Merge tag 'bcachefs-2024-10-05' of git://evilpiepirate.org/bcachefs
[linux-block.git] / drivers / net / ethernet / sfc / efx_channels.c
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 <linux/filter.h>
14 #include "efx_channels.h"
15 #include "efx.h"
16 #include "efx_common.h"
17 #include "tx_common.h"
18 #include "rx_common.h"
19 #include "nic.h"
20 #include "sriov.h"
21 #include "workarounds.h"
22
23 /* This is the first interrupt mode to try out of:
24  * 0 => MSI-X
25  * 1 => MSI
26  * 2 => legacy
27  */
28 unsigned int efx_interrupt_mode = EFX_INT_MODE_MSIX;
29
30 /* This is the requested number of CPUs to use for Receive-Side Scaling (RSS),
31  * i.e. the number of CPUs among which we may distribute simultaneous
32  * interrupt handling.
33  *
34  * Cards without MSI-X will only target one CPU via legacy or MSI interrupt.
35  * The default (0) means to assign an interrupt to each core.
36  */
37 unsigned int rss_cpus;
38
39 static unsigned int irq_adapt_low_thresh = 8000;
40 module_param(irq_adapt_low_thresh, uint, 0644);
41 MODULE_PARM_DESC(irq_adapt_low_thresh,
42                  "Threshold score for reducing IRQ moderation");
43
44 static unsigned int irq_adapt_high_thresh = 16000;
45 module_param(irq_adapt_high_thresh, uint, 0644);
46 MODULE_PARM_DESC(irq_adapt_high_thresh,
47                  "Threshold score for increasing IRQ moderation");
48
49 static const struct efx_channel_type efx_default_channel_type;
50
51 /*************
52  * INTERRUPTS
53  *************/
54
55 static unsigned int count_online_cores(struct efx_nic *efx, bool local_node)
56 {
57         cpumask_var_t filter_mask;
58         unsigned int count;
59         int cpu;
60
61         if (unlikely(!zalloc_cpumask_var(&filter_mask, GFP_KERNEL))) {
62                 netif_warn(efx, probe, efx->net_dev,
63                            "RSS disabled due to allocation failure\n");
64                 return 1;
65         }
66
67         cpumask_copy(filter_mask, cpu_online_mask);
68         if (local_node)
69                 cpumask_and(filter_mask, filter_mask,
70                             cpumask_of_pcibus(efx->pci_dev->bus));
71
72         count = 0;
73         for_each_cpu(cpu, filter_mask) {
74                 ++count;
75                 cpumask_andnot(filter_mask, filter_mask, topology_sibling_cpumask(cpu));
76         }
77
78         free_cpumask_var(filter_mask);
79
80         return count;
81 }
82
83 static unsigned int efx_wanted_parallelism(struct efx_nic *efx)
84 {
85         unsigned int count;
86
87         if (rss_cpus) {
88                 count = rss_cpus;
89         } else {
90                 count = count_online_cores(efx, true);
91
92                 /* If no online CPUs in local node, fallback to any online CPUs */
93                 if (count == 0)
94                         count = count_online_cores(efx, false);
95         }
96
97         if (count > EFX_MAX_RX_QUEUES) {
98                 netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn,
99                                "Reducing number of rx queues from %u to %u.\n",
100                                count, EFX_MAX_RX_QUEUES);
101                 count = EFX_MAX_RX_QUEUES;
102         }
103
104         /* If RSS is requested for the PF *and* VFs then we can't write RSS
105          * table entries that are inaccessible to VFs
106          */
107 #ifdef CONFIG_SFC_SRIOV
108         if (efx->type->sriov_wanted) {
109                 if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 &&
110                     count > efx_vf_size(efx)) {
111                         netif_warn(efx, probe, efx->net_dev,
112                                    "Reducing number of RSS channels from %u to %u for "
113                                    "VF support. Increase vf-msix-limit to use more "
114                                    "channels on the PF.\n",
115                                    count, efx_vf_size(efx));
116                         count = efx_vf_size(efx);
117                 }
118         }
119 #endif
120
121         return count;
122 }
123
124 static int efx_allocate_msix_channels(struct efx_nic *efx,
125                                       unsigned int max_channels,
126                                       unsigned int extra_channels,
127                                       unsigned int parallelism)
128 {
129         unsigned int n_channels = parallelism;
130         int vec_count;
131         int tx_per_ev;
132         int n_xdp_tx;
133         int n_xdp_ev;
134
135         if (efx_separate_tx_channels)
136                 n_channels *= 2;
137         n_channels += extra_channels;
138
139         /* To allow XDP transmit to happen from arbitrary NAPI contexts
140          * we allocate a TX queue per CPU. We share event queues across
141          * multiple tx queues, assuming tx and ev queues are both
142          * maximum size.
143          */
144         tx_per_ev = EFX_MAX_EVQ_SIZE / EFX_TXQ_MAX_ENT(efx);
145         tx_per_ev = min(tx_per_ev, EFX_MAX_TXQ_PER_CHANNEL);
146         n_xdp_tx = num_possible_cpus();
147         n_xdp_ev = DIV_ROUND_UP(n_xdp_tx, tx_per_ev);
148
149         vec_count = pci_msix_vec_count(efx->pci_dev);
150         if (vec_count < 0)
151                 return vec_count;
152
153         max_channels = min_t(unsigned int, vec_count, max_channels);
154
155         /* Check resources.
156          * We need a channel per event queue, plus a VI per tx queue.
157          * This may be more pessimistic than it needs to be.
158          */
159         if (n_channels >= max_channels) {
160                 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED;
161                 netif_warn(efx, drv, efx->net_dev,
162                            "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
163                            n_xdp_ev, n_channels, max_channels);
164                 netif_warn(efx, drv, efx->net_dev,
165                            "XDP_TX and XDP_REDIRECT might decrease device's performance\n");
166         } else if (n_channels + n_xdp_tx > efx->max_vis) {
167                 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED;
168                 netif_warn(efx, drv, efx->net_dev,
169                            "Insufficient resources for %d XDP TX queues (%d other channels, max VIs %d)\n",
170                            n_xdp_tx, n_channels, efx->max_vis);
171                 netif_warn(efx, drv, efx->net_dev,
172                            "XDP_TX and XDP_REDIRECT might decrease device's performance\n");
173         } else if (n_channels + n_xdp_ev > max_channels) {
174                 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_SHARED;
175                 netif_warn(efx, drv, efx->net_dev,
176                            "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n",
177                            n_xdp_ev, n_channels, max_channels);
178
179                 n_xdp_ev = max_channels - n_channels;
180                 netif_warn(efx, drv, efx->net_dev,
181                            "XDP_TX and XDP_REDIRECT will work with reduced performance (%d cpus/tx_queue)\n",
182                            DIV_ROUND_UP(n_xdp_tx, tx_per_ev * n_xdp_ev));
183         } else {
184                 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_DEDICATED;
185         }
186
187         if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_BORROWED) {
188                 efx->n_xdp_channels = n_xdp_ev;
189                 efx->xdp_tx_per_channel = tx_per_ev;
190                 efx->xdp_tx_queue_count = n_xdp_tx;
191                 n_channels += n_xdp_ev;
192                 netif_dbg(efx, drv, efx->net_dev,
193                           "Allocating %d TX and %d event queues for XDP\n",
194                           n_xdp_ev * tx_per_ev, n_xdp_ev);
195         } else {
196                 efx->n_xdp_channels = 0;
197                 efx->xdp_tx_per_channel = 0;
198                 efx->xdp_tx_queue_count = n_xdp_tx;
199         }
200
201         if (vec_count < n_channels) {
202                 netif_err(efx, drv, efx->net_dev,
203                           "WARNING: Insufficient MSI-X vectors available (%d < %u).\n",
204                           vec_count, n_channels);
205                 netif_err(efx, drv, efx->net_dev,
206                           "WARNING: Performance may be reduced.\n");
207                 n_channels = vec_count;
208         }
209
210         n_channels = min(n_channels, max_channels);
211
212         efx->n_channels = n_channels;
213
214         /* Ignore XDP tx channels when creating rx channels. */
215         n_channels -= efx->n_xdp_channels;
216
217         if (efx_separate_tx_channels) {
218                 efx->n_tx_channels =
219                         min(max(n_channels / 2, 1U),
220                             efx->max_tx_channels);
221                 efx->tx_channel_offset =
222                         n_channels - efx->n_tx_channels;
223                 efx->n_rx_channels =
224                         max(n_channels -
225                             efx->n_tx_channels, 1U);
226         } else {
227                 efx->n_tx_channels = min(n_channels, efx->max_tx_channels);
228                 efx->tx_channel_offset = 0;
229                 efx->n_rx_channels = n_channels;
230         }
231
232         efx->n_rx_channels = min(efx->n_rx_channels, parallelism);
233         efx->n_tx_channels = min(efx->n_tx_channels, parallelism);
234
235         efx->xdp_channel_offset = n_channels;
236
237         netif_dbg(efx, drv, efx->net_dev,
238                   "Allocating %u RX channels\n",
239                   efx->n_rx_channels);
240
241         return efx->n_channels;
242 }
243
244 /* Probe the number and type of interrupts we are able to obtain, and
245  * the resulting numbers of channels and RX queues.
246  */
247 int efx_probe_interrupts(struct efx_nic *efx)
248 {
249         unsigned int extra_channels = 0;
250         unsigned int rss_spread;
251         unsigned int i, j;
252         int rc;
253
254         for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++)
255                 if (efx->extra_channel_type[i])
256                         ++extra_channels;
257
258         if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
259                 unsigned int parallelism = efx_wanted_parallelism(efx);
260                 struct msix_entry xentries[EFX_MAX_CHANNELS];
261                 unsigned int n_channels;
262
263                 rc = efx_allocate_msix_channels(efx, efx->max_channels,
264                                                 extra_channels, parallelism);
265                 if (rc >= 0) {
266                         n_channels = rc;
267                         for (i = 0; i < n_channels; i++)
268                                 xentries[i].entry = i;
269                         rc = pci_enable_msix_range(efx->pci_dev, xentries, 1,
270                                                    n_channels);
271                 }
272                 if (rc < 0) {
273                         /* Fall back to single channel MSI */
274                         netif_err(efx, drv, efx->net_dev,
275                                   "could not enable MSI-X\n");
276                         if (efx->type->min_interrupt_mode >= EFX_INT_MODE_MSI)
277                                 efx->interrupt_mode = EFX_INT_MODE_MSI;
278                         else
279                                 return rc;
280                 } else if (rc < n_channels) {
281                         netif_err(efx, drv, efx->net_dev,
282                                   "WARNING: Insufficient MSI-X vectors"
283                                   " available (%d < %u).\n", rc, n_channels);
284                         netif_err(efx, drv, efx->net_dev,
285                                   "WARNING: Performance may be reduced.\n");
286                         n_channels = rc;
287                 }
288
289                 if (rc > 0) {
290                         for (i = 0; i < efx->n_channels; i++)
291                                 efx_get_channel(efx, i)->irq =
292                                         xentries[i].vector;
293                 }
294         }
295
296         /* Try single interrupt MSI */
297         if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
298                 efx->n_channels = 1;
299                 efx->n_rx_channels = 1;
300                 efx->n_tx_channels = 1;
301                 efx->tx_channel_offset = 0;
302                 efx->n_xdp_channels = 0;
303                 efx->xdp_channel_offset = efx->n_channels;
304                 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED;
305                 rc = pci_enable_msi(efx->pci_dev);
306                 if (rc == 0) {
307                         efx_get_channel(efx, 0)->irq = efx->pci_dev->irq;
308                 } else {
309                         netif_err(efx, drv, efx->net_dev,
310                                   "could not enable MSI\n");
311                         if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY)
312                                 efx->interrupt_mode = EFX_INT_MODE_LEGACY;
313                         else
314                                 return rc;
315                 }
316         }
317
318         /* Assume legacy interrupts */
319         if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
320                 efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0);
321                 efx->n_rx_channels = 1;
322                 efx->n_tx_channels = 1;
323                 efx->tx_channel_offset = efx_separate_tx_channels ? 1 : 0;
324                 efx->n_xdp_channels = 0;
325                 efx->xdp_channel_offset = efx->n_channels;
326                 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED;
327                 efx->legacy_irq = efx->pci_dev->irq;
328         }
329
330         /* Assign extra channels if possible, before XDP channels */
331         efx->n_extra_tx_channels = 0;
332         j = efx->xdp_channel_offset;
333         for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) {
334                 if (!efx->extra_channel_type[i])
335                         continue;
336                 if (j <= efx->tx_channel_offset + efx->n_tx_channels) {
337                         efx->extra_channel_type[i]->handle_no_channel(efx);
338                 } else {
339                         --j;
340                         efx_get_channel(efx, j)->type =
341                                 efx->extra_channel_type[i];
342                         if (efx_channel_has_tx_queues(efx_get_channel(efx, j)))
343                                 efx->n_extra_tx_channels++;
344                 }
345         }
346
347         rss_spread = efx->n_rx_channels;
348         /* RSS might be usable on VFs even if it is disabled on the PF */
349 #ifdef CONFIG_SFC_SRIOV
350         if (efx->type->sriov_wanted) {
351                 efx->rss_spread = ((rss_spread > 1 ||
352                                     !efx->type->sriov_wanted(efx)) ?
353                                    rss_spread : efx_vf_size(efx));
354                 return 0;
355         }
356 #endif
357         efx->rss_spread = rss_spread;
358
359         return 0;
360 }
361
362 #if defined(CONFIG_SMP)
363 void efx_set_interrupt_affinity(struct efx_nic *efx)
364 {
365         const struct cpumask *numa_mask = cpumask_of_pcibus(efx->pci_dev->bus);
366         struct efx_channel *channel;
367         unsigned int cpu;
368
369         /* If no online CPUs in local node, fallback to any online CPU */
370         if (cpumask_first_and(cpu_online_mask, numa_mask) >= nr_cpu_ids)
371                 numa_mask = cpu_online_mask;
372
373         cpu = -1;
374         efx_for_each_channel(channel, efx) {
375                 cpu = cpumask_next_and(cpu, cpu_online_mask, numa_mask);
376                 if (cpu >= nr_cpu_ids)
377                         cpu = cpumask_first_and(cpu_online_mask, numa_mask);
378                 irq_set_affinity_hint(channel->irq, cpumask_of(cpu));
379         }
380 }
381
382 void efx_clear_interrupt_affinity(struct efx_nic *efx)
383 {
384         struct efx_channel *channel;
385
386         efx_for_each_channel(channel, efx)
387                 irq_set_affinity_hint(channel->irq, NULL);
388 }
389 #else
390 void
391 efx_set_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
392 {
393 }
394
395 void
396 efx_clear_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused)))
397 {
398 }
399 #endif /* CONFIG_SMP */
400
401 void efx_remove_interrupts(struct efx_nic *efx)
402 {
403         struct efx_channel *channel;
404
405         /* Remove MSI/MSI-X interrupts */
406         efx_for_each_channel(channel, efx)
407                 channel->irq = 0;
408         pci_disable_msi(efx->pci_dev);
409         pci_disable_msix(efx->pci_dev);
410
411         /* Remove legacy interrupt */
412         efx->legacy_irq = 0;
413 }
414
415 /***************
416  * EVENT QUEUES
417  ***************/
418
419 /* Create event queue
420  * Event queue memory allocations are done only once.  If the channel
421  * is reset, the memory buffer will be reused; this guards against
422  * errors during channel reset and also simplifies interrupt handling.
423  */
424 int efx_probe_eventq(struct efx_channel *channel)
425 {
426         struct efx_nic *efx = channel->efx;
427         unsigned long entries;
428
429         netif_dbg(efx, probe, efx->net_dev,
430                   "chan %d create event queue\n", channel->channel);
431
432         /* Build an event queue with room for one event per tx and rx buffer,
433          * plus some extra for link state events and MCDI completions.
434          */
435         entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128);
436         EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE);
437         channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1;
438
439         return efx_nic_probe_eventq(channel);
440 }
441
442 /* Prepare channel's event queue */
443 int efx_init_eventq(struct efx_channel *channel)
444 {
445         struct efx_nic *efx = channel->efx;
446         int rc;
447
448         EFX_WARN_ON_PARANOID(channel->eventq_init);
449
450         netif_dbg(efx, drv, efx->net_dev,
451                   "chan %d init event queue\n", channel->channel);
452
453         rc = efx_nic_init_eventq(channel);
454         if (rc == 0) {
455                 efx->type->push_irq_moderation(channel);
456                 channel->eventq_read_ptr = 0;
457                 channel->eventq_init = true;
458         }
459         return rc;
460 }
461
462 /* Enable event queue processing and NAPI */
463 void efx_start_eventq(struct efx_channel *channel)
464 {
465         netif_dbg(channel->efx, ifup, channel->efx->net_dev,
466                   "chan %d start event queue\n", channel->channel);
467
468         /* Make sure the NAPI handler sees the enabled flag set */
469         channel->enabled = true;
470         smp_wmb();
471
472         napi_enable(&channel->napi_str);
473         efx_nic_eventq_read_ack(channel);
474 }
475
476 /* Disable event queue processing and NAPI */
477 void efx_stop_eventq(struct efx_channel *channel)
478 {
479         if (!channel->enabled)
480                 return;
481
482         napi_disable(&channel->napi_str);
483         channel->enabled = false;
484 }
485
486 void efx_fini_eventq(struct efx_channel *channel)
487 {
488         if (!channel->eventq_init)
489                 return;
490
491         netif_dbg(channel->efx, drv, channel->efx->net_dev,
492                   "chan %d fini event queue\n", channel->channel);
493
494         efx_nic_fini_eventq(channel);
495         channel->eventq_init = false;
496 }
497
498 void efx_remove_eventq(struct efx_channel *channel)
499 {
500         netif_dbg(channel->efx, drv, channel->efx->net_dev,
501                   "chan %d remove event queue\n", channel->channel);
502
503         efx_nic_remove_eventq(channel);
504 }
505
506 /**************************************************************************
507  *
508  * Channel handling
509  *
510  *************************************************************************/
511
512 #ifdef CONFIG_RFS_ACCEL
513 static void efx_filter_rfs_expire(struct work_struct *data)
514 {
515         struct delayed_work *dwork = to_delayed_work(data);
516         struct efx_channel *channel;
517         unsigned int time, quota;
518
519         channel = container_of(dwork, struct efx_channel, filter_work);
520         time = jiffies - channel->rfs_last_expiry;
521         quota = channel->rfs_filter_count * time / (30 * HZ);
522         if (quota >= 20 && __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, quota)))
523                 channel->rfs_last_expiry += time;
524         /* Ensure we do more work eventually even if NAPI poll is not happening */
525         schedule_delayed_work(dwork, 30 * HZ);
526 }
527 #endif
528
529 /* Allocate and initialise a channel structure. */
530 static struct efx_channel *efx_alloc_channel(struct efx_nic *efx, int i)
531 {
532         struct efx_rx_queue *rx_queue;
533         struct efx_tx_queue *tx_queue;
534         struct efx_channel *channel;
535         int j;
536
537         channel = kzalloc(sizeof(*channel), GFP_KERNEL);
538         if (!channel)
539                 return NULL;
540
541         channel->efx = efx;
542         channel->channel = i;
543         channel->type = &efx_default_channel_type;
544
545         for (j = 0; j < EFX_MAX_TXQ_PER_CHANNEL; j++) {
546                 tx_queue = &channel->tx_queue[j];
547                 tx_queue->efx = efx;
548                 tx_queue->queue = -1;
549                 tx_queue->label = j;
550                 tx_queue->channel = channel;
551         }
552
553 #ifdef CONFIG_RFS_ACCEL
554         INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire);
555 #endif
556
557         rx_queue = &channel->rx_queue;
558         rx_queue->efx = efx;
559         timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
560
561         return channel;
562 }
563
564 int efx_init_channels(struct efx_nic *efx)
565 {
566         unsigned int i;
567
568         for (i = 0; i < EFX_MAX_CHANNELS; i++) {
569                 efx->channel[i] = efx_alloc_channel(efx, i);
570                 if (!efx->channel[i])
571                         return -ENOMEM;
572                 efx->msi_context[i].efx = efx;
573                 efx->msi_context[i].index = i;
574         }
575
576         /* Higher numbered interrupt modes are less capable! */
577         efx->interrupt_mode = min(efx->type->min_interrupt_mode,
578                                   efx_interrupt_mode);
579
580         efx->max_channels = EFX_MAX_CHANNELS;
581         efx->max_tx_channels = EFX_MAX_CHANNELS;
582
583         return 0;
584 }
585
586 void efx_fini_channels(struct efx_nic *efx)
587 {
588         unsigned int i;
589
590         for (i = 0; i < EFX_MAX_CHANNELS; i++)
591                 if (efx->channel[i]) {
592                         kfree(efx->channel[i]);
593                         efx->channel[i] = NULL;
594                 }
595 }
596
597 /* Allocate and initialise a channel structure, copying parameters
598  * (but not resources) from an old channel structure.
599  */
600 struct efx_channel *efx_copy_channel(const struct efx_channel *old_channel)
601 {
602         struct efx_rx_queue *rx_queue;
603         struct efx_tx_queue *tx_queue;
604         struct efx_channel *channel;
605         int j;
606
607         channel = kmalloc(sizeof(*channel), GFP_KERNEL);
608         if (!channel)
609                 return NULL;
610
611         *channel = *old_channel;
612
613         channel->napi_dev = NULL;
614         INIT_HLIST_NODE(&channel->napi_str.napi_hash_node);
615         channel->napi_str.napi_id = 0;
616         channel->napi_str.state = 0;
617         memset(&channel->eventq, 0, sizeof(channel->eventq));
618
619         for (j = 0; j < EFX_MAX_TXQ_PER_CHANNEL; j++) {
620                 tx_queue = &channel->tx_queue[j];
621                 if (tx_queue->channel)
622                         tx_queue->channel = channel;
623                 tx_queue->buffer = NULL;
624                 tx_queue->cb_page = NULL;
625                 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd));
626         }
627
628         rx_queue = &channel->rx_queue;
629         rx_queue->buffer = NULL;
630         memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
631         timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
632 #ifdef CONFIG_RFS_ACCEL
633         INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire);
634 #endif
635
636         return channel;
637 }
638
639 static int efx_probe_channel(struct efx_channel *channel)
640 {
641         struct efx_tx_queue *tx_queue;
642         struct efx_rx_queue *rx_queue;
643         int rc;
644
645         netif_dbg(channel->efx, probe, channel->efx->net_dev,
646                   "creating channel %d\n", channel->channel);
647
648         rc = channel->type->pre_probe(channel);
649         if (rc)
650                 goto fail;
651
652         rc = efx_probe_eventq(channel);
653         if (rc)
654                 goto fail;
655
656         efx_for_each_channel_tx_queue(tx_queue, channel) {
657                 rc = efx_probe_tx_queue(tx_queue);
658                 if (rc)
659                         goto fail;
660         }
661
662         efx_for_each_channel_rx_queue(rx_queue, channel) {
663                 rc = efx_probe_rx_queue(rx_queue);
664                 if (rc)
665                         goto fail;
666         }
667
668         channel->rx_list = NULL;
669
670         return 0;
671
672 fail:
673         efx_remove_channel(channel);
674         return rc;
675 }
676
677 static void efx_get_channel_name(struct efx_channel *channel, char *buf,
678                                  size_t len)
679 {
680         struct efx_nic *efx = channel->efx;
681         const char *type;
682         int number;
683
684         number = channel->channel;
685
686         if (number >= efx->xdp_channel_offset &&
687             !WARN_ON_ONCE(!efx->n_xdp_channels)) {
688                 type = "-xdp";
689                 number -= efx->xdp_channel_offset;
690         } else if (efx->tx_channel_offset == 0) {
691                 type = "";
692         } else if (number < efx->tx_channel_offset) {
693                 type = "-rx";
694         } else {
695                 type = "-tx";
696                 number -= efx->tx_channel_offset;
697         }
698         snprintf(buf, len, "%s%s-%d", efx->name, type, number);
699 }
700
701 void efx_set_channel_names(struct efx_nic *efx)
702 {
703         struct efx_channel *channel;
704
705         efx_for_each_channel(channel, efx)
706                 channel->type->get_name(channel,
707                                         efx->msi_context[channel->channel].name,
708                                         sizeof(efx->msi_context[0].name));
709 }
710
711 int efx_probe_channels(struct efx_nic *efx)
712 {
713         struct efx_channel *channel;
714         int rc;
715
716         /* Probe channels in reverse, so that any 'extra' channels
717          * use the start of the buffer table. This allows the traffic
718          * channels to be resized without moving them or wasting the
719          * entries before them.
720          */
721         efx_for_each_channel_rev(channel, efx) {
722                 rc = efx_probe_channel(channel);
723                 if (rc) {
724                         netif_err(efx, probe, efx->net_dev,
725                                   "failed to create channel %d\n",
726                                   channel->channel);
727                         goto fail;
728                 }
729         }
730         efx_set_channel_names(efx);
731
732         return 0;
733
734 fail:
735         efx_remove_channels(efx);
736         return rc;
737 }
738
739 void efx_remove_channel(struct efx_channel *channel)
740 {
741         struct efx_tx_queue *tx_queue;
742         struct efx_rx_queue *rx_queue;
743
744         netif_dbg(channel->efx, drv, channel->efx->net_dev,
745                   "destroy chan %d\n", channel->channel);
746
747         efx_for_each_channel_rx_queue(rx_queue, channel)
748                 efx_remove_rx_queue(rx_queue);
749         efx_for_each_channel_tx_queue(tx_queue, channel)
750                 efx_remove_tx_queue(tx_queue);
751         efx_remove_eventq(channel);
752         channel->type->post_remove(channel);
753 }
754
755 void efx_remove_channels(struct efx_nic *efx)
756 {
757         struct efx_channel *channel;
758
759         efx_for_each_channel(channel, efx)
760                 efx_remove_channel(channel);
761
762         kfree(efx->xdp_tx_queues);
763 }
764
765 static int efx_set_xdp_tx_queue(struct efx_nic *efx, int xdp_queue_number,
766                                 struct efx_tx_queue *tx_queue)
767 {
768         if (xdp_queue_number >= efx->xdp_tx_queue_count)
769                 return -EINVAL;
770
771         netif_dbg(efx, drv, efx->net_dev,
772                   "Channel %u TXQ %u is XDP %u, HW %u\n",
773                   tx_queue->channel->channel, tx_queue->label,
774                   xdp_queue_number, tx_queue->queue);
775         efx->xdp_tx_queues[xdp_queue_number] = tx_queue;
776         return 0;
777 }
778
779 static void efx_set_xdp_channels(struct efx_nic *efx)
780 {
781         struct efx_tx_queue *tx_queue;
782         struct efx_channel *channel;
783         unsigned int next_queue = 0;
784         int xdp_queue_number = 0;
785         int rc;
786
787         /* We need to mark which channels really have RX and TX
788          * queues, and adjust the TX queue numbers if we have separate
789          * RX-only and TX-only channels.
790          */
791         efx_for_each_channel(channel, efx) {
792                 if (channel->channel < efx->tx_channel_offset)
793                         continue;
794
795                 if (efx_channel_is_xdp_tx(channel)) {
796                         efx_for_each_channel_tx_queue(tx_queue, channel) {
797                                 tx_queue->queue = next_queue++;
798                                 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number,
799                                                           tx_queue);
800                                 if (rc == 0)
801                                         xdp_queue_number++;
802                         }
803                 } else {
804                         efx_for_each_channel_tx_queue(tx_queue, channel) {
805                                 tx_queue->queue = next_queue++;
806                                 netif_dbg(efx, drv, efx->net_dev,
807                                           "Channel %u TXQ %u is HW %u\n",
808                                           channel->channel, tx_queue->label,
809                                           tx_queue->queue);
810                         }
811
812                         /* If XDP is borrowing queues from net stack, it must
813                          * use the queue with no csum offload, which is the
814                          * first one of the channel
815                          * (note: tx_queue_by_type is not initialized yet)
816                          */
817                         if (efx->xdp_txq_queues_mode ==
818                             EFX_XDP_TX_QUEUES_BORROWED) {
819                                 tx_queue = &channel->tx_queue[0];
820                                 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number,
821                                                           tx_queue);
822                                 if (rc == 0)
823                                         xdp_queue_number++;
824                         }
825                 }
826         }
827         WARN_ON(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DEDICATED &&
828                 xdp_queue_number != efx->xdp_tx_queue_count);
829         WARN_ON(efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED &&
830                 xdp_queue_number > efx->xdp_tx_queue_count);
831
832         /* If we have more CPUs than assigned XDP TX queues, assign the already
833          * existing queues to the exceeding CPUs
834          */
835         next_queue = 0;
836         while (xdp_queue_number < efx->xdp_tx_queue_count) {
837                 tx_queue = efx->xdp_tx_queues[next_queue++];
838                 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue);
839                 if (rc == 0)
840                         xdp_queue_number++;
841         }
842 }
843
844 int efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries)
845 {
846         struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel,
847                            *ptp_channel = efx_ptp_channel(efx);
848         struct efx_ptp_data *ptp_data = efx->ptp_data;
849         u32 old_rxq_entries, old_txq_entries;
850         unsigned int i;
851         int rc, rc2;
852
853         rc = efx_check_disabled(efx);
854         if (rc)
855                 return rc;
856
857         efx_device_detach_sync(efx);
858         efx_stop_all(efx);
859         efx_soft_disable_interrupts(efx);
860
861         /* Clone channels (where possible) */
862         memset(other_channel, 0, sizeof(other_channel));
863         for (i = 0; i < efx->n_channels; i++) {
864                 channel = efx->channel[i];
865                 if (channel->type->copy)
866                         channel = channel->type->copy(channel);
867                 if (!channel) {
868                         rc = -ENOMEM;
869                         goto out;
870                 }
871                 other_channel[i] = channel;
872         }
873
874         /* Swap entry counts and channel pointers */
875         old_rxq_entries = efx->rxq_entries;
876         old_txq_entries = efx->txq_entries;
877         efx->rxq_entries = rxq_entries;
878         efx->txq_entries = txq_entries;
879         for (i = 0; i < efx->n_channels; i++)
880                 swap(efx->channel[i], other_channel[i]);
881
882         for (i = 0; i < efx->n_channels; i++) {
883                 channel = efx->channel[i];
884                 if (!channel->type->copy)
885                         continue;
886                 rc = efx_probe_channel(channel);
887                 if (rc)
888                         goto rollback;
889                 efx_init_napi_channel(efx->channel[i]);
890         }
891
892         efx_set_xdp_channels(efx);
893 out:
894         efx->ptp_data = NULL;
895         /* Destroy unused channel structures */
896         for (i = 0; i < efx->n_channels; i++) {
897                 channel = other_channel[i];
898                 if (channel && channel->type->copy) {
899                         efx_fini_napi_channel(channel);
900                         efx_remove_channel(channel);
901                         kfree(channel);
902                 }
903         }
904
905         efx->ptp_data = ptp_data;
906         rc2 = efx_soft_enable_interrupts(efx);
907         if (rc2) {
908                 rc = rc ? rc : rc2;
909                 netif_err(efx, drv, efx->net_dev,
910                           "unable to restart interrupts on channel reallocation\n");
911                 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
912         } else {
913                 efx_start_all(efx);
914                 efx_device_attach_if_not_resetting(efx);
915         }
916         return rc;
917
918 rollback:
919         /* Swap back */
920         efx->rxq_entries = old_rxq_entries;
921         efx->txq_entries = old_txq_entries;
922         for (i = 0; i < efx->n_channels; i++)
923                 swap(efx->channel[i], other_channel[i]);
924         efx_ptp_update_channel(efx, ptp_channel);
925         goto out;
926 }
927
928 int efx_set_channels(struct efx_nic *efx)
929 {
930         struct efx_channel *channel;
931         int rc;
932
933         if (efx->xdp_tx_queue_count) {
934                 EFX_WARN_ON_PARANOID(efx->xdp_tx_queues);
935
936                 /* Allocate array for XDP TX queue lookup. */
937                 efx->xdp_tx_queues = kcalloc(efx->xdp_tx_queue_count,
938                                              sizeof(*efx->xdp_tx_queues),
939                                              GFP_KERNEL);
940                 if (!efx->xdp_tx_queues)
941                         return -ENOMEM;
942         }
943
944         efx_for_each_channel(channel, efx) {
945                 if (channel->channel < efx->n_rx_channels)
946                         channel->rx_queue.core_index = channel->channel;
947                 else
948                         channel->rx_queue.core_index = -1;
949         }
950
951         efx_set_xdp_channels(efx);
952
953         rc = netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels);
954         if (rc)
955                 return rc;
956         return netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels);
957 }
958
959 static bool efx_default_channel_want_txqs(struct efx_channel *channel)
960 {
961         return channel->channel - channel->efx->tx_channel_offset <
962                 channel->efx->n_tx_channels;
963 }
964
965 /*************
966  * START/STOP
967  *************/
968
969 int efx_soft_enable_interrupts(struct efx_nic *efx)
970 {
971         struct efx_channel *channel, *end_channel;
972         int rc;
973
974         BUG_ON(efx->state == STATE_DISABLED);
975
976         efx->irq_soft_enabled = true;
977         smp_wmb();
978
979         efx_for_each_channel(channel, efx) {
980                 if (!channel->type->keep_eventq) {
981                         rc = efx_init_eventq(channel);
982                         if (rc)
983                                 goto fail;
984                 }
985                 efx_start_eventq(channel);
986         }
987
988         efx_mcdi_mode_event(efx);
989
990         return 0;
991 fail:
992         end_channel = channel;
993         efx_for_each_channel(channel, efx) {
994                 if (channel == end_channel)
995                         break;
996                 efx_stop_eventq(channel);
997                 if (!channel->type->keep_eventq)
998                         efx_fini_eventq(channel);
999         }
1000
1001         return rc;
1002 }
1003
1004 void efx_soft_disable_interrupts(struct efx_nic *efx)
1005 {
1006         struct efx_channel *channel;
1007
1008         if (efx->state == STATE_DISABLED)
1009                 return;
1010
1011         efx_mcdi_mode_poll(efx);
1012
1013         efx->irq_soft_enabled = false;
1014         smp_wmb();
1015
1016         if (efx->legacy_irq)
1017                 synchronize_irq(efx->legacy_irq);
1018
1019         efx_for_each_channel(channel, efx) {
1020                 if (channel->irq)
1021                         synchronize_irq(channel->irq);
1022
1023                 efx_stop_eventq(channel);
1024                 if (!channel->type->keep_eventq)
1025                         efx_fini_eventq(channel);
1026         }
1027
1028         /* Flush the asynchronous MCDI request queue */
1029         efx_mcdi_flush_async(efx);
1030 }
1031
1032 int efx_enable_interrupts(struct efx_nic *efx)
1033 {
1034         struct efx_channel *channel, *end_channel;
1035         int rc;
1036
1037         /* TODO: Is this really a bug? */
1038         BUG_ON(efx->state == STATE_DISABLED);
1039
1040         if (efx->eeh_disabled_legacy_irq) {
1041                 enable_irq(efx->legacy_irq);
1042                 efx->eeh_disabled_legacy_irq = false;
1043         }
1044
1045         efx->type->irq_enable_master(efx);
1046
1047         efx_for_each_channel(channel, efx) {
1048                 if (channel->type->keep_eventq) {
1049                         rc = efx_init_eventq(channel);
1050                         if (rc)
1051                                 goto fail;
1052                 }
1053         }
1054
1055         rc = efx_soft_enable_interrupts(efx);
1056         if (rc)
1057                 goto fail;
1058
1059         return 0;
1060
1061 fail:
1062         end_channel = channel;
1063         efx_for_each_channel(channel, efx) {
1064                 if (channel == end_channel)
1065                         break;
1066                 if (channel->type->keep_eventq)
1067                         efx_fini_eventq(channel);
1068         }
1069
1070         efx->type->irq_disable_non_ev(efx);
1071
1072         return rc;
1073 }
1074
1075 void efx_disable_interrupts(struct efx_nic *efx)
1076 {
1077         struct efx_channel *channel;
1078
1079         efx_soft_disable_interrupts(efx);
1080
1081         efx_for_each_channel(channel, efx) {
1082                 if (channel->type->keep_eventq)
1083                         efx_fini_eventq(channel);
1084         }
1085
1086         efx->type->irq_disable_non_ev(efx);
1087 }
1088
1089 void efx_start_channels(struct efx_nic *efx)
1090 {
1091         struct efx_tx_queue *tx_queue;
1092         struct efx_rx_queue *rx_queue;
1093         struct efx_channel *channel;
1094
1095         efx_for_each_channel_rev(channel, efx) {
1096                 if (channel->type->start)
1097                         channel->type->start(channel);
1098                 efx_for_each_channel_tx_queue(tx_queue, channel) {
1099                         efx_init_tx_queue(tx_queue);
1100                         atomic_inc(&efx->active_queues);
1101                 }
1102
1103                 efx_for_each_channel_rx_queue(rx_queue, channel) {
1104                         efx_init_rx_queue(rx_queue);
1105                         atomic_inc(&efx->active_queues);
1106                         efx_stop_eventq(channel);
1107                         efx_fast_push_rx_descriptors(rx_queue, false);
1108                         efx_start_eventq(channel);
1109                 }
1110
1111                 WARN_ON(channel->rx_pkt_n_frags);
1112         }
1113 }
1114
1115 void efx_stop_channels(struct efx_nic *efx)
1116 {
1117         struct efx_tx_queue *tx_queue;
1118         struct efx_rx_queue *rx_queue;
1119         struct efx_channel *channel;
1120         int rc = 0;
1121
1122         /* Stop special channels and RX refill.
1123          * The channel's stop has to be called first, since it might wait
1124          * for a sentinel RX to indicate the channel has fully drained.
1125          */
1126         efx_for_each_channel(channel, efx) {
1127                 if (channel->type->stop)
1128                         channel->type->stop(channel);
1129                 efx_for_each_channel_rx_queue(rx_queue, channel)
1130                         rx_queue->refill_enabled = false;
1131         }
1132
1133         efx_for_each_channel(channel, efx) {
1134                 /* RX packet processing is pipelined, so wait for the
1135                  * NAPI handler to complete.  At least event queue 0
1136                  * might be kept active by non-data events, so don't
1137                  * use napi_synchronize() but actually disable NAPI
1138                  * temporarily.
1139                  */
1140                 if (efx_channel_has_rx_queue(channel)) {
1141                         efx_stop_eventq(channel);
1142                         efx_start_eventq(channel);
1143                 }
1144         }
1145
1146         if (efx->type->fini_dmaq)
1147                 rc = efx->type->fini_dmaq(efx);
1148
1149         if (rc) {
1150                 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n");
1151         } else {
1152                 netif_dbg(efx, drv, efx->net_dev,
1153                           "successfully flushed all queues\n");
1154         }
1155
1156         efx_for_each_channel(channel, efx) {
1157                 efx_for_each_channel_rx_queue(rx_queue, channel)
1158                         efx_fini_rx_queue(rx_queue);
1159                 efx_for_each_channel_tx_queue(tx_queue, channel)
1160                         efx_fini_tx_queue(tx_queue);
1161         }
1162 }
1163
1164 /**************************************************************************
1165  *
1166  * NAPI interface
1167  *
1168  *************************************************************************/
1169
1170 /* Process channel's event queue
1171  *
1172  * This function is responsible for processing the event queue of a
1173  * single channel.  The caller must guarantee that this function will
1174  * never be concurrently called more than once on the same channel,
1175  * though different channels may be being processed concurrently.
1176  */
1177 static int efx_process_channel(struct efx_channel *channel, int budget)
1178 {
1179         struct efx_tx_queue *tx_queue;
1180         struct list_head rx_list;
1181         int spent;
1182
1183         if (unlikely(!channel->enabled))
1184                 return 0;
1185
1186         /* Prepare the batch receive list */
1187         EFX_WARN_ON_PARANOID(channel->rx_list != NULL);
1188         INIT_LIST_HEAD(&rx_list);
1189         channel->rx_list = &rx_list;
1190
1191         efx_for_each_channel_tx_queue(tx_queue, channel) {
1192                 tx_queue->pkts_compl = 0;
1193                 tx_queue->bytes_compl = 0;
1194         }
1195
1196         spent = efx_nic_process_eventq(channel, budget);
1197         if (spent && efx_channel_has_rx_queue(channel)) {
1198                 struct efx_rx_queue *rx_queue =
1199                         efx_channel_get_rx_queue(channel);
1200
1201                 efx_rx_flush_packet(channel);
1202                 efx_fast_push_rx_descriptors(rx_queue, true);
1203         }
1204
1205         /* Update BQL */
1206         efx_for_each_channel_tx_queue(tx_queue, channel) {
1207                 if (tx_queue->bytes_compl) {
1208                         netdev_tx_completed_queue(tx_queue->core_txq,
1209                                                   tx_queue->pkts_compl,
1210                                                   tx_queue->bytes_compl);
1211                 }
1212         }
1213
1214         /* Receive any packets we queued up */
1215         netif_receive_skb_list(channel->rx_list);
1216         channel->rx_list = NULL;
1217
1218         return spent;
1219 }
1220
1221 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel)
1222 {
1223         int step = efx->irq_mod_step_us;
1224
1225         if (channel->irq_mod_score < irq_adapt_low_thresh) {
1226                 if (channel->irq_moderation_us > step) {
1227                         channel->irq_moderation_us -= step;
1228                         efx->type->push_irq_moderation(channel);
1229                 }
1230         } else if (channel->irq_mod_score > irq_adapt_high_thresh) {
1231                 if (channel->irq_moderation_us <
1232                     efx->irq_rx_moderation_us) {
1233                         channel->irq_moderation_us += step;
1234                         efx->type->push_irq_moderation(channel);
1235                 }
1236         }
1237
1238         channel->irq_count = 0;
1239         channel->irq_mod_score = 0;
1240 }
1241
1242 /* NAPI poll handler
1243  *
1244  * NAPI guarantees serialisation of polls of the same device, which
1245  * provides the guarantee required by efx_process_channel().
1246  */
1247 static int efx_poll(struct napi_struct *napi, int budget)
1248 {
1249         struct efx_channel *channel =
1250                 container_of(napi, struct efx_channel, napi_str);
1251         struct efx_nic *efx = channel->efx;
1252 #ifdef CONFIG_RFS_ACCEL
1253         unsigned int time;
1254 #endif
1255         int spent;
1256
1257         netif_vdbg(efx, intr, efx->net_dev,
1258                    "channel %d NAPI poll executing on CPU %d\n",
1259                    channel->channel, raw_smp_processor_id());
1260
1261         spent = efx_process_channel(channel, budget);
1262
1263         xdp_do_flush();
1264
1265         if (spent < budget) {
1266                 if (efx_channel_has_rx_queue(channel) &&
1267                     efx->irq_rx_adaptive &&
1268                     unlikely(++channel->irq_count == 1000)) {
1269                         efx_update_irq_mod(efx, channel);
1270                 }
1271
1272 #ifdef CONFIG_RFS_ACCEL
1273                 /* Perhaps expire some ARFS filters */
1274                 time = jiffies - channel->rfs_last_expiry;
1275                 /* Would our quota be >= 20? */
1276                 if (channel->rfs_filter_count * time >= 600 * HZ)
1277                         mod_delayed_work(system_wq, &channel->filter_work, 0);
1278 #endif
1279
1280                 /* There is no race here; although napi_disable() will
1281                  * only wait for napi_complete(), this isn't a problem
1282                  * since efx_nic_eventq_read_ack() will have no effect if
1283                  * interrupts have already been disabled.
1284                  */
1285                 if (napi_complete_done(napi, spent))
1286                         efx_nic_eventq_read_ack(channel);
1287         }
1288
1289         return spent;
1290 }
1291
1292 void efx_init_napi_channel(struct efx_channel *channel)
1293 {
1294         struct efx_nic *efx = channel->efx;
1295
1296         channel->napi_dev = efx->net_dev;
1297         netif_napi_add(channel->napi_dev, &channel->napi_str, efx_poll);
1298 }
1299
1300 void efx_init_napi(struct efx_nic *efx)
1301 {
1302         struct efx_channel *channel;
1303
1304         efx_for_each_channel(channel, efx)
1305                 efx_init_napi_channel(channel);
1306 }
1307
1308 void efx_fini_napi_channel(struct efx_channel *channel)
1309 {
1310         if (channel->napi_dev)
1311                 netif_napi_del(&channel->napi_str);
1312
1313         channel->napi_dev = NULL;
1314 }
1315
1316 void efx_fini_napi(struct efx_nic *efx)
1317 {
1318         struct efx_channel *channel;
1319
1320         efx_for_each_channel(channel, efx)
1321                 efx_fini_napi_channel(channel);
1322 }
1323
1324 /***************
1325  * Housekeeping
1326  ***************/
1327
1328 static int efx_channel_dummy_op_int(struct efx_channel *channel)
1329 {
1330         return 0;
1331 }
1332
1333 void efx_channel_dummy_op_void(struct efx_channel *channel)
1334 {
1335 }
1336
1337 static const struct efx_channel_type efx_default_channel_type = {
1338         .pre_probe              = efx_channel_dummy_op_int,
1339         .post_remove            = efx_channel_dummy_op_void,
1340         .get_name               = efx_get_channel_name,
1341         .copy                   = efx_copy_channel,
1342         .want_txqs              = efx_default_channel_want_txqs,
1343         .keep_eventq            = false,
1344         .want_pio               = true,
1345 };