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 /* This is the weight assigned to each of the (per-channel) virtual 50 * NAPI devices. 51 */ 52 static int napi_weight = 64; 53 54 /*************** 55 * Housekeeping 56 ***************/ 57 58 int efx_channel_dummy_op_int(struct efx_channel *channel) 59 { 60 return 0; 61 } 62 63 void efx_channel_dummy_op_void(struct efx_channel *channel) 64 { 65 } 66 67 static const struct efx_channel_type efx_default_channel_type = { 68 .pre_probe = efx_channel_dummy_op_int, 69 .post_remove = efx_channel_dummy_op_void, 70 .get_name = efx_get_channel_name, 71 .copy = efx_copy_channel, 72 .want_txqs = efx_default_channel_want_txqs, 73 .keep_eventq = false, 74 .want_pio = true, 75 }; 76 77 /************* 78 * INTERRUPTS 79 *************/ 80 81 static unsigned int count_online_cores(struct efx_nic *efx, bool local_node) 82 { 83 cpumask_var_t filter_mask; 84 unsigned int count; 85 int cpu; 86 87 if (unlikely(!zalloc_cpumask_var(&filter_mask, GFP_KERNEL))) { 88 netif_warn(efx, probe, efx->net_dev, 89 "RSS disabled due to allocation failure\n"); 90 return 1; 91 } 92 93 cpumask_copy(filter_mask, cpu_online_mask); 94 if (local_node) 95 cpumask_and(filter_mask, filter_mask, 96 cpumask_of_pcibus(efx->pci_dev->bus)); 97 98 count = 0; 99 for_each_cpu(cpu, filter_mask) { 100 ++count; 101 cpumask_andnot(filter_mask, filter_mask, topology_sibling_cpumask(cpu)); 102 } 103 104 free_cpumask_var(filter_mask); 105 106 return count; 107 } 108 109 static unsigned int efx_wanted_parallelism(struct efx_nic *efx) 110 { 111 unsigned int count; 112 113 if (rss_cpus) { 114 count = rss_cpus; 115 } else { 116 count = count_online_cores(efx, true); 117 118 /* If no online CPUs in local node, fallback to any online CPUs */ 119 if (count == 0) 120 count = count_online_cores(efx, false); 121 } 122 123 if (count > EFX_MAX_RX_QUEUES) { 124 netif_cond_dbg(efx, probe, efx->net_dev, !rss_cpus, warn, 125 "Reducing number of rx queues from %u to %u.\n", 126 count, EFX_MAX_RX_QUEUES); 127 count = EFX_MAX_RX_QUEUES; 128 } 129 130 /* If RSS is requested for the PF *and* VFs then we can't write RSS 131 * table entries that are inaccessible to VFs 132 */ 133 #ifdef CONFIG_SFC_SRIOV 134 if (efx->type->sriov_wanted) { 135 if (efx->type->sriov_wanted(efx) && efx_vf_size(efx) > 1 && 136 count > efx_vf_size(efx)) { 137 netif_warn(efx, probe, efx->net_dev, 138 "Reducing number of RSS channels from %u to %u for " 139 "VF support. Increase vf-msix-limit to use more " 140 "channels on the PF.\n", 141 count, efx_vf_size(efx)); 142 count = efx_vf_size(efx); 143 } 144 } 145 #endif 146 147 return count; 148 } 149 150 static int efx_allocate_msix_channels(struct efx_nic *efx, 151 unsigned int max_channels, 152 unsigned int extra_channels, 153 unsigned int parallelism) 154 { 155 unsigned int n_channels = parallelism; 156 int vec_count; 157 int tx_per_ev; 158 int n_xdp_tx; 159 int n_xdp_ev; 160 161 if (efx_separate_tx_channels) 162 n_channels *= 2; 163 n_channels += extra_channels; 164 165 /* To allow XDP transmit to happen from arbitrary NAPI contexts 166 * we allocate a TX queue per CPU. We share event queues across 167 * multiple tx queues, assuming tx and ev queues are both 168 * maximum size. 169 */ 170 tx_per_ev = EFX_MAX_EVQ_SIZE / EFX_TXQ_MAX_ENT(efx); 171 tx_per_ev = min(tx_per_ev, EFX_MAX_TXQ_PER_CHANNEL); 172 n_xdp_tx = num_possible_cpus(); 173 n_xdp_ev = DIV_ROUND_UP(n_xdp_tx, tx_per_ev); 174 175 vec_count = pci_msix_vec_count(efx->pci_dev); 176 if (vec_count < 0) 177 return vec_count; 178 179 max_channels = min_t(unsigned int, vec_count, max_channels); 180 181 /* Check resources. 182 * We need a channel per event queue, plus a VI per tx queue. 183 * This may be more pessimistic than it needs to be. 184 */ 185 if (n_channels >= max_channels) { 186 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED; 187 netif_warn(efx, drv, efx->net_dev, 188 "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n", 189 n_xdp_ev, n_channels, max_channels); 190 netif_warn(efx, drv, efx->net_dev, 191 "XDP_TX and XDP_REDIRECT might decrease device's performance\n"); 192 } else if (n_channels + n_xdp_tx > efx->max_vis) { 193 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_BORROWED; 194 netif_warn(efx, drv, efx->net_dev, 195 "Insufficient resources for %d XDP TX queues (%d other channels, max VIs %d)\n", 196 n_xdp_tx, n_channels, efx->max_vis); 197 netif_warn(efx, drv, efx->net_dev, 198 "XDP_TX and XDP_REDIRECT might decrease device's performance\n"); 199 } else if (n_channels + n_xdp_ev > max_channels) { 200 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_SHARED; 201 netif_warn(efx, drv, efx->net_dev, 202 "Insufficient resources for %d XDP event queues (%d other channels, max %d)\n", 203 n_xdp_ev, n_channels, max_channels); 204 205 n_xdp_ev = max_channels - n_channels; 206 netif_warn(efx, drv, efx->net_dev, 207 "XDP_TX and XDP_REDIRECT will work with reduced performance (%d cpus/tx_queue)\n", 208 DIV_ROUND_UP(n_xdp_tx, tx_per_ev * n_xdp_ev)); 209 } else { 210 efx->xdp_txq_queues_mode = EFX_XDP_TX_QUEUES_DEDICATED; 211 } 212 213 if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_BORROWED) { 214 efx->n_xdp_channels = n_xdp_ev; 215 efx->xdp_tx_per_channel = tx_per_ev; 216 efx->xdp_tx_queue_count = n_xdp_tx; 217 n_channels += n_xdp_ev; 218 netif_dbg(efx, drv, efx->net_dev, 219 "Allocating %d TX and %d event queues for XDP\n", 220 n_xdp_ev * tx_per_ev, n_xdp_ev); 221 } else { 222 efx->n_xdp_channels = 0; 223 efx->xdp_tx_per_channel = 0; 224 efx->xdp_tx_queue_count = n_xdp_tx; 225 } 226 227 if (vec_count < n_channels) { 228 netif_err(efx, drv, efx->net_dev, 229 "WARNING: Insufficient MSI-X vectors available (%d < %u).\n", 230 vec_count, n_channels); 231 netif_err(efx, drv, efx->net_dev, 232 "WARNING: Performance may be reduced.\n"); 233 n_channels = vec_count; 234 } 235 236 n_channels = min(n_channels, max_channels); 237 238 efx->n_channels = n_channels; 239 240 /* Ignore XDP tx channels when creating rx channels. */ 241 n_channels -= efx->n_xdp_channels; 242 243 if (efx_separate_tx_channels) { 244 efx->n_tx_channels = 245 min(max(n_channels / 2, 1U), 246 efx->max_tx_channels); 247 efx->tx_channel_offset = 248 n_channels - efx->n_tx_channels; 249 efx->n_rx_channels = 250 max(n_channels - 251 efx->n_tx_channels, 1U); 252 } else { 253 efx->n_tx_channels = min(n_channels, efx->max_tx_channels); 254 efx->tx_channel_offset = 0; 255 efx->n_rx_channels = n_channels; 256 } 257 258 efx->n_rx_channels = min(efx->n_rx_channels, parallelism); 259 efx->n_tx_channels = min(efx->n_tx_channels, parallelism); 260 261 efx->xdp_channel_offset = n_channels; 262 263 netif_dbg(efx, drv, efx->net_dev, 264 "Allocating %u RX channels\n", 265 efx->n_rx_channels); 266 267 return efx->n_channels; 268 } 269 270 /* Probe the number and type of interrupts we are able to obtain, and 271 * the resulting numbers of channels and RX queues. 272 */ 273 int efx_probe_interrupts(struct efx_nic *efx) 274 { 275 unsigned int extra_channels = 0; 276 unsigned int rss_spread; 277 unsigned int i, j; 278 int rc; 279 280 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) 281 if (efx->extra_channel_type[i]) 282 ++extra_channels; 283 284 if (efx->interrupt_mode == EFX_INT_MODE_MSIX) { 285 unsigned int parallelism = efx_wanted_parallelism(efx); 286 struct msix_entry xentries[EFX_MAX_CHANNELS]; 287 unsigned int n_channels; 288 289 rc = efx_allocate_msix_channels(efx, efx->max_channels, 290 extra_channels, parallelism); 291 if (rc >= 0) { 292 n_channels = rc; 293 for (i = 0; i < n_channels; i++) 294 xentries[i].entry = i; 295 rc = pci_enable_msix_range(efx->pci_dev, xentries, 1, 296 n_channels); 297 } 298 if (rc < 0) { 299 /* Fall back to single channel MSI */ 300 netif_err(efx, drv, efx->net_dev, 301 "could not enable MSI-X\n"); 302 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_MSI) 303 efx->interrupt_mode = EFX_INT_MODE_MSI; 304 else 305 return rc; 306 } else if (rc < n_channels) { 307 netif_err(efx, drv, efx->net_dev, 308 "WARNING: Insufficient MSI-X vectors" 309 " available (%d < %u).\n", rc, n_channels); 310 netif_err(efx, drv, efx->net_dev, 311 "WARNING: Performance may be reduced.\n"); 312 n_channels = rc; 313 } 314 315 if (rc > 0) { 316 for (i = 0; i < efx->n_channels; i++) 317 efx_get_channel(efx, i)->irq = 318 xentries[i].vector; 319 } 320 } 321 322 /* Try single interrupt MSI */ 323 if (efx->interrupt_mode == EFX_INT_MODE_MSI) { 324 efx->n_channels = 1; 325 efx->n_rx_channels = 1; 326 efx->n_tx_channels = 1; 327 efx->n_xdp_channels = 0; 328 efx->xdp_channel_offset = efx->n_channels; 329 rc = pci_enable_msi(efx->pci_dev); 330 if (rc == 0) { 331 efx_get_channel(efx, 0)->irq = efx->pci_dev->irq; 332 } else { 333 netif_err(efx, drv, efx->net_dev, 334 "could not enable MSI\n"); 335 if (efx->type->min_interrupt_mode >= EFX_INT_MODE_LEGACY) 336 efx->interrupt_mode = EFX_INT_MODE_LEGACY; 337 else 338 return rc; 339 } 340 } 341 342 /* Assume legacy interrupts */ 343 if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) { 344 efx->n_channels = 1 + (efx_separate_tx_channels ? 1 : 0); 345 efx->n_rx_channels = 1; 346 efx->n_tx_channels = 1; 347 efx->n_xdp_channels = 0; 348 efx->xdp_channel_offset = efx->n_channels; 349 efx->legacy_irq = efx->pci_dev->irq; 350 } 351 352 /* Assign extra channels if possible, before XDP channels */ 353 efx->n_extra_tx_channels = 0; 354 j = efx->xdp_channel_offset; 355 for (i = 0; i < EFX_MAX_EXTRA_CHANNELS; i++) { 356 if (!efx->extra_channel_type[i]) 357 continue; 358 if (j <= efx->tx_channel_offset + efx->n_tx_channels) { 359 efx->extra_channel_type[i]->handle_no_channel(efx); 360 } else { 361 --j; 362 efx_get_channel(efx, j)->type = 363 efx->extra_channel_type[i]; 364 if (efx_channel_has_tx_queues(efx_get_channel(efx, j))) 365 efx->n_extra_tx_channels++; 366 } 367 } 368 369 rss_spread = efx->n_rx_channels; 370 /* RSS might be usable on VFs even if it is disabled on the PF */ 371 #ifdef CONFIG_SFC_SRIOV 372 if (efx->type->sriov_wanted) { 373 efx->rss_spread = ((rss_spread > 1 || 374 !efx->type->sriov_wanted(efx)) ? 375 rss_spread : efx_vf_size(efx)); 376 return 0; 377 } 378 #endif 379 efx->rss_spread = rss_spread; 380 381 return 0; 382 } 383 384 #if defined(CONFIG_SMP) 385 void efx_set_interrupt_affinity(struct efx_nic *efx) 386 { 387 const struct cpumask *numa_mask = cpumask_of_pcibus(efx->pci_dev->bus); 388 struct efx_channel *channel; 389 unsigned int cpu; 390 391 /* If no online CPUs in local node, fallback to any online CPU */ 392 if (cpumask_first_and(cpu_online_mask, numa_mask) >= nr_cpu_ids) 393 numa_mask = cpu_online_mask; 394 395 cpu = -1; 396 efx_for_each_channel(channel, efx) { 397 cpu = cpumask_next_and(cpu, cpu_online_mask, numa_mask); 398 if (cpu >= nr_cpu_ids) 399 cpu = cpumask_first_and(cpu_online_mask, numa_mask); 400 irq_set_affinity_hint(channel->irq, cpumask_of(cpu)); 401 } 402 } 403 404 void efx_clear_interrupt_affinity(struct efx_nic *efx) 405 { 406 struct efx_channel *channel; 407 408 efx_for_each_channel(channel, efx) 409 irq_set_affinity_hint(channel->irq, NULL); 410 } 411 #else 412 void 413 efx_set_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused))) 414 { 415 } 416 417 void 418 efx_clear_interrupt_affinity(struct efx_nic *efx __attribute__ ((unused))) 419 { 420 } 421 #endif /* CONFIG_SMP */ 422 423 void efx_remove_interrupts(struct efx_nic *efx) 424 { 425 struct efx_channel *channel; 426 427 /* Remove MSI/MSI-X interrupts */ 428 efx_for_each_channel(channel, efx) 429 channel->irq = 0; 430 pci_disable_msi(efx->pci_dev); 431 pci_disable_msix(efx->pci_dev); 432 433 /* Remove legacy interrupt */ 434 efx->legacy_irq = 0; 435 } 436 437 /*************** 438 * EVENT QUEUES 439 ***************/ 440 441 /* Create event queue 442 * Event queue memory allocations are done only once. If the channel 443 * is reset, the memory buffer will be reused; this guards against 444 * errors during channel reset and also simplifies interrupt handling. 445 */ 446 int efx_probe_eventq(struct efx_channel *channel) 447 { 448 struct efx_nic *efx = channel->efx; 449 unsigned long entries; 450 451 netif_dbg(efx, probe, efx->net_dev, 452 "chan %d create event queue\n", channel->channel); 453 454 /* Build an event queue with room for one event per tx and rx buffer, 455 * plus some extra for link state events and MCDI completions. 456 */ 457 entries = roundup_pow_of_two(efx->rxq_entries + efx->txq_entries + 128); 458 EFX_WARN_ON_PARANOID(entries > EFX_MAX_EVQ_SIZE); 459 channel->eventq_mask = max(entries, EFX_MIN_EVQ_SIZE) - 1; 460 461 return efx_nic_probe_eventq(channel); 462 } 463 464 /* Prepare channel's event queue */ 465 int efx_init_eventq(struct efx_channel *channel) 466 { 467 struct efx_nic *efx = channel->efx; 468 int rc; 469 470 EFX_WARN_ON_PARANOID(channel->eventq_init); 471 472 netif_dbg(efx, drv, efx->net_dev, 473 "chan %d init event queue\n", channel->channel); 474 475 rc = efx_nic_init_eventq(channel); 476 if (rc == 0) { 477 efx->type->push_irq_moderation(channel); 478 channel->eventq_read_ptr = 0; 479 channel->eventq_init = true; 480 } 481 return rc; 482 } 483 484 /* Enable event queue processing and NAPI */ 485 void efx_start_eventq(struct efx_channel *channel) 486 { 487 netif_dbg(channel->efx, ifup, channel->efx->net_dev, 488 "chan %d start event queue\n", channel->channel); 489 490 /* Make sure the NAPI handler sees the enabled flag set */ 491 channel->enabled = true; 492 smp_wmb(); 493 494 napi_enable(&channel->napi_str); 495 efx_nic_eventq_read_ack(channel); 496 } 497 498 /* Disable event queue processing and NAPI */ 499 void efx_stop_eventq(struct efx_channel *channel) 500 { 501 if (!channel->enabled) 502 return; 503 504 napi_disable(&channel->napi_str); 505 channel->enabled = false; 506 } 507 508 void efx_fini_eventq(struct efx_channel *channel) 509 { 510 if (!channel->eventq_init) 511 return; 512 513 netif_dbg(channel->efx, drv, channel->efx->net_dev, 514 "chan %d fini event queue\n", channel->channel); 515 516 efx_nic_fini_eventq(channel); 517 channel->eventq_init = false; 518 } 519 520 void efx_remove_eventq(struct efx_channel *channel) 521 { 522 netif_dbg(channel->efx, drv, channel->efx->net_dev, 523 "chan %d remove event queue\n", channel->channel); 524 525 efx_nic_remove_eventq(channel); 526 } 527 528 /************************************************************************** 529 * 530 * Channel handling 531 * 532 *************************************************************************/ 533 534 #ifdef CONFIG_RFS_ACCEL 535 static void efx_filter_rfs_expire(struct work_struct *data) 536 { 537 struct delayed_work *dwork = to_delayed_work(data); 538 struct efx_channel *channel; 539 unsigned int time, quota; 540 541 channel = container_of(dwork, struct efx_channel, filter_work); 542 time = jiffies - channel->rfs_last_expiry; 543 quota = channel->rfs_filter_count * time / (30 * HZ); 544 if (quota >= 20 && __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, quota))) 545 channel->rfs_last_expiry += time; 546 /* Ensure we do more work eventually even if NAPI poll is not happening */ 547 schedule_delayed_work(dwork, 30 * HZ); 548 } 549 #endif 550 551 /* Allocate and initialise a channel structure. */ 552 static struct efx_channel *efx_alloc_channel(struct efx_nic *efx, int i) 553 { 554 struct efx_rx_queue *rx_queue; 555 struct efx_tx_queue *tx_queue; 556 struct efx_channel *channel; 557 int j; 558 559 channel = kzalloc(sizeof(*channel), GFP_KERNEL); 560 if (!channel) 561 return NULL; 562 563 channel->efx = efx; 564 channel->channel = i; 565 channel->type = &efx_default_channel_type; 566 567 for (j = 0; j < EFX_MAX_TXQ_PER_CHANNEL; j++) { 568 tx_queue = &channel->tx_queue[j]; 569 tx_queue->efx = efx; 570 tx_queue->queue = -1; 571 tx_queue->label = j; 572 tx_queue->channel = channel; 573 } 574 575 #ifdef CONFIG_RFS_ACCEL 576 INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire); 577 #endif 578 579 rx_queue = &channel->rx_queue; 580 rx_queue->efx = efx; 581 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0); 582 583 return channel; 584 } 585 586 int efx_init_channels(struct efx_nic *efx) 587 { 588 unsigned int i; 589 590 for (i = 0; i < EFX_MAX_CHANNELS; i++) { 591 efx->channel[i] = efx_alloc_channel(efx, i); 592 if (!efx->channel[i]) 593 return -ENOMEM; 594 efx->msi_context[i].efx = efx; 595 efx->msi_context[i].index = i; 596 } 597 598 /* Higher numbered interrupt modes are less capable! */ 599 efx->interrupt_mode = min(efx->type->min_interrupt_mode, 600 efx_interrupt_mode); 601 602 efx->max_channels = EFX_MAX_CHANNELS; 603 efx->max_tx_channels = EFX_MAX_CHANNELS; 604 605 return 0; 606 } 607 608 void efx_fini_channels(struct efx_nic *efx) 609 { 610 unsigned int i; 611 612 for (i = 0; i < EFX_MAX_CHANNELS; i++) 613 if (efx->channel[i]) { 614 kfree(efx->channel[i]); 615 efx->channel[i] = NULL; 616 } 617 } 618 619 /* Allocate and initialise a channel structure, copying parameters 620 * (but not resources) from an old channel structure. 621 */ 622 struct efx_channel *efx_copy_channel(const struct efx_channel *old_channel) 623 { 624 struct efx_rx_queue *rx_queue; 625 struct efx_tx_queue *tx_queue; 626 struct efx_channel *channel; 627 int j; 628 629 channel = kmalloc(sizeof(*channel), GFP_KERNEL); 630 if (!channel) 631 return NULL; 632 633 *channel = *old_channel; 634 635 channel->napi_dev = NULL; 636 INIT_HLIST_NODE(&channel->napi_str.napi_hash_node); 637 channel->napi_str.napi_id = 0; 638 channel->napi_str.state = 0; 639 memset(&channel->eventq, 0, sizeof(channel->eventq)); 640 641 for (j = 0; j < EFX_MAX_TXQ_PER_CHANNEL; j++) { 642 tx_queue = &channel->tx_queue[j]; 643 if (tx_queue->channel) 644 tx_queue->channel = channel; 645 tx_queue->buffer = NULL; 646 tx_queue->cb_page = NULL; 647 memset(&tx_queue->txd, 0, sizeof(tx_queue->txd)); 648 } 649 650 rx_queue = &channel->rx_queue; 651 rx_queue->buffer = NULL; 652 memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd)); 653 timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0); 654 #ifdef CONFIG_RFS_ACCEL 655 INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire); 656 #endif 657 658 return channel; 659 } 660 661 static int efx_probe_channel(struct efx_channel *channel) 662 { 663 struct efx_tx_queue *tx_queue; 664 struct efx_rx_queue *rx_queue; 665 int rc; 666 667 netif_dbg(channel->efx, probe, channel->efx->net_dev, 668 "creating channel %d\n", channel->channel); 669 670 rc = channel->type->pre_probe(channel); 671 if (rc) 672 goto fail; 673 674 rc = efx_probe_eventq(channel); 675 if (rc) 676 goto fail; 677 678 efx_for_each_channel_tx_queue(tx_queue, channel) { 679 rc = efx_probe_tx_queue(tx_queue); 680 if (rc) 681 goto fail; 682 } 683 684 efx_for_each_channel_rx_queue(rx_queue, channel) { 685 rc = efx_probe_rx_queue(rx_queue); 686 if (rc) 687 goto fail; 688 } 689 690 channel->rx_list = NULL; 691 692 return 0; 693 694 fail: 695 efx_remove_channel(channel); 696 return rc; 697 } 698 699 void efx_get_channel_name(struct efx_channel *channel, char *buf, size_t len) 700 { 701 struct efx_nic *efx = channel->efx; 702 const char *type; 703 int number; 704 705 number = channel->channel; 706 707 if (number >= efx->xdp_channel_offset && 708 !WARN_ON_ONCE(!efx->n_xdp_channels)) { 709 type = "-xdp"; 710 number -= efx->xdp_channel_offset; 711 } else if (efx->tx_channel_offset == 0) { 712 type = ""; 713 } else if (number < efx->tx_channel_offset) { 714 type = "-rx"; 715 } else { 716 type = "-tx"; 717 number -= efx->tx_channel_offset; 718 } 719 snprintf(buf, len, "%s%s-%d", efx->name, type, number); 720 } 721 722 void efx_set_channel_names(struct efx_nic *efx) 723 { 724 struct efx_channel *channel; 725 726 efx_for_each_channel(channel, efx) 727 channel->type->get_name(channel, 728 efx->msi_context[channel->channel].name, 729 sizeof(efx->msi_context[0].name)); 730 } 731 732 int efx_probe_channels(struct efx_nic *efx) 733 { 734 struct efx_channel *channel; 735 int rc; 736 737 /* Restart special buffer allocation */ 738 efx->next_buffer_table = 0; 739 740 /* Probe channels in reverse, so that any 'extra' channels 741 * use the start of the buffer table. This allows the traffic 742 * channels to be resized without moving them or wasting the 743 * entries before them. 744 */ 745 efx_for_each_channel_rev(channel, efx) { 746 rc = efx_probe_channel(channel); 747 if (rc) { 748 netif_err(efx, probe, efx->net_dev, 749 "failed to create channel %d\n", 750 channel->channel); 751 goto fail; 752 } 753 } 754 efx_set_channel_names(efx); 755 756 return 0; 757 758 fail: 759 efx_remove_channels(efx); 760 return rc; 761 } 762 763 void efx_remove_channel(struct efx_channel *channel) 764 { 765 struct efx_tx_queue *tx_queue; 766 struct efx_rx_queue *rx_queue; 767 768 netif_dbg(channel->efx, drv, channel->efx->net_dev, 769 "destroy chan %d\n", channel->channel); 770 771 efx_for_each_channel_rx_queue(rx_queue, channel) 772 efx_remove_rx_queue(rx_queue); 773 efx_for_each_channel_tx_queue(tx_queue, channel) 774 efx_remove_tx_queue(tx_queue); 775 efx_remove_eventq(channel); 776 channel->type->post_remove(channel); 777 } 778 779 void efx_remove_channels(struct efx_nic *efx) 780 { 781 struct efx_channel *channel; 782 783 efx_for_each_channel(channel, efx) 784 efx_remove_channel(channel); 785 786 kfree(efx->xdp_tx_queues); 787 } 788 789 static int efx_set_xdp_tx_queue(struct efx_nic *efx, int xdp_queue_number, 790 struct efx_tx_queue *tx_queue) 791 { 792 if (xdp_queue_number >= efx->xdp_tx_queue_count) 793 return -EINVAL; 794 795 netif_dbg(efx, drv, efx->net_dev, 796 "Channel %u TXQ %u is XDP %u, HW %u\n", 797 tx_queue->channel->channel, tx_queue->label, 798 xdp_queue_number, tx_queue->queue); 799 efx->xdp_tx_queues[xdp_queue_number] = tx_queue; 800 return 0; 801 } 802 803 static void efx_set_xdp_channels(struct efx_nic *efx) 804 { 805 struct efx_tx_queue *tx_queue; 806 struct efx_channel *channel; 807 unsigned int next_queue = 0; 808 int xdp_queue_number = 0; 809 int rc; 810 811 /* We need to mark which channels really have RX and TX 812 * queues, and adjust the TX queue numbers if we have separate 813 * RX-only and TX-only channels. 814 */ 815 efx_for_each_channel(channel, efx) { 816 if (channel->channel < efx->tx_channel_offset) 817 continue; 818 819 if (efx_channel_is_xdp_tx(channel)) { 820 efx_for_each_channel_tx_queue(tx_queue, channel) { 821 tx_queue->queue = next_queue++; 822 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, 823 tx_queue); 824 if (rc == 0) 825 xdp_queue_number++; 826 } 827 } else { 828 efx_for_each_channel_tx_queue(tx_queue, channel) { 829 tx_queue->queue = next_queue++; 830 netif_dbg(efx, drv, efx->net_dev, 831 "Channel %u TXQ %u is HW %u\n", 832 channel->channel, tx_queue->label, 833 tx_queue->queue); 834 } 835 836 /* If XDP is borrowing queues from net stack, it must 837 * use the queue with no csum offload, which is the 838 * first one of the channel 839 * (note: tx_queue_by_type is not initialized yet) 840 */ 841 if (efx->xdp_txq_queues_mode == 842 EFX_XDP_TX_QUEUES_BORROWED) { 843 tx_queue = &channel->tx_queue[0]; 844 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, 845 tx_queue); 846 if (rc == 0) 847 xdp_queue_number++; 848 } 849 } 850 } 851 WARN_ON(efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_DEDICATED && 852 xdp_queue_number != efx->xdp_tx_queue_count); 853 WARN_ON(efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED && 854 xdp_queue_number > efx->xdp_tx_queue_count); 855 856 /* If we have more CPUs than assigned XDP TX queues, assign the already 857 * existing queues to the exceeding CPUs 858 */ 859 next_queue = 0; 860 while (xdp_queue_number < efx->xdp_tx_queue_count) { 861 tx_queue = efx->xdp_tx_queues[next_queue++]; 862 rc = efx_set_xdp_tx_queue(efx, xdp_queue_number, tx_queue); 863 if (rc == 0) 864 xdp_queue_number++; 865 } 866 } 867 868 int efx_realloc_channels(struct efx_nic *efx, u32 rxq_entries, u32 txq_entries) 869 { 870 struct efx_channel *other_channel[EFX_MAX_CHANNELS], *channel; 871 unsigned int i, next_buffer_table = 0; 872 u32 old_rxq_entries, old_txq_entries; 873 int rc, rc2; 874 875 rc = efx_check_disabled(efx); 876 if (rc) 877 return rc; 878 879 /* Not all channels should be reallocated. We must avoid 880 * reallocating their buffer table entries. 881 */ 882 efx_for_each_channel(channel, efx) { 883 struct efx_rx_queue *rx_queue; 884 struct efx_tx_queue *tx_queue; 885 886 if (channel->type->copy) 887 continue; 888 next_buffer_table = max(next_buffer_table, 889 channel->eventq.index + 890 channel->eventq.entries); 891 efx_for_each_channel_rx_queue(rx_queue, channel) 892 next_buffer_table = max(next_buffer_table, 893 rx_queue->rxd.index + 894 rx_queue->rxd.entries); 895 efx_for_each_channel_tx_queue(tx_queue, channel) 896 next_buffer_table = max(next_buffer_table, 897 tx_queue->txd.index + 898 tx_queue->txd.entries); 899 } 900 901 efx_device_detach_sync(efx); 902 efx_stop_all(efx); 903 efx_soft_disable_interrupts(efx); 904 905 /* Clone channels (where possible) */ 906 memset(other_channel, 0, sizeof(other_channel)); 907 for (i = 0; i < efx->n_channels; i++) { 908 channel = efx->channel[i]; 909 if (channel->type->copy) 910 channel = channel->type->copy(channel); 911 if (!channel) { 912 rc = -ENOMEM; 913 goto out; 914 } 915 other_channel[i] = channel; 916 } 917 918 /* Swap entry counts and channel pointers */ 919 old_rxq_entries = efx->rxq_entries; 920 old_txq_entries = efx->txq_entries; 921 efx->rxq_entries = rxq_entries; 922 efx->txq_entries = txq_entries; 923 for (i = 0; i < efx->n_channels; i++) 924 swap(efx->channel[i], other_channel[i]); 925 926 /* Restart buffer table allocation */ 927 efx->next_buffer_table = next_buffer_table; 928 929 for (i = 0; i < efx->n_channels; i++) { 930 channel = efx->channel[i]; 931 if (!channel->type->copy) 932 continue; 933 rc = efx_probe_channel(channel); 934 if (rc) 935 goto rollback; 936 efx_init_napi_channel(efx->channel[i]); 937 } 938 939 efx_set_xdp_channels(efx); 940 out: 941 /* Destroy unused channel structures */ 942 for (i = 0; i < efx->n_channels; i++) { 943 channel = other_channel[i]; 944 if (channel && channel->type->copy) { 945 efx_fini_napi_channel(channel); 946 efx_remove_channel(channel); 947 kfree(channel); 948 } 949 } 950 951 rc2 = efx_soft_enable_interrupts(efx); 952 if (rc2) { 953 rc = rc ? rc : rc2; 954 netif_err(efx, drv, efx->net_dev, 955 "unable to restart interrupts on channel reallocation\n"); 956 efx_schedule_reset(efx, RESET_TYPE_DISABLE); 957 } else { 958 efx_start_all(efx); 959 efx_device_attach_if_not_resetting(efx); 960 } 961 return rc; 962 963 rollback: 964 /* Swap back */ 965 efx->rxq_entries = old_rxq_entries; 966 efx->txq_entries = old_txq_entries; 967 for (i = 0; i < efx->n_channels; i++) 968 swap(efx->channel[i], other_channel[i]); 969 goto out; 970 } 971 972 int efx_set_channels(struct efx_nic *efx) 973 { 974 struct efx_channel *channel; 975 int rc; 976 977 efx->tx_channel_offset = 978 efx_separate_tx_channels ? 979 efx->n_channels - efx->n_tx_channels : 0; 980 981 if (efx->xdp_tx_queue_count) { 982 EFX_WARN_ON_PARANOID(efx->xdp_tx_queues); 983 984 /* Allocate array for XDP TX queue lookup. */ 985 efx->xdp_tx_queues = kcalloc(efx->xdp_tx_queue_count, 986 sizeof(*efx->xdp_tx_queues), 987 GFP_KERNEL); 988 if (!efx->xdp_tx_queues) 989 return -ENOMEM; 990 } 991 992 efx_for_each_channel(channel, efx) { 993 if (channel->channel < efx->n_rx_channels) 994 channel->rx_queue.core_index = channel->channel; 995 else 996 channel->rx_queue.core_index = -1; 997 } 998 999 efx_set_xdp_channels(efx); 1000 1001 rc = netif_set_real_num_tx_queues(efx->net_dev, efx->n_tx_channels); 1002 if (rc) 1003 return rc; 1004 return netif_set_real_num_rx_queues(efx->net_dev, efx->n_rx_channels); 1005 } 1006 1007 bool efx_default_channel_want_txqs(struct efx_channel *channel) 1008 { 1009 return channel->channel - channel->efx->tx_channel_offset < 1010 channel->efx->n_tx_channels; 1011 } 1012 1013 /************* 1014 * START/STOP 1015 *************/ 1016 1017 int efx_soft_enable_interrupts(struct efx_nic *efx) 1018 { 1019 struct efx_channel *channel, *end_channel; 1020 int rc; 1021 1022 BUG_ON(efx->state == STATE_DISABLED); 1023 1024 efx->irq_soft_enabled = true; 1025 smp_wmb(); 1026 1027 efx_for_each_channel(channel, efx) { 1028 if (!channel->type->keep_eventq) { 1029 rc = efx_init_eventq(channel); 1030 if (rc) 1031 goto fail; 1032 } 1033 efx_start_eventq(channel); 1034 } 1035 1036 efx_mcdi_mode_event(efx); 1037 1038 return 0; 1039 fail: 1040 end_channel = channel; 1041 efx_for_each_channel(channel, efx) { 1042 if (channel == end_channel) 1043 break; 1044 efx_stop_eventq(channel); 1045 if (!channel->type->keep_eventq) 1046 efx_fini_eventq(channel); 1047 } 1048 1049 return rc; 1050 } 1051 1052 void efx_soft_disable_interrupts(struct efx_nic *efx) 1053 { 1054 struct efx_channel *channel; 1055 1056 if (efx->state == STATE_DISABLED) 1057 return; 1058 1059 efx_mcdi_mode_poll(efx); 1060 1061 efx->irq_soft_enabled = false; 1062 smp_wmb(); 1063 1064 if (efx->legacy_irq) 1065 synchronize_irq(efx->legacy_irq); 1066 1067 efx_for_each_channel(channel, efx) { 1068 if (channel->irq) 1069 synchronize_irq(channel->irq); 1070 1071 efx_stop_eventq(channel); 1072 if (!channel->type->keep_eventq) 1073 efx_fini_eventq(channel); 1074 } 1075 1076 /* Flush the asynchronous MCDI request queue */ 1077 efx_mcdi_flush_async(efx); 1078 } 1079 1080 int efx_enable_interrupts(struct efx_nic *efx) 1081 { 1082 struct efx_channel *channel, *end_channel; 1083 int rc; 1084 1085 /* TODO: Is this really a bug? */ 1086 BUG_ON(efx->state == STATE_DISABLED); 1087 1088 if (efx->eeh_disabled_legacy_irq) { 1089 enable_irq(efx->legacy_irq); 1090 efx->eeh_disabled_legacy_irq = false; 1091 } 1092 1093 efx->type->irq_enable_master(efx); 1094 1095 efx_for_each_channel(channel, efx) { 1096 if (channel->type->keep_eventq) { 1097 rc = efx_init_eventq(channel); 1098 if (rc) 1099 goto fail; 1100 } 1101 } 1102 1103 rc = efx_soft_enable_interrupts(efx); 1104 if (rc) 1105 goto fail; 1106 1107 return 0; 1108 1109 fail: 1110 end_channel = channel; 1111 efx_for_each_channel(channel, efx) { 1112 if (channel == end_channel) 1113 break; 1114 if (channel->type->keep_eventq) 1115 efx_fini_eventq(channel); 1116 } 1117 1118 efx->type->irq_disable_non_ev(efx); 1119 1120 return rc; 1121 } 1122 1123 void efx_disable_interrupts(struct efx_nic *efx) 1124 { 1125 struct efx_channel *channel; 1126 1127 efx_soft_disable_interrupts(efx); 1128 1129 efx_for_each_channel(channel, efx) { 1130 if (channel->type->keep_eventq) 1131 efx_fini_eventq(channel); 1132 } 1133 1134 efx->type->irq_disable_non_ev(efx); 1135 } 1136 1137 void efx_start_channels(struct efx_nic *efx) 1138 { 1139 struct efx_tx_queue *tx_queue; 1140 struct efx_rx_queue *rx_queue; 1141 struct efx_channel *channel; 1142 1143 efx_for_each_channel_rev(channel, efx) { 1144 efx_for_each_channel_tx_queue(tx_queue, channel) { 1145 efx_init_tx_queue(tx_queue); 1146 atomic_inc(&efx->active_queues); 1147 } 1148 1149 efx_for_each_channel_rx_queue(rx_queue, channel) { 1150 efx_init_rx_queue(rx_queue); 1151 atomic_inc(&efx->active_queues); 1152 efx_stop_eventq(channel); 1153 efx_fast_push_rx_descriptors(rx_queue, false); 1154 efx_start_eventq(channel); 1155 } 1156 1157 WARN_ON(channel->rx_pkt_n_frags); 1158 } 1159 } 1160 1161 void efx_stop_channels(struct efx_nic *efx) 1162 { 1163 struct efx_tx_queue *tx_queue; 1164 struct efx_rx_queue *rx_queue; 1165 struct efx_channel *channel; 1166 int rc = 0; 1167 1168 /* Stop RX refill */ 1169 efx_for_each_channel(channel, efx) { 1170 efx_for_each_channel_rx_queue(rx_queue, channel) 1171 rx_queue->refill_enabled = false; 1172 } 1173 1174 efx_for_each_channel(channel, efx) { 1175 /* RX packet processing is pipelined, so wait for the 1176 * NAPI handler to complete. At least event queue 0 1177 * might be kept active by non-data events, so don't 1178 * use napi_synchronize() but actually disable NAPI 1179 * temporarily. 1180 */ 1181 if (efx_channel_has_rx_queue(channel)) { 1182 efx_stop_eventq(channel); 1183 efx_start_eventq(channel); 1184 } 1185 } 1186 1187 if (efx->type->fini_dmaq) 1188 rc = efx->type->fini_dmaq(efx); 1189 1190 if (rc) { 1191 netif_err(efx, drv, efx->net_dev, "failed to flush queues\n"); 1192 } else { 1193 netif_dbg(efx, drv, efx->net_dev, 1194 "successfully flushed all queues\n"); 1195 } 1196 1197 efx_for_each_channel(channel, efx) { 1198 efx_for_each_channel_rx_queue(rx_queue, channel) 1199 efx_fini_rx_queue(rx_queue); 1200 efx_for_each_channel_tx_queue(tx_queue, channel) 1201 efx_fini_tx_queue(tx_queue); 1202 } 1203 } 1204 1205 /************************************************************************** 1206 * 1207 * NAPI interface 1208 * 1209 *************************************************************************/ 1210 1211 /* Process channel's event queue 1212 * 1213 * This function is responsible for processing the event queue of a 1214 * single channel. The caller must guarantee that this function will 1215 * never be concurrently called more than once on the same channel, 1216 * though different channels may be being processed concurrently. 1217 */ 1218 static int efx_process_channel(struct efx_channel *channel, int budget) 1219 { 1220 struct efx_tx_queue *tx_queue; 1221 struct list_head rx_list; 1222 int spent; 1223 1224 if (unlikely(!channel->enabled)) 1225 return 0; 1226 1227 /* Prepare the batch receive list */ 1228 EFX_WARN_ON_PARANOID(channel->rx_list != NULL); 1229 INIT_LIST_HEAD(&rx_list); 1230 channel->rx_list = &rx_list; 1231 1232 efx_for_each_channel_tx_queue(tx_queue, channel) { 1233 tx_queue->pkts_compl = 0; 1234 tx_queue->bytes_compl = 0; 1235 } 1236 1237 spent = efx_nic_process_eventq(channel, budget); 1238 if (spent && efx_channel_has_rx_queue(channel)) { 1239 struct efx_rx_queue *rx_queue = 1240 efx_channel_get_rx_queue(channel); 1241 1242 efx_rx_flush_packet(channel); 1243 efx_fast_push_rx_descriptors(rx_queue, true); 1244 } 1245 1246 /* Update BQL */ 1247 efx_for_each_channel_tx_queue(tx_queue, channel) { 1248 if (tx_queue->bytes_compl) { 1249 netdev_tx_completed_queue(tx_queue->core_txq, 1250 tx_queue->pkts_compl, 1251 tx_queue->bytes_compl); 1252 } 1253 } 1254 1255 /* Receive any packets we queued up */ 1256 netif_receive_skb_list(channel->rx_list); 1257 channel->rx_list = NULL; 1258 1259 return spent; 1260 } 1261 1262 static void efx_update_irq_mod(struct efx_nic *efx, struct efx_channel *channel) 1263 { 1264 int step = efx->irq_mod_step_us; 1265 1266 if (channel->irq_mod_score < irq_adapt_low_thresh) { 1267 if (channel->irq_moderation_us > step) { 1268 channel->irq_moderation_us -= step; 1269 efx->type->push_irq_moderation(channel); 1270 } 1271 } else if (channel->irq_mod_score > irq_adapt_high_thresh) { 1272 if (channel->irq_moderation_us < 1273 efx->irq_rx_moderation_us) { 1274 channel->irq_moderation_us += step; 1275 efx->type->push_irq_moderation(channel); 1276 } 1277 } 1278 1279 channel->irq_count = 0; 1280 channel->irq_mod_score = 0; 1281 } 1282 1283 /* NAPI poll handler 1284 * 1285 * NAPI guarantees serialisation of polls of the same device, which 1286 * provides the guarantee required by efx_process_channel(). 1287 */ 1288 static int efx_poll(struct napi_struct *napi, int budget) 1289 { 1290 struct efx_channel *channel = 1291 container_of(napi, struct efx_channel, napi_str); 1292 struct efx_nic *efx = channel->efx; 1293 #ifdef CONFIG_RFS_ACCEL 1294 unsigned int time; 1295 #endif 1296 int spent; 1297 1298 netif_vdbg(efx, intr, efx->net_dev, 1299 "channel %d NAPI poll executing on CPU %d\n", 1300 channel->channel, raw_smp_processor_id()); 1301 1302 spent = efx_process_channel(channel, budget); 1303 1304 xdp_do_flush_map(); 1305 1306 if (spent < budget) { 1307 if (efx_channel_has_rx_queue(channel) && 1308 efx->irq_rx_adaptive && 1309 unlikely(++channel->irq_count == 1000)) { 1310 efx_update_irq_mod(efx, channel); 1311 } 1312 1313 #ifdef CONFIG_RFS_ACCEL 1314 /* Perhaps expire some ARFS filters */ 1315 time = jiffies - channel->rfs_last_expiry; 1316 /* Would our quota be >= 20? */ 1317 if (channel->rfs_filter_count * time >= 600 * HZ) 1318 mod_delayed_work(system_wq, &channel->filter_work, 0); 1319 #endif 1320 1321 /* There is no race here; although napi_disable() will 1322 * only wait for napi_complete(), this isn't a problem 1323 * since efx_nic_eventq_read_ack() will have no effect if 1324 * interrupts have already been disabled. 1325 */ 1326 if (napi_complete_done(napi, spent)) 1327 efx_nic_eventq_read_ack(channel); 1328 } 1329 1330 return spent; 1331 } 1332 1333 void efx_init_napi_channel(struct efx_channel *channel) 1334 { 1335 struct efx_nic *efx = channel->efx; 1336 1337 channel->napi_dev = efx->net_dev; 1338 netif_napi_add(channel->napi_dev, &channel->napi_str, 1339 efx_poll, napi_weight); 1340 } 1341 1342 void efx_init_napi(struct efx_nic *efx) 1343 { 1344 struct efx_channel *channel; 1345 1346 efx_for_each_channel(channel, efx) 1347 efx_init_napi_channel(channel); 1348 } 1349 1350 void efx_fini_napi_channel(struct efx_channel *channel) 1351 { 1352 if (channel->napi_dev) 1353 netif_napi_del(&channel->napi_str); 1354 1355 channel->napi_dev = NULL; 1356 } 1357 1358 void efx_fini_napi(struct efx_nic *efx) 1359 { 1360 struct efx_channel *channel; 1361 1362 efx_for_each_channel(channel, efx) 1363 efx_fini_napi_channel(channel); 1364 } 1365