1 /* 2 * Copyright (C) 2015-2017 Netronome Systems, Inc. 3 * 4 * This software is dual licensed under the GNU General License Version 2, 5 * June 1991 as shown in the file COPYING in the top-level directory of this 6 * source tree or the BSD 2-Clause License provided below. You have the 7 * option to license this software under the complete terms of either license. 8 * 9 * The BSD 2-Clause License: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * 1. Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * 2. Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34 /* 35 * nfp_net_common.c 36 * Netronome network device driver: Common functions between PF and VF 37 * Authors: Jakub Kicinski <jakub.kicinski@netronome.com> 38 * Jason McMullan <jason.mcmullan@netronome.com> 39 * Rolf Neugebauer <rolf.neugebauer@netronome.com> 40 * Brad Petrus <brad.petrus@netronome.com> 41 * Chris Telfer <chris.telfer@netronome.com> 42 */ 43 44 #include <linux/bitfield.h> 45 #include <linux/bpf.h> 46 #include <linux/bpf_trace.h> 47 #include <linux/module.h> 48 #include <linux/kernel.h> 49 #include <linux/init.h> 50 #include <linux/fs.h> 51 #include <linux/netdevice.h> 52 #include <linux/etherdevice.h> 53 #include <linux/interrupt.h> 54 #include <linux/ip.h> 55 #include <linux/ipv6.h> 56 #include <linux/page_ref.h> 57 #include <linux/pci.h> 58 #include <linux/pci_regs.h> 59 #include <linux/msi.h> 60 #include <linux/ethtool.h> 61 #include <linux/log2.h> 62 #include <linux/if_vlan.h> 63 #include <linux/random.h> 64 #include <linux/vmalloc.h> 65 #include <linux/ktime.h> 66 67 #include <net/switchdev.h> 68 #include <net/vxlan.h> 69 70 #include "nfpcore/nfp_nsp.h" 71 #include "nfp_app.h" 72 #include "nfp_net_ctrl.h" 73 #include "nfp_net.h" 74 #include "nfp_net_sriov.h" 75 #include "nfp_port.h" 76 77 /** 78 * nfp_net_get_fw_version() - Read and parse the FW version 79 * @fw_ver: Output fw_version structure to read to 80 * @ctrl_bar: Mapped address of the control BAR 81 */ 82 void nfp_net_get_fw_version(struct nfp_net_fw_version *fw_ver, 83 void __iomem *ctrl_bar) 84 { 85 u32 reg; 86 87 reg = readl(ctrl_bar + NFP_NET_CFG_VERSION); 88 put_unaligned_le32(reg, fw_ver); 89 } 90 91 static dma_addr_t nfp_net_dma_map_rx(struct nfp_net_dp *dp, void *frag) 92 { 93 return dma_map_single_attrs(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM, 94 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA, 95 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC); 96 } 97 98 static void 99 nfp_net_dma_sync_dev_rx(const struct nfp_net_dp *dp, dma_addr_t dma_addr) 100 { 101 dma_sync_single_for_device(dp->dev, dma_addr, 102 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA, 103 dp->rx_dma_dir); 104 } 105 106 static void nfp_net_dma_unmap_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr) 107 { 108 dma_unmap_single_attrs(dp->dev, dma_addr, 109 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA, 110 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC); 111 } 112 113 static void nfp_net_dma_sync_cpu_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr, 114 unsigned int len) 115 { 116 dma_sync_single_for_cpu(dp->dev, dma_addr - NFP_NET_RX_BUF_HEADROOM, 117 len, dp->rx_dma_dir); 118 } 119 120 /* Firmware reconfig 121 * 122 * Firmware reconfig may take a while so we have two versions of it - 123 * synchronous and asynchronous (posted). All synchronous callers are holding 124 * RTNL so we don't have to worry about serializing them. 125 */ 126 static void nfp_net_reconfig_start(struct nfp_net *nn, u32 update) 127 { 128 nn_writel(nn, NFP_NET_CFG_UPDATE, update); 129 /* ensure update is written before pinging HW */ 130 nn_pci_flush(nn); 131 nfp_qcp_wr_ptr_add(nn->qcp_cfg, 1); 132 } 133 134 /* Pass 0 as update to run posted reconfigs. */ 135 static void nfp_net_reconfig_start_async(struct nfp_net *nn, u32 update) 136 { 137 update |= nn->reconfig_posted; 138 nn->reconfig_posted = 0; 139 140 nfp_net_reconfig_start(nn, update); 141 142 nn->reconfig_timer_active = true; 143 mod_timer(&nn->reconfig_timer, jiffies + NFP_NET_POLL_TIMEOUT * HZ); 144 } 145 146 static bool nfp_net_reconfig_check_done(struct nfp_net *nn, bool last_check) 147 { 148 u32 reg; 149 150 reg = nn_readl(nn, NFP_NET_CFG_UPDATE); 151 if (reg == 0) 152 return true; 153 if (reg & NFP_NET_CFG_UPDATE_ERR) { 154 nn_err(nn, "Reconfig error: 0x%08x\n", reg); 155 return true; 156 } else if (last_check) { 157 nn_err(nn, "Reconfig timeout: 0x%08x\n", reg); 158 return true; 159 } 160 161 return false; 162 } 163 164 static int nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline) 165 { 166 bool timed_out = false; 167 168 /* Poll update field, waiting for NFP to ack the config */ 169 while (!nfp_net_reconfig_check_done(nn, timed_out)) { 170 msleep(1); 171 timed_out = time_is_before_eq_jiffies(deadline); 172 } 173 174 if (nn_readl(nn, NFP_NET_CFG_UPDATE) & NFP_NET_CFG_UPDATE_ERR) 175 return -EIO; 176 177 return timed_out ? -EIO : 0; 178 } 179 180 static void nfp_net_reconfig_timer(struct timer_list *t) 181 { 182 struct nfp_net *nn = from_timer(nn, t, reconfig_timer); 183 184 spin_lock_bh(&nn->reconfig_lock); 185 186 nn->reconfig_timer_active = false; 187 188 /* If sync caller is present it will take over from us */ 189 if (nn->reconfig_sync_present) 190 goto done; 191 192 /* Read reconfig status and report errors */ 193 nfp_net_reconfig_check_done(nn, true); 194 195 if (nn->reconfig_posted) 196 nfp_net_reconfig_start_async(nn, 0); 197 done: 198 spin_unlock_bh(&nn->reconfig_lock); 199 } 200 201 /** 202 * nfp_net_reconfig_post() - Post async reconfig request 203 * @nn: NFP Net device to reconfigure 204 * @update: The value for the update field in the BAR config 205 * 206 * Record FW reconfiguration request. Reconfiguration will be kicked off 207 * whenever reconfiguration machinery is idle. Multiple requests can be 208 * merged together! 209 */ 210 static void nfp_net_reconfig_post(struct nfp_net *nn, u32 update) 211 { 212 spin_lock_bh(&nn->reconfig_lock); 213 214 /* Sync caller will kick off async reconf when it's done, just post */ 215 if (nn->reconfig_sync_present) { 216 nn->reconfig_posted |= update; 217 goto done; 218 } 219 220 /* Opportunistically check if the previous command is done */ 221 if (!nn->reconfig_timer_active || 222 nfp_net_reconfig_check_done(nn, false)) 223 nfp_net_reconfig_start_async(nn, update); 224 else 225 nn->reconfig_posted |= update; 226 done: 227 spin_unlock_bh(&nn->reconfig_lock); 228 } 229 230 /** 231 * nfp_net_reconfig() - Reconfigure the firmware 232 * @nn: NFP Net device to reconfigure 233 * @update: The value for the update field in the BAR config 234 * 235 * Write the update word to the BAR and ping the reconfig queue. The 236 * poll until the firmware has acknowledged the update by zeroing the 237 * update word. 238 * 239 * Return: Negative errno on error, 0 on success 240 */ 241 int nfp_net_reconfig(struct nfp_net *nn, u32 update) 242 { 243 bool cancelled_timer = false; 244 u32 pre_posted_requests; 245 int ret; 246 247 spin_lock_bh(&nn->reconfig_lock); 248 249 nn->reconfig_sync_present = true; 250 251 if (nn->reconfig_timer_active) { 252 del_timer(&nn->reconfig_timer); 253 nn->reconfig_timer_active = false; 254 cancelled_timer = true; 255 } 256 pre_posted_requests = nn->reconfig_posted; 257 nn->reconfig_posted = 0; 258 259 spin_unlock_bh(&nn->reconfig_lock); 260 261 if (cancelled_timer) 262 nfp_net_reconfig_wait(nn, nn->reconfig_timer.expires); 263 264 /* Run the posted reconfigs which were issued before we started */ 265 if (pre_posted_requests) { 266 nfp_net_reconfig_start(nn, pre_posted_requests); 267 nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT); 268 } 269 270 nfp_net_reconfig_start(nn, update); 271 ret = nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT); 272 273 spin_lock_bh(&nn->reconfig_lock); 274 275 if (nn->reconfig_posted) 276 nfp_net_reconfig_start_async(nn, 0); 277 278 nn->reconfig_sync_present = false; 279 280 spin_unlock_bh(&nn->reconfig_lock); 281 282 return ret; 283 } 284 285 /** 286 * nfp_net_reconfig_mbox() - Reconfigure the firmware via the mailbox 287 * @nn: NFP Net device to reconfigure 288 * @mbox_cmd: The value for the mailbox command 289 * 290 * Helper function for mailbox updates 291 * 292 * Return: Negative errno on error, 0 on success 293 */ 294 static int nfp_net_reconfig_mbox(struct nfp_net *nn, u32 mbox_cmd) 295 { 296 u32 mbox = nn->tlv_caps.mbox_off; 297 int ret; 298 299 if (!nfp_net_has_mbox(&nn->tlv_caps)) { 300 nn_err(nn, "no mailbox present, command: %u\n", mbox_cmd); 301 return -EIO; 302 } 303 304 nn_writeq(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_CMD, mbox_cmd); 305 306 ret = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MBOX); 307 if (ret) { 308 nn_err(nn, "Mailbox update error\n"); 309 return ret; 310 } 311 312 return -nn_readl(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_RET); 313 } 314 315 /* Interrupt configuration and handling 316 */ 317 318 /** 319 * nfp_net_irq_unmask() - Unmask automasked interrupt 320 * @nn: NFP Network structure 321 * @entry_nr: MSI-X table entry 322 * 323 * Clear the ICR for the IRQ entry. 324 */ 325 static void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr) 326 { 327 nn_writeb(nn, NFP_NET_CFG_ICR(entry_nr), NFP_NET_CFG_ICR_UNMASKED); 328 nn_pci_flush(nn); 329 } 330 331 /** 332 * nfp_net_irqs_alloc() - allocates MSI-X irqs 333 * @pdev: PCI device structure 334 * @irq_entries: Array to be initialized and used to hold the irq entries 335 * @min_irqs: Minimal acceptable number of interrupts 336 * @wanted_irqs: Target number of interrupts to allocate 337 * 338 * Return: Number of irqs obtained or 0 on error. 339 */ 340 unsigned int 341 nfp_net_irqs_alloc(struct pci_dev *pdev, struct msix_entry *irq_entries, 342 unsigned int min_irqs, unsigned int wanted_irqs) 343 { 344 unsigned int i; 345 int got_irqs; 346 347 for (i = 0; i < wanted_irqs; i++) 348 irq_entries[i].entry = i; 349 350 got_irqs = pci_enable_msix_range(pdev, irq_entries, 351 min_irqs, wanted_irqs); 352 if (got_irqs < 0) { 353 dev_err(&pdev->dev, "Failed to enable %d-%d MSI-X (err=%d)\n", 354 min_irqs, wanted_irqs, got_irqs); 355 return 0; 356 } 357 358 if (got_irqs < wanted_irqs) 359 dev_warn(&pdev->dev, "Unable to allocate %d IRQs got only %d\n", 360 wanted_irqs, got_irqs); 361 362 return got_irqs; 363 } 364 365 /** 366 * nfp_net_irqs_assign() - Assign interrupts allocated externally to netdev 367 * @nn: NFP Network structure 368 * @irq_entries: Table of allocated interrupts 369 * @n: Size of @irq_entries (number of entries to grab) 370 * 371 * After interrupts are allocated with nfp_net_irqs_alloc() this function 372 * should be called to assign them to a specific netdev (port). 373 */ 374 void 375 nfp_net_irqs_assign(struct nfp_net *nn, struct msix_entry *irq_entries, 376 unsigned int n) 377 { 378 struct nfp_net_dp *dp = &nn->dp; 379 380 nn->max_r_vecs = n - NFP_NET_NON_Q_VECTORS; 381 dp->num_r_vecs = nn->max_r_vecs; 382 383 memcpy(nn->irq_entries, irq_entries, sizeof(*irq_entries) * n); 384 385 if (dp->num_rx_rings > dp->num_r_vecs || 386 dp->num_tx_rings > dp->num_r_vecs) 387 dev_warn(nn->dp.dev, "More rings (%d,%d) than vectors (%d).\n", 388 dp->num_rx_rings, dp->num_tx_rings, 389 dp->num_r_vecs); 390 391 dp->num_rx_rings = min(dp->num_r_vecs, dp->num_rx_rings); 392 dp->num_tx_rings = min(dp->num_r_vecs, dp->num_tx_rings); 393 dp->num_stack_tx_rings = dp->num_tx_rings; 394 } 395 396 /** 397 * nfp_net_irqs_disable() - Disable interrupts 398 * @pdev: PCI device structure 399 * 400 * Undoes what @nfp_net_irqs_alloc() does. 401 */ 402 void nfp_net_irqs_disable(struct pci_dev *pdev) 403 { 404 pci_disable_msix(pdev); 405 } 406 407 /** 408 * nfp_net_irq_rxtx() - Interrupt service routine for RX/TX rings. 409 * @irq: Interrupt 410 * @data: Opaque data structure 411 * 412 * Return: Indicate if the interrupt has been handled. 413 */ 414 static irqreturn_t nfp_net_irq_rxtx(int irq, void *data) 415 { 416 struct nfp_net_r_vector *r_vec = data; 417 418 napi_schedule_irqoff(&r_vec->napi); 419 420 /* The FW auto-masks any interrupt, either via the MASK bit in 421 * the MSI-X table or via the per entry ICR field. So there 422 * is no need to disable interrupts here. 423 */ 424 return IRQ_HANDLED; 425 } 426 427 static irqreturn_t nfp_ctrl_irq_rxtx(int irq, void *data) 428 { 429 struct nfp_net_r_vector *r_vec = data; 430 431 tasklet_schedule(&r_vec->tasklet); 432 433 return IRQ_HANDLED; 434 } 435 436 /** 437 * nfp_net_read_link_status() - Reread link status from control BAR 438 * @nn: NFP Network structure 439 */ 440 static void nfp_net_read_link_status(struct nfp_net *nn) 441 { 442 unsigned long flags; 443 bool link_up; 444 u32 sts; 445 446 spin_lock_irqsave(&nn->link_status_lock, flags); 447 448 sts = nn_readl(nn, NFP_NET_CFG_STS); 449 link_up = !!(sts & NFP_NET_CFG_STS_LINK); 450 451 if (nn->link_up == link_up) 452 goto out; 453 454 nn->link_up = link_up; 455 if (nn->port) 456 set_bit(NFP_PORT_CHANGED, &nn->port->flags); 457 458 if (nn->link_up) { 459 netif_carrier_on(nn->dp.netdev); 460 netdev_info(nn->dp.netdev, "NIC Link is Up\n"); 461 } else { 462 netif_carrier_off(nn->dp.netdev); 463 netdev_info(nn->dp.netdev, "NIC Link is Down\n"); 464 } 465 out: 466 spin_unlock_irqrestore(&nn->link_status_lock, flags); 467 } 468 469 /** 470 * nfp_net_irq_lsc() - Interrupt service routine for link state changes 471 * @irq: Interrupt 472 * @data: Opaque data structure 473 * 474 * Return: Indicate if the interrupt has been handled. 475 */ 476 static irqreturn_t nfp_net_irq_lsc(int irq, void *data) 477 { 478 struct nfp_net *nn = data; 479 struct msix_entry *entry; 480 481 entry = &nn->irq_entries[NFP_NET_IRQ_LSC_IDX]; 482 483 nfp_net_read_link_status(nn); 484 485 nfp_net_irq_unmask(nn, entry->entry); 486 487 return IRQ_HANDLED; 488 } 489 490 /** 491 * nfp_net_irq_exn() - Interrupt service routine for exceptions 492 * @irq: Interrupt 493 * @data: Opaque data structure 494 * 495 * Return: Indicate if the interrupt has been handled. 496 */ 497 static irqreturn_t nfp_net_irq_exn(int irq, void *data) 498 { 499 struct nfp_net *nn = data; 500 501 nn_err(nn, "%s: UNIMPLEMENTED.\n", __func__); 502 /* XXX TO BE IMPLEMENTED */ 503 return IRQ_HANDLED; 504 } 505 506 /** 507 * nfp_net_tx_ring_init() - Fill in the boilerplate for a TX ring 508 * @tx_ring: TX ring structure 509 * @r_vec: IRQ vector servicing this ring 510 * @idx: Ring index 511 * @is_xdp: Is this an XDP TX ring? 512 */ 513 static void 514 nfp_net_tx_ring_init(struct nfp_net_tx_ring *tx_ring, 515 struct nfp_net_r_vector *r_vec, unsigned int idx, 516 bool is_xdp) 517 { 518 struct nfp_net *nn = r_vec->nfp_net; 519 520 tx_ring->idx = idx; 521 tx_ring->r_vec = r_vec; 522 tx_ring->is_xdp = is_xdp; 523 u64_stats_init(&tx_ring->r_vec->tx_sync); 524 525 tx_ring->qcidx = tx_ring->idx * nn->stride_tx; 526 tx_ring->qcp_q = nn->tx_bar + NFP_QCP_QUEUE_OFF(tx_ring->qcidx); 527 } 528 529 /** 530 * nfp_net_rx_ring_init() - Fill in the boilerplate for a RX ring 531 * @rx_ring: RX ring structure 532 * @r_vec: IRQ vector servicing this ring 533 * @idx: Ring index 534 */ 535 static void 536 nfp_net_rx_ring_init(struct nfp_net_rx_ring *rx_ring, 537 struct nfp_net_r_vector *r_vec, unsigned int idx) 538 { 539 struct nfp_net *nn = r_vec->nfp_net; 540 541 rx_ring->idx = idx; 542 rx_ring->r_vec = r_vec; 543 u64_stats_init(&rx_ring->r_vec->rx_sync); 544 545 rx_ring->fl_qcidx = rx_ring->idx * nn->stride_rx; 546 rx_ring->qcp_fl = nn->rx_bar + NFP_QCP_QUEUE_OFF(rx_ring->fl_qcidx); 547 } 548 549 /** 550 * nfp_net_aux_irq_request() - Request an auxiliary interrupt (LSC or EXN) 551 * @nn: NFP Network structure 552 * @ctrl_offset: Control BAR offset where IRQ configuration should be written 553 * @format: printf-style format to construct the interrupt name 554 * @name: Pointer to allocated space for interrupt name 555 * @name_sz: Size of space for interrupt name 556 * @vector_idx: Index of MSI-X vector used for this interrupt 557 * @handler: IRQ handler to register for this interrupt 558 */ 559 static int 560 nfp_net_aux_irq_request(struct nfp_net *nn, u32 ctrl_offset, 561 const char *format, char *name, size_t name_sz, 562 unsigned int vector_idx, irq_handler_t handler) 563 { 564 struct msix_entry *entry; 565 int err; 566 567 entry = &nn->irq_entries[vector_idx]; 568 569 snprintf(name, name_sz, format, nfp_net_name(nn)); 570 err = request_irq(entry->vector, handler, 0, name, nn); 571 if (err) { 572 nn_err(nn, "Failed to request IRQ %d (err=%d).\n", 573 entry->vector, err); 574 return err; 575 } 576 nn_writeb(nn, ctrl_offset, entry->entry); 577 nfp_net_irq_unmask(nn, entry->entry); 578 579 return 0; 580 } 581 582 /** 583 * nfp_net_aux_irq_free() - Free an auxiliary interrupt (LSC or EXN) 584 * @nn: NFP Network structure 585 * @ctrl_offset: Control BAR offset where IRQ configuration should be written 586 * @vector_idx: Index of MSI-X vector used for this interrupt 587 */ 588 static void nfp_net_aux_irq_free(struct nfp_net *nn, u32 ctrl_offset, 589 unsigned int vector_idx) 590 { 591 nn_writeb(nn, ctrl_offset, 0xff); 592 nn_pci_flush(nn); 593 free_irq(nn->irq_entries[vector_idx].vector, nn); 594 } 595 596 /* Transmit 597 * 598 * One queue controller peripheral queue is used for transmit. The 599 * driver en-queues packets for transmit by advancing the write 600 * pointer. The device indicates that packets have transmitted by 601 * advancing the read pointer. The driver maintains a local copy of 602 * the read and write pointer in @struct nfp_net_tx_ring. The driver 603 * keeps @wr_p in sync with the queue controller write pointer and can 604 * determine how many packets have been transmitted by comparing its 605 * copy of the read pointer @rd_p with the read pointer maintained by 606 * the queue controller peripheral. 607 */ 608 609 /** 610 * nfp_net_tx_full() - Check if the TX ring is full 611 * @tx_ring: TX ring to check 612 * @dcnt: Number of descriptors that need to be enqueued (must be >= 1) 613 * 614 * This function checks, based on the *host copy* of read/write 615 * pointer if a given TX ring is full. The real TX queue may have 616 * some newly made available slots. 617 * 618 * Return: True if the ring is full. 619 */ 620 static int nfp_net_tx_full(struct nfp_net_tx_ring *tx_ring, int dcnt) 621 { 622 return (tx_ring->wr_p - tx_ring->rd_p) >= (tx_ring->cnt - dcnt); 623 } 624 625 /* Wrappers for deciding when to stop and restart TX queues */ 626 static int nfp_net_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring) 627 { 628 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4); 629 } 630 631 static int nfp_net_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring) 632 { 633 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1); 634 } 635 636 /** 637 * nfp_net_tx_ring_stop() - stop tx ring 638 * @nd_q: netdev queue 639 * @tx_ring: driver tx queue structure 640 * 641 * Safely stop TX ring. Remember that while we are running .start_xmit() 642 * someone else may be cleaning the TX ring completions so we need to be 643 * extra careful here. 644 */ 645 static void nfp_net_tx_ring_stop(struct netdev_queue *nd_q, 646 struct nfp_net_tx_ring *tx_ring) 647 { 648 netif_tx_stop_queue(nd_q); 649 650 /* We can race with the TX completion out of NAPI so recheck */ 651 smp_mb(); 652 if (unlikely(nfp_net_tx_ring_should_wake(tx_ring))) 653 netif_tx_start_queue(nd_q); 654 } 655 656 /** 657 * nfp_net_tx_tso() - Set up Tx descriptor for LSO 658 * @r_vec: per-ring structure 659 * @txbuf: Pointer to driver soft TX descriptor 660 * @txd: Pointer to HW TX descriptor 661 * @skb: Pointer to SKB 662 * 663 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs. 664 * Return error on packet header greater than maximum supported LSO header size. 665 */ 666 static void nfp_net_tx_tso(struct nfp_net_r_vector *r_vec, 667 struct nfp_net_tx_buf *txbuf, 668 struct nfp_net_tx_desc *txd, struct sk_buff *skb) 669 { 670 u32 hdrlen; 671 u16 mss; 672 673 if (!skb_is_gso(skb)) 674 return; 675 676 if (!skb->encapsulation) { 677 txd->l3_offset = skb_network_offset(skb); 678 txd->l4_offset = skb_transport_offset(skb); 679 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb); 680 } else { 681 txd->l3_offset = skb_inner_network_offset(skb); 682 txd->l4_offset = skb_inner_transport_offset(skb); 683 hdrlen = skb_inner_transport_header(skb) - skb->data + 684 inner_tcp_hdrlen(skb); 685 } 686 687 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs; 688 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1); 689 690 mss = skb_shinfo(skb)->gso_size & PCIE_DESC_TX_MSS_MASK; 691 txd->lso_hdrlen = hdrlen; 692 txd->mss = cpu_to_le16(mss); 693 txd->flags |= PCIE_DESC_TX_LSO; 694 695 u64_stats_update_begin(&r_vec->tx_sync); 696 r_vec->tx_lso++; 697 u64_stats_update_end(&r_vec->tx_sync); 698 } 699 700 /** 701 * nfp_net_tx_csum() - Set TX CSUM offload flags in TX descriptor 702 * @dp: NFP Net data path struct 703 * @r_vec: per-ring structure 704 * @txbuf: Pointer to driver soft TX descriptor 705 * @txd: Pointer to TX descriptor 706 * @skb: Pointer to SKB 707 * 708 * This function sets the TX checksum flags in the TX descriptor based 709 * on the configuration and the protocol of the packet to be transmitted. 710 */ 711 static void nfp_net_tx_csum(struct nfp_net_dp *dp, 712 struct nfp_net_r_vector *r_vec, 713 struct nfp_net_tx_buf *txbuf, 714 struct nfp_net_tx_desc *txd, struct sk_buff *skb) 715 { 716 struct ipv6hdr *ipv6h; 717 struct iphdr *iph; 718 u8 l4_hdr; 719 720 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM)) 721 return; 722 723 if (skb->ip_summed != CHECKSUM_PARTIAL) 724 return; 725 726 txd->flags |= PCIE_DESC_TX_CSUM; 727 if (skb->encapsulation) 728 txd->flags |= PCIE_DESC_TX_ENCAP; 729 730 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb); 731 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb); 732 733 if (iph->version == 4) { 734 txd->flags |= PCIE_DESC_TX_IP4_CSUM; 735 l4_hdr = iph->protocol; 736 } else if (ipv6h->version == 6) { 737 l4_hdr = ipv6h->nexthdr; 738 } else { 739 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version); 740 return; 741 } 742 743 switch (l4_hdr) { 744 case IPPROTO_TCP: 745 txd->flags |= PCIE_DESC_TX_TCP_CSUM; 746 break; 747 case IPPROTO_UDP: 748 txd->flags |= PCIE_DESC_TX_UDP_CSUM; 749 break; 750 default: 751 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr); 752 return; 753 } 754 755 u64_stats_update_begin(&r_vec->tx_sync); 756 if (skb->encapsulation) 757 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt; 758 else 759 r_vec->hw_csum_tx += txbuf->pkt_cnt; 760 u64_stats_update_end(&r_vec->tx_sync); 761 } 762 763 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring) 764 { 765 wmb(); 766 nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add); 767 tx_ring->wr_ptr_add = 0; 768 } 769 770 static int nfp_net_prep_port_id(struct sk_buff *skb) 771 { 772 struct metadata_dst *md_dst = skb_metadata_dst(skb); 773 unsigned char *data; 774 775 if (likely(!md_dst)) 776 return 0; 777 if (unlikely(md_dst->type != METADATA_HW_PORT_MUX)) 778 return 0; 779 780 if (unlikely(skb_cow_head(skb, 8))) 781 return -ENOMEM; 782 783 data = skb_push(skb, 8); 784 put_unaligned_be32(NFP_NET_META_PORTID, data); 785 put_unaligned_be32(md_dst->u.port_info.port_id, data + 4); 786 787 return 8; 788 } 789 790 /** 791 * nfp_net_tx() - Main transmit entry point 792 * @skb: SKB to transmit 793 * @netdev: netdev structure 794 * 795 * Return: NETDEV_TX_OK on success. 796 */ 797 static int nfp_net_tx(struct sk_buff *skb, struct net_device *netdev) 798 { 799 struct nfp_net *nn = netdev_priv(netdev); 800 const struct skb_frag_struct *frag; 801 struct nfp_net_tx_desc *txd, txdg; 802 int f, nr_frags, wr_idx, md_bytes; 803 struct nfp_net_tx_ring *tx_ring; 804 struct nfp_net_r_vector *r_vec; 805 struct nfp_net_tx_buf *txbuf; 806 struct netdev_queue *nd_q; 807 struct nfp_net_dp *dp; 808 dma_addr_t dma_addr; 809 unsigned int fsize; 810 u16 qidx; 811 812 dp = &nn->dp; 813 qidx = skb_get_queue_mapping(skb); 814 tx_ring = &dp->tx_rings[qidx]; 815 r_vec = tx_ring->r_vec; 816 nd_q = netdev_get_tx_queue(dp->netdev, qidx); 817 818 nr_frags = skb_shinfo(skb)->nr_frags; 819 820 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) { 821 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n", 822 qidx, tx_ring->wr_p, tx_ring->rd_p); 823 netif_tx_stop_queue(nd_q); 824 nfp_net_tx_xmit_more_flush(tx_ring); 825 u64_stats_update_begin(&r_vec->tx_sync); 826 r_vec->tx_busy++; 827 u64_stats_update_end(&r_vec->tx_sync); 828 return NETDEV_TX_BUSY; 829 } 830 831 md_bytes = nfp_net_prep_port_id(skb); 832 if (unlikely(md_bytes < 0)) { 833 nfp_net_tx_xmit_more_flush(tx_ring); 834 dev_kfree_skb_any(skb); 835 return NETDEV_TX_OK; 836 } 837 838 /* Start with the head skbuf */ 839 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 840 DMA_TO_DEVICE); 841 if (dma_mapping_error(dp->dev, dma_addr)) 842 goto err_free; 843 844 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 845 846 /* Stash the soft descriptor of the head then initialize it */ 847 txbuf = &tx_ring->txbufs[wr_idx]; 848 txbuf->skb = skb; 849 txbuf->dma_addr = dma_addr; 850 txbuf->fidx = -1; 851 txbuf->pkt_cnt = 1; 852 txbuf->real_len = skb->len; 853 854 /* Build TX descriptor */ 855 txd = &tx_ring->txds[wr_idx]; 856 txd->offset_eop = (nr_frags ? 0 : PCIE_DESC_TX_EOP) | md_bytes; 857 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 858 nfp_desc_set_dma_addr(txd, dma_addr); 859 txd->data_len = cpu_to_le16(skb->len); 860 861 txd->flags = 0; 862 txd->mss = 0; 863 txd->lso_hdrlen = 0; 864 865 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */ 866 nfp_net_tx_tso(r_vec, txbuf, txd, skb); 867 nfp_net_tx_csum(dp, r_vec, txbuf, txd, skb); 868 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) { 869 txd->flags |= PCIE_DESC_TX_VLAN; 870 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 871 } 872 873 /* Gather DMA */ 874 if (nr_frags > 0) { 875 /* all descs must match except for in addr, length and eop */ 876 txdg = *txd; 877 878 for (f = 0; f < nr_frags; f++) { 879 frag = &skb_shinfo(skb)->frags[f]; 880 fsize = skb_frag_size(frag); 881 882 dma_addr = skb_frag_dma_map(dp->dev, frag, 0, 883 fsize, DMA_TO_DEVICE); 884 if (dma_mapping_error(dp->dev, dma_addr)) 885 goto err_unmap; 886 887 wr_idx = D_IDX(tx_ring, wr_idx + 1); 888 tx_ring->txbufs[wr_idx].skb = skb; 889 tx_ring->txbufs[wr_idx].dma_addr = dma_addr; 890 tx_ring->txbufs[wr_idx].fidx = f; 891 892 txd = &tx_ring->txds[wr_idx]; 893 *txd = txdg; 894 txd->dma_len = cpu_to_le16(fsize); 895 nfp_desc_set_dma_addr(txd, dma_addr); 896 txd->offset_eop |= 897 (f == nr_frags - 1) ? PCIE_DESC_TX_EOP : 0; 898 } 899 900 u64_stats_update_begin(&r_vec->tx_sync); 901 r_vec->tx_gather++; 902 u64_stats_update_end(&r_vec->tx_sync); 903 } 904 905 netdev_tx_sent_queue(nd_q, txbuf->real_len); 906 907 skb_tx_timestamp(skb); 908 909 tx_ring->wr_p += nr_frags + 1; 910 if (nfp_net_tx_ring_should_stop(tx_ring)) 911 nfp_net_tx_ring_stop(nd_q, tx_ring); 912 913 tx_ring->wr_ptr_add += nr_frags + 1; 914 if (!skb->xmit_more || netif_xmit_stopped(nd_q)) 915 nfp_net_tx_xmit_more_flush(tx_ring); 916 917 return NETDEV_TX_OK; 918 919 err_unmap: 920 while (--f >= 0) { 921 frag = &skb_shinfo(skb)->frags[f]; 922 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 923 skb_frag_size(frag), DMA_TO_DEVICE); 924 tx_ring->txbufs[wr_idx].skb = NULL; 925 tx_ring->txbufs[wr_idx].dma_addr = 0; 926 tx_ring->txbufs[wr_idx].fidx = -2; 927 wr_idx = wr_idx - 1; 928 if (wr_idx < 0) 929 wr_idx += tx_ring->cnt; 930 } 931 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 932 skb_headlen(skb), DMA_TO_DEVICE); 933 tx_ring->txbufs[wr_idx].skb = NULL; 934 tx_ring->txbufs[wr_idx].dma_addr = 0; 935 tx_ring->txbufs[wr_idx].fidx = -2; 936 err_free: 937 nn_dp_warn(dp, "Failed to map DMA TX buffer\n"); 938 nfp_net_tx_xmit_more_flush(tx_ring); 939 u64_stats_update_begin(&r_vec->tx_sync); 940 r_vec->tx_errors++; 941 u64_stats_update_end(&r_vec->tx_sync); 942 dev_kfree_skb_any(skb); 943 return NETDEV_TX_OK; 944 } 945 946 /** 947 * nfp_net_tx_complete() - Handled completed TX packets 948 * @tx_ring: TX ring structure 949 * @budget: NAPI budget (only used as bool to determine if in NAPI context) 950 * 951 * Return: Number of completed TX descriptors 952 */ 953 static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget) 954 { 955 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 956 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 957 const struct skb_frag_struct *frag; 958 struct netdev_queue *nd_q; 959 u32 done_pkts = 0, done_bytes = 0; 960 struct sk_buff *skb; 961 int todo, nr_frags; 962 u32 qcp_rd_p; 963 int fidx; 964 int idx; 965 966 if (tx_ring->wr_p == tx_ring->rd_p) 967 return; 968 969 /* Work out how many descriptors have been transmitted */ 970 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q); 971 972 if (qcp_rd_p == tx_ring->qcp_rd_p) 973 return; 974 975 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 976 977 while (todo--) { 978 idx = D_IDX(tx_ring, tx_ring->rd_p++); 979 980 skb = tx_ring->txbufs[idx].skb; 981 if (!skb) 982 continue; 983 984 nr_frags = skb_shinfo(skb)->nr_frags; 985 fidx = tx_ring->txbufs[idx].fidx; 986 987 if (fidx == -1) { 988 /* unmap head */ 989 dma_unmap_single(dp->dev, tx_ring->txbufs[idx].dma_addr, 990 skb_headlen(skb), DMA_TO_DEVICE); 991 992 done_pkts += tx_ring->txbufs[idx].pkt_cnt; 993 done_bytes += tx_ring->txbufs[idx].real_len; 994 } else { 995 /* unmap fragment */ 996 frag = &skb_shinfo(skb)->frags[fidx]; 997 dma_unmap_page(dp->dev, tx_ring->txbufs[idx].dma_addr, 998 skb_frag_size(frag), DMA_TO_DEVICE); 999 } 1000 1001 /* check for last gather fragment */ 1002 if (fidx == nr_frags - 1) 1003 napi_consume_skb(skb, budget); 1004 1005 tx_ring->txbufs[idx].dma_addr = 0; 1006 tx_ring->txbufs[idx].skb = NULL; 1007 tx_ring->txbufs[idx].fidx = -2; 1008 } 1009 1010 tx_ring->qcp_rd_p = qcp_rd_p; 1011 1012 u64_stats_update_begin(&r_vec->tx_sync); 1013 r_vec->tx_bytes += done_bytes; 1014 r_vec->tx_pkts += done_pkts; 1015 u64_stats_update_end(&r_vec->tx_sync); 1016 1017 if (!dp->netdev) 1018 return; 1019 1020 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 1021 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes); 1022 if (nfp_net_tx_ring_should_wake(tx_ring)) { 1023 /* Make sure TX thread will see updated tx_ring->rd_p */ 1024 smp_mb(); 1025 1026 if (unlikely(netif_tx_queue_stopped(nd_q))) 1027 netif_tx_wake_queue(nd_q); 1028 } 1029 1030 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 1031 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 1032 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 1033 } 1034 1035 static bool nfp_net_xdp_complete(struct nfp_net_tx_ring *tx_ring) 1036 { 1037 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 1038 u32 done_pkts = 0, done_bytes = 0; 1039 bool done_all; 1040 int idx, todo; 1041 u32 qcp_rd_p; 1042 1043 /* Work out how many descriptors have been transmitted */ 1044 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q); 1045 1046 if (qcp_rd_p == tx_ring->qcp_rd_p) 1047 return true; 1048 1049 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 1050 1051 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE; 1052 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE); 1053 1054 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo); 1055 1056 done_pkts = todo; 1057 while (todo--) { 1058 idx = D_IDX(tx_ring, tx_ring->rd_p); 1059 tx_ring->rd_p++; 1060 1061 done_bytes += tx_ring->txbufs[idx].real_len; 1062 } 1063 1064 u64_stats_update_begin(&r_vec->tx_sync); 1065 r_vec->tx_bytes += done_bytes; 1066 r_vec->tx_pkts += done_pkts; 1067 u64_stats_update_end(&r_vec->tx_sync); 1068 1069 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 1070 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 1071 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 1072 1073 return done_all; 1074 } 1075 1076 /** 1077 * nfp_net_tx_ring_reset() - Free any untransmitted buffers and reset pointers 1078 * @dp: NFP Net data path struct 1079 * @tx_ring: TX ring structure 1080 * 1081 * Assumes that the device is stopped 1082 */ 1083 static void 1084 nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring) 1085 { 1086 const struct skb_frag_struct *frag; 1087 struct netdev_queue *nd_q; 1088 1089 while (!tx_ring->is_xdp && tx_ring->rd_p != tx_ring->wr_p) { 1090 struct nfp_net_tx_buf *tx_buf; 1091 struct sk_buff *skb; 1092 int idx, nr_frags; 1093 1094 idx = D_IDX(tx_ring, tx_ring->rd_p); 1095 tx_buf = &tx_ring->txbufs[idx]; 1096 1097 skb = tx_ring->txbufs[idx].skb; 1098 nr_frags = skb_shinfo(skb)->nr_frags; 1099 1100 if (tx_buf->fidx == -1) { 1101 /* unmap head */ 1102 dma_unmap_single(dp->dev, tx_buf->dma_addr, 1103 skb_headlen(skb), DMA_TO_DEVICE); 1104 } else { 1105 /* unmap fragment */ 1106 frag = &skb_shinfo(skb)->frags[tx_buf->fidx]; 1107 dma_unmap_page(dp->dev, tx_buf->dma_addr, 1108 skb_frag_size(frag), DMA_TO_DEVICE); 1109 } 1110 1111 /* check for last gather fragment */ 1112 if (tx_buf->fidx == nr_frags - 1) 1113 dev_kfree_skb_any(skb); 1114 1115 tx_buf->dma_addr = 0; 1116 tx_buf->skb = NULL; 1117 tx_buf->fidx = -2; 1118 1119 tx_ring->qcp_rd_p++; 1120 tx_ring->rd_p++; 1121 } 1122 1123 memset(tx_ring->txds, 0, sizeof(*tx_ring->txds) * tx_ring->cnt); 1124 tx_ring->wr_p = 0; 1125 tx_ring->rd_p = 0; 1126 tx_ring->qcp_rd_p = 0; 1127 tx_ring->wr_ptr_add = 0; 1128 1129 if (tx_ring->is_xdp || !dp->netdev) 1130 return; 1131 1132 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 1133 netdev_tx_reset_queue(nd_q); 1134 } 1135 1136 static void nfp_net_tx_timeout(struct net_device *netdev) 1137 { 1138 struct nfp_net *nn = netdev_priv(netdev); 1139 int i; 1140 1141 for (i = 0; i < nn->dp.netdev->real_num_tx_queues; i++) { 1142 if (!netif_tx_queue_stopped(netdev_get_tx_queue(netdev, i))) 1143 continue; 1144 nn_warn(nn, "TX timeout on ring: %d\n", i); 1145 } 1146 nn_warn(nn, "TX watchdog timeout\n"); 1147 } 1148 1149 /* Receive processing 1150 */ 1151 static unsigned int 1152 nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp) 1153 { 1154 unsigned int fl_bufsz; 1155 1156 fl_bufsz = NFP_NET_RX_BUF_HEADROOM; 1157 fl_bufsz += dp->rx_dma_off; 1158 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1159 fl_bufsz += NFP_NET_MAX_PREPEND; 1160 else 1161 fl_bufsz += dp->rx_offset; 1162 fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu; 1163 1164 fl_bufsz = SKB_DATA_ALIGN(fl_bufsz); 1165 fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1166 1167 return fl_bufsz; 1168 } 1169 1170 static void 1171 nfp_net_free_frag(void *frag, bool xdp) 1172 { 1173 if (!xdp) 1174 skb_free_frag(frag); 1175 else 1176 __free_page(virt_to_page(frag)); 1177 } 1178 1179 /** 1180 * nfp_net_rx_alloc_one() - Allocate and map page frag for RX 1181 * @dp: NFP Net data path struct 1182 * @dma_addr: Pointer to storage for DMA address (output param) 1183 * 1184 * This function will allcate a new page frag, map it for DMA. 1185 * 1186 * Return: allocated page frag or NULL on failure. 1187 */ 1188 static void *nfp_net_rx_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 1189 { 1190 void *frag; 1191 1192 if (!dp->xdp_prog) { 1193 frag = netdev_alloc_frag(dp->fl_bufsz); 1194 } else { 1195 struct page *page; 1196 1197 page = alloc_page(GFP_KERNEL); 1198 frag = page ? page_address(page) : NULL; 1199 } 1200 if (!frag) { 1201 nn_dp_warn(dp, "Failed to alloc receive page frag\n"); 1202 return NULL; 1203 } 1204 1205 *dma_addr = nfp_net_dma_map_rx(dp, frag); 1206 if (dma_mapping_error(dp->dev, *dma_addr)) { 1207 nfp_net_free_frag(frag, dp->xdp_prog); 1208 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 1209 return NULL; 1210 } 1211 1212 return frag; 1213 } 1214 1215 static void *nfp_net_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 1216 { 1217 void *frag; 1218 1219 if (!dp->xdp_prog) { 1220 frag = napi_alloc_frag(dp->fl_bufsz); 1221 if (unlikely(!frag)) 1222 return NULL; 1223 } else { 1224 struct page *page; 1225 1226 page = dev_alloc_page(); 1227 if (unlikely(!page)) 1228 return NULL; 1229 frag = page_address(page); 1230 } 1231 1232 *dma_addr = nfp_net_dma_map_rx(dp, frag); 1233 if (dma_mapping_error(dp->dev, *dma_addr)) { 1234 nfp_net_free_frag(frag, dp->xdp_prog); 1235 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 1236 return NULL; 1237 } 1238 1239 return frag; 1240 } 1241 1242 /** 1243 * nfp_net_rx_give_one() - Put mapped skb on the software and hardware rings 1244 * @dp: NFP Net data path struct 1245 * @rx_ring: RX ring structure 1246 * @frag: page fragment buffer 1247 * @dma_addr: DMA address of skb mapping 1248 */ 1249 static void nfp_net_rx_give_one(const struct nfp_net_dp *dp, 1250 struct nfp_net_rx_ring *rx_ring, 1251 void *frag, dma_addr_t dma_addr) 1252 { 1253 unsigned int wr_idx; 1254 1255 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 1256 1257 nfp_net_dma_sync_dev_rx(dp, dma_addr); 1258 1259 /* Stash SKB and DMA address away */ 1260 rx_ring->rxbufs[wr_idx].frag = frag; 1261 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr; 1262 1263 /* Fill freelist descriptor */ 1264 rx_ring->rxds[wr_idx].fld.reserved = 0; 1265 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0; 1266 nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld, 1267 dma_addr + dp->rx_dma_off); 1268 1269 rx_ring->wr_p++; 1270 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) { 1271 /* Update write pointer of the freelist queue. Make 1272 * sure all writes are flushed before telling the hardware. 1273 */ 1274 wmb(); 1275 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH); 1276 } 1277 } 1278 1279 /** 1280 * nfp_net_rx_ring_reset() - Reflect in SW state of freelist after disable 1281 * @rx_ring: RX ring structure 1282 * 1283 * Warning: Do *not* call if ring buffers were never put on the FW freelist 1284 * (i.e. device was not enabled)! 1285 */ 1286 static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring) 1287 { 1288 unsigned int wr_idx, last_idx; 1289 1290 /* Move the empty entry to the end of the list */ 1291 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 1292 last_idx = rx_ring->cnt - 1; 1293 rx_ring->rxbufs[wr_idx].dma_addr = rx_ring->rxbufs[last_idx].dma_addr; 1294 rx_ring->rxbufs[wr_idx].frag = rx_ring->rxbufs[last_idx].frag; 1295 rx_ring->rxbufs[last_idx].dma_addr = 0; 1296 rx_ring->rxbufs[last_idx].frag = NULL; 1297 1298 memset(rx_ring->rxds, 0, sizeof(*rx_ring->rxds) * rx_ring->cnt); 1299 rx_ring->wr_p = 0; 1300 rx_ring->rd_p = 0; 1301 } 1302 1303 /** 1304 * nfp_net_rx_ring_bufs_free() - Free any buffers currently on the RX ring 1305 * @dp: NFP Net data path struct 1306 * @rx_ring: RX ring to remove buffers from 1307 * 1308 * Assumes that the device is stopped and buffers are in [0, ring->cnt - 1) 1309 * entries. After device is disabled nfp_net_rx_ring_reset() must be called 1310 * to restore required ring geometry. 1311 */ 1312 static void 1313 nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp, 1314 struct nfp_net_rx_ring *rx_ring) 1315 { 1316 unsigned int i; 1317 1318 for (i = 0; i < rx_ring->cnt - 1; i++) { 1319 /* NULL skb can only happen when initial filling of the ring 1320 * fails to allocate enough buffers and calls here to free 1321 * already allocated ones. 1322 */ 1323 if (!rx_ring->rxbufs[i].frag) 1324 continue; 1325 1326 nfp_net_dma_unmap_rx(dp, rx_ring->rxbufs[i].dma_addr); 1327 nfp_net_free_frag(rx_ring->rxbufs[i].frag, dp->xdp_prog); 1328 rx_ring->rxbufs[i].dma_addr = 0; 1329 rx_ring->rxbufs[i].frag = NULL; 1330 } 1331 } 1332 1333 /** 1334 * nfp_net_rx_ring_bufs_alloc() - Fill RX ring with buffers (don't give to FW) 1335 * @dp: NFP Net data path struct 1336 * @rx_ring: RX ring to remove buffers from 1337 */ 1338 static int 1339 nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp, 1340 struct nfp_net_rx_ring *rx_ring) 1341 { 1342 struct nfp_net_rx_buf *rxbufs; 1343 unsigned int i; 1344 1345 rxbufs = rx_ring->rxbufs; 1346 1347 for (i = 0; i < rx_ring->cnt - 1; i++) { 1348 rxbufs[i].frag = nfp_net_rx_alloc_one(dp, &rxbufs[i].dma_addr); 1349 if (!rxbufs[i].frag) { 1350 nfp_net_rx_ring_bufs_free(dp, rx_ring); 1351 return -ENOMEM; 1352 } 1353 } 1354 1355 return 0; 1356 } 1357 1358 /** 1359 * nfp_net_rx_ring_fill_freelist() - Give buffers from the ring to FW 1360 * @dp: NFP Net data path struct 1361 * @rx_ring: RX ring to fill 1362 */ 1363 static void 1364 nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp, 1365 struct nfp_net_rx_ring *rx_ring) 1366 { 1367 unsigned int i; 1368 1369 for (i = 0; i < rx_ring->cnt - 1; i++) 1370 nfp_net_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag, 1371 rx_ring->rxbufs[i].dma_addr); 1372 } 1373 1374 /** 1375 * nfp_net_rx_csum_has_errors() - group check if rxd has any csum errors 1376 * @flags: RX descriptor flags field in CPU byte order 1377 */ 1378 static int nfp_net_rx_csum_has_errors(u16 flags) 1379 { 1380 u16 csum_all_checked, csum_all_ok; 1381 1382 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL; 1383 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK; 1384 1385 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT); 1386 } 1387 1388 /** 1389 * nfp_net_rx_csum() - set SKB checksum field based on RX descriptor flags 1390 * @dp: NFP Net data path struct 1391 * @r_vec: per-ring structure 1392 * @rxd: Pointer to RX descriptor 1393 * @meta: Parsed metadata prepend 1394 * @skb: Pointer to SKB 1395 */ 1396 static void nfp_net_rx_csum(struct nfp_net_dp *dp, 1397 struct nfp_net_r_vector *r_vec, 1398 struct nfp_net_rx_desc *rxd, 1399 struct nfp_meta_parsed *meta, struct sk_buff *skb) 1400 { 1401 skb_checksum_none_assert(skb); 1402 1403 if (!(dp->netdev->features & NETIF_F_RXCSUM)) 1404 return; 1405 1406 if (meta->csum_type) { 1407 skb->ip_summed = meta->csum_type; 1408 skb->csum = meta->csum; 1409 u64_stats_update_begin(&r_vec->rx_sync); 1410 r_vec->hw_csum_rx_complete++; 1411 u64_stats_update_end(&r_vec->rx_sync); 1412 return; 1413 } 1414 1415 if (nfp_net_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) { 1416 u64_stats_update_begin(&r_vec->rx_sync); 1417 r_vec->hw_csum_rx_error++; 1418 u64_stats_update_end(&r_vec->rx_sync); 1419 return; 1420 } 1421 1422 /* Assume that the firmware will never report inner CSUM_OK unless outer 1423 * L4 headers were successfully parsed. FW will always report zero UDP 1424 * checksum as CSUM_OK. 1425 */ 1426 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK || 1427 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) { 1428 __skb_incr_checksum_unnecessary(skb); 1429 u64_stats_update_begin(&r_vec->rx_sync); 1430 r_vec->hw_csum_rx_ok++; 1431 u64_stats_update_end(&r_vec->rx_sync); 1432 } 1433 1434 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK || 1435 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) { 1436 __skb_incr_checksum_unnecessary(skb); 1437 u64_stats_update_begin(&r_vec->rx_sync); 1438 r_vec->hw_csum_rx_inner_ok++; 1439 u64_stats_update_end(&r_vec->rx_sync); 1440 } 1441 } 1442 1443 static void 1444 nfp_net_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta, 1445 unsigned int type, __be32 *hash) 1446 { 1447 if (!(netdev->features & NETIF_F_RXHASH)) 1448 return; 1449 1450 switch (type) { 1451 case NFP_NET_RSS_IPV4: 1452 case NFP_NET_RSS_IPV6: 1453 case NFP_NET_RSS_IPV6_EX: 1454 meta->hash_type = PKT_HASH_TYPE_L3; 1455 break; 1456 default: 1457 meta->hash_type = PKT_HASH_TYPE_L4; 1458 break; 1459 } 1460 1461 meta->hash = get_unaligned_be32(hash); 1462 } 1463 1464 static void 1465 nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta, 1466 void *data, struct nfp_net_rx_desc *rxd) 1467 { 1468 struct nfp_net_rx_hash *rx_hash = data; 1469 1470 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS)) 1471 return; 1472 1473 nfp_net_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type), 1474 &rx_hash->hash); 1475 } 1476 1477 static void * 1478 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta, 1479 void *data, int meta_len) 1480 { 1481 u32 meta_info; 1482 1483 meta_info = get_unaligned_be32(data); 1484 data += 4; 1485 1486 while (meta_info) { 1487 switch (meta_info & NFP_NET_META_FIELD_MASK) { 1488 case NFP_NET_META_HASH: 1489 meta_info >>= NFP_NET_META_FIELD_SIZE; 1490 nfp_net_set_hash(netdev, meta, 1491 meta_info & NFP_NET_META_FIELD_MASK, 1492 (__be32 *)data); 1493 data += 4; 1494 break; 1495 case NFP_NET_META_MARK: 1496 meta->mark = get_unaligned_be32(data); 1497 data += 4; 1498 break; 1499 case NFP_NET_META_PORTID: 1500 meta->portid = get_unaligned_be32(data); 1501 data += 4; 1502 break; 1503 case NFP_NET_META_CSUM: 1504 meta->csum_type = CHECKSUM_COMPLETE; 1505 meta->csum = 1506 (__force __wsum)__get_unaligned_cpu32(data); 1507 data += 4; 1508 break; 1509 default: 1510 return NULL; 1511 } 1512 1513 meta_info >>= NFP_NET_META_FIELD_SIZE; 1514 } 1515 1516 return data; 1517 } 1518 1519 static void 1520 nfp_net_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, 1521 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf, 1522 struct sk_buff *skb) 1523 { 1524 u64_stats_update_begin(&r_vec->rx_sync); 1525 r_vec->rx_drops++; 1526 /* If we have both skb and rxbuf the replacement buffer allocation 1527 * must have failed, count this as an alloc failure. 1528 */ 1529 if (skb && rxbuf) 1530 r_vec->rx_replace_buf_alloc_fail++; 1531 u64_stats_update_end(&r_vec->rx_sync); 1532 1533 /* skb is build based on the frag, free_skb() would free the frag 1534 * so to be able to reuse it we need an extra ref. 1535 */ 1536 if (skb && rxbuf && skb->head == rxbuf->frag) 1537 page_ref_inc(virt_to_head_page(rxbuf->frag)); 1538 if (rxbuf) 1539 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr); 1540 if (skb) 1541 dev_kfree_skb_any(skb); 1542 } 1543 1544 static bool 1545 nfp_net_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring, 1546 struct nfp_net_tx_ring *tx_ring, 1547 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off, 1548 unsigned int pkt_len, bool *completed) 1549 { 1550 struct nfp_net_tx_buf *txbuf; 1551 struct nfp_net_tx_desc *txd; 1552 int wr_idx; 1553 1554 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1555 if (!*completed) { 1556 nfp_net_xdp_complete(tx_ring); 1557 *completed = true; 1558 } 1559 1560 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1561 nfp_net_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf, 1562 NULL); 1563 return false; 1564 } 1565 } 1566 1567 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1568 1569 /* Stash the soft descriptor of the head then initialize it */ 1570 txbuf = &tx_ring->txbufs[wr_idx]; 1571 1572 nfp_net_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr); 1573 1574 txbuf->frag = rxbuf->frag; 1575 txbuf->dma_addr = rxbuf->dma_addr; 1576 txbuf->fidx = -1; 1577 txbuf->pkt_cnt = 1; 1578 txbuf->real_len = pkt_len; 1579 1580 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off, 1581 pkt_len, DMA_BIDIRECTIONAL); 1582 1583 /* Build TX descriptor */ 1584 txd = &tx_ring->txds[wr_idx]; 1585 txd->offset_eop = PCIE_DESC_TX_EOP; 1586 txd->dma_len = cpu_to_le16(pkt_len); 1587 nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off); 1588 txd->data_len = cpu_to_le16(pkt_len); 1589 1590 txd->flags = 0; 1591 txd->mss = 0; 1592 txd->lso_hdrlen = 0; 1593 1594 tx_ring->wr_p++; 1595 tx_ring->wr_ptr_add++; 1596 return true; 1597 } 1598 1599 /** 1600 * nfp_net_rx() - receive up to @budget packets on @rx_ring 1601 * @rx_ring: RX ring to receive from 1602 * @budget: NAPI budget 1603 * 1604 * Note, this function is separated out from the napi poll function to 1605 * more cleanly separate packet receive code from other bookkeeping 1606 * functions performed in the napi poll function. 1607 * 1608 * Return: Number of packets received. 1609 */ 1610 static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget) 1611 { 1612 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 1613 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 1614 struct nfp_net_tx_ring *tx_ring; 1615 struct bpf_prog *xdp_prog; 1616 bool xdp_tx_cmpl = false; 1617 unsigned int true_bufsz; 1618 struct sk_buff *skb; 1619 int pkts_polled = 0; 1620 struct xdp_buff xdp; 1621 int idx; 1622 1623 rcu_read_lock(); 1624 xdp_prog = READ_ONCE(dp->xdp_prog); 1625 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz; 1626 xdp.rxq = &rx_ring->xdp_rxq; 1627 tx_ring = r_vec->xdp_ring; 1628 1629 while (pkts_polled < budget) { 1630 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1631 struct nfp_net_rx_buf *rxbuf; 1632 struct nfp_net_rx_desc *rxd; 1633 struct nfp_meta_parsed meta; 1634 struct net_device *netdev; 1635 dma_addr_t new_dma_addr; 1636 u32 meta_len_xdp = 0; 1637 void *new_frag; 1638 1639 idx = D_IDX(rx_ring, rx_ring->rd_p); 1640 1641 rxd = &rx_ring->rxds[idx]; 1642 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 1643 break; 1644 1645 /* Memory barrier to ensure that we won't do other reads 1646 * before the DD bit. 1647 */ 1648 dma_rmb(); 1649 1650 memset(&meta, 0, sizeof(meta)); 1651 1652 rx_ring->rd_p++; 1653 pkts_polled++; 1654 1655 rxbuf = &rx_ring->rxbufs[idx]; 1656 /* < meta_len > 1657 * <-- [rx_offset] --> 1658 * --------------------------------------------------------- 1659 * | [XX] | metadata | packet | XXXX | 1660 * --------------------------------------------------------- 1661 * <---------------- data_len ---------------> 1662 * 1663 * The rx_offset is fixed for all packets, the meta_len can vary 1664 * on a packet by packet basis. If rx_offset is set to zero 1665 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the 1666 * buffer and is immediately followed by the packet (no [XX]). 1667 */ 1668 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 1669 data_len = le16_to_cpu(rxd->rxd.data_len); 1670 pkt_len = data_len - meta_len; 1671 1672 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 1673 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1674 pkt_off += meta_len; 1675 else 1676 pkt_off += dp->rx_offset; 1677 meta_off = pkt_off - meta_len; 1678 1679 /* Stats update */ 1680 u64_stats_update_begin(&r_vec->rx_sync); 1681 r_vec->rx_pkts++; 1682 r_vec->rx_bytes += pkt_len; 1683 u64_stats_update_end(&r_vec->rx_sync); 1684 1685 if (unlikely(meta_len > NFP_NET_MAX_PREPEND || 1686 (dp->rx_offset && meta_len > dp->rx_offset))) { 1687 nn_dp_warn(dp, "oversized RX packet metadata %u\n", 1688 meta_len); 1689 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1690 continue; 1691 } 1692 1693 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, 1694 data_len); 1695 1696 if (!dp->chained_metadata_format) { 1697 nfp_net_set_hash_desc(dp->netdev, &meta, 1698 rxbuf->frag + meta_off, rxd); 1699 } else if (meta_len) { 1700 void *end; 1701 1702 end = nfp_net_parse_meta(dp->netdev, &meta, 1703 rxbuf->frag + meta_off, 1704 meta_len); 1705 if (unlikely(end != rxbuf->frag + pkt_off)) { 1706 nn_dp_warn(dp, "invalid RX packet metadata\n"); 1707 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, 1708 NULL); 1709 continue; 1710 } 1711 } 1712 1713 if (xdp_prog && !(rxd->rxd.flags & PCIE_DESC_RX_BPF && 1714 dp->bpf_offload_xdp) && !meta.portid) { 1715 void *orig_data = rxbuf->frag + pkt_off; 1716 unsigned int dma_off; 1717 int act; 1718 1719 xdp.data_hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM; 1720 xdp.data = orig_data; 1721 xdp.data_meta = orig_data; 1722 xdp.data_end = orig_data + pkt_len; 1723 1724 act = bpf_prog_run_xdp(xdp_prog, &xdp); 1725 1726 pkt_len = xdp.data_end - xdp.data; 1727 pkt_off += xdp.data - orig_data; 1728 1729 switch (act) { 1730 case XDP_PASS: 1731 meta_len_xdp = xdp.data - xdp.data_meta; 1732 break; 1733 case XDP_TX: 1734 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM; 1735 if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring, 1736 tx_ring, rxbuf, 1737 dma_off, 1738 pkt_len, 1739 &xdp_tx_cmpl))) 1740 trace_xdp_exception(dp->netdev, 1741 xdp_prog, act); 1742 continue; 1743 default: 1744 bpf_warn_invalid_xdp_action(act); 1745 /* fall through */ 1746 case XDP_ABORTED: 1747 trace_xdp_exception(dp->netdev, xdp_prog, act); 1748 /* fall through */ 1749 case XDP_DROP: 1750 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, 1751 rxbuf->dma_addr); 1752 continue; 1753 } 1754 } 1755 1756 skb = build_skb(rxbuf->frag, true_bufsz); 1757 if (unlikely(!skb)) { 1758 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1759 continue; 1760 } 1761 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr); 1762 if (unlikely(!new_frag)) { 1763 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 1764 continue; 1765 } 1766 1767 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 1768 1769 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 1770 1771 if (likely(!meta.portid)) { 1772 netdev = dp->netdev; 1773 } else { 1774 struct nfp_net *nn; 1775 1776 nn = netdev_priv(dp->netdev); 1777 netdev = nfp_app_repr_get(nn->app, meta.portid); 1778 if (unlikely(!netdev)) { 1779 nfp_net_rx_drop(dp, r_vec, rx_ring, NULL, skb); 1780 continue; 1781 } 1782 nfp_repr_inc_rx_stats(netdev, pkt_len); 1783 } 1784 1785 skb_reserve(skb, pkt_off); 1786 skb_put(skb, pkt_len); 1787 1788 skb->mark = meta.mark; 1789 skb_set_hash(skb, meta.hash, meta.hash_type); 1790 1791 skb_record_rx_queue(skb, rx_ring->idx); 1792 skb->protocol = eth_type_trans(skb, netdev); 1793 1794 nfp_net_rx_csum(dp, r_vec, rxd, &meta, skb); 1795 1796 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN) 1797 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), 1798 le16_to_cpu(rxd->rxd.vlan)); 1799 if (meta_len_xdp) 1800 skb_metadata_set(skb, meta_len_xdp); 1801 1802 napi_gro_receive(&rx_ring->r_vec->napi, skb); 1803 } 1804 1805 if (xdp_prog) { 1806 if (tx_ring->wr_ptr_add) 1807 nfp_net_tx_xmit_more_flush(tx_ring); 1808 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) && 1809 !xdp_tx_cmpl) 1810 if (!nfp_net_xdp_complete(tx_ring)) 1811 pkts_polled = budget; 1812 } 1813 rcu_read_unlock(); 1814 1815 return pkts_polled; 1816 } 1817 1818 /** 1819 * nfp_net_poll() - napi poll function 1820 * @napi: NAPI structure 1821 * @budget: NAPI budget 1822 * 1823 * Return: number of packets polled. 1824 */ 1825 static int nfp_net_poll(struct napi_struct *napi, int budget) 1826 { 1827 struct nfp_net_r_vector *r_vec = 1828 container_of(napi, struct nfp_net_r_vector, napi); 1829 unsigned int pkts_polled = 0; 1830 1831 if (r_vec->tx_ring) 1832 nfp_net_tx_complete(r_vec->tx_ring, budget); 1833 if (r_vec->rx_ring) 1834 pkts_polled = nfp_net_rx(r_vec->rx_ring, budget); 1835 1836 if (pkts_polled < budget) 1837 if (napi_complete_done(napi, pkts_polled)) 1838 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 1839 1840 return pkts_polled; 1841 } 1842 1843 /* Control device data path 1844 */ 1845 1846 static bool 1847 nfp_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 1848 struct sk_buff *skb, bool old) 1849 { 1850 unsigned int real_len = skb->len, meta_len = 0; 1851 struct nfp_net_tx_ring *tx_ring; 1852 struct nfp_net_tx_buf *txbuf; 1853 struct nfp_net_tx_desc *txd; 1854 struct nfp_net_dp *dp; 1855 dma_addr_t dma_addr; 1856 int wr_idx; 1857 1858 dp = &r_vec->nfp_net->dp; 1859 tx_ring = r_vec->tx_ring; 1860 1861 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) { 1862 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n"); 1863 goto err_free; 1864 } 1865 1866 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1867 u64_stats_update_begin(&r_vec->tx_sync); 1868 r_vec->tx_busy++; 1869 u64_stats_update_end(&r_vec->tx_sync); 1870 if (!old) 1871 __skb_queue_tail(&r_vec->queue, skb); 1872 else 1873 __skb_queue_head(&r_vec->queue, skb); 1874 return true; 1875 } 1876 1877 if (nfp_app_ctrl_has_meta(nn->app)) { 1878 if (unlikely(skb_headroom(skb) < 8)) { 1879 nn_dp_warn(dp, "CTRL TX on skb without headroom\n"); 1880 goto err_free; 1881 } 1882 meta_len = 8; 1883 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4)); 1884 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4)); 1885 } 1886 1887 /* Start with the head skbuf */ 1888 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 1889 DMA_TO_DEVICE); 1890 if (dma_mapping_error(dp->dev, dma_addr)) 1891 goto err_dma_warn; 1892 1893 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1894 1895 /* Stash the soft descriptor of the head then initialize it */ 1896 txbuf = &tx_ring->txbufs[wr_idx]; 1897 txbuf->skb = skb; 1898 txbuf->dma_addr = dma_addr; 1899 txbuf->fidx = -1; 1900 txbuf->pkt_cnt = 1; 1901 txbuf->real_len = real_len; 1902 1903 /* Build TX descriptor */ 1904 txd = &tx_ring->txds[wr_idx]; 1905 txd->offset_eop = meta_len | PCIE_DESC_TX_EOP; 1906 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 1907 nfp_desc_set_dma_addr(txd, dma_addr); 1908 txd->data_len = cpu_to_le16(skb->len); 1909 1910 txd->flags = 0; 1911 txd->mss = 0; 1912 txd->lso_hdrlen = 0; 1913 1914 tx_ring->wr_p++; 1915 tx_ring->wr_ptr_add++; 1916 nfp_net_tx_xmit_more_flush(tx_ring); 1917 1918 return false; 1919 1920 err_dma_warn: 1921 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n"); 1922 err_free: 1923 u64_stats_update_begin(&r_vec->tx_sync); 1924 r_vec->tx_errors++; 1925 u64_stats_update_end(&r_vec->tx_sync); 1926 dev_kfree_skb_any(skb); 1927 return false; 1928 } 1929 1930 bool __nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb) 1931 { 1932 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0]; 1933 1934 return nfp_ctrl_tx_one(nn, r_vec, skb, false); 1935 } 1936 1937 bool nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb) 1938 { 1939 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0]; 1940 bool ret; 1941 1942 spin_lock_bh(&r_vec->lock); 1943 ret = nfp_ctrl_tx_one(nn, r_vec, skb, false); 1944 spin_unlock_bh(&r_vec->lock); 1945 1946 return ret; 1947 } 1948 1949 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec) 1950 { 1951 struct sk_buff *skb; 1952 1953 while ((skb = __skb_dequeue(&r_vec->queue))) 1954 if (nfp_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true)) 1955 return; 1956 } 1957 1958 static bool 1959 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len) 1960 { 1961 u32 meta_type, meta_tag; 1962 1963 if (!nfp_app_ctrl_has_meta(nn->app)) 1964 return !meta_len; 1965 1966 if (meta_len != 8) 1967 return false; 1968 1969 meta_type = get_unaligned_be32(data); 1970 meta_tag = get_unaligned_be32(data + 4); 1971 1972 return (meta_type == NFP_NET_META_PORTID && 1973 meta_tag == NFP_META_PORT_ID_CTRL); 1974 } 1975 1976 static bool 1977 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp, 1978 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring) 1979 { 1980 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1981 struct nfp_net_rx_buf *rxbuf; 1982 struct nfp_net_rx_desc *rxd; 1983 dma_addr_t new_dma_addr; 1984 struct sk_buff *skb; 1985 void *new_frag; 1986 int idx; 1987 1988 idx = D_IDX(rx_ring, rx_ring->rd_p); 1989 1990 rxd = &rx_ring->rxds[idx]; 1991 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 1992 return false; 1993 1994 /* Memory barrier to ensure that we won't do other reads 1995 * before the DD bit. 1996 */ 1997 dma_rmb(); 1998 1999 rx_ring->rd_p++; 2000 2001 rxbuf = &rx_ring->rxbufs[idx]; 2002 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 2003 data_len = le16_to_cpu(rxd->rxd.data_len); 2004 pkt_len = data_len - meta_len; 2005 2006 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 2007 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 2008 pkt_off += meta_len; 2009 else 2010 pkt_off += dp->rx_offset; 2011 meta_off = pkt_off - meta_len; 2012 2013 /* Stats update */ 2014 u64_stats_update_begin(&r_vec->rx_sync); 2015 r_vec->rx_pkts++; 2016 r_vec->rx_bytes += pkt_len; 2017 u64_stats_update_end(&r_vec->rx_sync); 2018 2019 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len); 2020 2021 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) { 2022 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n", 2023 meta_len); 2024 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 2025 return true; 2026 } 2027 2028 skb = build_skb(rxbuf->frag, dp->fl_bufsz); 2029 if (unlikely(!skb)) { 2030 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 2031 return true; 2032 } 2033 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr); 2034 if (unlikely(!new_frag)) { 2035 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 2036 return true; 2037 } 2038 2039 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 2040 2041 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 2042 2043 skb_reserve(skb, pkt_off); 2044 skb_put(skb, pkt_len); 2045 2046 nfp_app_ctrl_rx(nn->app, skb); 2047 2048 return true; 2049 } 2050 2051 static void nfp_ctrl_rx(struct nfp_net_r_vector *r_vec) 2052 { 2053 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring; 2054 struct nfp_net *nn = r_vec->nfp_net; 2055 struct nfp_net_dp *dp = &nn->dp; 2056 2057 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring)) 2058 continue; 2059 } 2060 2061 static void nfp_ctrl_poll(unsigned long arg) 2062 { 2063 struct nfp_net_r_vector *r_vec = (void *)arg; 2064 2065 spin_lock_bh(&r_vec->lock); 2066 nfp_net_tx_complete(r_vec->tx_ring, 0); 2067 __nfp_ctrl_tx_queued(r_vec); 2068 spin_unlock_bh(&r_vec->lock); 2069 2070 nfp_ctrl_rx(r_vec); 2071 2072 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 2073 } 2074 2075 /* Setup and Configuration 2076 */ 2077 2078 /** 2079 * nfp_net_vecs_init() - Assign IRQs and setup rvecs. 2080 * @nn: NFP Network structure 2081 */ 2082 static void nfp_net_vecs_init(struct nfp_net *nn) 2083 { 2084 struct nfp_net_r_vector *r_vec; 2085 int r; 2086 2087 nn->lsc_handler = nfp_net_irq_lsc; 2088 nn->exn_handler = nfp_net_irq_exn; 2089 2090 for (r = 0; r < nn->max_r_vecs; r++) { 2091 struct msix_entry *entry; 2092 2093 entry = &nn->irq_entries[NFP_NET_NON_Q_VECTORS + r]; 2094 2095 r_vec = &nn->r_vecs[r]; 2096 r_vec->nfp_net = nn; 2097 r_vec->irq_entry = entry->entry; 2098 r_vec->irq_vector = entry->vector; 2099 2100 if (nn->dp.netdev) { 2101 r_vec->handler = nfp_net_irq_rxtx; 2102 } else { 2103 r_vec->handler = nfp_ctrl_irq_rxtx; 2104 2105 __skb_queue_head_init(&r_vec->queue); 2106 spin_lock_init(&r_vec->lock); 2107 tasklet_init(&r_vec->tasklet, nfp_ctrl_poll, 2108 (unsigned long)r_vec); 2109 tasklet_disable(&r_vec->tasklet); 2110 } 2111 2112 cpumask_set_cpu(r, &r_vec->affinity_mask); 2113 } 2114 } 2115 2116 /** 2117 * nfp_net_tx_ring_free() - Free resources allocated to a TX ring 2118 * @tx_ring: TX ring to free 2119 */ 2120 static void nfp_net_tx_ring_free(struct nfp_net_tx_ring *tx_ring) 2121 { 2122 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 2123 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 2124 2125 kfree(tx_ring->txbufs); 2126 2127 if (tx_ring->txds) 2128 dma_free_coherent(dp->dev, tx_ring->size, 2129 tx_ring->txds, tx_ring->dma); 2130 2131 tx_ring->cnt = 0; 2132 tx_ring->txbufs = NULL; 2133 tx_ring->txds = NULL; 2134 tx_ring->dma = 0; 2135 tx_ring->size = 0; 2136 } 2137 2138 /** 2139 * nfp_net_tx_ring_alloc() - Allocate resource for a TX ring 2140 * @dp: NFP Net data path struct 2141 * @tx_ring: TX Ring structure to allocate 2142 * 2143 * Return: 0 on success, negative errno otherwise. 2144 */ 2145 static int 2146 nfp_net_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring) 2147 { 2148 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 2149 int sz; 2150 2151 tx_ring->cnt = dp->txd_cnt; 2152 2153 tx_ring->size = sizeof(*tx_ring->txds) * tx_ring->cnt; 2154 tx_ring->txds = dma_zalloc_coherent(dp->dev, tx_ring->size, 2155 &tx_ring->dma, GFP_KERNEL); 2156 if (!tx_ring->txds) 2157 goto err_alloc; 2158 2159 sz = sizeof(*tx_ring->txbufs) * tx_ring->cnt; 2160 tx_ring->txbufs = kzalloc(sz, GFP_KERNEL); 2161 if (!tx_ring->txbufs) 2162 goto err_alloc; 2163 2164 if (!tx_ring->is_xdp && dp->netdev) 2165 netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask, 2166 tx_ring->idx); 2167 2168 return 0; 2169 2170 err_alloc: 2171 nfp_net_tx_ring_free(tx_ring); 2172 return -ENOMEM; 2173 } 2174 2175 static void 2176 nfp_net_tx_ring_bufs_free(struct nfp_net_dp *dp, 2177 struct nfp_net_tx_ring *tx_ring) 2178 { 2179 unsigned int i; 2180 2181 if (!tx_ring->is_xdp) 2182 return; 2183 2184 for (i = 0; i < tx_ring->cnt; i++) { 2185 if (!tx_ring->txbufs[i].frag) 2186 return; 2187 2188 nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr); 2189 __free_page(virt_to_page(tx_ring->txbufs[i].frag)); 2190 } 2191 } 2192 2193 static int 2194 nfp_net_tx_ring_bufs_alloc(struct nfp_net_dp *dp, 2195 struct nfp_net_tx_ring *tx_ring) 2196 { 2197 struct nfp_net_tx_buf *txbufs = tx_ring->txbufs; 2198 unsigned int i; 2199 2200 if (!tx_ring->is_xdp) 2201 return 0; 2202 2203 for (i = 0; i < tx_ring->cnt; i++) { 2204 txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr); 2205 if (!txbufs[i].frag) { 2206 nfp_net_tx_ring_bufs_free(dp, tx_ring); 2207 return -ENOMEM; 2208 } 2209 } 2210 2211 return 0; 2212 } 2213 2214 static int nfp_net_tx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp) 2215 { 2216 unsigned int r; 2217 2218 dp->tx_rings = kcalloc(dp->num_tx_rings, sizeof(*dp->tx_rings), 2219 GFP_KERNEL); 2220 if (!dp->tx_rings) 2221 return -ENOMEM; 2222 2223 for (r = 0; r < dp->num_tx_rings; r++) { 2224 int bias = 0; 2225 2226 if (r >= dp->num_stack_tx_rings) 2227 bias = dp->num_stack_tx_rings; 2228 2229 nfp_net_tx_ring_init(&dp->tx_rings[r], &nn->r_vecs[r - bias], 2230 r, bias); 2231 2232 if (nfp_net_tx_ring_alloc(dp, &dp->tx_rings[r])) 2233 goto err_free_prev; 2234 2235 if (nfp_net_tx_ring_bufs_alloc(dp, &dp->tx_rings[r])) 2236 goto err_free_ring; 2237 } 2238 2239 return 0; 2240 2241 err_free_prev: 2242 while (r--) { 2243 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]); 2244 err_free_ring: 2245 nfp_net_tx_ring_free(&dp->tx_rings[r]); 2246 } 2247 kfree(dp->tx_rings); 2248 return -ENOMEM; 2249 } 2250 2251 static void nfp_net_tx_rings_free(struct nfp_net_dp *dp) 2252 { 2253 unsigned int r; 2254 2255 for (r = 0; r < dp->num_tx_rings; r++) { 2256 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]); 2257 nfp_net_tx_ring_free(&dp->tx_rings[r]); 2258 } 2259 2260 kfree(dp->tx_rings); 2261 } 2262 2263 /** 2264 * nfp_net_rx_ring_free() - Free resources allocated to a RX ring 2265 * @rx_ring: RX ring to free 2266 */ 2267 static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring) 2268 { 2269 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 2270 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 2271 2272 if (dp->netdev) 2273 xdp_rxq_info_unreg(&rx_ring->xdp_rxq); 2274 kfree(rx_ring->rxbufs); 2275 2276 if (rx_ring->rxds) 2277 dma_free_coherent(dp->dev, rx_ring->size, 2278 rx_ring->rxds, rx_ring->dma); 2279 2280 rx_ring->cnt = 0; 2281 rx_ring->rxbufs = NULL; 2282 rx_ring->rxds = NULL; 2283 rx_ring->dma = 0; 2284 rx_ring->size = 0; 2285 } 2286 2287 /** 2288 * nfp_net_rx_ring_alloc() - Allocate resource for a RX ring 2289 * @dp: NFP Net data path struct 2290 * @rx_ring: RX ring to allocate 2291 * 2292 * Return: 0 on success, negative errno otherwise. 2293 */ 2294 static int 2295 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring) 2296 { 2297 int sz, err; 2298 2299 if (dp->netdev) { 2300 err = xdp_rxq_info_reg(&rx_ring->xdp_rxq, dp->netdev, 2301 rx_ring->idx); 2302 if (err < 0) 2303 return err; 2304 } 2305 2306 rx_ring->cnt = dp->rxd_cnt; 2307 rx_ring->size = sizeof(*rx_ring->rxds) * rx_ring->cnt; 2308 rx_ring->rxds = dma_zalloc_coherent(dp->dev, rx_ring->size, 2309 &rx_ring->dma, GFP_KERNEL); 2310 if (!rx_ring->rxds) 2311 goto err_alloc; 2312 2313 sz = sizeof(*rx_ring->rxbufs) * rx_ring->cnt; 2314 rx_ring->rxbufs = kzalloc(sz, GFP_KERNEL); 2315 if (!rx_ring->rxbufs) 2316 goto err_alloc; 2317 2318 return 0; 2319 2320 err_alloc: 2321 nfp_net_rx_ring_free(rx_ring); 2322 return -ENOMEM; 2323 } 2324 2325 static int nfp_net_rx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp) 2326 { 2327 unsigned int r; 2328 2329 dp->rx_rings = kcalloc(dp->num_rx_rings, sizeof(*dp->rx_rings), 2330 GFP_KERNEL); 2331 if (!dp->rx_rings) 2332 return -ENOMEM; 2333 2334 for (r = 0; r < dp->num_rx_rings; r++) { 2335 nfp_net_rx_ring_init(&dp->rx_rings[r], &nn->r_vecs[r], r); 2336 2337 if (nfp_net_rx_ring_alloc(dp, &dp->rx_rings[r])) 2338 goto err_free_prev; 2339 2340 if (nfp_net_rx_ring_bufs_alloc(dp, &dp->rx_rings[r])) 2341 goto err_free_ring; 2342 } 2343 2344 return 0; 2345 2346 err_free_prev: 2347 while (r--) { 2348 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]); 2349 err_free_ring: 2350 nfp_net_rx_ring_free(&dp->rx_rings[r]); 2351 } 2352 kfree(dp->rx_rings); 2353 return -ENOMEM; 2354 } 2355 2356 static void nfp_net_rx_rings_free(struct nfp_net_dp *dp) 2357 { 2358 unsigned int r; 2359 2360 for (r = 0; r < dp->num_rx_rings; r++) { 2361 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]); 2362 nfp_net_rx_ring_free(&dp->rx_rings[r]); 2363 } 2364 2365 kfree(dp->rx_rings); 2366 } 2367 2368 static void 2369 nfp_net_vector_assign_rings(struct nfp_net_dp *dp, 2370 struct nfp_net_r_vector *r_vec, int idx) 2371 { 2372 r_vec->rx_ring = idx < dp->num_rx_rings ? &dp->rx_rings[idx] : NULL; 2373 r_vec->tx_ring = 2374 idx < dp->num_stack_tx_rings ? &dp->tx_rings[idx] : NULL; 2375 2376 r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ? 2377 &dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL; 2378 } 2379 2380 static int 2381 nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 2382 int idx) 2383 { 2384 int err; 2385 2386 /* Setup NAPI */ 2387 if (nn->dp.netdev) 2388 netif_napi_add(nn->dp.netdev, &r_vec->napi, 2389 nfp_net_poll, NAPI_POLL_WEIGHT); 2390 else 2391 tasklet_enable(&r_vec->tasklet); 2392 2393 snprintf(r_vec->name, sizeof(r_vec->name), 2394 "%s-rxtx-%d", nfp_net_name(nn), idx); 2395 err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name, 2396 r_vec); 2397 if (err) { 2398 if (nn->dp.netdev) 2399 netif_napi_del(&r_vec->napi); 2400 else 2401 tasklet_disable(&r_vec->tasklet); 2402 2403 nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector); 2404 return err; 2405 } 2406 disable_irq(r_vec->irq_vector); 2407 2408 irq_set_affinity_hint(r_vec->irq_vector, &r_vec->affinity_mask); 2409 2410 nn_dbg(nn, "RV%02d: irq=%03d/%03d\n", idx, r_vec->irq_vector, 2411 r_vec->irq_entry); 2412 2413 return 0; 2414 } 2415 2416 static void 2417 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec) 2418 { 2419 irq_set_affinity_hint(r_vec->irq_vector, NULL); 2420 if (nn->dp.netdev) 2421 netif_napi_del(&r_vec->napi); 2422 else 2423 tasklet_disable(&r_vec->tasklet); 2424 2425 free_irq(r_vec->irq_vector, r_vec); 2426 } 2427 2428 /** 2429 * nfp_net_rss_write_itbl() - Write RSS indirection table to device 2430 * @nn: NFP Net device to reconfigure 2431 */ 2432 void nfp_net_rss_write_itbl(struct nfp_net *nn) 2433 { 2434 int i; 2435 2436 for (i = 0; i < NFP_NET_CFG_RSS_ITBL_SZ; i += 4) 2437 nn_writel(nn, NFP_NET_CFG_RSS_ITBL + i, 2438 get_unaligned_le32(nn->rss_itbl + i)); 2439 } 2440 2441 /** 2442 * nfp_net_rss_write_key() - Write RSS hash key to device 2443 * @nn: NFP Net device to reconfigure 2444 */ 2445 void nfp_net_rss_write_key(struct nfp_net *nn) 2446 { 2447 int i; 2448 2449 for (i = 0; i < nfp_net_rss_key_sz(nn); i += 4) 2450 nn_writel(nn, NFP_NET_CFG_RSS_KEY + i, 2451 get_unaligned_le32(nn->rss_key + i)); 2452 } 2453 2454 /** 2455 * nfp_net_coalesce_write_cfg() - Write irq coalescence configuration to HW 2456 * @nn: NFP Net device to reconfigure 2457 */ 2458 void nfp_net_coalesce_write_cfg(struct nfp_net *nn) 2459 { 2460 u8 i; 2461 u32 factor; 2462 u32 value; 2463 2464 /* Compute factor used to convert coalesce '_usecs' parameters to 2465 * ME timestamp ticks. There are 16 ME clock cycles for each timestamp 2466 * count. 2467 */ 2468 factor = nn->tlv_caps.me_freq_mhz / 16; 2469 2470 /* copy RX interrupt coalesce parameters */ 2471 value = (nn->rx_coalesce_max_frames << 16) | 2472 (factor * nn->rx_coalesce_usecs); 2473 for (i = 0; i < nn->dp.num_rx_rings; i++) 2474 nn_writel(nn, NFP_NET_CFG_RXR_IRQ_MOD(i), value); 2475 2476 /* copy TX interrupt coalesce parameters */ 2477 value = (nn->tx_coalesce_max_frames << 16) | 2478 (factor * nn->tx_coalesce_usecs); 2479 for (i = 0; i < nn->dp.num_tx_rings; i++) 2480 nn_writel(nn, NFP_NET_CFG_TXR_IRQ_MOD(i), value); 2481 } 2482 2483 /** 2484 * nfp_net_write_mac_addr() - Write mac address to the device control BAR 2485 * @nn: NFP Net device to reconfigure 2486 * @addr: MAC address to write 2487 * 2488 * Writes the MAC address from the netdev to the device control BAR. Does not 2489 * perform the required reconfig. We do a bit of byte swapping dance because 2490 * firmware is LE. 2491 */ 2492 static void nfp_net_write_mac_addr(struct nfp_net *nn, const u8 *addr) 2493 { 2494 nn_writel(nn, NFP_NET_CFG_MACADDR + 0, get_unaligned_be32(addr)); 2495 nn_writew(nn, NFP_NET_CFG_MACADDR + 6, get_unaligned_be16(addr + 4)); 2496 } 2497 2498 static void nfp_net_vec_clear_ring_data(struct nfp_net *nn, unsigned int idx) 2499 { 2500 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), 0); 2501 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), 0); 2502 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), 0); 2503 2504 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), 0); 2505 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), 0); 2506 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), 0); 2507 } 2508 2509 /** 2510 * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP 2511 * @nn: NFP Net device to reconfigure 2512 */ 2513 static void nfp_net_clear_config_and_disable(struct nfp_net *nn) 2514 { 2515 u32 new_ctrl, update; 2516 unsigned int r; 2517 int err; 2518 2519 new_ctrl = nn->dp.ctrl; 2520 new_ctrl &= ~NFP_NET_CFG_CTRL_ENABLE; 2521 update = NFP_NET_CFG_UPDATE_GEN; 2522 update |= NFP_NET_CFG_UPDATE_MSIX; 2523 update |= NFP_NET_CFG_UPDATE_RING; 2524 2525 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG) 2526 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG; 2527 2528 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0); 2529 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0); 2530 2531 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2532 err = nfp_net_reconfig(nn, update); 2533 if (err) 2534 nn_err(nn, "Could not disable device: %d\n", err); 2535 2536 for (r = 0; r < nn->dp.num_rx_rings; r++) 2537 nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]); 2538 for (r = 0; r < nn->dp.num_tx_rings; r++) 2539 nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]); 2540 for (r = 0; r < nn->dp.num_r_vecs; r++) 2541 nfp_net_vec_clear_ring_data(nn, r); 2542 2543 nn->dp.ctrl = new_ctrl; 2544 } 2545 2546 static void 2547 nfp_net_rx_ring_hw_cfg_write(struct nfp_net *nn, 2548 struct nfp_net_rx_ring *rx_ring, unsigned int idx) 2549 { 2550 /* Write the DMA address, size and MSI-X info to the device */ 2551 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), rx_ring->dma); 2552 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), ilog2(rx_ring->cnt)); 2553 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), rx_ring->r_vec->irq_entry); 2554 } 2555 2556 static void 2557 nfp_net_tx_ring_hw_cfg_write(struct nfp_net *nn, 2558 struct nfp_net_tx_ring *tx_ring, unsigned int idx) 2559 { 2560 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), tx_ring->dma); 2561 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), ilog2(tx_ring->cnt)); 2562 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), tx_ring->r_vec->irq_entry); 2563 } 2564 2565 /** 2566 * nfp_net_set_config_and_enable() - Write control BAR and enable NFP 2567 * @nn: NFP Net device to reconfigure 2568 */ 2569 static int nfp_net_set_config_and_enable(struct nfp_net *nn) 2570 { 2571 u32 bufsz, new_ctrl, update = 0; 2572 unsigned int r; 2573 int err; 2574 2575 new_ctrl = nn->dp.ctrl; 2576 2577 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_RSS_ANY) { 2578 nfp_net_rss_write_key(nn); 2579 nfp_net_rss_write_itbl(nn); 2580 nn_writel(nn, NFP_NET_CFG_RSS_CTRL, nn->rss_cfg); 2581 update |= NFP_NET_CFG_UPDATE_RSS; 2582 } 2583 2584 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_IRQMOD) { 2585 nfp_net_coalesce_write_cfg(nn); 2586 update |= NFP_NET_CFG_UPDATE_IRQMOD; 2587 } 2588 2589 for (r = 0; r < nn->dp.num_tx_rings; r++) 2590 nfp_net_tx_ring_hw_cfg_write(nn, &nn->dp.tx_rings[r], r); 2591 for (r = 0; r < nn->dp.num_rx_rings; r++) 2592 nfp_net_rx_ring_hw_cfg_write(nn, &nn->dp.rx_rings[r], r); 2593 2594 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, nn->dp.num_tx_rings == 64 ? 2595 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_tx_rings) - 1); 2596 2597 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, nn->dp.num_rx_rings == 64 ? 2598 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_rx_rings) - 1); 2599 2600 if (nn->dp.netdev) 2601 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr); 2602 2603 nn_writel(nn, NFP_NET_CFG_MTU, nn->dp.mtu); 2604 2605 bufsz = nn->dp.fl_bufsz - nn->dp.rx_dma_off - NFP_NET_RX_BUF_NON_DATA; 2606 nn_writel(nn, NFP_NET_CFG_FLBUFSZ, bufsz); 2607 2608 /* Enable device */ 2609 new_ctrl |= NFP_NET_CFG_CTRL_ENABLE; 2610 update |= NFP_NET_CFG_UPDATE_GEN; 2611 update |= NFP_NET_CFG_UPDATE_MSIX; 2612 update |= NFP_NET_CFG_UPDATE_RING; 2613 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG) 2614 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG; 2615 2616 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2617 err = nfp_net_reconfig(nn, update); 2618 if (err) { 2619 nfp_net_clear_config_and_disable(nn); 2620 return err; 2621 } 2622 2623 nn->dp.ctrl = new_ctrl; 2624 2625 for (r = 0; r < nn->dp.num_rx_rings; r++) 2626 nfp_net_rx_ring_fill_freelist(&nn->dp, &nn->dp.rx_rings[r]); 2627 2628 /* Since reconfiguration requests while NFP is down are ignored we 2629 * have to wipe the entire VXLAN configuration and reinitialize it. 2630 */ 2631 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN) { 2632 memset(&nn->vxlan_ports, 0, sizeof(nn->vxlan_ports)); 2633 memset(&nn->vxlan_usecnt, 0, sizeof(nn->vxlan_usecnt)); 2634 udp_tunnel_get_rx_info(nn->dp.netdev); 2635 } 2636 2637 return 0; 2638 } 2639 2640 /** 2641 * nfp_net_close_stack() - Quiesce the stack (part of close) 2642 * @nn: NFP Net device to reconfigure 2643 */ 2644 static void nfp_net_close_stack(struct nfp_net *nn) 2645 { 2646 unsigned int r; 2647 2648 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2649 netif_carrier_off(nn->dp.netdev); 2650 nn->link_up = false; 2651 2652 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2653 disable_irq(nn->r_vecs[r].irq_vector); 2654 napi_disable(&nn->r_vecs[r].napi); 2655 } 2656 2657 netif_tx_disable(nn->dp.netdev); 2658 } 2659 2660 /** 2661 * nfp_net_close_free_all() - Free all runtime resources 2662 * @nn: NFP Net device to reconfigure 2663 */ 2664 static void nfp_net_close_free_all(struct nfp_net *nn) 2665 { 2666 unsigned int r; 2667 2668 nfp_net_tx_rings_free(&nn->dp); 2669 nfp_net_rx_rings_free(&nn->dp); 2670 2671 for (r = 0; r < nn->dp.num_r_vecs; r++) 2672 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 2673 2674 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX); 2675 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX); 2676 } 2677 2678 /** 2679 * nfp_net_netdev_close() - Called when the device is downed 2680 * @netdev: netdev structure 2681 */ 2682 static int nfp_net_netdev_close(struct net_device *netdev) 2683 { 2684 struct nfp_net *nn = netdev_priv(netdev); 2685 2686 /* Step 1: Disable RX and TX rings from the Linux kernel perspective 2687 */ 2688 nfp_net_close_stack(nn); 2689 2690 /* Step 2: Tell NFP 2691 */ 2692 nfp_net_clear_config_and_disable(nn); 2693 nfp_port_configure(netdev, false); 2694 2695 /* Step 3: Free resources 2696 */ 2697 nfp_net_close_free_all(nn); 2698 2699 nn_dbg(nn, "%s down", netdev->name); 2700 return 0; 2701 } 2702 2703 void nfp_ctrl_close(struct nfp_net *nn) 2704 { 2705 int r; 2706 2707 rtnl_lock(); 2708 2709 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2710 disable_irq(nn->r_vecs[r].irq_vector); 2711 tasklet_disable(&nn->r_vecs[r].tasklet); 2712 } 2713 2714 nfp_net_clear_config_and_disable(nn); 2715 2716 nfp_net_close_free_all(nn); 2717 2718 rtnl_unlock(); 2719 } 2720 2721 /** 2722 * nfp_net_open_stack() - Start the device from stack's perspective 2723 * @nn: NFP Net device to reconfigure 2724 */ 2725 static void nfp_net_open_stack(struct nfp_net *nn) 2726 { 2727 unsigned int r; 2728 2729 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2730 napi_enable(&nn->r_vecs[r].napi); 2731 enable_irq(nn->r_vecs[r].irq_vector); 2732 } 2733 2734 netif_tx_wake_all_queues(nn->dp.netdev); 2735 2736 enable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2737 nfp_net_read_link_status(nn); 2738 } 2739 2740 static int nfp_net_open_alloc_all(struct nfp_net *nn) 2741 { 2742 int err, r; 2743 2744 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_EXN, "%s-exn", 2745 nn->exn_name, sizeof(nn->exn_name), 2746 NFP_NET_IRQ_EXN_IDX, nn->exn_handler); 2747 if (err) 2748 return err; 2749 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_LSC, "%s-lsc", 2750 nn->lsc_name, sizeof(nn->lsc_name), 2751 NFP_NET_IRQ_LSC_IDX, nn->lsc_handler); 2752 if (err) 2753 goto err_free_exn; 2754 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2755 2756 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2757 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r); 2758 if (err) 2759 goto err_cleanup_vec_p; 2760 } 2761 2762 err = nfp_net_rx_rings_prepare(nn, &nn->dp); 2763 if (err) 2764 goto err_cleanup_vec; 2765 2766 err = nfp_net_tx_rings_prepare(nn, &nn->dp); 2767 if (err) 2768 goto err_free_rx_rings; 2769 2770 for (r = 0; r < nn->max_r_vecs; r++) 2771 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r); 2772 2773 return 0; 2774 2775 err_free_rx_rings: 2776 nfp_net_rx_rings_free(&nn->dp); 2777 err_cleanup_vec: 2778 r = nn->dp.num_r_vecs; 2779 err_cleanup_vec_p: 2780 while (r--) 2781 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 2782 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX); 2783 err_free_exn: 2784 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX); 2785 return err; 2786 } 2787 2788 static int nfp_net_netdev_open(struct net_device *netdev) 2789 { 2790 struct nfp_net *nn = netdev_priv(netdev); 2791 int err; 2792 2793 /* Step 1: Allocate resources for rings and the like 2794 * - Request interrupts 2795 * - Allocate RX and TX ring resources 2796 * - Setup initial RSS table 2797 */ 2798 err = nfp_net_open_alloc_all(nn); 2799 if (err) 2800 return err; 2801 2802 err = netif_set_real_num_tx_queues(netdev, nn->dp.num_stack_tx_rings); 2803 if (err) 2804 goto err_free_all; 2805 2806 err = netif_set_real_num_rx_queues(netdev, nn->dp.num_rx_rings); 2807 if (err) 2808 goto err_free_all; 2809 2810 /* Step 2: Configure the NFP 2811 * - Ifup the physical interface if it exists 2812 * - Enable rings from 0 to tx_rings/rx_rings - 1. 2813 * - Write MAC address (in case it changed) 2814 * - Set the MTU 2815 * - Set the Freelist buffer size 2816 * - Enable the FW 2817 */ 2818 err = nfp_port_configure(netdev, true); 2819 if (err) 2820 goto err_free_all; 2821 2822 err = nfp_net_set_config_and_enable(nn); 2823 if (err) 2824 goto err_port_disable; 2825 2826 /* Step 3: Enable for kernel 2827 * - put some freelist descriptors on each RX ring 2828 * - enable NAPI on each ring 2829 * - enable all TX queues 2830 * - set link state 2831 */ 2832 nfp_net_open_stack(nn); 2833 2834 return 0; 2835 2836 err_port_disable: 2837 nfp_port_configure(netdev, false); 2838 err_free_all: 2839 nfp_net_close_free_all(nn); 2840 return err; 2841 } 2842 2843 int nfp_ctrl_open(struct nfp_net *nn) 2844 { 2845 int err, r; 2846 2847 /* ring dumping depends on vNICs being opened/closed under rtnl */ 2848 rtnl_lock(); 2849 2850 err = nfp_net_open_alloc_all(nn); 2851 if (err) 2852 goto err_unlock; 2853 2854 err = nfp_net_set_config_and_enable(nn); 2855 if (err) 2856 goto err_free_all; 2857 2858 for (r = 0; r < nn->dp.num_r_vecs; r++) 2859 enable_irq(nn->r_vecs[r].irq_vector); 2860 2861 rtnl_unlock(); 2862 2863 return 0; 2864 2865 err_free_all: 2866 nfp_net_close_free_all(nn); 2867 err_unlock: 2868 rtnl_unlock(); 2869 return err; 2870 } 2871 2872 static void nfp_net_set_rx_mode(struct net_device *netdev) 2873 { 2874 struct nfp_net *nn = netdev_priv(netdev); 2875 u32 new_ctrl; 2876 2877 new_ctrl = nn->dp.ctrl; 2878 2879 if (!netdev_mc_empty(netdev) || netdev->flags & IFF_ALLMULTI) 2880 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_L2MC; 2881 else 2882 new_ctrl &= ~NFP_NET_CFG_CTRL_L2MC; 2883 2884 if (netdev->flags & IFF_PROMISC) { 2885 if (nn->cap & NFP_NET_CFG_CTRL_PROMISC) 2886 new_ctrl |= NFP_NET_CFG_CTRL_PROMISC; 2887 else 2888 nn_warn(nn, "FW does not support promiscuous mode\n"); 2889 } else { 2890 new_ctrl &= ~NFP_NET_CFG_CTRL_PROMISC; 2891 } 2892 2893 if (new_ctrl == nn->dp.ctrl) 2894 return; 2895 2896 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2897 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_GEN); 2898 2899 nn->dp.ctrl = new_ctrl; 2900 } 2901 2902 static void nfp_net_rss_init_itbl(struct nfp_net *nn) 2903 { 2904 int i; 2905 2906 for (i = 0; i < sizeof(nn->rss_itbl); i++) 2907 nn->rss_itbl[i] = 2908 ethtool_rxfh_indir_default(i, nn->dp.num_rx_rings); 2909 } 2910 2911 static void nfp_net_dp_swap(struct nfp_net *nn, struct nfp_net_dp *dp) 2912 { 2913 struct nfp_net_dp new_dp = *dp; 2914 2915 *dp = nn->dp; 2916 nn->dp = new_dp; 2917 2918 nn->dp.netdev->mtu = new_dp.mtu; 2919 2920 if (!netif_is_rxfh_configured(nn->dp.netdev)) 2921 nfp_net_rss_init_itbl(nn); 2922 } 2923 2924 static int nfp_net_dp_swap_enable(struct nfp_net *nn, struct nfp_net_dp *dp) 2925 { 2926 unsigned int r; 2927 int err; 2928 2929 nfp_net_dp_swap(nn, dp); 2930 2931 for (r = 0; r < nn->max_r_vecs; r++) 2932 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r); 2933 2934 err = netif_set_real_num_rx_queues(nn->dp.netdev, nn->dp.num_rx_rings); 2935 if (err) 2936 return err; 2937 2938 if (nn->dp.netdev->real_num_tx_queues != nn->dp.num_stack_tx_rings) { 2939 err = netif_set_real_num_tx_queues(nn->dp.netdev, 2940 nn->dp.num_stack_tx_rings); 2941 if (err) 2942 return err; 2943 } 2944 2945 return nfp_net_set_config_and_enable(nn); 2946 } 2947 2948 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn) 2949 { 2950 struct nfp_net_dp *new; 2951 2952 new = kmalloc(sizeof(*new), GFP_KERNEL); 2953 if (!new) 2954 return NULL; 2955 2956 *new = nn->dp; 2957 2958 /* Clear things which need to be recomputed */ 2959 new->fl_bufsz = 0; 2960 new->tx_rings = NULL; 2961 new->rx_rings = NULL; 2962 new->num_r_vecs = 0; 2963 new->num_stack_tx_rings = 0; 2964 2965 return new; 2966 } 2967 2968 static int 2969 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp, 2970 struct netlink_ext_ack *extack) 2971 { 2972 /* XDP-enabled tests */ 2973 if (!dp->xdp_prog) 2974 return 0; 2975 if (dp->fl_bufsz > PAGE_SIZE) { 2976 NL_SET_ERR_MSG_MOD(extack, "MTU too large w/ XDP enabled"); 2977 return -EINVAL; 2978 } 2979 if (dp->num_tx_rings > nn->max_tx_rings) { 2980 NL_SET_ERR_MSG_MOD(extack, "Insufficient number of TX rings w/ XDP enabled"); 2981 return -EINVAL; 2982 } 2983 2984 return 0; 2985 } 2986 2987 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp, 2988 struct netlink_ext_ack *extack) 2989 { 2990 int r, err; 2991 2992 dp->fl_bufsz = nfp_net_calc_fl_bufsz(dp); 2993 2994 dp->num_stack_tx_rings = dp->num_tx_rings; 2995 if (dp->xdp_prog) 2996 dp->num_stack_tx_rings -= dp->num_rx_rings; 2997 2998 dp->num_r_vecs = max(dp->num_rx_rings, dp->num_stack_tx_rings); 2999 3000 err = nfp_net_check_config(nn, dp, extack); 3001 if (err) 3002 goto exit_free_dp; 3003 3004 if (!netif_running(dp->netdev)) { 3005 nfp_net_dp_swap(nn, dp); 3006 err = 0; 3007 goto exit_free_dp; 3008 } 3009 3010 /* Prepare new rings */ 3011 for (r = nn->dp.num_r_vecs; r < dp->num_r_vecs; r++) { 3012 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r); 3013 if (err) { 3014 dp->num_r_vecs = r; 3015 goto err_cleanup_vecs; 3016 } 3017 } 3018 3019 err = nfp_net_rx_rings_prepare(nn, dp); 3020 if (err) 3021 goto err_cleanup_vecs; 3022 3023 err = nfp_net_tx_rings_prepare(nn, dp); 3024 if (err) 3025 goto err_free_rx; 3026 3027 /* Stop device, swap in new rings, try to start the firmware */ 3028 nfp_net_close_stack(nn); 3029 nfp_net_clear_config_and_disable(nn); 3030 3031 err = nfp_net_dp_swap_enable(nn, dp); 3032 if (err) { 3033 int err2; 3034 3035 nfp_net_clear_config_and_disable(nn); 3036 3037 /* Try with old configuration and old rings */ 3038 err2 = nfp_net_dp_swap_enable(nn, dp); 3039 if (err2) 3040 nn_err(nn, "Can't restore ring config - FW communication failed (%d,%d)\n", 3041 err, err2); 3042 } 3043 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--) 3044 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 3045 3046 nfp_net_rx_rings_free(dp); 3047 nfp_net_tx_rings_free(dp); 3048 3049 nfp_net_open_stack(nn); 3050 exit_free_dp: 3051 kfree(dp); 3052 3053 return err; 3054 3055 err_free_rx: 3056 nfp_net_rx_rings_free(dp); 3057 err_cleanup_vecs: 3058 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--) 3059 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 3060 kfree(dp); 3061 return err; 3062 } 3063 3064 static int nfp_net_change_mtu(struct net_device *netdev, int new_mtu) 3065 { 3066 struct nfp_net *nn = netdev_priv(netdev); 3067 struct nfp_net_dp *dp; 3068 int err; 3069 3070 err = nfp_app_check_mtu(nn->app, netdev, new_mtu); 3071 if (err) 3072 return err; 3073 3074 dp = nfp_net_clone_dp(nn); 3075 if (!dp) 3076 return -ENOMEM; 3077 3078 dp->mtu = new_mtu; 3079 3080 return nfp_net_ring_reconfig(nn, dp, NULL); 3081 } 3082 3083 static int 3084 nfp_net_vlan_rx_add_vid(struct net_device *netdev, __be16 proto, u16 vid) 3085 { 3086 struct nfp_net *nn = netdev_priv(netdev); 3087 3088 /* Priority tagged packets with vlan id 0 are processed by the 3089 * NFP as untagged packets 3090 */ 3091 if (!vid) 3092 return 0; 3093 3094 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid); 3095 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO, 3096 ETH_P_8021Q); 3097 3098 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_ADD); 3099 } 3100 3101 static int 3102 nfp_net_vlan_rx_kill_vid(struct net_device *netdev, __be16 proto, u16 vid) 3103 { 3104 struct nfp_net *nn = netdev_priv(netdev); 3105 3106 /* Priority tagged packets with vlan id 0 are processed by the 3107 * NFP as untagged packets 3108 */ 3109 if (!vid) 3110 return 0; 3111 3112 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid); 3113 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO, 3114 ETH_P_8021Q); 3115 3116 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_KILL); 3117 } 3118 3119 #ifdef CONFIG_NET_POLL_CONTROLLER 3120 static void nfp_net_netpoll(struct net_device *netdev) 3121 { 3122 struct nfp_net *nn = netdev_priv(netdev); 3123 int i; 3124 3125 /* nfp_net's NAPIs are statically allocated so even if there is a race 3126 * with reconfig path this will simply try to schedule some disabled 3127 * NAPI instances. 3128 */ 3129 for (i = 0; i < nn->dp.num_stack_tx_rings; i++) 3130 napi_schedule_irqoff(&nn->r_vecs[i].napi); 3131 } 3132 #endif 3133 3134 static void nfp_net_stat64(struct net_device *netdev, 3135 struct rtnl_link_stats64 *stats) 3136 { 3137 struct nfp_net *nn = netdev_priv(netdev); 3138 int r; 3139 3140 for (r = 0; r < nn->max_r_vecs; r++) { 3141 struct nfp_net_r_vector *r_vec = &nn->r_vecs[r]; 3142 u64 data[3]; 3143 unsigned int start; 3144 3145 do { 3146 start = u64_stats_fetch_begin(&r_vec->rx_sync); 3147 data[0] = r_vec->rx_pkts; 3148 data[1] = r_vec->rx_bytes; 3149 data[2] = r_vec->rx_drops; 3150 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start)); 3151 stats->rx_packets += data[0]; 3152 stats->rx_bytes += data[1]; 3153 stats->rx_dropped += data[2]; 3154 3155 do { 3156 start = u64_stats_fetch_begin(&r_vec->tx_sync); 3157 data[0] = r_vec->tx_pkts; 3158 data[1] = r_vec->tx_bytes; 3159 data[2] = r_vec->tx_errors; 3160 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start)); 3161 stats->tx_packets += data[0]; 3162 stats->tx_bytes += data[1]; 3163 stats->tx_errors += data[2]; 3164 } 3165 } 3166 3167 static int nfp_net_set_features(struct net_device *netdev, 3168 netdev_features_t features) 3169 { 3170 netdev_features_t changed = netdev->features ^ features; 3171 struct nfp_net *nn = netdev_priv(netdev); 3172 u32 new_ctrl; 3173 int err; 3174 3175 /* Assume this is not called with features we have not advertised */ 3176 3177 new_ctrl = nn->dp.ctrl; 3178 3179 if (changed & NETIF_F_RXCSUM) { 3180 if (features & NETIF_F_RXCSUM) 3181 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY; 3182 else 3183 new_ctrl &= ~NFP_NET_CFG_CTRL_RXCSUM_ANY; 3184 } 3185 3186 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3187 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) 3188 new_ctrl |= NFP_NET_CFG_CTRL_TXCSUM; 3189 else 3190 new_ctrl &= ~NFP_NET_CFG_CTRL_TXCSUM; 3191 } 3192 3193 if (changed & (NETIF_F_TSO | NETIF_F_TSO6)) { 3194 if (features & (NETIF_F_TSO | NETIF_F_TSO6)) 3195 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?: 3196 NFP_NET_CFG_CTRL_LSO; 3197 else 3198 new_ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY; 3199 } 3200 3201 if (changed & NETIF_F_HW_VLAN_CTAG_RX) { 3202 if (features & NETIF_F_HW_VLAN_CTAG_RX) 3203 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN; 3204 else 3205 new_ctrl &= ~NFP_NET_CFG_CTRL_RXVLAN; 3206 } 3207 3208 if (changed & NETIF_F_HW_VLAN_CTAG_TX) { 3209 if (features & NETIF_F_HW_VLAN_CTAG_TX) 3210 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN; 3211 else 3212 new_ctrl &= ~NFP_NET_CFG_CTRL_TXVLAN; 3213 } 3214 3215 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) { 3216 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) 3217 new_ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER; 3218 else 3219 new_ctrl &= ~NFP_NET_CFG_CTRL_CTAG_FILTER; 3220 } 3221 3222 if (changed & NETIF_F_SG) { 3223 if (features & NETIF_F_SG) 3224 new_ctrl |= NFP_NET_CFG_CTRL_GATHER; 3225 else 3226 new_ctrl &= ~NFP_NET_CFG_CTRL_GATHER; 3227 } 3228 3229 err = nfp_port_set_features(netdev, features); 3230 if (err) 3231 return err; 3232 3233 nn_dbg(nn, "Feature change 0x%llx -> 0x%llx (changed=0x%llx)\n", 3234 netdev->features, features, changed); 3235 3236 if (new_ctrl == nn->dp.ctrl) 3237 return 0; 3238 3239 nn_dbg(nn, "NIC ctrl: 0x%x -> 0x%x\n", nn->dp.ctrl, new_ctrl); 3240 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 3241 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN); 3242 if (err) 3243 return err; 3244 3245 nn->dp.ctrl = new_ctrl; 3246 3247 return 0; 3248 } 3249 3250 static netdev_features_t 3251 nfp_net_features_check(struct sk_buff *skb, struct net_device *dev, 3252 netdev_features_t features) 3253 { 3254 u8 l4_hdr; 3255 3256 /* We can't do TSO over double tagged packets (802.1AD) */ 3257 features &= vlan_features_check(skb, features); 3258 3259 if (!skb->encapsulation) 3260 return features; 3261 3262 /* Ensure that inner L4 header offset fits into TX descriptor field */ 3263 if (skb_is_gso(skb)) { 3264 u32 hdrlen; 3265 3266 hdrlen = skb_inner_transport_header(skb) - skb->data + 3267 inner_tcp_hdrlen(skb); 3268 3269 if (unlikely(hdrlen > NFP_NET_LSO_MAX_HDR_SZ)) 3270 features &= ~NETIF_F_GSO_MASK; 3271 } 3272 3273 /* VXLAN/GRE check */ 3274 switch (vlan_get_protocol(skb)) { 3275 case htons(ETH_P_IP): 3276 l4_hdr = ip_hdr(skb)->protocol; 3277 break; 3278 case htons(ETH_P_IPV6): 3279 l4_hdr = ipv6_hdr(skb)->nexthdr; 3280 break; 3281 default: 3282 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3283 } 3284 3285 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER || 3286 skb->inner_protocol != htons(ETH_P_TEB) || 3287 (l4_hdr != IPPROTO_UDP && l4_hdr != IPPROTO_GRE) || 3288 (l4_hdr == IPPROTO_UDP && 3289 (skb_inner_mac_header(skb) - skb_transport_header(skb) != 3290 sizeof(struct udphdr) + sizeof(struct vxlanhdr)))) 3291 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3292 3293 return features; 3294 } 3295 3296 static int 3297 nfp_net_get_phys_port_name(struct net_device *netdev, char *name, size_t len) 3298 { 3299 struct nfp_net *nn = netdev_priv(netdev); 3300 int n; 3301 3302 if (nn->port) 3303 return nfp_port_get_phys_port_name(netdev, name, len); 3304 3305 if (nn->dp.is_vf || nn->vnic_no_name) 3306 return -EOPNOTSUPP; 3307 3308 n = snprintf(name, len, "n%d", nn->id); 3309 if (n >= len) 3310 return -EINVAL; 3311 3312 return 0; 3313 } 3314 3315 /** 3316 * nfp_net_set_vxlan_port() - set vxlan port in SW and reconfigure HW 3317 * @nn: NFP Net device to reconfigure 3318 * @idx: Index into the port table where new port should be written 3319 * @port: UDP port to configure (pass zero to remove VXLAN port) 3320 */ 3321 static void nfp_net_set_vxlan_port(struct nfp_net *nn, int idx, __be16 port) 3322 { 3323 int i; 3324 3325 nn->vxlan_ports[idx] = port; 3326 3327 if (!(nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN)) 3328 return; 3329 3330 BUILD_BUG_ON(NFP_NET_N_VXLAN_PORTS & 1); 3331 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i += 2) 3332 nn_writel(nn, NFP_NET_CFG_VXLAN_PORT + i * sizeof(port), 3333 be16_to_cpu(nn->vxlan_ports[i + 1]) << 16 | 3334 be16_to_cpu(nn->vxlan_ports[i])); 3335 3336 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_VXLAN); 3337 } 3338 3339 /** 3340 * nfp_net_find_vxlan_idx() - find table entry of the port or a free one 3341 * @nn: NFP Network structure 3342 * @port: UDP port to look for 3343 * 3344 * Return: if the port is already in the table -- it's position; 3345 * if the port is not in the table -- free position to use; 3346 * if the table is full -- -ENOSPC. 3347 */ 3348 static int nfp_net_find_vxlan_idx(struct nfp_net *nn, __be16 port) 3349 { 3350 int i, free_idx = -ENOSPC; 3351 3352 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i++) { 3353 if (nn->vxlan_ports[i] == port) 3354 return i; 3355 if (!nn->vxlan_usecnt[i]) 3356 free_idx = i; 3357 } 3358 3359 return free_idx; 3360 } 3361 3362 static void nfp_net_add_vxlan_port(struct net_device *netdev, 3363 struct udp_tunnel_info *ti) 3364 { 3365 struct nfp_net *nn = netdev_priv(netdev); 3366 int idx; 3367 3368 if (ti->type != UDP_TUNNEL_TYPE_VXLAN) 3369 return; 3370 3371 idx = nfp_net_find_vxlan_idx(nn, ti->port); 3372 if (idx == -ENOSPC) 3373 return; 3374 3375 if (!nn->vxlan_usecnt[idx]++) 3376 nfp_net_set_vxlan_port(nn, idx, ti->port); 3377 } 3378 3379 static void nfp_net_del_vxlan_port(struct net_device *netdev, 3380 struct udp_tunnel_info *ti) 3381 { 3382 struct nfp_net *nn = netdev_priv(netdev); 3383 int idx; 3384 3385 if (ti->type != UDP_TUNNEL_TYPE_VXLAN) 3386 return; 3387 3388 idx = nfp_net_find_vxlan_idx(nn, ti->port); 3389 if (idx == -ENOSPC || !nn->vxlan_usecnt[idx]) 3390 return; 3391 3392 if (!--nn->vxlan_usecnt[idx]) 3393 nfp_net_set_vxlan_port(nn, idx, 0); 3394 } 3395 3396 static int 3397 nfp_net_xdp_setup_drv(struct nfp_net *nn, struct bpf_prog *prog, 3398 struct netlink_ext_ack *extack) 3399 { 3400 struct nfp_net_dp *dp; 3401 3402 if (!prog == !nn->dp.xdp_prog) { 3403 WRITE_ONCE(nn->dp.xdp_prog, prog); 3404 return 0; 3405 } 3406 3407 dp = nfp_net_clone_dp(nn); 3408 if (!dp) 3409 return -ENOMEM; 3410 3411 dp->xdp_prog = prog; 3412 dp->num_tx_rings += prog ? nn->dp.num_rx_rings : -nn->dp.num_rx_rings; 3413 dp->rx_dma_dir = prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE; 3414 dp->rx_dma_off = prog ? XDP_PACKET_HEADROOM - nn->dp.rx_offset : 0; 3415 3416 /* We need RX reconfig to remap the buffers (BIDIR vs FROM_DEV) */ 3417 return nfp_net_ring_reconfig(nn, dp, extack); 3418 } 3419 3420 static int 3421 nfp_net_xdp_setup(struct nfp_net *nn, struct bpf_prog *prog, u32 flags, 3422 struct netlink_ext_ack *extack) 3423 { 3424 struct bpf_prog *drv_prog, *offload_prog; 3425 int err; 3426 3427 if (nn->xdp_prog && (flags ^ nn->xdp_flags) & XDP_FLAGS_MODES) 3428 return -EBUSY; 3429 3430 /* Load both when no flags set to allow easy activation of driver path 3431 * when program is replaced by one which can't be offloaded. 3432 */ 3433 drv_prog = flags & XDP_FLAGS_HW_MODE ? NULL : prog; 3434 offload_prog = flags & XDP_FLAGS_DRV_MODE ? NULL : prog; 3435 3436 err = nfp_net_xdp_setup_drv(nn, drv_prog, extack); 3437 if (err) 3438 return err; 3439 3440 err = nfp_app_xdp_offload(nn->app, nn, offload_prog, extack); 3441 if (err && flags & XDP_FLAGS_HW_MODE) 3442 return err; 3443 3444 if (nn->xdp_prog) 3445 bpf_prog_put(nn->xdp_prog); 3446 nn->xdp_prog = prog; 3447 nn->xdp_flags = flags; 3448 3449 return 0; 3450 } 3451 3452 static int nfp_net_xdp(struct net_device *netdev, struct netdev_bpf *xdp) 3453 { 3454 struct nfp_net *nn = netdev_priv(netdev); 3455 3456 switch (xdp->command) { 3457 case XDP_SETUP_PROG: 3458 case XDP_SETUP_PROG_HW: 3459 return nfp_net_xdp_setup(nn, xdp->prog, xdp->flags, 3460 xdp->extack); 3461 case XDP_QUERY_PROG: 3462 xdp->prog_attached = !!nn->xdp_prog; 3463 if (nn->dp.bpf_offload_xdp) 3464 xdp->prog_attached = XDP_ATTACHED_HW; 3465 xdp->prog_id = nn->xdp_prog ? nn->xdp_prog->aux->id : 0; 3466 xdp->prog_flags = nn->xdp_prog ? nn->xdp_flags : 0; 3467 return 0; 3468 default: 3469 return nfp_app_bpf(nn->app, nn, xdp); 3470 } 3471 } 3472 3473 static int nfp_net_set_mac_address(struct net_device *netdev, void *addr) 3474 { 3475 struct nfp_net *nn = netdev_priv(netdev); 3476 struct sockaddr *saddr = addr; 3477 int err; 3478 3479 err = eth_prepare_mac_addr_change(netdev, addr); 3480 if (err) 3481 return err; 3482 3483 nfp_net_write_mac_addr(nn, saddr->sa_data); 3484 3485 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MACADDR); 3486 if (err) 3487 return err; 3488 3489 eth_commit_mac_addr_change(netdev, addr); 3490 3491 return 0; 3492 } 3493 3494 const struct net_device_ops nfp_net_netdev_ops = { 3495 .ndo_open = nfp_net_netdev_open, 3496 .ndo_stop = nfp_net_netdev_close, 3497 .ndo_start_xmit = nfp_net_tx, 3498 .ndo_get_stats64 = nfp_net_stat64, 3499 .ndo_vlan_rx_add_vid = nfp_net_vlan_rx_add_vid, 3500 .ndo_vlan_rx_kill_vid = nfp_net_vlan_rx_kill_vid, 3501 #ifdef CONFIG_NET_POLL_CONTROLLER 3502 .ndo_poll_controller = nfp_net_netpoll, 3503 #endif 3504 .ndo_set_vf_mac = nfp_app_set_vf_mac, 3505 .ndo_set_vf_vlan = nfp_app_set_vf_vlan, 3506 .ndo_set_vf_spoofchk = nfp_app_set_vf_spoofchk, 3507 .ndo_get_vf_config = nfp_app_get_vf_config, 3508 .ndo_set_vf_link_state = nfp_app_set_vf_link_state, 3509 .ndo_setup_tc = nfp_port_setup_tc, 3510 .ndo_tx_timeout = nfp_net_tx_timeout, 3511 .ndo_set_rx_mode = nfp_net_set_rx_mode, 3512 .ndo_change_mtu = nfp_net_change_mtu, 3513 .ndo_set_mac_address = nfp_net_set_mac_address, 3514 .ndo_set_features = nfp_net_set_features, 3515 .ndo_features_check = nfp_net_features_check, 3516 .ndo_get_phys_port_name = nfp_net_get_phys_port_name, 3517 .ndo_udp_tunnel_add = nfp_net_add_vxlan_port, 3518 .ndo_udp_tunnel_del = nfp_net_del_vxlan_port, 3519 .ndo_bpf = nfp_net_xdp, 3520 }; 3521 3522 /** 3523 * nfp_net_info() - Print general info about the NIC 3524 * @nn: NFP Net device to reconfigure 3525 */ 3526 void nfp_net_info(struct nfp_net *nn) 3527 { 3528 nn_info(nn, "Netronome NFP-6xxx %sNetdev: TxQs=%d/%d RxQs=%d/%d\n", 3529 nn->dp.is_vf ? "VF " : "", 3530 nn->dp.num_tx_rings, nn->max_tx_rings, 3531 nn->dp.num_rx_rings, nn->max_rx_rings); 3532 nn_info(nn, "VER: %d.%d.%d.%d, Maximum supported MTU: %d\n", 3533 nn->fw_ver.resv, nn->fw_ver.class, 3534 nn->fw_ver.major, nn->fw_ver.minor, 3535 nn->max_mtu); 3536 nn_info(nn, "CAP: %#x %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n", 3537 nn->cap, 3538 nn->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "", 3539 nn->cap & NFP_NET_CFG_CTRL_L2BC ? "L2BCFILT " : "", 3540 nn->cap & NFP_NET_CFG_CTRL_L2MC ? "L2MCFILT " : "", 3541 nn->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "", 3542 nn->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "", 3543 nn->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "", 3544 nn->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "", 3545 nn->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "", 3546 nn->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "", 3547 nn->cap & NFP_NET_CFG_CTRL_LSO ? "TSO1 " : "", 3548 nn->cap & NFP_NET_CFG_CTRL_LSO2 ? "TSO2 " : "", 3549 nn->cap & NFP_NET_CFG_CTRL_RSS ? "RSS1 " : "", 3550 nn->cap & NFP_NET_CFG_CTRL_RSS2 ? "RSS2 " : "", 3551 nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER ? "CTAG_FILTER " : "", 3552 nn->cap & NFP_NET_CFG_CTRL_L2SWITCH ? "L2SWITCH " : "", 3553 nn->cap & NFP_NET_CFG_CTRL_MSIXAUTO ? "AUTOMASK " : "", 3554 nn->cap & NFP_NET_CFG_CTRL_IRQMOD ? "IRQMOD " : "", 3555 nn->cap & NFP_NET_CFG_CTRL_VXLAN ? "VXLAN " : "", 3556 nn->cap & NFP_NET_CFG_CTRL_NVGRE ? "NVGRE " : "", 3557 nn->cap & NFP_NET_CFG_CTRL_CSUM_COMPLETE ? 3558 "RXCSUM_COMPLETE " : "", 3559 nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR ? "LIVE_ADDR " : "", 3560 nfp_app_extra_cap(nn->app, nn)); 3561 } 3562 3563 /** 3564 * nfp_net_alloc() - Allocate netdev and related structure 3565 * @pdev: PCI device 3566 * @needs_netdev: Whether to allocate a netdev for this vNIC 3567 * @max_tx_rings: Maximum number of TX rings supported by device 3568 * @max_rx_rings: Maximum number of RX rings supported by device 3569 * 3570 * This function allocates a netdev device and fills in the initial 3571 * part of the @struct nfp_net structure. In case of control device 3572 * nfp_net structure is allocated without the netdev. 3573 * 3574 * Return: NFP Net device structure, or ERR_PTR on error. 3575 */ 3576 struct nfp_net *nfp_net_alloc(struct pci_dev *pdev, bool needs_netdev, 3577 unsigned int max_tx_rings, 3578 unsigned int max_rx_rings) 3579 { 3580 struct nfp_net *nn; 3581 3582 if (needs_netdev) { 3583 struct net_device *netdev; 3584 3585 netdev = alloc_etherdev_mqs(sizeof(struct nfp_net), 3586 max_tx_rings, max_rx_rings); 3587 if (!netdev) 3588 return ERR_PTR(-ENOMEM); 3589 3590 SET_NETDEV_DEV(netdev, &pdev->dev); 3591 nn = netdev_priv(netdev); 3592 nn->dp.netdev = netdev; 3593 } else { 3594 nn = vzalloc(sizeof(*nn)); 3595 if (!nn) 3596 return ERR_PTR(-ENOMEM); 3597 } 3598 3599 nn->dp.dev = &pdev->dev; 3600 nn->pdev = pdev; 3601 3602 nn->max_tx_rings = max_tx_rings; 3603 nn->max_rx_rings = max_rx_rings; 3604 3605 nn->dp.num_tx_rings = min_t(unsigned int, 3606 max_tx_rings, num_online_cpus()); 3607 nn->dp.num_rx_rings = min_t(unsigned int, max_rx_rings, 3608 netif_get_num_default_rss_queues()); 3609 3610 nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings); 3611 nn->dp.num_r_vecs = min_t(unsigned int, 3612 nn->dp.num_r_vecs, num_online_cpus()); 3613 3614 nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT; 3615 nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT; 3616 3617 spin_lock_init(&nn->reconfig_lock); 3618 spin_lock_init(&nn->link_status_lock); 3619 3620 timer_setup(&nn->reconfig_timer, nfp_net_reconfig_timer, 0); 3621 3622 return nn; 3623 } 3624 3625 /** 3626 * nfp_net_free() - Undo what @nfp_net_alloc() did 3627 * @nn: NFP Net device to reconfigure 3628 */ 3629 void nfp_net_free(struct nfp_net *nn) 3630 { 3631 if (nn->dp.netdev) 3632 free_netdev(nn->dp.netdev); 3633 else 3634 vfree(nn); 3635 } 3636 3637 /** 3638 * nfp_net_rss_key_sz() - Get current size of the RSS key 3639 * @nn: NFP Net device instance 3640 * 3641 * Return: size of the RSS key for currently selected hash function. 3642 */ 3643 unsigned int nfp_net_rss_key_sz(struct nfp_net *nn) 3644 { 3645 switch (nn->rss_hfunc) { 3646 case ETH_RSS_HASH_TOP: 3647 return NFP_NET_CFG_RSS_KEY_SZ; 3648 case ETH_RSS_HASH_XOR: 3649 return 0; 3650 case ETH_RSS_HASH_CRC32: 3651 return 4; 3652 } 3653 3654 nn_warn(nn, "Unknown hash function: %u\n", nn->rss_hfunc); 3655 return 0; 3656 } 3657 3658 /** 3659 * nfp_net_rss_init() - Set the initial RSS parameters 3660 * @nn: NFP Net device to reconfigure 3661 */ 3662 static void nfp_net_rss_init(struct nfp_net *nn) 3663 { 3664 unsigned long func_bit, rss_cap_hfunc; 3665 u32 reg; 3666 3667 /* Read the RSS function capability and select first supported func */ 3668 reg = nn_readl(nn, NFP_NET_CFG_RSS_CAP); 3669 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, reg); 3670 if (!rss_cap_hfunc) 3671 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, 3672 NFP_NET_CFG_RSS_TOEPLITZ); 3673 3674 func_bit = find_first_bit(&rss_cap_hfunc, NFP_NET_CFG_RSS_HFUNCS); 3675 if (func_bit == NFP_NET_CFG_RSS_HFUNCS) { 3676 dev_warn(nn->dp.dev, 3677 "Bad RSS config, defaulting to Toeplitz hash\n"); 3678 func_bit = ETH_RSS_HASH_TOP_BIT; 3679 } 3680 nn->rss_hfunc = 1 << func_bit; 3681 3682 netdev_rss_key_fill(nn->rss_key, nfp_net_rss_key_sz(nn)); 3683 3684 nfp_net_rss_init_itbl(nn); 3685 3686 /* Enable IPv4/IPv6 TCP by default */ 3687 nn->rss_cfg = NFP_NET_CFG_RSS_IPV4_TCP | 3688 NFP_NET_CFG_RSS_IPV6_TCP | 3689 FIELD_PREP(NFP_NET_CFG_RSS_HFUNC, nn->rss_hfunc) | 3690 NFP_NET_CFG_RSS_MASK; 3691 } 3692 3693 /** 3694 * nfp_net_irqmod_init() - Set the initial IRQ moderation parameters 3695 * @nn: NFP Net device to reconfigure 3696 */ 3697 static void nfp_net_irqmod_init(struct nfp_net *nn) 3698 { 3699 nn->rx_coalesce_usecs = 50; 3700 nn->rx_coalesce_max_frames = 64; 3701 nn->tx_coalesce_usecs = 50; 3702 nn->tx_coalesce_max_frames = 64; 3703 } 3704 3705 static void nfp_net_netdev_init(struct nfp_net *nn) 3706 { 3707 struct net_device *netdev = nn->dp.netdev; 3708 3709 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr); 3710 3711 netdev->mtu = nn->dp.mtu; 3712 3713 /* Advertise/enable offloads based on capabilities 3714 * 3715 * Note: netdev->features show the currently enabled features 3716 * and netdev->hw_features advertises which features are 3717 * supported. By default we enable most features. 3718 */ 3719 if (nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR) 3720 netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 3721 3722 netdev->hw_features = NETIF_F_HIGHDMA; 3723 if (nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY) { 3724 netdev->hw_features |= NETIF_F_RXCSUM; 3725 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY; 3726 } 3727 if (nn->cap & NFP_NET_CFG_CTRL_TXCSUM) { 3728 netdev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 3729 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXCSUM; 3730 } 3731 if (nn->cap & NFP_NET_CFG_CTRL_GATHER) { 3732 netdev->hw_features |= NETIF_F_SG; 3733 nn->dp.ctrl |= NFP_NET_CFG_CTRL_GATHER; 3734 } 3735 if ((nn->cap & NFP_NET_CFG_CTRL_LSO && nn->fw_ver.major > 2) || 3736 nn->cap & NFP_NET_CFG_CTRL_LSO2) { 3737 netdev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 3738 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?: 3739 NFP_NET_CFG_CTRL_LSO; 3740 } 3741 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) 3742 netdev->hw_features |= NETIF_F_RXHASH; 3743 if (nn->cap & NFP_NET_CFG_CTRL_VXLAN && 3744 nn->cap & NFP_NET_CFG_CTRL_NVGRE) { 3745 if (nn->cap & NFP_NET_CFG_CTRL_LSO) 3746 netdev->hw_features |= NETIF_F_GSO_GRE | 3747 NETIF_F_GSO_UDP_TUNNEL; 3748 nn->dp.ctrl |= NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE; 3749 3750 netdev->hw_enc_features = netdev->hw_features; 3751 } 3752 3753 netdev->vlan_features = netdev->hw_features; 3754 3755 if (nn->cap & NFP_NET_CFG_CTRL_RXVLAN) { 3756 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX; 3757 nn->dp.ctrl |= NFP_NET_CFG_CTRL_RXVLAN; 3758 } 3759 if (nn->cap & NFP_NET_CFG_CTRL_TXVLAN) { 3760 if (nn->cap & NFP_NET_CFG_CTRL_LSO2) { 3761 nn_warn(nn, "Device advertises both TSO2 and TXVLAN. Refusing to enable TXVLAN.\n"); 3762 } else { 3763 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX; 3764 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXVLAN; 3765 } 3766 } 3767 if (nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER) { 3768 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER; 3769 nn->dp.ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER; 3770 } 3771 3772 netdev->features = netdev->hw_features; 3773 3774 if (nfp_app_has_tc(nn->app) && nn->port) 3775 netdev->hw_features |= NETIF_F_HW_TC; 3776 3777 /* Advertise but disable TSO by default. */ 3778 netdev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6); 3779 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY; 3780 3781 /* Finalise the netdev setup */ 3782 netdev->netdev_ops = &nfp_net_netdev_ops; 3783 netdev->watchdog_timeo = msecs_to_jiffies(5 * 1000); 3784 3785 SWITCHDEV_SET_OPS(netdev, &nfp_port_switchdev_ops); 3786 3787 /* MTU range: 68 - hw-specific max */ 3788 netdev->min_mtu = ETH_MIN_MTU; 3789 netdev->max_mtu = nn->max_mtu; 3790 3791 netdev->gso_max_segs = NFP_NET_LSO_MAX_SEGS; 3792 3793 netif_carrier_off(netdev); 3794 3795 nfp_net_set_ethtool_ops(netdev); 3796 } 3797 3798 static int nfp_net_read_caps(struct nfp_net *nn) 3799 { 3800 /* Get some of the read-only fields from the BAR */ 3801 nn->cap = nn_readl(nn, NFP_NET_CFG_CAP); 3802 nn->max_mtu = nn_readl(nn, NFP_NET_CFG_MAX_MTU); 3803 3804 /* ABI 4.x and ctrl vNIC always use chained metadata, in other cases 3805 * we allow use of non-chained metadata if RSS(v1) is the only 3806 * advertised capability requiring metadata. 3807 */ 3808 nn->dp.chained_metadata_format = nn->fw_ver.major == 4 || 3809 !nn->dp.netdev || 3810 !(nn->cap & NFP_NET_CFG_CTRL_RSS) || 3811 nn->cap & NFP_NET_CFG_CTRL_CHAIN_META; 3812 /* RSS(v1) uses non-chained metadata format, except in ABI 4.x where 3813 * it has the same meaning as RSSv2. 3814 */ 3815 if (nn->dp.chained_metadata_format && nn->fw_ver.major != 4) 3816 nn->cap &= ~NFP_NET_CFG_CTRL_RSS; 3817 3818 /* Determine RX packet/metadata boundary offset */ 3819 if (nn->fw_ver.major >= 2) { 3820 u32 reg; 3821 3822 reg = nn_readl(nn, NFP_NET_CFG_RX_OFFSET); 3823 if (reg > NFP_NET_MAX_PREPEND) { 3824 nn_err(nn, "Invalid rx offset: %d\n", reg); 3825 return -EINVAL; 3826 } 3827 nn->dp.rx_offset = reg; 3828 } else { 3829 nn->dp.rx_offset = NFP_NET_RX_OFFSET; 3830 } 3831 3832 /* For control vNICs mask out the capabilities app doesn't want. */ 3833 if (!nn->dp.netdev) 3834 nn->cap &= nn->app->type->ctrl_cap_mask; 3835 3836 return 0; 3837 } 3838 3839 /** 3840 * nfp_net_init() - Initialise/finalise the nfp_net structure 3841 * @nn: NFP Net device structure 3842 * 3843 * Return: 0 on success or negative errno on error. 3844 */ 3845 int nfp_net_init(struct nfp_net *nn) 3846 { 3847 int err; 3848 3849 nn->dp.rx_dma_dir = DMA_FROM_DEVICE; 3850 3851 err = nfp_net_read_caps(nn); 3852 if (err) 3853 return err; 3854 3855 /* Set default MTU and Freelist buffer size */ 3856 if (nn->max_mtu < NFP_NET_DEFAULT_MTU) 3857 nn->dp.mtu = nn->max_mtu; 3858 else 3859 nn->dp.mtu = NFP_NET_DEFAULT_MTU; 3860 nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp); 3861 3862 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) { 3863 nfp_net_rss_init(nn); 3864 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?: 3865 NFP_NET_CFG_CTRL_RSS; 3866 } 3867 3868 /* Allow L2 Broadcast and Multicast through by default, if supported */ 3869 if (nn->cap & NFP_NET_CFG_CTRL_L2BC) 3870 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2BC; 3871 3872 /* Allow IRQ moderation, if supported */ 3873 if (nn->cap & NFP_NET_CFG_CTRL_IRQMOD) { 3874 nfp_net_irqmod_init(nn); 3875 nn->dp.ctrl |= NFP_NET_CFG_CTRL_IRQMOD; 3876 } 3877 3878 err = nfp_net_tlv_caps_parse(&nn->pdev->dev, nn->dp.ctrl_bar, 3879 &nn->tlv_caps); 3880 if (err) 3881 return err; 3882 3883 if (nn->dp.netdev) 3884 nfp_net_netdev_init(nn); 3885 3886 /* Stash the re-configuration queue away. First odd queue in TX Bar */ 3887 nn->qcp_cfg = nn->tx_bar + NFP_QCP_QUEUE_ADDR_SZ; 3888 3889 /* Make sure the FW knows the netdev is supposed to be disabled here */ 3890 nn_writel(nn, NFP_NET_CFG_CTRL, 0); 3891 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0); 3892 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0); 3893 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_RING | 3894 NFP_NET_CFG_UPDATE_GEN); 3895 if (err) 3896 return err; 3897 3898 nfp_net_vecs_init(nn); 3899 3900 if (!nn->dp.netdev) 3901 return 0; 3902 return register_netdev(nn->dp.netdev); 3903 } 3904 3905 /** 3906 * nfp_net_clean() - Undo what nfp_net_init() did. 3907 * @nn: NFP Net device structure 3908 */ 3909 void nfp_net_clean(struct nfp_net *nn) 3910 { 3911 if (!nn->dp.netdev) 3912 return; 3913 3914 unregister_netdev(nn->dp.netdev); 3915 } 3916