 */
1876 WARN_ON(netif_running(efx->net_dev) &&
1877 netif_device_present(efx->net_dev));
1878 netif_tx_disable(efx->net_dev);
1879
1880 ef4_stop_datapath(efx);
1881 }
1882
1883 static void ef4_remove_all(struct ef4_nic *efx)
1884 {
1885 ef4_remove_channels(efx);
1886 ef4_remove_filters(efx);
1887 ef4_remove_port(efx);
1888 ef4_remove_nic(efx);
1889 }
1890
1891 /**************************************************************************
1892 *
1893 * Interrupt moderation
1894 *
1895 **************************************************************************/
1896 unsigned int ef4_usecs_to_ticks(struct ef4_nic *efx, unsigned int usecs)
1897 {
1898 if (usecs == 0)
1899 return 0;
1900 if (usecs * 1000 < efx->timer_quantum_ns)
1901 return 1; /* never round down to 0 */
1902 return usecs * 1000 / efx->timer_quantum_ns;
1903 }
1904
1905 unsigned int ef4_ticks_to_usecs(struct ef4_nic *efx, unsigned int ticks)
1906 {
1907 /* We must round up when converting ticks to microseconds
1908 * because we round down when converting the other way.
1909 */
1910 return DIV_ROUND_UP(ticks * efx->timer_quantum_ns, 1000);
1911 }
1912
1913 /* Set interrupt moderation parameters */
1914 int ef4_init_irq_moderation(struct ef4_nic *efx, unsigned int tx_usecs,
1915 unsigned int rx_usecs, bool rx_adaptive,
1916 bool rx_may_override_tx)
1917 {
1918 struct ef4_channel *channel;
1919 unsigned int timer_max_us;
1920
1921 EF4_ASSERT_RESET_SERIALISED(efx);
1922
1923 timer_max_us = efx->timer_max_ns / 1000;
1924
1925 if (tx_usecs > timer_max_us || rx_usecs > timer_max_us)
1926 return -EINVAL;
1927
1928 if (tx_usecs != rx_usecs && efx->tx_channel_offset == 0 &&
1929 !rx_may_override_tx) {
1930 netif_err(efx, drv, efx->net_dev, "Channels are shared. "
1931 "RX and TX IRQ moderation must be equal\n");
1932 return -EINVAL;
1933 }
1934
1935 efx->irq_rx_adaptive = rx_adaptive;
1936 efx->irq_rx_moderation_us = rx_usecs;
1937 ef4_for_each_channel(channel, efx) {
1938 if (ef4_channel_has_rx_queue(channel))
1939 channel->irq_moderation_us = rx_usecs;
1940 else if (ef4_channel_has_tx_queues(channel))
1941 channel->irq_moderation_us = tx_usecs;
1942 }
1943
1944 return 0;
1945 }
1946
1947 void ef4_get_irq_moderation(struct ef4_nic *efx, unsigned int *tx_usecs,
1948 unsigned int *rx_usecs, bool *rx_adaptive)
1949 {
1950 *rx_adaptive = efx->irq_rx_adaptive;
1951 *rx_usecs = efx->irq_rx_moderation_us;
1952
1953 /* If channels are shared between RX and TX, so is IRQ
1954 * moderation. Otherwise, IRQ moderation is the same for all
1955 * TX channels and is not adaptive.
1956 */
1957 if (efx->tx_channel_offset == 0) {
1958 *tx_usecs = *rx_usecs;
1959 } else {
1960 struct ef4_channel *tx_channel;
1961
1962 tx_channel = efx->channel[efx->tx_channel_offset];
1963 *tx_usecs = tx_channel->irq_moderation_us;
1964 }
1965 }
1966
1967 /**************************************************************************
1968 *
1969 * Hardware monitor
1970 *
1971 **************************************************************************/
1972
1973 /* Run periodically off the general workqueue */
1974 static void ef4_monitor(struct work_struct *data)
1975 {
1976 struct ef4_nic *efx = container_of(data, struct ef4_nic,
1977 monitor_work.work);
1978
1979 netif_vdbg(efx, timer, efx->net_dev,
1980 "hardware monitor executing on CPU %d\n",
1981 raw_smp_processor_id());
1982 BUG_ON(efx->type->monitor == NULL);
1983
1984 /* If the mac_lock is already held then it is likely a port
1985 * reconfiguration is already in place, which will likely do
1986 * most of the work of monitor() anyway. */
1987 if (mutex_trylock(&efx->mac_lock)) {
1988 if (efx->port_enabled)
1989 efx->type->monitor(efx);
1990 mutex_unlock(&efx->mac_lock);
1991 }
1992
1993 queue_delayed_work(efx->workqueue, &efx->monitor_work,
1994 ef4_monitor_interval);
1995 }
1996
1997 /**************************************************************************
1998 *
1999 * ioctls
2000 *
2001 *************************************************************************/
2002
2003 /* Net device ioctl
2004 * Context: process, rtnl_lock() held.
2005 */
2006 static int ef4_ioctl(struct net_device *net_dev, struct ifreq *ifr, int cmd)
2007 {
2008 struct ef4_nic *efx = netdev_priv(net_dev);
2009 struct mii_ioctl_data *data = if_mii(ifr);
2010
2011 /* Convert phy_id from older PRTAD/DEVAD format */
2012 if ((cmd == SIOCGMIIREG || cmd == SIOCSMIIREG) &&
2013 (data->phy_id & 0xfc00) == 0x0400)
2014 data->phy_id ^= MDIO_PHY_ID_C45 | 0x0400;
2015
2016 return mdio_mii_ioctl(&efx->mdio, data, cmd);
2017 }
2018
2019 /**************************************************************************
2020 *
2021 * NAPI interface
2022 *
2023 **************************************************************************/
2024
2025 static void ef4_init_napi_channel(struct ef4_channel *channel)
2026 {
2027 struct ef4_nic *efx = channel->efx;
2028
2029 channel->napi_dev = efx->net_dev;
2030 netif_napi_add(channel->napi_dev, &channel->napi_str,
2031 ef4_poll, napi_weight);
2032 }
2033
2034 static void ef4_init_napi(struct ef4_nic *efx)
2035 {
2036 struct ef4_channel *channel;
2037
2038 ef4_for_each_channel(channel, efx)
2039 ef4_init_napi_channel(channel);
2040 }
2041
2042 static void ef4_fini_napi_channel(struct ef4_channel *channel)
2043 {
2044 if (channel->napi_dev)
2045 netif_napi_del(&channel->napi_str);
2046
2047 channel->napi_dev = NULL;
2048 }
2049
2050 static void ef4_fini_napi(struct ef4_nic *efx)
2051 {
2052 struct ef4_channel *channel;
2053
2054 ef4_for_each_channel(channel, efx)
2055 ef4_fini_napi_channel(channel);
2056 }
2057
2058 /**************************************************************************
2059 *
2060 * Kernel netpoll interface
2061 *
2062 *************************************************************************/
2063
2064 #ifdef CONFIG_NET_POLL_CONTROLLER
2065
2066 /* Although in the common case interrupts will be disabled, this is not
2067 * guaranteed. However, all our work happens inside the NAPI callback,
2068 * so no locking is required.
2069 */
2070 static void ef4_netpoll(struct net_device *net_dev)
2071 {
2072 struct ef4_nic *efx = netdev_priv(net_dev);
2073 struct ef4_channel *channel;
2074
2075 ef4_for_each_channel(channel, efx)
2076 ef4_schedule_channel(channel);
2077 }
2078
2079 #endif
2080
2081 /**************************************************************************
2082 *
2083 * Kernel net device interface
2084 *
2085 *************************************************************************/
2086
2087 /* Context: process, rtnl_lock() held. */
2088 int ef4_net_open(struct net_device *net_dev)
2089 {
2090 struct ef4_nic *efx = netdev_priv(net_dev);
2091 int rc;
2092
2093 netif_dbg(efx, ifup, efx->net_dev, "opening device on CPU %d\n",
2094 raw_smp_processor_id());
2095
2096 rc = ef4_check_disabled(efx);
2097 if (rc)
2098 return rc;
2099 if (efx->phy_mode & PHY_MODE_SPECIAL)
2100 return -EBUSY;
2101
2102 /* Notify the kernel of the link state polled during driver load,
2103 * before the monitor starts running */
2104 ef4_link_status_changed(efx);
2105
2106 ef4_start_all(efx);
2107 ef4_selftest_async_start(efx);
2108 return 0;
2109 }
2110
2111 /* Context: process, rtnl_lock() held.
2112 * Note that the kernel will ignore our return code; this method
2113 * should really be a void.
2114 */
2115 int ef4_net_stop(struct net_device *net_dev)
2116 {
2117 struct ef4_nic *efx = netdev_priv(net_dev);
2118
2119 netif_dbg(efx, ifdown, efx->net_dev, "closing on CPU %d\n",
2120 raw_smp_processor_id());
2121
2122 /* Stop the device and flush all the channels */
2123 ef4_stop_all(efx);
2124
2125 return 0;
2126 }
2127
2128 /* Context: process, dev_base_lock or RTNL held, non-blocking. */
2129 static void ef4_net_stats(struct net_device *net_dev,
2130 struct rtnl_link_stats64 *stats)
2131 {
2132 struct ef4_nic *efx = netdev_priv(net_dev);
2133
2134 spin_lock_bh(&efx->stats_lock);
2135 efx->type->update_stats(efx, NULL, stats);
2136 spin_unlock_bh(&efx->stats_lock);
2137 }
2138
2139 /* Context: netif_tx_lock held, BHs disabled. */
2140 static void ef4_watchdog(struct net_device *net_dev)
2141 {
2142 struct ef4_nic *efx = netdev_priv(net_dev);
2143
2144 netif_err(efx, tx_err, efx->net_dev,
2145 "TX stuck with port_enabled=%d: resetting channels\n",
2146 efx->port_enabled);
2147
2148 ef4_schedule_reset(efx, RESET_TYPE_TX_WATCHDOG);
2149 }
2150
2151
2152 /* Context: process, rtnl_lock() held. */
2153 static int ef4_change_mtu(struct net_device *net_dev, int new_mtu)
2154 {
2155 struct ef4_nic *efx = netdev_priv(net_dev);
2156 int rc;
2157
2158 rc = ef4_check_disabled(efx);
2159 if (rc)
2160 return rc;
2161
2162 netif_dbg(efx, drv, efx->net_dev, "changing MTU to %d\n", new_mtu);
2163
2164 ef4_device_detach_sync(efx);
2165 ef4_stop_all(efx);
2166
2167 mutex_lock(&efx->mac_lock);
2168 net_dev->mtu = new_mtu;
2169 ef4_mac_reconfigure(efx);
2170 mutex_unlock(&efx->mac_lock);
2171
2172 ef4_start_all(efx);
2173 netif_device_attach(efx->net_dev);
2174 return 0;
2175 }
2176
2177 static int ef4_set_mac_address(struct net_device *net_dev, void *data)
2178 {
2179 struct ef4_nic *efx = netdev_priv(net_dev);
2180 struct sockaddr *addr = data;
2181 u8 *new_addr = addr->sa_data;
2182 u8 old_addr[6];
2183 int rc;
2184
2185 if (!is_valid_ether_addr(new_addr)) {
2186 netif_err(efx, drv, efx->net_dev,
2187 "invalid ethernet MAC address requested: %pM\n",
2188 new_addr);
2189 return -EADDRNOTAVAIL;
2190 }
2191
2192 /* save old address */
2193 ether_addr_copy(old_addr, net_dev->dev_addr);
2194 ether_addr_copy(net_dev->dev_addr, new_addr);
2195 if (efx->type->set_mac_address) {
2196 rc = efx->type->set_mac_address(efx);
2197 if (rc) {
2198 ether_addr_copy(net_dev->dev_addr, old_addr);
2199 return rc;
2200 }
2201 }
2202
2203 /* Reconfigure the MAC */
2204 mutex_lock(&efx->mac_lock);
2205 ef4_mac_reconfigure(efx);
2206 mutex_unlock(&efx->mac_lock);
2207
2208 return 0;
2209 }
2210
2211 /* Context: netif_addr_lock held, BHs disabled. */
2212 static void ef4_set_rx_mode(struct net_device *net_dev)
2213 {
2214 struct ef4_nic *efx = netdev_priv(net_dev);
2215
2216 if (efx->port_enabled)
2217 queue_work(efx->workqueue, &efx->mac_work);
2218 /* Otherwise ef4_start_port() will do this */
2219 }
2220
2221 static int ef4_set_features(struct net_device *net_dev, netdev_features_t data)
2222 {
2223 struct ef4_nic *efx = netdev_priv(net_dev);
2224 int rc;
2225
2226 /* If disabling RX n-tuple filtering, clear existing filters */
2227 if (net_dev->features & ~data & NETIF_F_NTUPLE) {
2228 rc = efx->type->filter_clear_rx(efx, EF4_FILTER_PRI_MANUAL);
2229 if (rc)
2230 return rc;
2231 }
2232
2233 /* If Rx VLAN filter is changed, update filters via mac_reconfigure */
2234 if ((net_dev->features ^ data) & NETIF_F_HW_VLAN_CTAG_FILTER) {
2235 /* ef4_set_rx_mode() will schedule MAC work to update filters
2236 * when a new features are finally set in net_dev.
2237 */
2238 ef4_set_rx_mode(net_dev);
2239 }
2240
2241 return 0;
2242 }
2243
2244 static const struct net_device_ops ef4_netdev_ops = {
2245 .ndo_open = ef4_net_open,
2246 .ndo_stop = ef4_net_stop,
2247 .ndo_get_stats64 = ef4_net_stats,
2248 .ndo_tx_timeout = ef4_watchdog,
2249 .ndo_start_xmit = ef4_hard_start_xmit,
2250 .ndo_validate_addr = eth_validate_addr,
2251 .ndo_do_ioctl = ef4_ioctl,
2252 .ndo_change_mtu = ef4_change_mtu,
2253 .ndo_set_mac_address = ef4_set_mac_address,
2254 .ndo_set_rx_mode = ef4_set_rx_mode,
2255 .ndo_set_features = ef4_set_features,
2256 #ifdef CONFIG_NET_POLL_CONTROLLER
2257 .ndo_poll_controller = ef4_netpoll,
2258 #endif
2259 .ndo_setup_tc = ef4_setup_tc,
2260 #ifdef CONFIG_RFS_ACCEL
2261 .ndo_rx_flow_steer = ef4_filter_rfs,
2262 #endif
2263 };
2264
2265 static void ef4_update_name(struct ef4_nic *efx)
2266 {
2267 strcpy(efx->name, efx->net_dev->name);
2268 ef4_mtd_rename(efx);
2269 ef4_set_channel_names(efx);
2270 }
2271
2272 static int ef4_netdev_event(struct notifier_block *this,
2273 unsigned long event, void *ptr)
2274 {
2275 struct net_device *net_dev = netdev_notifier_info_to_dev(ptr);
2276
2277 if ((net_dev->netdev_ops == &ef4_netdev_ops) &&
2278 event == NETDEV_CHANGENAME)
2279 ef4_update_name(netdev_priv(net_dev));
2280
2281 return NOTIFY_DONE;
2282 }
2283
2284 static struct notifier_block ef4_netdev_notifier = {
2285 .notifier_call = ef4_netdev_event,
2286 };
2287
2288 static ssize_t
2289 show_phy_type(struct device *dev, struct device_attribute *attr, char *buf)
2290 {
2291 struct ef4_nic *efx = pci_get_drvdata(to_pci_dev(dev));
2292 return sprintf(buf, "%d\n", efx->phy_type);
2293 }
2294 static DEVICE_ATTR(phy_type, 0444, show_phy_type, NULL);
2295
2296 static int ef4_register_netdev(struct ef4_nic *efx)
2297 {
2298 struct net_device *net_dev = efx->net_dev;
2299 struct ef4_channel *channel;
2300 int rc;
2301
2302 net_dev->watchdog_timeo = 5 * HZ;
2303 net_dev->irq = efx->pci_dev->irq;
2304 net_dev->netdev_ops = &ef4_netdev_ops;
2305 net_dev->ethtool_ops = &ef4_ethtool_ops;
2306 net_dev->gso_max_segs = EF4_TSO_MAX_SEGS;
2307 net_dev->min_mtu = EF4_MIN_MTU;
2308 net_dev->max_mtu = EF4_MAX_MTU;
2309
2310 rtnl_lock();
2311
2312 /* Enable resets to be scheduled and check whether any were
2313 * already requested. If so, the NIC is probably hosed so we
2314 * abort.
2315 */
2316 efx->state = STATE_READY;
2317 smp_mb(); /* ensure we change state before checking reset_pending */
2318 if (efx->reset_pending) {
2319 netif_err(efx, probe, efx->net_dev,
2320 "aborting probe due to scheduled reset\n");
2321 rc = -EIO;
2322 goto fail_locked;
2323 }
2324
2325 rc = dev_alloc_name(net_dev, net_dev->name);
2326 if (rc < 0)
2327 goto fail_locked;
2328 ef4_update_name(efx);
2329
2330 /* Always start with carrier off; PHY events will detect the link */
2331 netif_carrier_off(net_dev);
2332
2333 rc = register_netdevice(net_dev);
2334 if (rc)
2335 goto fail_locked;
2336
2337 ef4_for_each_channel(channel, efx) {
2338 struct ef4_tx_queue *tx_queue;
2339 ef4_for_each_channel_tx_queue(tx_queue, channel)
2340 ef4_init_tx_queue_core_txq(tx_queue);
2341 }
2342
2343 ef4_associate(efx);
2344
2345 rtnl_unlock();
2346
2347 rc = device_create_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2348 if (rc) {
2349 netif_err(efx, drv, efx->net_dev,
2350 "failed to init net dev attributes\n");
2351 goto fail_registered;
2352 }
2353 return 0;
2354
2355 fail_registered:
2356 rtnl_lock();
2357 ef4_dissociate(efx);
2358 unregister_netdevice(net_dev);
2359 fail_locked:
2360 efx->state = STATE_UNINIT;
2361 rtnl_unlock();
2362 netif_err(efx, drv, efx->net_dev, "could not register net dev\n");
2363 return rc;
2364 }
2365
2366 static void ef4_unregister_netdev(struct ef4_nic *efx)
2367 {
2368 if (!efx->net_dev)
2369 return;
2370
2371 BUG_ON(netdev_priv(efx->net_dev) != efx);
2372
2373 if (ef4_dev_registered(efx)) {
2374 strlcpy(efx->name, pci_name(efx->pci_dev), sizeof(efx->name));
2375 device_remove_file(&efx->pci_dev->dev, &dev_attr_phy_type);
2376 unregister_netdev(efx->net_dev);
2377 }
2378 }
2379
2380 /**************************************************************************
2381 *
2382 * Device reset and suspend
2383 *
2384 **************************************************************************/
2385
2386 /* Tears down the entire software state and most of the hardware state
2387 * before reset. */
2388 void ef4_reset_down(struct ef4_nic *efx, enum reset_type method)
2389 {
2390 EF4_ASSERT_RESET_SERIALISED(efx);
2391
2392 ef4_stop_all(efx);
2393 ef4_disable_interrupts(efx);
2394
2395 mutex_lock(&efx->mac_lock);
2396 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2397 method != RESET_TYPE_DATAPATH)
2398 efx->phy_op->fini(efx);
2399 efx->type->fini(efx);
2400 }
2401
2402 /* This function will always ensure that the locks acquired in
2403 * ef4_reset_down() are released. A failure return code indicates
2404 * that we were unable to reinitialise the hardware, and the
2405 * driver should be disabled. If ok is false, then the rx and tx
2406 * engines are not restarted, pending a RESET_DISABLE. */
2407 int ef4_reset_up(struct ef4_nic *efx, enum reset_type method, bool ok)
2408 {
2409 int rc;
2410
2411 EF4_ASSERT_RESET_SERIALISED(efx);
2412
2413 /* Ensure that SRAM is initialised even if we're disabling the device */
2414 rc = efx->type->init(efx);
2415 if (rc) {
2416 netif_err(efx, drv, efx->net_dev, "failed to initialise NIC\n");
2417 goto fail;
2418 }
2419
2420 if (!ok)
2421 goto fail;
2422
2423 if (efx->port_initialized && method != RESET_TYPE_INVISIBLE &&
2424 method != RESET_TYPE_DATAPATH) {
2425 rc = efx->phy_op->init(efx);
2426 if (rc)
2427 goto fail;
2428 rc = efx->phy_op->reconfigure(efx);
2429 if (rc && rc != -EPERM)
2430 netif_err(efx, drv, efx->net_dev,
2431 "could not restore PHY settings\n");
2432 }
2433
2434 rc = ef4_enable_interrupts(efx);
2435 if (rc)
2436 goto fail;
2437
2438 down_read(&efx->filter_sem);
2439 ef4_restore_filters(efx);
2440 up_read(&efx->filter_sem);
2441
2442 mutex_unlock(&efx->mac_lock);
2443
2444 ef4_start_all(efx);
2445
2446 return 0;
2447
2448 fail:
2449 efx->port_initialized = false;
2450
2451 mutex_unlock(&efx->mac_lock);
2452
2453 return rc;
2454 }
2455
2456 /* Reset the NIC using the specified method. Note that the reset may
2457 * fail, in which case the card will be left in an unusable state.
2458 *
2459 * Caller must hold the rtnl_lock.
2460 */
2461 int ef4_reset(struct ef4_nic *efx, enum reset_type method)
2462 {
2463 int rc, rc2;
2464 bool disabled;
2465
2466 netif_info(efx, drv, efx->net_dev, "resetting (%s)\n",
2467 RESET_TYPE(method));
2468
2469 ef4_device_detach_sync(efx);
2470 ef4_reset_down(efx, method);
2471
2472 rc = efx->type->reset(efx, method);
2473 if (rc) {
2474 netif_err(efx, drv, efx->net_dev, "failed to reset hardware\n");
2475 goto out;
2476 }
2477
2478 /* Clear flags for the scopes we covered. We assume the NIC and
2479 * driver are now quiescent so that there is no race here.
2480 */
2481 if (method < RESET_TYPE_MAX_METHOD)
2482 efx->reset_pending &= -(1 << (method + 1));
2483 else /* it doesn't fit into the well-ordered scope hierarchy */
2484 __clear_bit(method, &efx->reset_pending);
2485
2486 /* Reinitialise bus-mastering, which may have been turned off before
2487 * the reset was scheduled. This is still appropriate, even in the
2488 * RESET_TYPE_DISABLE since this driver generally assumes the hardware
2489 * can respond to requests. */
2490 pci_set_master(efx->pci_dev);
2491
2492 out:
2493 /* Leave device stopped if necessary */
2494 disabled = rc ||
2495 method == RESET_TYPE_DISABLE ||
2496 method == RESET_TYPE_RECOVER_OR_DISABLE;
2497 rc2 = ef4_reset_up(efx, method, !disabled);
2498 if (rc2) {
2499 disabled = true;
2500 if (!rc)
2501 rc = rc2;
2502 }
2503
2504 if (disabled) {
2505 dev_close(efx->net_dev);
2506 netif_err(efx, drv, efx->net_dev, "has been disabled\n");
2507 efx->state = STATE_DISABLED;
2508 } else {
2509 netif_dbg(efx, drv, efx->net_dev, "reset complete\n");
2510 netif_device_attach(efx->net_dev);
2511 }
2512 return rc;
2513 }
2514
2515 /* Try recovery mechanisms.
2516 * For now only EEH is supported.
2517 * Returns 0 if the recovery mechanisms are unsuccessful.
2518 * Returns a non-zero value otherwise.
2519 */
2520 int ef4_try_recovery(struct ef4_nic *efx)
2521 {
2522 #ifdef CONFIG_EEH
2523 /* A PCI error can occur and not be seen by EEH because nothing
2524 * happens on the PCI bus. In this case the driver may fail and
2525 * schedule a 'recover or reset', leading to this recovery handler.
2526 * Manually call the eeh failure check function.
2527 */
2528 struct eeh_dev *eehdev = pci_dev_to_eeh_dev(efx->pci_dev);
2529 if (eeh_dev_check_failure(eehdev)) {
2530 /* The EEH mechanisms will handle the error and reset the
2531 * device if necessary.
2532 */
2533 return 1;
2534 }
2535 #endif
2536 return 0;
2537 }
2538
2539 /* The worker thread exists so that code that cannot sleep can
2540 * schedule a reset for later.
2541 */
2542 static void ef4_reset_work(struct work_struct *data)
2543 {
2544 struct ef4_nic *efx = container_of(data, struct ef4_nic, reset_work);
2545 unsigned long pending;
2546 enum reset_type method;
2547
2548 pending = ACCESS_ONCE(efx->reset_pending);
2549 method = fls(pending) - 1;
2550
2551 if ((method == RESET_TYPE_RECOVER_OR_DISABLE ||
2552 method == RESET_TYPE_RECOVER_OR_ALL) &&
2553 ef4_try_recovery(efx))
2554 return;
2555
2556 if (!pending)
2557 return;
2558
2559 rtnl_lock();
2560
2561 /* We checked the state in ef4_schedule_reset() but it may
2562 * have changed by now. Now that we have the RTNL lock,
2563 * it cannot change again.
2564 */
2565 if (efx->state == STATE_READY)
2566 (void)ef4_reset(efx, method);
2567
2568 rtnl_unlock();
2569 }
2570
2571 void ef4_schedule_reset(struct ef4_nic *efx, enum reset_type type)
2572 {
2573 enum reset_type method;
2574
2575 if (efx->state == STATE_RECOVERY) {
2576 netif_dbg(efx, drv, efx->net_dev,
2577 "recovering: skip scheduling %s reset\n",
2578 RESET_TYPE(type));
2579 return;
2580 }
2581
2582 switch (type) {
2583 case RESET_TYPE_INVISIBLE:
2584 case RESET_TYPE_ALL:
2585 case RESET_TYPE_RECOVER_OR_ALL:
2586 case RESET_TYPE_WORLD:
2587 case RESET_TYPE_DISABLE:
2588 case RESET_TYPE_RECOVER_OR_DISABLE:
2589 case RESET_TYPE_DATAPATH:
2590 method = type;
2591 netif_dbg(efx, drv, efx->net_dev, "scheduling %s reset\n",
2592 RESET_TYPE(method));
2593 break;
2594 default:
2595 method = efx->type->map_reset_reason(type);
2596 netif_dbg(efx, drv, efx->net_dev,
2597 "scheduling %s reset for %s\n",
2598 RESET_TYPE(method), RESET_TYPE(type));
2599 break;
2600 }
2601
2602 set_bit(method, &efx->reset_pending);
2603 smp_mb(); /* ensure we change reset_pending before checking state */
2604
2605 /* If we're not READY then just leave the flags set as the cue
2606 * to abort probing or reschedule the reset later.
2607 */
2608 if (ACCESS_ONCE(efx->state) != STATE_READY)
2609 return;
2610
2611 queue_work(reset_workqueue, &efx->reset_work);
2612 }
2613
2614 /**************************************************************************
2615 *
2616 * List of NICs we support
2617 *
2618 **************************************************************************/
2619
2620 /* PCI device ID table */
2621 static const struct pci_device_id ef4_pci_table[] = {
2622 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2623 PCI_DEVICE_ID_SOLARFLARE_SFC4000A_0),
2624 .driver_data = (unsigned long) &falcon_a1_nic_type},
2625 {PCI_DEVICE(PCI_VENDOR_ID_SOLARFLARE,
2626 PCI_DEVICE_ID_SOLARFLARE_SFC4000B),
2627 .driver_data = (unsigned long) &falcon_b0_nic_type},
2628 {0} /* end of list */
2629 };
2630
2631 /**************************************************************************
2632 *
2633 * Dummy PHY/MAC operations
2634 *
2635 * Can be used for some unimplemented operations
2636 * Needed so all function pointers are valid and do not have to be tested
2637 * before use
2638 *
2639 **************************************************************************/
2640 int ef4_port_dummy_op_int(struct ef4_nic *efx)
2641 {
2642 return 0;
2643 }
2644 void ef4_port_dummy_op_void(struct ef4_nic *efx) {}
2645
2646 static bool ef4_port_dummy_op_poll(struct ef4_nic *efx)
2647 {
2648 return false;
2649 }
2650
2651 static const struct ef4_phy_operations ef4_dummy_phy_operations = {
2652 .init = ef4_port_dummy_op_int,
2653 .reconfigure = ef4_port_dummy_op_int,
2654 .poll = ef4_port_dummy_op_poll,
2655 .fini = ef4_port_dummy_op_void,
2656 };
2657
2658 /**************************************************************************
2659 *
2660 * Data housekeeping
2661 *
2662 **************************************************************************/
2663
2664 /* This zeroes out and then fills in the invariants in a struct
2665 * ef4_nic (including all sub-structures).
2666 */
2667 static int ef4_init_struct(struct ef4_nic *efx,
2668 struct pci_dev *pci_dev, struct net_device *net_dev)
2669 {
2670 int i;
2671
2672 /* Initialise common structures */
2673 INIT_LIST_HEAD(&efx->node);
2674 INIT_LIST_HEAD(&efx->secondary_list);
2675 spin_lock_init(&efx->biu_lock);
2676 #ifdef CONFIG_SFC_FALCON_MTD
2677 INIT_LIST_HEAD(&efx->mtd_list);
2678 #endif
2679 INIT_WORK(&efx->reset_work, ef4_reset_work);
2680 INIT_DELAYED_WORK(&efx->monitor_work, ef4_monitor);
2681 INIT_DELAYED_WORK(&efx->selftest_work, ef4_selftest_async_work);
2682 efx->pci_dev = pci_dev;
2683 efx->msg_enable = debug;
2684 efx->state = STATE_UNINIT;
2685 strlcpy(efx->name, pci_name(pci_dev), sizeof(efx->name));
2686
2687 efx->net_dev = net_dev;
2688 efx->rx_prefix_size = efx->type->rx_prefix_size;
2689 efx->rx_ip_align =
2690 NET_IP_ALIGN ? (efx->rx_prefix_size + NET_IP_ALIGN) % 4 : 0;
2691 efx->rx_packet_hash_offset =
2692 efx->type->rx_hash_offset - efx->type->rx_prefix_size;
2693 efx->rx_packet_ts_offset =
2694 efx->type->rx_ts_offset - efx->type->rx_prefix_size;
2695 spin_lock_init(&efx->stats_lock);
2696 mutex_init(&efx->mac_lock);
2697 efx->phy_op = &ef4_dummy_phy_operations;
2698 efx->mdio.dev = net_dev;
2699 INIT_WORK(&efx->mac_work, ef4_mac_work);
2700 init_waitqueue_head(&efx->flush_wq);
2701
2702 for (i = 0; i < EF4_MAX_CHANNELS; i++) {
2703 efx->channel[i] = ef4_alloc_channel(efx, i, NULL);
2704 if (!efx->channel[i])
2705 goto fail;
2706 efx->msi_context[i].efx = efx;
2707 efx->msi_context[i].index = i;
2708 }
2709
2710 /* Higher numbered interrupt modes are less capable! */
2711 efx->interrupt_mode = max(efx->type->max_interrupt_mode,
2712 interrupt_mode);
2713
2714 /* Would be good to use the net_dev name, but we're too early */
2715 snprintf(efx->workqueue_name, sizeof(efx->workqueue_name), "sfc%s",
2716 pci_name(pci_dev));
2717 efx->workqueue = create_singlethread_workqueue(efx->workqueue_name);
2718 if (!efx->workqueue)
2719 goto fail;
2720
2721 return 0;
2722
2723 fail:
2724 ef4_fini_struct(efx);
2725 return -ENOMEM;
2726 }
2727
2728 static void ef4_fini_struct(struct ef4_nic *efx)
2729 {
2730 int i;
2731
2732 for (i = 0; i < EF4_MAX_CHANNELS; i++)
2733 kfree(efx->channel[i]);
2734
2735 kfree(efx->vpd_sn);
2736
2737 if (efx->workqueue) {
2738 destroy_workqueue(efx->workqueue);
2739 efx->workqueue = NULL;
2740 }
2741 }
2742
2743 void ef4_update_sw_stats(struct ef4_nic *efx, u64 *stats)
2744 {
2745 u64 n_rx_nodesc_trunc = 0;
2746 struct ef4_channel *channel;
2747
2748 ef4_for_each_channel(channel, efx)
2749 n_rx_nodesc_trunc += channel->n_rx_nodesc_trunc;
2750 stats[GENERIC_STAT_rx_nodesc_trunc] = n_rx_nodesc_trunc;
2751 stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
2752 }
2753
2754 /**************************************************************************
2755 *
2756 * PCI interface
2757 *
2758 **************************************************************************/
2759
2760 /* Main body of final NIC shutdown code
2761 * This is called only at module unload (or hotplug removal).
2762 */
2763 static void ef4_pci_remove_main(struct ef4_nic *efx)
2764 {
2765 /* Flush reset_work. It can no longer be scheduled since we
2766 * are not READY.
2767 */
2768 BUG_ON(efx->state == STATE_READY);
2769 cancel_work_sync(&efx->reset_work);
2770
2771 ef4_disable_interrupts(efx);
2772 ef4_nic_fini_interrupt(efx);
2773 ef4_fini_port(efx);
2774 efx->type->fini(efx);
2775 ef4_fini_napi(efx);
2776 ef4_remove_all(efx);
2777 }
2778
2779 /* Final NIC shutdown
2780 * This is called only at module unload (or hotplug removal). A PF can call
2781 * this on its VFs to ensure they are unbound first.
2782 */
2783 static void ef4_pci_remove(struct pci_dev *pci_dev)
2784 {
2785 struct ef4_nic *efx;
2786
2787 efx = pci_get_drvdata(pci_dev);
2788 if (!efx)
2789 return;
2790
2791 /* Mark the NIC as fini, then stop the interface */
2792 rtnl_lock();
2793 ef4_dissociate(efx);
2794 dev_close(efx->net_dev);
2795 ef4_disable_interrupts(efx);
2796 efx->state = STATE_UNINIT;
2797 rtnl_unlock();
2798
2799 ef4_unregister_netdev(efx);
2800
2801 ef4_mtd_remove(efx);
2802
2803 ef4_pci_remove_main(efx);
2804
2805 ef4_fini_io(efx);
2806 netif_dbg(efx, drv, efx->net_dev, "shutdown successful\n");
2807
2808 ef4_fini_struct(efx);
2809 free_netdev(efx->net_dev);
2810
2811 pci_disable_pcie_error_reporting(pci_dev);
2812 };
2813
2814 /* NIC VPD information
2815 * Called during probe to display the part number of the
2816 * installed NIC. VPD is potentially very large but this should
2817 * always appear within the first 512 bytes.
2818 */
2819 #define SFC_VPD_LEN 512
2820 static void ef4_probe_vpd_strings(struct ef4_nic *efx)
2821 {
2822 struct pci_dev *dev = efx->pci_dev;
2823 char vpd_data[SFC_VPD_LEN];
2824 ssize_t vpd_size;
2825 int ro_start, ro_size, i, j;
2826
2827 /* Get the vpd data from the device */
2828 vpd_size = pci_read_vpd(dev, 0, sizeof(vpd_data), vpd_data);
2829 if (vpd_size <= 0) {
2830 netif_err(efx, drv, efx->net_dev, "Unable to read VPD\n");
2831 return;
2832 }
2833
2834 /* Get the Read only section */
2835 ro_start = pci_vpd_find_tag(vpd_data, 0, vpd_size, PCI_VPD_LRDT_RO_DATA);
2836 if (ro_start < 0) {
2837 netif_err(efx, drv, efx->net_dev, "VPD Read-only not found\n");
2838 return;
2839 }
2840
2841 ro_size = pci_vpd_lrdt_size(&vpd_data[ro_start]);
2842 j = ro_size;
2843 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2844 if (i + j > vpd_size)
2845 j = vpd_size - i;
2846
2847 /* Get the Part number */
2848 i = pci_vpd_find_info_keyword(vpd_data, i, j, "PN");
2849 if (i < 0) {
2850 netif_err(efx, drv, efx->net_dev, "Part number not found\n");
2851 return;
2852 }
2853
2854 j = pci_vpd_info_field_size(&vpd_data[i]);
2855 i += PCI_VPD_INFO_FLD_HDR_SIZE;
2856 if (i + j > vpd_size) {
2857 netif_err(efx, drv, efx->net_dev, "Incomplete part number\n");
2858 return;
2859 }
2860
2861 netif_info(efx, drv, efx->net_dev,
2862 "Part Number : %.*s\n", j, &vpd_data[i]);
2863
2864 i = ro_start + PCI_VPD_LRDT_TAG_SIZE;
2865 j = ro_size;
2866 i = pci_vpd_find_info_keyword(vpd_data, i, j, "SN");
2867 if (i < 0) {
2868 netif_err(efx, drv, efx->net_dev, "Serial number not found\n");
2869 return;
2870 }
2871
2872 j = pci_vpd_info_field_size(&vpd_data[i]);
2873 i += PCI_VPD_INFO_FLD_HDR_SIZE;
2874 if (i + j > vpd_size) {
2875 netif_err(efx, drv, efx->net_dev, "Incomplete serial number\n");
2876 return;
2877 }
2878
2879 efx->vpd_sn = kmalloc(j + 1, GFP_KERNEL);
2880 if (!efx->vpd_sn)
2881 return;
2882
2883 snprintf(efx->vpd_sn, j + 1, "%s", &vpd_data[i]);
2884 }
2885
2886
2887 /* Main body of NIC initialisation
2888 * This is called at module load (or hotplug insertion, theoretically).
2889 */
2890 static int ef4_pci_probe_main(struct ef4_nic *efx)
2891 {
2892 int rc;
2893
2894 /* Do start-of-day initialisation */
2895 rc = ef4_probe_all(efx);
2896 if (rc)
2897 goto fail1;
2898
2899 ef4_init_napi(efx);
2900
2901 rc = efx->type->init(efx);
2902 if (rc) {
2903 netif_err(efx, probe, efx->net_dev,
2904 "failed to initialise NIC\n");
2905 goto fail3;
2906 }
2907
2908 rc = ef4_init_port(efx);
2909 if (rc) {
2910 netif_err(efx, probe, efx->net_dev,
2911 "failed to initialise port\n");
2912 goto fail4;
2913 }
2914
2915 rc = ef4_nic_init_interrupt(efx);
2916 if (rc)
2917 goto fail5;
2918 rc = ef4_enable_interrupts(efx);
2919 if (rc)
2920 goto fail6;
2921
2922 return 0;
2923
2924 fail6:
2925 ef4_nic_fini_interrupt(efx);
2926 fail5:
2927 ef4_fini_port(efx);
2928 fail4:
2929 efx->type->fini(efx);
2930 fail3:
2931 ef4_fini_napi(efx);
2932 ef4_remove_all(efx);
2933 fail1:
2934 return rc;
2935 }
2936
2937 /* NIC initialisation
2938 *
2939 * This is called at module load (or hotplug insertion,
2940 * theoretically). It sets up PCI mappings, resets the NIC,
2941 * sets up and registers the network devices with the kernel and hooks
2942 * the interrupt service routine. It does not prepare the device for
2943 * transmission; this is left to the first time one of the network
2944 * interfaces is brought up (i.e. ef4_net_open).
2945 */
2946 static int ef4_pci_probe(struct pci_dev *pci_dev,
2947 const struct pci_device_id *entry)
2948 {
2949 struct net_device *net_dev;
2950 struct ef4_nic *efx;
2951 int rc;
2952
2953 /* Allocate and initialise a struct net_device and struct ef4_nic */
2954 net_dev = alloc_etherdev_mqs(sizeof(*efx), EF4_MAX_CORE_TX_QUEUES,
2955 EF4_MAX_RX_QUEUES);
2956 if (!net_dev)
2957 return -ENOMEM;
2958 efx = netdev_priv(net_dev);
2959 efx->type = (const struct ef4_nic_type *) entry->driver_data;
2960 efx->fixed_features |= NETIF_F_HIGHDMA;
2961
2962 pci_set_drvdata(pci_dev, efx);
2963 SET_NETDEV_DEV(net_dev, &pci_dev->dev);
2964 rc = ef4_init_struct(efx, pci_dev, net_dev);
2965 if (rc)
2966 goto fail1;
2967
2968 netif_info(efx, probe, efx->net_dev,
2969 "Solarflare NIC detected\n");
2970
2971 ef4_probe_vpd_strings(efx);
2972
2973 /* Set up basic I/O (BAR mappings etc) */
2974 rc = ef4_init_io(efx);
2975 if (rc)
2976 goto fail2;
2977
2978 rc = ef4_pci_probe_main(efx);
2979 if (rc)
2980 goto fail3;
2981
2982 net_dev->features |= (efx->type->offload_features | NETIF_F_SG |
2983 NETIF_F_RXCSUM);
2984 /* Mask for features that also apply to VLAN devices */
2985 net_dev->vlan_features |= (NETIF_F_HW_CSUM | NETIF_F_SG |
2986 NETIF_F_HIGHDMA | NETIF_F_RXCSUM);
2987
2988 net_dev->hw_features = net_dev->features & ~efx->fixed_features;
2989
2990 /* Disable VLAN filtering by default. It may be enforced if
2991 * the feature is fixed (i.e. VLAN filters are required to
2992 * receive VLAN tagged packets due to vPort restrictions).
2993 */
2994 net_dev->features &= ~NETIF_F_HW_VLAN_CTAG_FILTER;
2995 net_dev->features |= efx->fixed_features;
2996
2997 rc = ef4_register_netdev(efx);
2998 if (rc)
2999 goto fail4;
3000
3001 netif_dbg(efx, probe, efx->net_dev, "initialisation successful\n");
3002
3003 /* Try to create MTDs, but allow this to fail */
3004 rtnl_lock();
3005 rc = ef4_mtd_probe(efx);
3006 rtnl_unlock();
3007 if (rc && rc != -EPERM)
3008 netif_warn(efx, probe, efx->net_dev,
3009 "failed to create MTDs (%d)\n", rc);
3010
3011 rc = pci_enable_pcie_error_reporting(pci_dev);
3012 if (rc && rc != -EINVAL)
3013 netif_notice(efx, probe, efx->net_dev,
3014 "PCIE error reporting unavailable (%d).\n",
3015 rc);
3016
3017 return 0;
3018
3019 fail4:
3020 ef4_pci_remove_main(efx);
3021 fail3:
3022 ef4_fini_io(efx);
3023 fail2:
3024 ef4_fini_struct(efx);
3025 fail1:
3026 WARN_ON(rc > 0);
3027 netif_dbg(efx, drv, efx->net_dev, "initialisation failed. rc=%d\n", rc);
3028 free_netdev(net_dev);
3029 return rc;
3030 }
3031
3032 static int ef4_pm_freeze(struct device *dev)
3033 {
3034 struct ef4_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3035
3036 rtnl_lock();
3037
3038 if (efx->state != STATE_DISABLED) {
3039 efx->state = STATE_UNINIT;
3040
3041 ef4_device_detach_sync(efx);
3042
3043 ef4_stop_all(efx);
3044 ef4_disable_interrupts(efx);
3045 }
3046
3047 rtnl_unlock();
3048
3049 return 0;
3050 }
3051
3052 static int ef4_pm_thaw(struct device *dev)
3053 {
3054 int rc;
3055 struct ef4_nic *efx = pci_get_drvdata(to_pci_dev(dev));
3056
3057 rtnl_lock();
3058
3059 if (efx->state != STATE_DISABLED) {
3060 rc = ef4_enable_interrupts(efx);
3061 if (rc)
3062 goto fail;
3063
3064 mutex_lock(&efx->mac_lock);
3065 efx->phy_op->reconfigure(efx);
3066 mutex_unlock(&efx->mac_lock);
3067
3068 ef4_start_all(efx);
3069
3070 netif_device_attach(efx->net_dev);
3071
3072 efx->state = STATE_READY;
3073
3074 efx->type->resume_wol(efx);
3075 }
3076
3077 rtnl_unlock();
3078
3079 /* Reschedule any quenched resets scheduled during ef4_pm_freeze() */
3080 queue_work(reset_workqueue, &efx->reset_work);
3081
3082 return 0;
3083
3084 fail:
3085 rtnl_unlock();
3086
3087 return rc;
3088 }
3089
3090 static int ef4_pm_poweroff(struct device *dev)
3091 {
3092 struct pci_dev *pci_dev = to_pci_dev(dev);
3093 struct ef4_nic *efx = pci_get_drvdata(pci_dev);
3094
3095 efx->type->fini(efx);
3096
3097 efx->reset_pending = 0;
3098
3099 pci_save_state(pci_dev);
3100 return pci_set_power_state(pci_dev, PCI_D3hot);
3101 }
3102
3103 /* Used for both resume and restore */
3104 static int ef4_pm_resume(struct device *dev)
3105 {
3106 struct pci_dev *pci_dev = to_pci_dev(dev);
3107 struct ef4_nic *efx = pci_get_drvdata(pci_dev);
3108 int rc;
3109
3110 rc = pci_set_power_state(pci_dev, PCI_D0);
3111 if (rc)
3112 return rc;
3113 pci_restore_state(pci_dev);
3114 rc = pci_enable_device(pci_dev);
3115 if (rc)
3116 return rc;
3117 pci_set_master(efx->pci_dev);
3118 rc = efx->type->reset(efx, RESET_TYPE_ALL);
3119 if (rc)
3120 return rc;
3121 rc = efx->type->init(efx);
3122 if (rc)
3123 return rc;
3124 rc = ef4_pm_thaw(dev);
3125 return rc;
3126 }
3127
3128 static int ef4_pm_suspend(struct device *dev)
3129 {
3130 int rc;
3131
3132 ef4_pm_freeze(dev);
3133 rc = ef4_pm_poweroff(dev);
3134 if (rc)
3135 ef4_pm_resume(dev);
3136 return rc;
3137 }
3138
3139 static const struct dev_pm_ops ef4_pm_ops = {
3140 .suspend = ef4_pm_suspend,
3141 .resume = ef4_pm_resume,
3142 .freeze = ef4_pm_freeze,
3143 .thaw = ef4_pm_thaw,
3144 .poweroff = ef4_pm_poweroff,
3145 .restore = ef4_pm_resume,
3146 };
3147
3148 /* A PCI error affecting this device was detected.
3149 * At this point MMIO and DMA may be disabled.
3150 * Stop the software path and request a slot reset.
3151 */
3152 static pci_ers_result_t ef4_io_error_detected(struct pci_dev *pdev,
3153 enum pci_channel_state state)
3154 {
3155 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3156 struct ef4_nic *efx = pci_get_drvdata(pdev);
3157
3158 if (state == pci_channel_io_perm_failure)
3159 return PCI_ERS_RESULT_DISCONNECT;
3160
3161 rtnl_lock();
3162
3163 if (efx->state != STATE_DISABLED) {
3164 efx->state = STATE_RECOVERY;
3165 efx->reset_pending = 0;
3166
3167 ef4_device_detach_sync(efx);
3168
3169 ef4_stop_all(efx);
3170 ef4_disable_interrupts(efx);
3171
3172 status = PCI_ERS_RESULT_NEED_RESET;
3173 } else {
3174 /* If the interface is disabled we don't want to do anything
3175 * with it.
3176 */
3177 status = PCI_ERS_RESULT_RECOVERED;
3178 }
3179
3180 rtnl_unlock();
3181
3182 pci_disable_device(pdev);
3183
3184 return status;
3185 }
3186
3187 /* Fake a successful reset, which will be performed later in ef4_io_resume. */
3188 static pci_ers_result_t ef4_io_slot_reset(struct pci_dev *pdev)
3189 {
3190 struct ef4_nic *efx = pci_get_drvdata(pdev);
3191 pci_ers_result_t status = PCI_ERS_RESULT_RECOVERED;
3192 int rc;
3193
3194 if (pci_enable_device(pdev)) {
3195 netif_err(efx, hw, efx->net_dev,
3196 "Cannot re-enable PCI device after reset.\n");
3197 status = PCI_ERS_RESULT_DISCONNECT;
3198 }
3199
3200 rc = pci_cleanup_aer_uncorrect_error_status(pdev);
3201 if (rc) {
3202 netif_err(efx, hw, efx->net_dev,
3203 "pci_cleanup_aer_uncorrect_error_status failed (%d)\n", rc);
3204 /* Non-fatal error. Continue. */
3205 }
3206
3207 return status;
3208 }
3209
3210 /* Perform the actual reset and resume I/O operations. */
3211 static void ef4_io_resume(struct pci_dev *pdev)
3212 {
3213 struct ef4_nic *efx = pci_get_drvdata(pdev);
3214 int rc;
3215
3216 rtnl_lock();
3217
3218 if (efx->state == STATE_DISABLED)
3219 goto out;
3220
3221 rc = ef4_reset(efx, RESET_TYPE_ALL);
3222 if (rc) {
3223 netif_err(efx, hw, efx->net_dev,
3224 "ef4_reset failed after PCI error (%d)\n", rc);
3225 } else {
3226 efx->state = STATE_READY;
3227 netif_dbg(efx, hw, efx->net_dev,
3228 "Done resetting and resuming IO after PCI error.\n");
3229 }
3230
3231 out:
3232 rtnl_unlock();
3233 }
3234
3235 /* For simplicity and reliability, we always require a slot reset and try to
3236 * reset the hardware when a pci error affecting the device is detected.
3237 * We leave both the link_reset and mmio_enabled callback unimplemented:
3238 * with our request for slot reset the mmio_enabled callback will never be
3239 * called, and the link_reset callback is not used by AER or EEH mechanisms.
3240 */
3241 static const struct pci_error_handlers ef4_err_handlers = {
3242 .error_detected = ef4_io_error_detected,
3243 .slot_reset = ef4_io_slot_reset,
3244 .resume = ef4_io_resume,
3245 };
3246
3247 static struct pci_driver ef4_pci_driver = {
3248 .name = KBUILD_MODNAME,
3249 .id_table = ef4_pci_table,
3250 .probe = ef4_pci_probe,
3251 .remove = ef4_pci_remove,
3252 .driver.pm = &ef4_pm_ops,
3253 .err_handler = &ef4_err_handlers,
3254 };
3255
3256 /**************************************************************************
3257 *
3258 * Kernel module interface
3259 *
3260 *************************************************************************/
3261
3262 module_param(interrupt_mode, uint, 0444);
3263 MODULE_PARM_DESC(interrupt_mode,
3264 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
3265
3266 static int __init ef4_init_module(void)
3267 {
3268 int rc;
3269
3270 printk(KERN_INFO "Solarflare Falcon driver v" EF4_DRIVER_VERSION "\n");
3271
3272 rc = register_netdevice_notifier(&ef4_netdev_notifier);
3273 if (rc)
3274 goto err_notifier;
3275
3276 reset_workqueue = create_singlethread_workqueue("sfc_reset");
3277 if (!reset_workqueue) {
3278 rc = -ENOMEM;
3279 goto err_reset;
3280 }
3281
3282 rc = pci_register_driver(&ef4_pci_driver);
3283 if (rc < 0)
3284 goto err_pci;
3285
3286 return 0;
3287
3288 err_pci:
3289 destroy_workqueue(reset_workqueue);
3290 err_reset:
3291 unregister_netdevice_notifier(&ef4_netdev_notifier);
3292 err_notifier:
3293 return rc;
3294 }
3295
3296 static void __exit ef4_exit_module(void)
3297 {
3298 printk(KERN_INFO "Solarflare Falcon driver unloading\n");
3299
3300 pci_unregister_driver(&ef4_pci_driver);
3301 destroy_workqueue(reset_workqueue);
3302 unregister_netdevice_notifier(&ef4_netdev_notifier);
3303
3304 }
3305
3306 module_init(ef4_init_module);
3307 module_exit(ef4_exit_module);
3308
3309 MODULE_AUTHOR("Solarflare Communications and "
3310 "Michael Brown <mbrown@fensystems.co.uk>");
3311 MODULE_DESCRIPTION("Solarflare Falcon network driver");
3312 MODULE_LICENSE("GPL");
3313 MODULE_DEVICE_TABLE(pci, ef4_pci_table);
3314 MODULE_VERSION(EF4_DRIVER_VERSION);
Linux with added FPC-III support