xref: /linux/drivers/net/ethernet/sfc/efx_channels.c (revision 3494bec0f6ac8ac06e0ad7c35933db345b2c5a83)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3  * Driver for Solarflare network controllers and boards
4  * Copyright 2018 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 "net_driver.h"
12 #include <linux/module.h>
13 #include "efx_channels.h"
14 #include "efx.h"
15 #include "efx_common.h"
16 #include "tx_common.h"
17 #include "rx_common.h"
18 #include "nic.h"
19 #include "sriov.h"
20 
21 /* This is the first interrupt mode to try out of:
22  * 0 => MSI-X
23  * 1 => MSI
24  * 2 => legacy
25  */
26 static unsigned int interrupt_mode;
27 module_param(interrupt_mode, uint, 0444);
28 MODULE_PARM_DESC(interrupt_mode,
29 		 "Interrupt mode (0=>MSIX 1=>MSI 2=>legacy)");
30 
31 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
32  * i.e. the number of CPUs among which we may distribute simultaneous
33  * interrupt handling.
34  *
35  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
36  * The default (0) means to assign an interrupt to each core.
37  */
38 static unsigned int rss_cpus;
39 module_param(rss_cpus, uint, 0444);
40 MODULE_PARM_DESC(rss_cpus, "Number of CPUs to use for Receive-Side Scaling");
41 
42 static unsigned int irq_adapt_low_thresh = 8000;
43 module_param(irq_adapt_low_thresh, uint, 0644);
44 MODULE_PARM_DESC(irq_adapt_low_thresh,
45 		 "Threshold score for reducing IRQ moderation");
46 
47 static unsigned int irq_adapt_high_thresh = 16000;
48 module_param(irq_adapt_high_thresh, uint, 0644);
49 MODULE_PARM_DESC(irq_adapt_high_thresh,
50 		 "Threshold score for increasing IRQ moderation");
51 
52 /* This is the weight assigned to each of the (per-channel) virtual
53  * NAPI devices.
54  */
55 static int napi_weight = 64;
56 
57 /***************
58  * Housekeeping
59  ***************/
60 
61 int efx_channel_dummy_op_int(struct efx_channel *channel)
62 {
63 	return 0;
64 }
65 
66 void efx_channel_dummy_op_void(struct efx_channel *channel)
67 {
68 }
69 
70 static const struct efx_channel_type efx_default_channel_type = {
71 	.pre_probe		= efx_channel_dummy_op_int,
72 	.post_remove		= efx_channel_dummy_op_void,
73 	.get_name		= efx_get_channel_name,
74 	.copy			= efx_copy_channel,
75 	.want_txqs		= efx_default_channel_want_txqs,
76 	.keep_eventq		= false,
77 	.want_pio		= true,
78 };
79 
80 /*************
81  * INTERRUPTS
82  *************/
83 
84 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
85 {
86 	cpumask_var_t thread_mask;
87 	unsigned int count;
88 	int cpu;
89 
90 	if (rss_cpus) {
91 		count = rss_cpus;
92 	} else {
93 		if (unlikely(!zalloc_cpumask_var(&thread_mask, GFP_KERNEL))) {
94 			netif_warn(efx, probe, efx->net_dev,
95 				   "RSS disabled due to allocation failure\n");
96 			return 1;
97 		}
98 
99 		count = 0;
100 		for_each_online_cpu(cpu) {
101 			if (!cpumask_test_cpu(cpu, thread_mask)) {
102 				++count;
103 				cpumask_or(thread_mask, thread_mask,
104 					   topology_sibling_cpumask(cpu));
105 			}
106 		}
107 
108 		free_cpumask_var(thread_mask);
109 	}
110 
111 	if (count > EFX_MAX_RX_QUEUES) {
112 		netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
113 			       "Reducing number of rx queues from %u to %u.\n",
114 			       count, EFX_MAX_RX_QUEUES);
115 		count = EFX_MAX_RX_QUEUES;
116 	}
117 
118 	/* If RSS is requested for the PF *and* VFs then we can't write RSS
119 	 * table entries that are inaccessible to VFs
120 	 */
121 #ifdef CONFIG_SFC_SRIOV
122 	if (efx->type->sriov_wanted) {
123 		if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
124 		    count > efx_vf_size(efx)) {
125 			netif_warn(efx, probe, efx->net_dev,
126 				   "Reducing number of RSS channels from %u to %u for "
127 				   "VF support. Increase vf-msix-limit to use more "
128 				   "channels on the PF.\n",
129 				   count, efx_vf_size(efx));
130 			count = efx_vf_size(efx);
131 		}
132 	}
133 #endif
134 
135 	return count;
136 }
137 
138 static int efx_allocate_msix_channels(struct efx_nic *efx,
139 				      unsigned int max_channels,
140 				      unsigned int extra_channels,
141 				      unsigned int parallelism)
142 {
143 	unsigned int n_channels = parallelism;
144 	int vec_count;
145 	int n_xdp_tx;
146 	int n_xdp_ev;
147 
148 	if (efx_separate_tx_channels)
149 		n_channels *= 2;
150 	n_channels += extra_channels;
151 
152 	/* To allow XDP transmit to happen from arbitrary NAPI contexts
153 	 * we allocate a TX queue per CPU. We share event queues across
154 	 * multiple tx queues, assuming tx and ev queues are both
155 	 * maximum size.
156 	 */
157 
158 	n_xdp_tx = num_possible_cpus();
159 	n_xdp_ev = DIV_ROUND_UP(n_xdp_tx, EFX_TXQ_TYPES);
160 
161 	vec_count = pci_msix_vec_count(efx->pci_dev);
162 	if (vec_count < 0)
163 		return vec_count;
164 
165 	max_channels = min_t(unsigned int, vec_count, max_channels);
166 
167 	/* Check resources.
168 	 * We need a channel per event queue, plus a VI per tx queue.
169 	 * This may be more pessimistic than it needs to be.
170 	 */
171 	if (n_channels + n_xdp_ev > max_channels) {
172 		netif_err(efx, drv, efx->net_dev,
173 			  "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
174 			  n_xdp_ev, n_channels, max_channels);
175 		efx->n_xdp_channels = 0;
176 		efx->xdp_tx_per_channel = 0;
177 		efx->xdp_tx_queue_count = 0;
178 	} else {
179 		efx->n_xdp_channels = n_xdp_ev;
180 		efx->xdp_tx_per_channel = EFX_TXQ_TYPES;
181 		efx->xdp_tx_queue_count = n_xdp_tx;
182 		n_channels += n_xdp_ev;
183 		netif_dbg(efx, drv, efx->net_dev,
184 			  "Allocating %d TX and %d event queues for XDP\n",
185 			  n_xdp_tx, n_xdp_ev);
186 	}
187 
188 	if (vec_count < n_channels) {
189 		netif_err(efx, drv, efx->net_dev,
190 			  "WARNING: Insufficient MSI-X vectors available (%d < %u).\n",
191 			  vec_count, n_channels);
192 		netif_err(efx, drv, efx->net_dev,
193 			  "WARNING: Performance may be reduced.\n");
194 		n_channels = vec_count;
195 	}
196 
197 	n_channels = min(n_channels, max_channels);
198 
199 	efx->n_channels = n_channels;
200 
201 	/* Ignore XDP tx channels when creating rx channels. */
202 	n_channels -= efx->n_xdp_channels;
203 
204 	if (efx_separate_tx_channels) {
205 		efx->n_tx_channels =
206 			min(max(n_channels / 2, 1U),
207 			    efx->max_tx_channels);
208 		efx->tx_channel_offset =
209 			n_channels - efx->n_tx_channels;
210 		efx->n_rx_channels =
211 			max(n_channels -
212 			    efx->n_tx_channels, 1U);
213 	} else {
214 		efx->n_tx_channels = min(n_channels, efx->max_tx_channels);
215 		efx->tx_channel_offset = 0;
216 		efx->n_rx_channels = n_channels;
217 	}
218 
219 	efx->n_rx_channels = min(efx->n_rx_channels, parallelism);
220 	efx->n_tx_channels = min(efx->n_tx_channels, parallelism);
221 
222 	efx->xdp_channel_offset = n_channels;
223 
224 	netif_dbg(efx, drv, efx->net_dev,
225 		  "Allocating %u RX channels\n",
226 		  efx->n_rx_channels);
227 
228 	return efx->n_channels;
229 }
230 
231 /* Probe the number and type of interrupts we are able to obtain, and
232  * the resulting numbers of channels and RX queues.
233  */
234 int efx_probe_interrupts(struct efx_nic *efx)
235 {
236 	unsigned int extra_channels = 0;
237 	unsigned int rss_spread;
238 	unsigned int i, j;
239 	int rc;
240 
241 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
242 		if (efx->extra_channel_type[i])
243 			++extra_channels;
244 
245 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
246 		unsigned int parallelism = efx_wanted_parallelism(efx);
247 		struct msix_entry xentries[EFX_MAX_CHANNELS];
248 		unsigned int n_channels;
249 
250 		rc = efx_allocate_msix_channels(efx, efx->max_channels,
251 						extra_channels, parallelism);
252 		if (rc >= 0) {
253 			n_channels = rc;
254 			for (i = 0; i < n_channels; i++)
255 				xentries[i].entry = i;
256 			rc = pci_enable_msix_range(efx->pci_dev, xentries, 1,
257 						   n_channels);
258 		}
259 		if (rc < 0) {
260 			/* Fall back to single channel MSI */
261 			netif_err(efx, drv, efx->net_dev,
262 				  "could not enable MSI-X\n");
263 			if (efx->type->min_interrupt_mode >= EFX_INT_MODE_MSI)
264 				efx->interrupt_mode = EFX_INT_MODE_MSI;
265 			else
266 				return rc;
267 		} else if (rc < n_channels) {
268 			netif_err(efx, drv, efx->net_dev,
269 				  "WARNING: Insufficient MSI-X vectors"
270 				  " available (%d < %u).\n", rc, n_channels);
271 			netif_err(efx, drv, efx->net_dev,
272 				  "WARNING: Performance may be reduced.\n");
273 			n_channels = rc;
274 		}
275 
276 		if (rc > 0) {
277 			for (i = 0; i < efx->n_channels; i++)
278 				efx_get_channel(efx, i)->irq =
279 					xentries[i].vector;
280 		}
281 	}
282 
283 	/* Try single interrupt MSI */
284 	if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
285 		efx->n_channels = 1;
286 		efx->n_rx_channels = 1;
287 		efx->n_tx_channels = 1;
288 		efx->n_xdp_channels = 0;
289 		efx->xdp_channel_offset = efx->n_channels;
290 		rc = pci_enable_msi(efx->pci_dev);
291 		if (rc == 0) {
292 			efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
293 		} else {
294 			netif_err(efx, drv, efx->net_dev,
295 				  "could not enable MSI\n");
296 			if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY)
297 				efx->interrupt_mode = EFX_INT_MODE_LEGACY;
298 			else
299 				return rc;
300 		}
301 	}
302 
303 	/* Assume legacy interrupts */
304 	if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
305 		efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
306 		efx->n_rx_channels = 1;
307 		efx->n_tx_channels = 1;
308 		efx->n_xdp_channels = 0;
309 		efx->xdp_channel_offset = efx->n_channels;
310 		efx->legacy_irq = efx->pci_dev->irq;
311 	}
312 
313 	/* Assign extra channels if possible, before XDP channels */
314 	efx->n_extra_tx_channels = 0;
315 	j = efx->xdp_channel_offset;
316 	for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
317 		if (!efx->extra_channel_type[i])
318 			continue;
319 		if (j <= efx->tx_channel_offset + efx->n_tx_channels) {
320 			efx->extra_channel_type[i]->handle_no_channel(efx);
321 		} else {
322 			--j;
323 			efx_get_channel(efx, j)->type =
324 				efx->extra_channel_type[i];
325 			if (efx_channel_has_tx_queues(efx_get_channel(efx, j)))
326 				efx->n_extra_tx_channels++;
327 		}
328 	}
329 
330 	rss_spread = efx->n_rx_channels;
331 	/* RSS might be usable on VFs even if it is disabled on the PF */
332 #ifdef CONFIG_SFC_SRIOV
333 	if (efx->type->sriov_wanted) {
334 		efx->rss_spread = ((rss_spread > 1 ||
335 				    !efx->type->sriov_wanted(efx)) ?
336 				   rss_spread : efx_vf_size(efx));
337 		return 0;
338 	}
339 #endif
340 	efx->rss_spread = rss_spread;
341 
342 	return 0;
343 }
344 
345 #if defined(CONFIG_SMP)
346 void efx_set_interrupt_affinity(struct efx_nic *efx)
347 {
348 	struct efx_channel *channel;
349 	unsigned int cpu;
350 
351 	efx_for_each_channel(channel, efx) {
352 		cpu = cpumask_local_spread(channel->channel,
353 					   pcibus_to_node(efx->pci_dev->bus));
354 		irq_set_affinity_hint(channel->irq, cpumask_of(cpu));
355 	}
356 }
357 
358 void efx_clear_interrupt_affinity(struct efx_nic *efx)
359 {
360 	struct efx_channel *channel;
361 
362 	efx_for_each_channel(channel, efx)
363 		irq_set_affinity_hint(channel->irq, NULL);
364 }
365 #else
366 void
367 efx_set_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
368 {
369 }
370 
371 void
372 efx_clear_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
373 {
374 }
375 #endif /* CONFIG_SMP */
376 
377 void efx_remove_interrupts(struct efx_nic *efx)
378 {
379 	struct efx_channel *channel;
380 
381 	/* Remove MSI/MSI-X interrupts */
382 	efx_for_each_channel(channel, efx)
383 		channel->irq = 0;
384 	pci_disable_msi(efx->pci_dev);
385 	pci_disable_msix(efx->pci_dev);
386 
387 	/* Remove legacy interrupt */
388 	efx->legacy_irq = 0;
389 }
390 
391 /***************
392  * EVENT QUEUES
393  ***************/
394 
395 /* Create event queue
396  * Event queue memory allocations are done only once.  If the channel
397  * is reset, the memory buffer will be reused; this guards against
398  * errors during channel reset and also simplifies interrupt handling.
399  */
400 int efx_probe_eventq(struct efx_channel *channel)
401 {
402 	struct efx_nic *efx = channel->efx;
403 	unsigned long entries;
404 
405 	netif_dbg(efx, probe, efx->net_dev,
406 		  "chan %d create event queue\n", channel->channel);
407 
408 	/* Build an event queue with room for one event per tx and rx buffer,
409 	 * plus some extra for link state events and MCDI completions.
410 	 */
411 	entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
412 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
413 	channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
414 
415 	return efx_nic_probe_eventq(channel);
416 }
417 
418 /* Prepare channel's event queue */
419 int efx_init_eventq(struct efx_channel *channel)
420 {
421 	struct efx_nic *efx = channel->efx;
422 	int rc;
423 
424 	EFX_WARN_ON_PARANOID(channel->eventq_init);
425 
426 	netif_dbg(efx, drv, efx->net_dev,
427 		  "chan %d init event queue\n", channel->channel);
428 
429 	rc = efx_nic_init_eventq(channel);
430 	if (rc == 0) {
431 		efx->type->push_irq_moderation(channel);
432 		channel->eventq_read_ptr = 0;
433 		channel->eventq_init = true;
434 	}
435 	return rc;
436 }
437 
438 /* Enable event queue processing and NAPI */
439 void efx_start_eventq(struct efx_channel *channel)
440 {
441 	netif_dbg(channel->efx, ifup, channel->efx->net_dev,
442 		  "chan %d start event queue\n", channel->channel);
443 
444 	/* Make sure the NAPI handler sees the enabled flag set */
445 	channel->enabled = true;
446 	smp_wmb();
447 
448 	napi_enable(&channel->napi_str);
449 	efx_nic_eventq_read_ack(channel);
450 }
451 
452 /* Disable event queue processing and NAPI */
453 void efx_stop_eventq(struct efx_channel *channel)
454 {
455 	if (!channel->enabled)
456 		return;
457 
458 	napi_disable(&channel->napi_str);
459 	channel->enabled = false;
460 }
461 
462 void efx_fini_eventq(struct efx_channel *channel)
463 {
464 	if (!channel->eventq_init)
465 		return;
466 
467 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
468 		  "chan %d fini event queue\n", channel->channel);
469 
470 	efx_nic_fini_eventq(channel);
471 	channel->eventq_init = false;
472 }
473 
474 void efx_remove_eventq(struct efx_channel *channel)
475 {
476 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
477 		  "chan %d remove event queue\n", channel->channel);
478 
479 	efx_nic_remove_eventq(channel);
480 }
481 
482 /**************************************************************************
483  *
484  * Channel handling
485  *
486  *************************************************************************/
487 
488 #ifdef CONFIG_RFS_ACCEL
489 static void efx_filter_rfs_expire(struct work_struct *data)
490 {
491 	struct delayed_work *dwork = to_delayed_work(data);
492 	struct efx_channel *channel;
493 	unsigned int time, quota;
494 
495 	channel = container_of(dwork, struct efx_channel, filter_work);
496 	time = jiffies - channel->rfs_last_expiry;
497 	quota = channel->rfs_filter_count * time / (30 * HZ);
498 	if (quota >= 20 && __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, quota)))
499 		channel->rfs_last_expiry += time;
500 	/* Ensure we do more work eventually even if NAPI poll is not happening */
501 	schedule_delayed_work(dwork, 30 * HZ);
502 }
503 #endif
504 
505 /* Allocate and initialise a channel structure. */
506 struct efx_channel *
507 efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
508 {
509 	struct efx_rx_queue *rx_queue;
510 	struct efx_tx_queue *tx_queue;
511 	struct efx_channel *channel;
512 	int j;
513 
514 	channel = kzalloc(sizeof(*channel), GFP_KERNEL);
515 	if (!channel)
516 		return NULL;
517 
518 	channel->efx = efx;
519 	channel->channel = i;
520 	channel->type = &efx_default_channel_type;
521 
522 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
523 		tx_queue = &channel->tx_queue[j];
524 		tx_queue->efx = efx;
525 		tx_queue->queue = i * EFX_TXQ_TYPES + j;
526 		tx_queue->channel = channel;
527 	}
528 
529 #ifdef CONFIG_RFS_ACCEL
530 	INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire);
531 #endif
532 
533 	rx_queue = &channel->rx_queue;
534 	rx_queue->efx = efx;
535 	timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
536 
537 	return channel;
538 }
539 
540 int efx_init_channels(struct efx_nic *efx)
541 {
542 	unsigned int i;
543 
544 	for (i = 0; i < EFX_MAX_CHANNELS; i++) {
545 		efx->channel[i] = efx_alloc_channel(efx, i, NULL);
546 		if (!efx->channel[i])
547 			return -ENOMEM;
548 		efx->msi_context[i].efx = efx;
549 		efx->msi_context[i].index = i;
550 	}
551 
552 	/* Higher numbered interrupt modes are less capable! */
553 	if (WARN_ON_ONCE(efx->type->max_interrupt_mode >
554 			 efx->type->min_interrupt_mode)) {
555 		return -EIO;
556 	}
557 	efx->interrupt_mode = max(efx->type->max_interrupt_mode,
558 				  interrupt_mode);
559 	efx->interrupt_mode = min(efx->type->min_interrupt_mode,
560 				  interrupt_mode);
561 
562 	return 0;
563 }
564 
565 void efx_fini_channels(struct efx_nic *efx)
566 {
567 	unsigned int i;
568 
569 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
570 		if (efx->channel[i]) {
571 			kfree(efx->channel[i]);
572 			efx->channel[i] = NULL;
573 		}
574 }
575 
576 /* Allocate and initialise a channel structure, copying parameters
577  * (but not resources) from an old channel structure.
578  */
579 struct efx_channel *efx_copy_channel(const struct efx_channel *old_channel)
580 {
581 	struct efx_rx_queue *rx_queue;
582 	struct efx_tx_queue *tx_queue;
583 	struct efx_channel *channel;
584 	int j;
585 
586 	channel = kmalloc(sizeof(*channel), GFP_KERNEL);
587 	if (!channel)
588 		return NULL;
589 
590 	*channel = *old_channel;
591 
592 	channel->napi_dev = NULL;
593 	INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
594 	channel->napi_str.napi_id = 0;
595 	channel->napi_str.state = 0;
596 	memset(&channel->eventq, 0, sizeof(channel->eventq));
597 
598 	for (j = 0; j < EFX_TXQ_TYPES; j++) {
599 		tx_queue = &channel->tx_queue[j];
600 		if (tx_queue->channel)
601 			tx_queue->channel = channel;
602 		tx_queue->buffer = NULL;
603 		memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
604 	}
605 
606 	rx_queue = &channel->rx_queue;
607 	rx_queue->buffer = NULL;
608 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
609 	timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
610 #ifdef CONFIG_RFS_ACCEL
611 	INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire);
612 #endif
613 
614 	return channel;
615 }
616 
617 static int efx_probe_channel(struct efx_channel *channel)
618 {
619 	struct efx_tx_queue *tx_queue;
620 	struct efx_rx_queue *rx_queue;
621 	int rc;
622 
623 	netif_dbg(channel->efx, probe, channel->efx->net_dev,
624 		  "creating channel %d\n", channel->channel);
625 
626 	rc = channel->type->pre_probe(channel);
627 	if (rc)
628 		goto fail;
629 
630 	rc = efx_probe_eventq(channel);
631 	if (rc)
632 		goto fail;
633 
634 	efx_for_each_channel_tx_queue(tx_queue, channel) {
635 		rc = efx_probe_tx_queue(tx_queue);
636 		if (rc)
637 			goto fail;
638 	}
639 
640 	efx_for_each_channel_rx_queue(rx_queue, channel) {
641 		rc = efx_probe_rx_queue(rx_queue);
642 		if (rc)
643 			goto fail;
644 	}
645 
646 	channel->rx_list = NULL;
647 
648 	return 0;
649 
650 fail:
651 	efx_remove_channel(channel);
652 	return rc;
653 }
654 
655 void efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
656 {
657 	struct efx_nic *efx = channel->efx;
658 	const char *type;
659 	int number;
660 
661 	number = channel->channel;
662 
663 	if (number >= efx->xdp_channel_offset &&
664 	    !WARN_ON_ONCE(!efx->n_xdp_channels)) {
665 		type = "-xdp";
666 		number -= efx->xdp_channel_offset;
667 	} else if (efx->tx_channel_offset == 0) {
668 		type = "";
669 	} else if (number < efx->tx_channel_offset) {
670 		type = "-rx";
671 	} else {
672 		type = "-tx";
673 		number -= efx->tx_channel_offset;
674 	}
675 	snprintf(buf, len, "%s%s-%d", efx->name, type, number);
676 }
677 
678 void efx_set_channel_names(struct efx_nic *efx)
679 {
680 	struct efx_channel *channel;
681 
682 	efx_for_each_channel(channel, efx)
683 		channel->type->get_name(channel,
684 					efx->msi_context[channel->channel].name,
685 					sizeof(efx->msi_context[0].name));
686 }
687 
688 int efx_probe_channels(struct efx_nic *efx)
689 {
690 	struct efx_channel *channel;
691 	int rc;
692 
693 	/* Restart special buffer allocation */
694 	efx->next_buffer_table = 0;
695 
696 	/* Probe channels in reverse, so that any 'extra' channels
697 	 * use the start of the buffer table. This allows the traffic
698 	 * channels to be resized without moving them or wasting the
699 	 * entries before them.
700 	 */
701 	efx_for_each_channel_rev(channel, efx) {
702 		rc = efx_probe_channel(channel);
703 		if (rc) {
704 			netif_err(efx, probe, efx->net_dev,
705 				  "failed to create channel %d\n",
706 				  channel->channel);
707 			goto fail;
708 		}
709 	}
710 	efx_set_channel_names(efx);
711 
712 	return 0;
713 
714 fail:
715 	efx_remove_channels(efx);
716 	return rc;
717 }
718 
719 void efx_remove_channel(struct efx_channel *channel)
720 {
721 	struct efx_tx_queue *tx_queue;
722 	struct efx_rx_queue *rx_queue;
723 
724 	netif_dbg(channel->efx, drv, channel->efx->net_dev,
725 		  "destroy chan %d\n", channel->channel);
726 
727 	efx_for_each_channel_rx_queue(rx_queue, channel)
728 		efx_remove_rx_queue(rx_queue);
729 	efx_for_each_possible_channel_tx_queue(tx_queue, channel)
730 		efx_remove_tx_queue(tx_queue);
731 	efx_remove_eventq(channel);
732 	channel->type->post_remove(channel);
733 }
734 
735 void efx_remove_channels(struct efx_nic *efx)
736 {
737 	struct efx_channel *channel;
738 
739 	efx_for_each_channel(channel, efx)
740 		efx_remove_channel(channel);
741 
742 	kfree(efx->xdp_tx_queues);
743 }
744 
745 int efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
746 {
747 	struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel;
748 	unsigned int i, next_buffer_table = 0;
749 	u32 old_rxq_entries, old_txq_entries;
750 	int rc, rc2;
751 
752 	rc = efx_check_disabled(efx);
753 	if (rc)
754 		return rc;
755 
756 	/* Not all channels should be reallocated. We must avoid
757 	 * reallocating their buffer table entries.
758 	 */
759 	efx_for_each_channel(channel, efx) {
760 		struct efx_rx_queue *rx_queue;
761 		struct efx_tx_queue *tx_queue;
762 
763 		if (channel->type->copy)
764 			continue;
765 		next_buffer_table = max(next_buffer_table,
766 					channel->eventq.index +
767 					channel->eventq.entries);
768 		efx_for_each_channel_rx_queue(rx_queue, channel)
769 			next_buffer_table = max(next_buffer_table,
770 						rx_queue->rxd.index +
771 						rx_queue->rxd.entries);
772 		efx_for_each_channel_tx_queue(tx_queue, channel)
773 			next_buffer_table = max(next_buffer_table,
774 						tx_queue->txd.index +
775 						tx_queue->txd.entries);
776 	}
777 
778 	efx_device_detach_sync(efx);
779 	efx_stop_all(efx);
780 	efx_soft_disable_interrupts(efx);
781 
782 	/* Clone channels (where possible) */
783 	memset(other_channel, 0, sizeof(other_channel));
784 	for (i = 0; i < efx->n_channels; i++) {
785 		channel = efx->channel[i];
786 		if (channel->type->copy)
787 			channel = channel->type->copy(channel);
788 		if (!channel) {
789 			rc = -ENOMEM;
790 			goto out;
791 		}
792 		other_channel[i] = channel;
793 	}
794 
795 	/* Swap entry counts and channel pointers */
796 	old_rxq_entries = efx->rxq_entries;
797 	old_txq_entries = efx->txq_entries;
798 	efx->rxq_entries = rxq_entries;
799 	efx->txq_entries = txq_entries;
800 	for (i = 0; i < efx->n_channels; i++) {
801 		channel = efx->channel[i];
802 		efx->channel[i] = other_channel[i];
803 		other_channel[i] = channel;
804 	}
805 
806 	/* Restart buffer table allocation */
807 	efx->next_buffer_table = next_buffer_table;
808 
809 	for (i = 0; i < efx->n_channels; i++) {
810 		channel = efx->channel[i];
811 		if (!channel->type->copy)
812 			continue;
813 		rc = efx_probe_channel(channel);
814 		if (rc)
815 			goto rollback;
816 		efx_init_napi_channel(efx->channel[i]);
817 	}
818 
819 out:
820 	/* Destroy unused channel structures */
821 	for (i = 0; i < efx->n_channels; i++) {
822 		channel = other_channel[i];
823 		if (channel && channel->type->copy) {
824 			efx_fini_napi_channel(channel);
825 			efx_remove_channel(channel);
826 			kfree(channel);
827 		}
828 	}
829 
830 	rc2 = efx_soft_enable_interrupts(efx);
831 	if (rc2) {
832 		rc = rc ? rc : rc2;
833 		netif_err(efx, drv, efx->net_dev,
834 			  "unable to restart interrupts on channel reallocation\n");
835 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
836 	} else {
837 		efx_start_all(efx);
838 		efx_device_attach_if_not_resetting(efx);
839 	}
840 	return rc;
841 
842 rollback:
843 	/* Swap back */
844 	efx->rxq_entries = old_rxq_entries;
845 	efx->txq_entries = old_txq_entries;
846 	for (i = 0; i < efx->n_channels; i++) {
847 		channel = efx->channel[i];
848 		efx->channel[i] = other_channel[i];
849 		other_channel[i] = channel;
850 	}
851 	goto out;
852 }
853 
854 int efx_set_channels(struct efx_nic *efx)
855 {
856 	struct efx_channel *channel;
857 	struct efx_tx_queue *tx_queue;
858 	int xdp_queue_number;
859 
860 	efx->tx_channel_offset =
861 		efx_separate_tx_channels ?
862 		efx->n_channels - efx->n_tx_channels : 0;
863 
864 	if (efx->xdp_tx_queue_count) {
865 		EFX_WARN_ON_PARANOID(efx->xdp_tx_queues);
866 
867 		/* Allocate array for XDP TX queue lookup. */
868 		efx->xdp_tx_queues = kcalloc(efx->xdp_tx_queue_count,
869 					     sizeof(*efx->xdp_tx_queues),
870 					     GFP_KERNEL);
871 		if (!efx->xdp_tx_queues)
872 			return -ENOMEM;
873 	}
874 
875 	/* We need to mark which channels really have RX and TX
876 	 * queues, and adjust the TX queue numbers if we have separate
877 	 * RX-only and TX-only channels.
878 	 */
879 	xdp_queue_number = 0;
880 	efx_for_each_channel(channel, efx) {
881 		if (channel->channel < efx->n_rx_channels)
882 			channel->rx_queue.core_index = channel->channel;
883 		else
884 			channel->rx_queue.core_index = -1;
885 
886 		efx_for_each_channel_tx_queue(tx_queue, channel) {
887 			tx_queue->queue -= (efx->tx_channel_offset *
888 					    EFX_TXQ_TYPES);
889 
890 			if (efx_channel_is_xdp_tx(channel) &&
891 			    xdp_queue_number < efx->xdp_tx_queue_count) {
892 				efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
893 				xdp_queue_number++;
894 			}
895 		}
896 	}
897 	return 0;
898 }
899 
900 bool efx_default_channel_want_txqs(struct efx_channel *channel)
901 {
902 	return channel->channel - channel->efx->tx_channel_offset <
903 		channel->efx->n_tx_channels;
904 }
905 
906 /*************
907  * START/STOP
908  *************/
909 
910 int efx_soft_enable_interrupts(struct efx_nic *efx)
911 {
912 	struct efx_channel *channel, *end_channel;
913 	int rc;
914 
915 	BUG_ON(efx->state == STATE_DISABLED);
916 
917 	efx->irq_soft_enabled = true;
918 	smp_wmb();
919 
920 	efx_for_each_channel(channel, efx) {
921 		if (!channel->type->keep_eventq) {
922 			rc = efx_init_eventq(channel);
923 			if (rc)
924 				goto fail;
925 		}
926 		efx_start_eventq(channel);
927 	}
928 
929 	efx_mcdi_mode_event(efx);
930 
931 	return 0;
932 fail:
933 	end_channel = channel;
934 	efx_for_each_channel(channel, efx) {
935 		if (channel == end_channel)
936 			break;
937 		efx_stop_eventq(channel);
938 		if (!channel->type->keep_eventq)
939 			efx_fini_eventq(channel);
940 	}
941 
942 	return rc;
943 }
944 
945 void efx_soft_disable_interrupts(struct efx_nic *efx)
946 {
947 	struct efx_channel *channel;
948 
949 	if (efx->state == STATE_DISABLED)
950 		return;
951 
952 	efx_mcdi_mode_poll(efx);
953 
954 	efx->irq_soft_enabled = false;
955 	smp_wmb();
956 
957 	if (efx->legacy_irq)
958 		synchronize_irq(efx->legacy_irq);
959 
960 	efx_for_each_channel(channel, efx) {
961 		if (channel->irq)
962 			synchronize_irq(channel->irq);
963 
964 		efx_stop_eventq(channel);
965 		if (!channel->type->keep_eventq)
966 			efx_fini_eventq(channel);
967 	}
968 
969 	/* Flush the asynchronous MCDI request queue */
970 	efx_mcdi_flush_async(efx);
971 }
972 
973 int efx_enable_interrupts(struct efx_nic *efx)
974 {
975 	struct efx_channel *channel, *end_channel;
976 	int rc;
977 
978 	/* TODO: Is this really a bug? */
979 	BUG_ON(efx->state == STATE_DISABLED);
980 
981 	if (efx->eeh_disabled_legacy_irq) {
982 		enable_irq(efx->legacy_irq);
983 		efx->eeh_disabled_legacy_irq = false;
984 	}
985 
986 	efx->type->irq_enable_master(efx);
987 
988 	efx_for_each_channel(channel, efx) {
989 		if (channel->type->keep_eventq) {
990 			rc = efx_init_eventq(channel);
991 			if (rc)
992 				goto fail;
993 		}
994 	}
995 
996 	rc = efx_soft_enable_interrupts(efx);
997 	if (rc)
998 		goto fail;
999 
1000 	return 0;
1001 
1002 fail:
1003 	end_channel = channel;
1004 	efx_for_each_channel(channel, efx) {
1005 		if (channel == end_channel)
1006 			break;
1007 		if (channel->type->keep_eventq)
1008 			efx_fini_eventq(channel);
1009 	}
1010 
1011 	efx->type->irq_disable_non_ev(efx);
1012 
1013 	return rc;
1014 }
1015 
1016 void efx_disable_interrupts(struct efx_nic *efx)
1017 {
1018 	struct efx_channel *channel;
1019 
1020 	efx_soft_disable_interrupts(efx);
1021 
1022 	efx_for_each_channel(channel, efx) {
1023 		if (channel->type->keep_eventq)
1024 			efx_fini_eventq(channel);
1025 	}
1026 
1027 	efx->type->irq_disable_non_ev(efx);
1028 }
1029 
1030 void efx_start_channels(struct efx_nic *efx)
1031 {
1032 	struct efx_tx_queue *tx_queue;
1033 	struct efx_rx_queue *rx_queue;
1034 	struct efx_channel *channel;
1035 
1036 	efx_for_each_channel(channel, efx) {
1037 		efx_for_each_channel_tx_queue(tx_queue, channel) {
1038 			efx_init_tx_queue(tx_queue);
1039 			atomic_inc(&efx->active_queues);
1040 		}
1041 
1042 		efx_for_each_channel_rx_queue(rx_queue, channel) {
1043 			efx_init_rx_queue(rx_queue);
1044 			atomic_inc(&efx->active_queues);
1045 			efx_stop_eventq(channel);
1046 			efx_fast_push_rx_descriptors(rx_queue, false);
1047 			efx_start_eventq(channel);
1048 		}
1049 
1050 		WARN_ON(channel->rx_pkt_n_frags);
1051 	}
1052 }
1053 
1054 void efx_stop_channels(struct efx_nic *efx)
1055 {
1056 	struct efx_tx_queue *tx_queue;
1057 	struct efx_rx_queue *rx_queue;
1058 	struct efx_channel *channel;
1059 	int rc = 0;
1060 
1061 	/* Stop RX refill */
1062 	efx_for_each_channel(channel, efx) {
1063 		efx_for_each_channel_rx_queue(rx_queue, channel)
1064 			rx_queue->refill_enabled = false;
1065 	}
1066 
1067 	efx_for_each_channel(channel, efx) {
1068 		/* RX packet processing is pipelined, so wait for the
1069 		 * NAPI handler to complete.  At least event queue 0
1070 		 * might be kept active by non-data events, so don't
1071 		 * use napi_synchronize() but actually disable NAPI
1072 		 * temporarily.
1073 		 */
1074 		if (efx_channel_has_rx_queue(channel)) {
1075 			efx_stop_eventq(channel);
1076 			efx_start_eventq(channel);
1077 		}
1078 	}
1079 
1080 	if (efx->type->fini_dmaq)
1081 		rc = efx->type->fini_dmaq(efx);
1082 
1083 	if (rc) {
1084 		netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
1085 	} else {
1086 		netif_dbg(efx, drv, efx->net_dev,
1087 			  "successfully flushed all queues\n");
1088 	}
1089 
1090 	efx_for_each_channel(channel, efx) {
1091 		efx_for_each_channel_rx_queue(rx_queue, channel)
1092 			efx_fini_rx_queue(rx_queue);
1093 		efx_for_each_possible_channel_tx_queue(tx_queue, channel)
1094 			efx_fini_tx_queue(tx_queue);
1095 	}
1096 }
1097 
1098 /**************************************************************************
1099  *
1100  * NAPI interface
1101  *
1102  *************************************************************************/
1103 
1104 /* Process channel's event queue
1105  *
1106  * This function is responsible for processing the event queue of a
1107  * single channel.  The caller must guarantee that this function will
1108  * never be concurrently called more than once on the same channel,
1109  * though different channels may be being processed concurrently.
1110  */
1111 static int efx_process_channel(struct efx_channel *channel, int budget)
1112 {
1113 	struct efx_tx_queue *tx_queue;
1114 	struct list_head rx_list;
1115 	int spent;
1116 
1117 	if (unlikely(!channel->enabled))
1118 		return 0;
1119 
1120 	/* Prepare the batch receive list */
1121 	EFX_WARN_ON_PARANOID(channel->rx_list != NULL);
1122 	INIT_LIST_HEAD(&rx_list);
1123 	channel->rx_list = &rx_list;
1124 
1125 	efx_for_each_channel_tx_queue(tx_queue, channel) {
1126 		tx_queue->pkts_compl = 0;
1127 		tx_queue->bytes_compl = 0;
1128 	}
1129 
1130 	spent = efx_nic_process_eventq(channel, budget);
1131 	if (spent && efx_channel_has_rx_queue(channel)) {
1132 		struct efx_rx_queue *rx_queue =
1133 			efx_channel_get_rx_queue(channel);
1134 
1135 		efx_rx_flush_packet(channel);
1136 		efx_fast_push_rx_descriptors(rx_queue, true);
1137 	}
1138 
1139 	/* Update BQL */
1140 	efx_for_each_channel_tx_queue(tx_queue, channel) {
1141 		if (tx_queue->bytes_compl) {
1142 			netdev_tx_completed_queue(tx_queue->core_txq,
1143 						  tx_queue->pkts_compl,
1144 						  tx_queue->bytes_compl);
1145 		}
1146 	}
1147 
1148 	/* Receive any packets we queued up */
1149 	netif_receive_skb_list(channel->rx_list);
1150 	channel->rx_list = NULL;
1151 
1152 	return spent;
1153 }
1154 
1155 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
1156 {
1157 	int step = efx->irq_mod_step_us;
1158 
1159 	if (channel->irq_mod_score < irq_adapt_low_thresh) {
1160 		if (channel->irq_moderation_us > step) {
1161 			channel->irq_moderation_us -= step;
1162 			efx->type->push_irq_moderation(channel);
1163 		}
1164 	} else if (channel->irq_mod_score > irq_adapt_high_thresh) {
1165 		if (channel->irq_moderation_us <
1166 		    efx->irq_rx_moderation_us) {
1167 			channel->irq_moderation_us += step;
1168 			efx->type->push_irq_moderation(channel);
1169 		}
1170 	}
1171 
1172 	channel->irq_count = 0;
1173 	channel->irq_mod_score = 0;
1174 }
1175 
1176 /* NAPI poll handler
1177  *
1178  * NAPI guarantees serialisation of polls of the same device, which
1179  * provides the guarantee required by efx_process_channel().
1180  */
1181 static int efx_poll(struct napi_struct *napi, int budget)
1182 {
1183 	struct efx_channel *channel =
1184 		container_of(napi, struct efx_channel, napi_str);
1185 	struct efx_nic *efx = channel->efx;
1186 #ifdef CONFIG_RFS_ACCEL
1187 	unsigned int time;
1188 #endif
1189 	int spent;
1190 
1191 	netif_vdbg(efx, intr, efx->net_dev,
1192 		   "channel %d NAPI poll executing on CPU %d\n",
1193 		   channel->channel, raw_smp_processor_id());
1194 
1195 	spent = efx_process_channel(channel, budget);
1196 
1197 	xdp_do_flush_map();
1198 
1199 	if (spent < budget) {
1200 		if (efx_channel_has_rx_queue(channel) &&
1201 		    efx->irq_rx_adaptive &&
1202 		    unlikely(++channel->irq_count == 1000)) {
1203 			efx_update_irq_mod(efx, channel);
1204 		}
1205 
1206 #ifdef CONFIG_RFS_ACCEL
1207 		/* Perhaps expire some ARFS filters */
1208 		time = jiffies - channel->rfs_last_expiry;
1209 		/* Would our quota be >= 20? */
1210 		if (channel->rfs_filter_count * time >= 600 * HZ)
1211 			mod_delayed_work(system_wq, &channel->filter_work, 0);
1212 #endif
1213 
1214 		/* There is no race here; although napi_disable() will
1215 		 * only wait for napi_complete(), this isn't a problem
1216 		 * since efx_nic_eventq_read_ack() will have no effect if
1217 		 * interrupts have already been disabled.
1218 		 */
1219 		if (napi_complete_done(napi, spent))
1220 			efx_nic_eventq_read_ack(channel);
1221 	}
1222 
1223 	return spent;
1224 }
1225 
1226 void efx_init_napi_channel(struct efx_channel *channel)
1227 {
1228 	struct efx_nic *efx = channel->efx;
1229 
1230 	channel->napi_dev = efx->net_dev;
1231 	netif_napi_add(channel->napi_dev, &channel->napi_str,
1232 		       efx_poll, napi_weight);
1233 }
1234 
1235 void efx_init_napi(struct efx_nic *efx)
1236 {
1237 	struct efx_channel *channel;
1238 
1239 	efx_for_each_channel(channel, efx)
1240 		efx_init_napi_channel(channel);
1241 }
1242 
1243 void efx_fini_napi_channel(struct efx_channel *channel)
1244 {
1245 	if (channel->napi_dev)
1246 		netif_napi_del(&channel->napi_str);
1247 
1248 	channel->napi_dev = NULL;
1249 }
1250 
1251 void efx_fini_napi(struct efx_nic *efx)
1252 {
1253 	struct efx_channel *channel;
1254 
1255 	efx_for_each_channel(channel, efx)
1256 		efx_fini_napi_channel(channel);
1257 }
1258