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 rc = pci_enable_msi(efx->pci_dev); 305 if (rc == 0) { 306 efx_get_channel(efx, 0)->irq = efx->pci_dev->irq; 307 } else { 308 netif_err(efx, drv, efx->net_dev, 309 "could not enable MSI\n"); 310 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY) 311 efx->interrupt_mode = EFX_INT_MODE_LEGACY; 312 else 313 return rc; 314 } 315 } 316 317 /* Assume legacy interrupts */ 318 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) { 319 efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0); 320 efx->n_rx_channels = 1; 321 efx->n_tx_channels = 1; 322 efx->tx_channel_offset = 1; 323 efx->n_xdp_channels = 0; 324 efx->xdp_channel_offset = efx->n_channels; 325 efx->legacy_irq = efx->pci_dev->irq; 326 } 327 328 /* Assign extra channels if possible, before XDP channels */ 329 efx->n_extra_tx_channels = 0; 330 j = efx->xdp_channel_offset; 331 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) { 332 if (!efx->extra_channel_type[i]) 333 continue; 334 if (j <= efx->tx_channel_offset + efx->n_tx_channels) { 335 efx->extra_channel_type[i]->handle_no_channel(efx); 336 } else { 337 --j; 338 efx_get_channel(efx, j)->type = 339 efx->extra_channel_type[i]; 340 if (efx_channel_has_tx_queues(efx_get_channel(efx, j))) 341 efx->n_extra_tx_channels++; 342 } 343 } 344 345 rss_spread = efx->n_rx_channels; 346 /* RSS might be usable on VFs even if it is disabled on the PF */ 347 #ifdef CONFIG_SFC_SRIOV 348 if (efx->type->sriov_wanted) { 349 efx->rss_spread = ((rss_spread > 1 || 350 !efx->type->sriov_wanted(efx)) ? 351 rss_spread : efx_vf_size(efx)); 352 return 0; 353 } 354 #endif 355 efx->rss_spread = rss_spread; 356 357 return 0; 358 } 359 360 #if defined(CONFIG_SMP) 361 void efx_set_interrupt_affinity(struct efx_nic *efx) 362 { 363 const struct cpumask *numa_mask = cpumask_of_pcibus(efx->pci_dev->bus); 364 struct efx_channel *channel; 365 unsigned int cpu; 366 367 /* If no online CPUs in local node, fallback to any online CPU */ 368 if (cpumask_first_and(cpu_online_mask, numa_mask) >= nr_cpu_ids) 369 numa_mask = cpu_online_mask; 370 371 cpu = -1; 372 efx_for_each_channel(channel, efx) { 373 cpu = cpumask_next_and(cpu, cpu_online_mask, numa_mask); 374 if (cpu >= nr_cpu_ids) 375 cpu = cpumask_first_and(cpu_online_mask, numa_mask); 376 irq_set_affinity_hint(channel->irq, cpumask_of(cpu)); 377 } 378 } 379 380 void efx_clear_interrupt_affinity(struct efx_nic *efx) 381 { 382 struct efx_channel *channel; 383 384 efx_for_each_channel(channel, efx) 385 irq_set_affinity_hint(channel->irq, NULL); 386 } 387 #else 388 void 389 efx_set_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused))) 390 { 391 } 392 393 void 394 efx_clear_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused))) 395 { 396 } 397 #endif /* CONFIG_SMP */ 398 399 void efx_remove_interrupts(struct efx_nic *efx) 400 { 401 struct efx_channel *channel; 402 403 /* Remove MSI/MSI-X interrupts */ 404 efx_for_each_channel(channel, efx) 405 channel->irq = 0; 406 pci_disable_msi(efx->pci_dev); 407 pci_disable_msix(efx->pci_dev); 408 409 /* Remove legacy interrupt */ 410 efx->legacy_irq = 0; 411 } 412 413 /*************** 414 * EVENT QUEUES 415 ***************/ 416 417 /* Create event queue 418 * Event queue memory allocations are done only once. If the channel 419 * is reset, the memory buffer will be reused; this guards against 420 * errors during channel reset and also simplifies interrupt handling. 421 */ 422 int efx_probe_eventq(struct efx_channel *channel) 423 { 424 struct efx_nic *efx = channel->efx; 425 unsigned long entries; 426 427 netif_dbg(efx, probe, efx->net_dev, 428 "chan %d create event queue\n", channel->channel); 429 430 /* Build an event queue with room for one event per tx and rx buffer, 431 * plus some extra for link state events and MCDI completions. 432 */ 433 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128); 434 EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE); 435 channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1; 436 437 return efx_nic_probe_eventq(channel); 438 } 439 440 /* Prepare channel's event queue */ 441 int efx_init_eventq(struct efx_channel *channel) 442 { 443 struct efx_nic *efx = channel->efx; 444 int rc; 445 446 EFX_WARN_ON_PARANOID(channel->eventq_init); 447 448 netif_dbg(efx, drv, efx->net_dev, 449 "chan %d init event queue\n", channel->channel); 450 451 rc = efx_nic_init_eventq(channel); 452 if (rc == 0) { 453 efx->type->push_irq_moderation(channel); 454 channel->eventq_read_ptr = 0; 455 channel->eventq_init = true; 456 } 457 return rc; 458 } 459 460 /* Enable event queue processing and NAPI */ 461 void efx_start_eventq(struct efx_channel *channel) 462 { 463 netif_dbg(channel->efx, ifup, channel->efx->net_dev, 464 "chan %d start event queue\n", channel->channel); 465 466 /* Make sure the NAPI handler sees the enabled flag set */ 467 channel->enabled = true; 468 smp_wmb(); 469 470 napi_enable(&channel->napi_str); 471 efx_nic_eventq_read_ack(channel); 472 } 473 474 /* Disable event queue processing and NAPI */ 475 void efx_stop_eventq(struct efx_channel *channel) 476 { 477 if (!channel->enabled) 478 return; 479 480 napi_disable(&channel->napi_str); 481 channel->enabled = false; 482 } 483 484 void efx_fini_eventq(struct efx_channel *channel) 485 { 486 if (!channel->eventq_init) 487 return; 488 489 netif_dbg(channel->efx, drv, channel->efx->net_dev, 490 "chan %d fini event queue\n", channel->channel); 491 492 efx_nic_fini_eventq(channel); 493 channel->eventq_init = false; 494 } 495 496 void efx_remove_eventq(struct efx_channel *channel) 497 { 498 netif_dbg(channel->efx, drv, channel->efx->net_dev, 499 "chan %d remove event queue\n", channel->channel); 500 501 efx_nic_remove_eventq(channel); 502 } 503 504 /************************************************************************** 505 * 506 * Channel handling 507 * 508 *************************************************************************/ 509 510 #ifdef CONFIG_RFS_ACCEL 511 static void efx_filter_rfs_expire(struct work_struct *data) 512 { 513 struct delayed_work *dwork = to_delayed_work(data); 514 struct efx_channel *channel; 515 unsigned int time, quota; 516 517 channel = container_of(dwork, struct efx_channel, filter_work); 518 time = jiffies - channel->rfs_last_expiry; 519 quota = channel->rfs_filter_count * time / (30 * HZ); 520 if (quota >= 20 && __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, quota))) 521 channel->rfs_last_expiry += time; 522 /* Ensure we do more work eventually even if NAPI poll is not happening */ 523 schedule_delayed_work(dwork, 30 * HZ); 524 } 525 #endif 526 527 /* Allocate and initialise a channel structure. */ 528 static struct efx_channel *efx_alloc_channel(struct efx_nic *efx, int i) 529 { 530 struct efx_rx_queue *rx_queue; 531 struct efx_tx_queue *tx_queue; 532 struct efx_channel *channel; 533 int j; 534 535 channel = kzalloc(sizeof(*channel), GFP_KERNEL); 536 if (!channel) 537 return NULL; 538 539 channel->efx = efx; 540 channel->channel = i; 541 channel->type = &efx_default_channel_type; 542 543 for (j = 0; j < EFX_MAX_TXQ_PER_CHANNEL; j++) { 544 tx_queue = &channel->tx_queue[j]; 545 tx_queue->efx = efx; 546 tx_queue->queue = -1; 547 tx_queue->label = j; 548 tx_queue->channel = channel; 549 } 550 551 #ifdef CONFIG_RFS_ACCEL 552 INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire); 553 #endif 554 555 rx_queue = &channel->rx_queue; 556 rx_queue->efx = efx; 557 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0); 558 559 return channel; 560 } 561 562 int efx_init_channels(struct efx_nic *efx) 563 { 564 unsigned int i; 565 566 for (i = 0; i < EFX_MAX_CHANNELS; i++) { 567 efx->channel[i] = efx_alloc_channel(efx, i); 568 if (!efx->channel[i]) 569 return -ENOMEM; 570 efx->msi_context[i].efx = efx; 571 efx->msi_context[i].index = i; 572 } 573 574 /* Higher numbered interrupt modes are less capable! */ 575 efx->interrupt_mode = min(efx->type->min_interrupt_mode, 576 efx_interrupt_mode); 577 578 efx->max_channels = EFX_MAX_CHANNELS; 579 efx->max_tx_channels = EFX_MAX_CHANNELS; 580 581 return 0; 582 } 583 584 void efx_fini_channels(struct efx_nic *efx) 585 { 586 unsigned int i; 587 588 for (i = 0; i < EFX_MAX_CHANNELS; i++) 589 if (efx->channel[i]) { 590 kfree(efx->channel[i]); 591 efx->channel[i] = NULL; 592 } 593 } 594 595 /* Allocate and initialise a channel structure, copying parameters 596 * (but not resources) from an old channel structure. 597 */ 598 struct efx_channel *efx_copy_channel(const struct efx_channel *old_channel) 599 { 600 struct efx_rx_queue *rx_queue; 601 struct efx_tx_queue *tx_queue; 602 struct efx_channel *channel; 603 int j; 604 605 channel = kmalloc(sizeof(*channel), GFP_KERNEL); 606 if (!channel) 607 return NULL; 608 609 *channel = *old_channel; 610 611 channel->napi_dev = NULL; 612 INIT_HLIST_NODE(&channel->napi_str.napi_hash_node); 613 channel->napi_str.napi_id = 0; 614 channel->napi_str.state = 0; 615 memset(&channel->eventq, 0, sizeof(channel->eventq)); 616 617 for (j = 0; j < EFX_MAX_TXQ_PER_CHANNEL; j++) { 618 tx_queue = &channel->tx_queue[j]; 619 if (tx_queue->channel) 620 tx_queue->channel = channel; 621 tx_queue->buffer = NULL; 622 tx_queue->cb_page = NULL; 623 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd)); 624 } 625 626 rx_queue = &channel->rx_queue; 627 rx_queue->buffer = NULL; 628 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd)); 629 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0); 630 #ifdef CONFIG_RFS_ACCEL 631 INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire); 632 #endif 633 634 return channel; 635 } 636 637 static int efx_probe_channel(struct efx_channel *channel) 638 { 639 struct efx_tx_queue *tx_queue; 640 struct efx_rx_queue *rx_queue; 641 int rc; 642 643 netif_dbg(channel->efx, probe, channel->efx->net_dev, 644 "creating channel %d\n", channel->channel); 645 646 rc = channel->type->pre_probe(channel); 647 if (rc) 648 goto fail; 649 650 rc = efx_probe_eventq(channel); 651 if (rc) 652 goto fail; 653 654 efx_for_each_channel_tx_queue(tx_queue, channel) { 655 rc = efx_probe_tx_queue(tx_queue); 656 if (rc) 657 goto fail; 658 } 659 660 efx_for_each_channel_rx_queue(rx_queue, channel) { 661 rc = efx_probe_rx_queue(rx_queue); 662 if (rc) 663 goto fail; 664 } 665 666 channel->rx_list = NULL; 667 668 return 0; 669 670 fail: 671 efx_remove_channel(channel); 672 return rc; 673 } 674 675 static void efx_get_channel_name(struct efx_channel *channel, char *buf, 676 size_t len) 677 { 678 struct efx_nic *efx = channel->efx; 679 const char *type; 680 int number; 681 682 number = channel->channel; 683 684 if (number >= efx->xdp_channel_offset && 685 !WARN_ON_ONCE(!efx->n_xdp_channels)) { 686 type = "-xdp"; 687 number -= efx->xdp_channel_offset; 688 } else if (efx->tx_channel_offset == 0) { 689 type = ""; 690 } else if (number < efx->tx_channel_offset) { 691 type = "-rx"; 692 } else { 693 type = "-tx"; 694 number -= efx->tx_channel_offset; 695 } 696 snprintf(buf, len, "%s%s-%d", efx->name, type, number); 697 } 698 699 void efx_set_channel_names(struct efx_nic *efx) 700 { 701 struct efx_channel *channel; 702 703 efx_for_each_channel(channel, efx) 704 channel->type->get_name(channel, 705 efx->msi_context[channel->channel].name, 706 sizeof(efx->msi_context[0].name)); 707 } 708 709 int efx_probe_channels(struct efx_nic *efx) 710 { 711 struct efx_channel *channel; 712 int rc; 713 714 /* Restart special buffer allocation */ 715 efx->next_buffer_table = 0; 716 717 /* Probe channels in reverse, so that any 'extra' channels 718 * use the start of the buffer table. This allows the traffic 719 * channels to be resized without moving them or wasting the 720 * entries before them. 721 */ 722 efx_for_each_channel_rev(channel, efx) { 723 rc = efx_probe_channel(channel); 724 if (rc) { 725 netif_err(efx, probe, efx->net_dev, 726 "failed to create channel %d\n", 727 channel->channel); 728 goto fail; 729 } 730 } 731 efx_set_channel_names(efx); 732 733 return 0; 734 735 fail: 736 efx_remove_channels(efx); 737 return rc; 738 } 739 740 void efx_remove_channel(struct efx_channel *channel) 741 { 742 struct efx_tx_queue *tx_queue; 743 struct efx_rx_queue *rx_queue; 744 745 netif_dbg(channel->efx, drv, channel->efx->net_dev, 746 "destroy chan %d\n", channel->channel); 747 748 efx_for_each_channel_rx_queue(rx_queue, channel) 749 efx_remove_rx_queue(rx_queue); 750 efx_for_each_channel_tx_queue(tx_queue, channel) 751 efx_remove_tx_queue(tx_queue); 752 efx_remove_eventq(channel); 753 channel->type->post_remove(channel); 754 } 755 756 void efx_remove_channels(struct efx_nic *efx) 757 { 758 struct efx_channel *channel; 759 760 efx_for_each_channel(channel, efx) 761 efx_remove_channel(channel); 762 763 kfree(efx->xdp_tx_queues); 764 } 765 766 static int efx_set_xdp_tx_queue(struct efx_nic *efx, int xdp_queue_number, 767 struct efx_tx_queue *tx_queue) 768 { 769 if (xdp_queue_number >= efx->xdp_tx_queue_count) 770 return -EINVAL; 771 772 netif_dbg(efx, drv, efx->net_dev, 773 "Channel %u TXQ %u is XDP %u, HW %u\n", 774 tx_queue->channel->channel, tx_queue->label, 775 xdp_queue_number, tx_queue->queue); 776 efx->xdp_tx_queues[xdp_queue_number] = tx_queue; 777 return 0; 778 } 779 780 static void efx_set_xdp_channels(struct efx_nic *efx) 781 { 782 struct efx_tx_queue *tx_queue; 783 struct efx_channel *channel; 784 unsigned int next_queue = 0; 785 int xdp_queue_number = 0; 786 int rc; 787 788 /* We need to mark which channels really have RX and TX 789 * queues, and adjust the TX queue numbers if we have separate 790 * RX-only and TX-only channels. 791 */ 792 efx_for_each_channel(channel, efx) { 793 if (channel->channel < efx->tx_channel_offset) 794 continue; 795 796 if (efx_channel_is_xdp_tx(channel)) { 797 efx_for_each_channel_tx_queue(tx_queue, channel) { 798 tx_queue->queue = next_queue++; 799 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, 800 tx_queue); 801 if (rc == 0) 802 xdp_queue_number++; 803 } 804 } else { 805 efx_for_each_channel_tx_queue(tx_queue, channel) { 806 tx_queue->queue = next_queue++; 807 netif_dbg(efx, drv, efx->net_dev, 808 "Channel %u TXQ %u is HW %u\n", 809 channel->channel, tx_queue->label, 810 tx_queue->queue); 811 } 812 813 /* If XDP is borrowing queues from net stack, it must 814 * use the queue with no csum offload, which is the 815 * first one of the channel 816 * (note: tx_queue_by_type is not initialized yet) 817 */ 818 if (efx->xdp_txq_queues_mode == 819 EFX_XDP_TX_QUEUES_BORROWED) { 820 tx_queue = &channel->tx_queue[0]; 821 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, 822 tx_queue); 823 if (rc == 0) 824 xdp_queue_number++; 825 } 826 } 827 } 828 WARN_ON(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DEDICATED && 829 xdp_queue_number != efx->xdp_tx_queue_count); 830 WARN_ON(efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED && 831 xdp_queue_number > efx->xdp_tx_queue_count); 832 833 /* If we have more CPUs than assigned XDP TX queues, assign the already 834 * existing queues to the exceeding CPUs 835 */ 836 next_queue = 0; 837 while (xdp_queue_number < efx->xdp_tx_queue_count) { 838 tx_queue = efx->xdp_tx_queues[next_queue++]; 839 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue); 840 if (rc == 0) 841 xdp_queue_number++; 842 } 843 } 844 845 int efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries) 846 { 847 struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel, 848 *ptp_channel = efx_ptp_channel(efx); 849 struct efx_ptp_data *ptp_data = efx->ptp_data; 850 unsigned int i, next_buffer_table = 0; 851 u32 old_rxq_entries, old_txq_entries; 852 int rc, rc2; 853 854 rc = efx_check_disabled(efx); 855 if (rc) 856 return rc; 857 858 /* Not all channels should be reallocated. We must avoid 859 * reallocating their buffer table entries. 860 */ 861 efx_for_each_channel(channel, efx) { 862 struct efx_rx_queue *rx_queue; 863 struct efx_tx_queue *tx_queue; 864 865 if (channel->type->copy) 866 continue; 867 next_buffer_table = max(next_buffer_table, 868 channel->eventq.index + 869 channel->eventq.entries); 870 efx_for_each_channel_rx_queue(rx_queue, channel) 871 next_buffer_table = max(next_buffer_table, 872 rx_queue->rxd.index + 873 rx_queue->rxd.entries); 874 efx_for_each_channel_tx_queue(tx_queue, channel) 875 next_buffer_table = max(next_buffer_table, 876 tx_queue->txd.index + 877 tx_queue->txd.entries); 878 } 879 880 efx_device_detach_sync(efx); 881 efx_stop_all(efx); 882 efx_soft_disable_interrupts(efx); 883 884 /* Clone channels (where possible) */ 885 memset(other_channel, 0, sizeof(other_channel)); 886 for (i = 0; i < efx->n_channels; i++) { 887 channel = efx->channel[i]; 888 if (channel->type->copy) 889 channel = channel->type->copy(channel); 890 if (!channel) { 891 rc = -ENOMEM; 892 goto out; 893 } 894 other_channel[i] = channel; 895 } 896 897 /* Swap entry counts and channel pointers */ 898 old_rxq_entries = efx->rxq_entries; 899 old_txq_entries = efx->txq_entries; 900 efx->rxq_entries = rxq_entries; 901 efx->txq_entries = txq_entries; 902 for (i = 0; i < efx->n_channels; i++) 903 swap(efx->channel[i], other_channel[i]); 904 905 /* Restart buffer table allocation */ 906 efx->next_buffer_table = next_buffer_table; 907 908 for (i = 0; i < efx->n_channels; i++) { 909 channel = efx->channel[i]; 910 if (!channel->type->copy) 911 continue; 912 rc = efx_probe_channel(channel); 913 if (rc) 914 goto rollback; 915 efx_init_napi_channel(efx->channel[i]); 916 } 917 918 efx_set_xdp_channels(efx); 919 out: 920 efx->ptp_data = NULL; 921 /* Destroy unused channel structures */ 922 for (i = 0; i < efx->n_channels; i++) { 923 channel = other_channel[i]; 924 if (channel && channel->type->copy) { 925 efx_fini_napi_channel(channel); 926 efx_remove_channel(channel); 927 kfree(channel); 928 } 929 } 930 931 efx->ptp_data = ptp_data; 932 rc2 = efx_soft_enable_interrupts(efx); 933 if (rc2) { 934 rc = rc ? rc : rc2; 935 netif_err(efx, drv, efx->net_dev, 936 "unable to restart interrupts on channel reallocation\n"); 937 efx_schedule_reset(efx, RESET_TYPE_DISABLE); 938 } else { 939 efx_start_all(efx); 940 efx_device_attach_if_not_resetting(efx); 941 } 942 return rc; 943 944 rollback: 945 /* Swap back */ 946 efx->rxq_entries = old_rxq_entries; 947 efx->txq_entries = old_txq_entries; 948 for (i = 0; i < efx->n_channels; i++) 949 swap(efx->channel[i], other_channel[i]); 950 efx_ptp_update_channel(efx, ptp_channel); 951 goto out; 952 } 953 954 int efx_set_channels(struct efx_nic *efx) 955 { 956 struct efx_channel *channel; 957 int rc; 958 959 if (efx->xdp_tx_queue_count) { 960 EFX_WARN_ON_PARANOID(efx->xdp_tx_queues); 961 962 /* Allocate array for XDP TX queue lookup. */ 963 efx->xdp_tx_queues = kcalloc(efx->xdp_tx_queue_count, 964 sizeof(*efx->xdp_tx_queues), 965 GFP_KERNEL); 966 if (!efx->xdp_tx_queues) 967 return -ENOMEM; 968 } 969 970 efx_for_each_channel(channel, efx) { 971 if (channel->channel < efx->n_rx_channels) 972 channel->rx_queue.core_index = channel->channel; 973 else 974 channel->rx_queue.core_index = -1; 975 } 976 977 efx_set_xdp_channels(efx); 978 979 rc = netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels); 980 if (rc) 981 return rc; 982 return netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels); 983 } 984 985 static bool efx_default_channel_want_txqs(struct efx_channel *channel) 986 { 987 return channel->channel - channel->efx->tx_channel_offset < 988 channel->efx->n_tx_channels; 989 } 990 991 /************* 992 * START/STOP 993 *************/ 994 995 int efx_soft_enable_interrupts(struct efx_nic *efx) 996 { 997 struct efx_channel *channel, *end_channel; 998 int rc; 999 1000 BUG_ON(efx->state == STATE_DISABLED); 1001 1002 efx->irq_soft_enabled = true; 1003 smp_wmb(); 1004 1005 efx_for_each_channel(channel, efx) { 1006 if (!channel->type->keep_eventq) { 1007 rc = efx_init_eventq(channel); 1008 if (rc) 1009 goto fail; 1010 } 1011 efx_start_eventq(channel); 1012 } 1013 1014 efx_mcdi_mode_event(efx); 1015 1016 return 0; 1017 fail: 1018 end_channel = channel; 1019 efx_for_each_channel(channel, efx) { 1020 if (channel == end_channel) 1021 break; 1022 efx_stop_eventq(channel); 1023 if (!channel->type->keep_eventq) 1024 efx_fini_eventq(channel); 1025 } 1026 1027 return rc; 1028 } 1029 1030 void efx_soft_disable_interrupts(struct efx_nic *efx) 1031 { 1032 struct efx_channel *channel; 1033 1034 if (efx->state == STATE_DISABLED) 1035 return; 1036 1037 efx_mcdi_mode_poll(efx); 1038 1039 efx->irq_soft_enabled = false; 1040 smp_wmb(); 1041 1042 if (efx->legacy_irq) 1043 synchronize_irq(efx->legacy_irq); 1044 1045 efx_for_each_channel(channel, efx) { 1046 if (channel->irq) 1047 synchronize_irq(channel->irq); 1048 1049 efx_stop_eventq(channel); 1050 if (!channel->type->keep_eventq) 1051 efx_fini_eventq(channel); 1052 } 1053 1054 /* Flush the asynchronous MCDI request queue */ 1055 efx_mcdi_flush_async(efx); 1056 } 1057 1058 int efx_enable_interrupts(struct efx_nic *efx) 1059 { 1060 struct efx_channel *channel, *end_channel; 1061 int rc; 1062 1063 /* TODO: Is this really a bug? */ 1064 BUG_ON(efx->state == STATE_DISABLED); 1065 1066 if (efx->eeh_disabled_legacy_irq) { 1067 enable_irq(efx->legacy_irq); 1068 efx->eeh_disabled_legacy_irq = false; 1069 } 1070 1071 efx->type->irq_enable_master(efx); 1072 1073 efx_for_each_channel(channel, efx) { 1074 if (channel->type->keep_eventq) { 1075 rc = efx_init_eventq(channel); 1076 if (rc) 1077 goto fail; 1078 } 1079 } 1080 1081 rc = efx_soft_enable_interrupts(efx); 1082 if (rc) 1083 goto fail; 1084 1085 return 0; 1086 1087 fail: 1088 end_channel = channel; 1089 efx_for_each_channel(channel, efx) { 1090 if (channel == end_channel) 1091 break; 1092 if (channel->type->keep_eventq) 1093 efx_fini_eventq(channel); 1094 } 1095 1096 efx->type->irq_disable_non_ev(efx); 1097 1098 return rc; 1099 } 1100 1101 void efx_disable_interrupts(struct efx_nic *efx) 1102 { 1103 struct efx_channel *channel; 1104 1105 efx_soft_disable_interrupts(efx); 1106 1107 efx_for_each_channel(channel, efx) { 1108 if (channel->type->keep_eventq) 1109 efx_fini_eventq(channel); 1110 } 1111 1112 efx->type->irq_disable_non_ev(efx); 1113 } 1114 1115 void efx_start_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 1121 efx_for_each_channel_rev(channel, efx) { 1122 efx_for_each_channel_tx_queue(tx_queue, channel) { 1123 efx_init_tx_queue(tx_queue); 1124 atomic_inc(&efx->active_queues); 1125 } 1126 1127 efx_for_each_channel_rx_queue(rx_queue, channel) { 1128 efx_init_rx_queue(rx_queue); 1129 atomic_inc(&efx->active_queues); 1130 efx_stop_eventq(channel); 1131 efx_fast_push_rx_descriptors(rx_queue, false); 1132 efx_start_eventq(channel); 1133 } 1134 1135 WARN_ON(channel->rx_pkt_n_frags); 1136 } 1137 } 1138 1139 void efx_stop_channels(struct efx_nic *efx) 1140 { 1141 struct efx_tx_queue *tx_queue; 1142 struct efx_rx_queue *rx_queue; 1143 struct efx_channel *channel; 1144 int rc = 0; 1145 1146 /* Stop RX refill */ 1147 efx_for_each_channel(channel, efx) { 1148 efx_for_each_channel_rx_queue(rx_queue, channel) 1149 rx_queue->refill_enabled = false; 1150 } 1151 1152 efx_for_each_channel(channel, efx) { 1153 /* RX packet processing is pipelined, so wait for the 1154 * NAPI handler to complete. At least event queue 0 1155 * might be kept active by non-data events, so don't 1156 * use napi_synchronize() but actually disable NAPI 1157 * temporarily. 1158 */ 1159 if (efx_channel_has_rx_queue(channel)) { 1160 efx_stop_eventq(channel); 1161 efx_start_eventq(channel); 1162 } 1163 } 1164 1165 if (efx->type->fini_dmaq) 1166 rc = efx->type->fini_dmaq(efx); 1167 1168 if (rc) { 1169 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n"); 1170 } else { 1171 netif_dbg(efx, drv, efx->net_dev, 1172 "successfully flushed all queues\n"); 1173 } 1174 1175 efx_for_each_channel(channel, efx) { 1176 efx_for_each_channel_rx_queue(rx_queue, channel) 1177 efx_fini_rx_queue(rx_queue); 1178 efx_for_each_channel_tx_queue(tx_queue, channel) 1179 efx_fini_tx_queue(tx_queue); 1180 } 1181 } 1182 1183 /************************************************************************** 1184 * 1185 * NAPI interface 1186 * 1187 *************************************************************************/ 1188 1189 /* Process channel's event queue 1190 * 1191 * This function is responsible for processing the event queue of a 1192 * single channel. The caller must guarantee that this function will 1193 * never be concurrently called more than once on the same channel, 1194 * though different channels may be being processed concurrently. 1195 */ 1196 static int efx_process_channel(struct efx_channel *channel, int budget) 1197 { 1198 struct efx_tx_queue *tx_queue; 1199 struct list_head rx_list; 1200 int spent; 1201 1202 if (unlikely(!channel->enabled)) 1203 return 0; 1204 1205 /* Prepare the batch receive list */ 1206 EFX_WARN_ON_PARANOID(channel->rx_list != NULL); 1207 INIT_LIST_HEAD(&rx_list); 1208 channel->rx_list = &rx_list; 1209 1210 efx_for_each_channel_tx_queue(tx_queue, channel) { 1211 tx_queue->pkts_compl = 0; 1212 tx_queue->bytes_compl = 0; 1213 } 1214 1215 spent = efx_nic_process_eventq(channel, budget); 1216 if (spent && efx_channel_has_rx_queue(channel)) { 1217 struct efx_rx_queue *rx_queue = 1218 efx_channel_get_rx_queue(channel); 1219 1220 efx_rx_flush_packet(channel); 1221 efx_fast_push_rx_descriptors(rx_queue, true); 1222 } 1223 1224 /* Update BQL */ 1225 efx_for_each_channel_tx_queue(tx_queue, channel) { 1226 if (tx_queue->bytes_compl) { 1227 netdev_tx_completed_queue(tx_queue->core_txq, 1228 tx_queue->pkts_compl, 1229 tx_queue->bytes_compl); 1230 } 1231 } 1232 1233 /* Receive any packets we queued up */ 1234 netif_receive_skb_list(channel->rx_list); 1235 channel->rx_list = NULL; 1236 1237 return spent; 1238 } 1239 1240 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel) 1241 { 1242 int step = efx->irq_mod_step_us; 1243 1244 if (channel->irq_mod_score < irq_adapt_low_thresh) { 1245 if (channel->irq_moderation_us > step) { 1246 channel->irq_moderation_us -= step; 1247 efx->type->push_irq_moderation(channel); 1248 } 1249 } else if (channel->irq_mod_score > irq_adapt_high_thresh) { 1250 if (channel->irq_moderation_us < 1251 efx->irq_rx_moderation_us) { 1252 channel->irq_moderation_us += step; 1253 efx->type->push_irq_moderation(channel); 1254 } 1255 } 1256 1257 channel->irq_count = 0; 1258 channel->irq_mod_score = 0; 1259 } 1260 1261 /* NAPI poll handler 1262 * 1263 * NAPI guarantees serialisation of polls of the same device, which 1264 * provides the guarantee required by efx_process_channel(). 1265 */ 1266 static int efx_poll(struct napi_struct *napi, int budget) 1267 { 1268 struct efx_channel *channel = 1269 container_of(napi, struct efx_channel, napi_str); 1270 struct efx_nic *efx = channel->efx; 1271 #ifdef CONFIG_RFS_ACCEL 1272 unsigned int time; 1273 #endif 1274 int spent; 1275 1276 netif_vdbg(efx, intr, efx->net_dev, 1277 "channel %d NAPI poll executing on CPU %d\n", 1278 channel->channel, raw_smp_processor_id()); 1279 1280 spent = efx_process_channel(channel, budget); 1281 1282 xdp_do_flush_map(); 1283 1284 if (spent < budget) { 1285 if (efx_channel_has_rx_queue(channel) && 1286 efx->irq_rx_adaptive && 1287 unlikely(++channel->irq_count == 1000)) { 1288 efx_update_irq_mod(efx, channel); 1289 } 1290 1291 #ifdef CONFIG_RFS_ACCEL 1292 /* Perhaps expire some ARFS filters */ 1293 time = jiffies - channel->rfs_last_expiry; 1294 /* Would our quota be >= 20? */ 1295 if (channel->rfs_filter_count * time >= 600 * HZ) 1296 mod_delayed_work(system_wq, &channel->filter_work, 0); 1297 #endif 1298 1299 /* There is no race here; although napi_disable() will 1300 * only wait for napi_complete(), this isn't a problem 1301 * since efx_nic_eventq_read_ack() will have no effect if 1302 * interrupts have already been disabled. 1303 */ 1304 if (napi_complete_done(napi, spent)) 1305 efx_nic_eventq_read_ack(channel); 1306 } 1307 1308 return spent; 1309 } 1310 1311 void efx_init_napi_channel(struct efx_channel *channel) 1312 { 1313 struct efx_nic *efx = channel->efx; 1314 1315 channel->napi_dev = efx->net_dev; 1316 netif_napi_add(channel->napi_dev, &channel->napi_str, efx_poll, 64); 1317 } 1318 1319 void efx_init_napi(struct efx_nic *efx) 1320 { 1321 struct efx_channel *channel; 1322 1323 efx_for_each_channel(channel, efx) 1324 efx_init_napi_channel(channel); 1325 } 1326 1327 void efx_fini_napi_channel(struct efx_channel *channel) 1328 { 1329 if (channel->napi_dev) 1330 netif_napi_del(&channel->napi_str); 1331 1332 channel->napi_dev = NULL; 1333 } 1334 1335 void efx_fini_napi(struct efx_nic *efx) 1336 { 1337 struct efx_channel *channel; 1338 1339 efx_for_each_channel(channel, efx) 1340 efx_fini_napi_channel(channel); 1341 } 1342 1343 /*************** 1344 * Housekeeping 1345 ***************/ 1346 1347 static int efx_channel_dummy_op_int(struct efx_channel *channel) 1348 { 1349 return 0; 1350 } 1351 1352 void efx_channel_dummy_op_void(struct efx_channel *channel) 1353 { 1354 } 1355 1356 static const struct efx_channel_type efx_default_channel_type = { 1357 .pre_probe = efx_channel_dummy_op_int, 1358 .post_remove = efx_channel_dummy_op_void, 1359 .get_name = efx_get_channel_name, 1360 .copy = efx_copy_channel, 1361 .want_txqs = efx_default_channel_want_txqs, 1362 .keep_eventq = false, 1363 .want_pio = true, 1364 }; 1365