1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) 2 /* Copyright (C) 2015-2018 Netronome Systems, Inc. */ 3 4 /* 5 * nfp_net_common.c 6 * Netronome network device driver: Common functions between PF and VF 7 * Authors: Jakub Kicinski <jakub.kicinski@netronome.com> 8 * Jason McMullan <jason.mcmullan@netronome.com> 9 * Rolf Neugebauer <rolf.neugebauer@netronome.com> 10 * Brad Petrus <brad.petrus@netronome.com> 11 * Chris Telfer <chris.telfer@netronome.com> 12 */ 13 14 #include <linux/bitfield.h> 15 #include <linux/bpf.h> 16 #include <linux/bpf_trace.h> 17 #include <linux/module.h> 18 #include <linux/kernel.h> 19 #include <linux/init.h> 20 #include <linux/fs.h> 21 #include <linux/netdevice.h> 22 #include <linux/etherdevice.h> 23 #include <linux/interrupt.h> 24 #include <linux/ip.h> 25 #include <linux/ipv6.h> 26 #include <linux/mm.h> 27 #include <linux/overflow.h> 28 #include <linux/page_ref.h> 29 #include <linux/pci.h> 30 #include <linux/pci_regs.h> 31 #include <linux/msi.h> 32 #include <linux/ethtool.h> 33 #include <linux/log2.h> 34 #include <linux/if_vlan.h> 35 #include <linux/random.h> 36 #include <linux/vmalloc.h> 37 #include <linux/ktime.h> 38 39 #include <net/switchdev.h> 40 #include <net/vxlan.h> 41 42 #include "nfpcore/nfp_nsp.h" 43 #include "nfp_app.h" 44 #include "nfp_net_ctrl.h" 45 #include "nfp_net.h" 46 #include "nfp_net_sriov.h" 47 #include "nfp_port.h" 48 49 /** 50 * nfp_net_get_fw_version() - Read and parse the FW version 51 * @fw_ver: Output fw_version structure to read to 52 * @ctrl_bar: Mapped address of the control BAR 53 */ 54 void nfp_net_get_fw_version(struct nfp_net_fw_version *fw_ver, 55 void __iomem *ctrl_bar) 56 { 57 u32 reg; 58 59 reg = readl(ctrl_bar + NFP_NET_CFG_VERSION); 60 put_unaligned_le32(reg, fw_ver); 61 } 62 63 static dma_addr_t nfp_net_dma_map_rx(struct nfp_net_dp *dp, void *frag) 64 { 65 return dma_map_single_attrs(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM, 66 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA, 67 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC); 68 } 69 70 static void 71 nfp_net_dma_sync_dev_rx(const struct nfp_net_dp *dp, dma_addr_t dma_addr) 72 { 73 dma_sync_single_for_device(dp->dev, dma_addr, 74 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA, 75 dp->rx_dma_dir); 76 } 77 78 static void nfp_net_dma_unmap_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr) 79 { 80 dma_unmap_single_attrs(dp->dev, dma_addr, 81 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA, 82 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC); 83 } 84 85 static void nfp_net_dma_sync_cpu_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr, 86 unsigned int len) 87 { 88 dma_sync_single_for_cpu(dp->dev, dma_addr - NFP_NET_RX_BUF_HEADROOM, 89 len, dp->rx_dma_dir); 90 } 91 92 /* Firmware reconfig 93 * 94 * Firmware reconfig may take a while so we have two versions of it - 95 * synchronous and asynchronous (posted). All synchronous callers are holding 96 * RTNL so we don't have to worry about serializing them. 97 */ 98 static void nfp_net_reconfig_start(struct nfp_net *nn, u32 update) 99 { 100 nn_writel(nn, NFP_NET_CFG_UPDATE, update); 101 /* ensure update is written before pinging HW */ 102 nn_pci_flush(nn); 103 nfp_qcp_wr_ptr_add(nn->qcp_cfg, 1); 104 } 105 106 /* Pass 0 as update to run posted reconfigs. */ 107 static void nfp_net_reconfig_start_async(struct nfp_net *nn, u32 update) 108 { 109 update |= nn->reconfig_posted; 110 nn->reconfig_posted = 0; 111 112 nfp_net_reconfig_start(nn, update); 113 114 nn->reconfig_timer_active = true; 115 mod_timer(&nn->reconfig_timer, jiffies + NFP_NET_POLL_TIMEOUT * HZ); 116 } 117 118 static bool nfp_net_reconfig_check_done(struct nfp_net *nn, bool last_check) 119 { 120 u32 reg; 121 122 reg = nn_readl(nn, NFP_NET_CFG_UPDATE); 123 if (reg == 0) 124 return true; 125 if (reg & NFP_NET_CFG_UPDATE_ERR) { 126 nn_err(nn, "Reconfig error: 0x%08x\n", reg); 127 return true; 128 } else if (last_check) { 129 nn_err(nn, "Reconfig timeout: 0x%08x\n", reg); 130 return true; 131 } 132 133 return false; 134 } 135 136 static int nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline) 137 { 138 bool timed_out = false; 139 140 /* Poll update field, waiting for NFP to ack the config */ 141 while (!nfp_net_reconfig_check_done(nn, timed_out)) { 142 msleep(1); 143 timed_out = time_is_before_eq_jiffies(deadline); 144 } 145 146 if (nn_readl(nn, NFP_NET_CFG_UPDATE) & NFP_NET_CFG_UPDATE_ERR) 147 return -EIO; 148 149 return timed_out ? -EIO : 0; 150 } 151 152 static void nfp_net_reconfig_timer(struct timer_list *t) 153 { 154 struct nfp_net *nn = from_timer(nn, t, reconfig_timer); 155 156 spin_lock_bh(&nn->reconfig_lock); 157 158 nn->reconfig_timer_active = false; 159 160 /* If sync caller is present it will take over from us */ 161 if (nn->reconfig_sync_present) 162 goto done; 163 164 /* Read reconfig status and report errors */ 165 nfp_net_reconfig_check_done(nn, true); 166 167 if (nn->reconfig_posted) 168 nfp_net_reconfig_start_async(nn, 0); 169 done: 170 spin_unlock_bh(&nn->reconfig_lock); 171 } 172 173 /** 174 * nfp_net_reconfig_post() - Post async reconfig request 175 * @nn: NFP Net device to reconfigure 176 * @update: The value for the update field in the BAR config 177 * 178 * Record FW reconfiguration request. Reconfiguration will be kicked off 179 * whenever reconfiguration machinery is idle. Multiple requests can be 180 * merged together! 181 */ 182 static void nfp_net_reconfig_post(struct nfp_net *nn, u32 update) 183 { 184 spin_lock_bh(&nn->reconfig_lock); 185 186 /* Sync caller will kick off async reconf when it's done, just post */ 187 if (nn->reconfig_sync_present) { 188 nn->reconfig_posted |= update; 189 goto done; 190 } 191 192 /* Opportunistically check if the previous command is done */ 193 if (!nn->reconfig_timer_active || 194 nfp_net_reconfig_check_done(nn, false)) 195 nfp_net_reconfig_start_async(nn, update); 196 else 197 nn->reconfig_posted |= update; 198 done: 199 spin_unlock_bh(&nn->reconfig_lock); 200 } 201 202 static void nfp_net_reconfig_sync_enter(struct nfp_net *nn) 203 { 204 bool cancelled_timer = false; 205 u32 pre_posted_requests; 206 207 spin_lock_bh(&nn->reconfig_lock); 208 209 nn->reconfig_sync_present = true; 210 211 if (nn->reconfig_timer_active) { 212 nn->reconfig_timer_active = false; 213 cancelled_timer = true; 214 } 215 pre_posted_requests = nn->reconfig_posted; 216 nn->reconfig_posted = 0; 217 218 spin_unlock_bh(&nn->reconfig_lock); 219 220 if (cancelled_timer) { 221 del_timer_sync(&nn->reconfig_timer); 222 nfp_net_reconfig_wait(nn, nn->reconfig_timer.expires); 223 } 224 225 /* Run the posted reconfigs which were issued before we started */ 226 if (pre_posted_requests) { 227 nfp_net_reconfig_start(nn, pre_posted_requests); 228 nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT); 229 } 230 } 231 232 static void nfp_net_reconfig_wait_posted(struct nfp_net *nn) 233 { 234 nfp_net_reconfig_sync_enter(nn); 235 236 spin_lock_bh(&nn->reconfig_lock); 237 nn->reconfig_sync_present = false; 238 spin_unlock_bh(&nn->reconfig_lock); 239 } 240 241 /** 242 * nfp_net_reconfig() - Reconfigure the firmware 243 * @nn: NFP Net device to reconfigure 244 * @update: The value for the update field in the BAR config 245 * 246 * Write the update word to the BAR and ping the reconfig queue. The 247 * poll until the firmware has acknowledged the update by zeroing the 248 * update word. 249 * 250 * Return: Negative errno on error, 0 on success 251 */ 252 int nfp_net_reconfig(struct nfp_net *nn, u32 update) 253 { 254 int ret; 255 256 nfp_net_reconfig_sync_enter(nn); 257 258 nfp_net_reconfig_start(nn, update); 259 ret = nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT); 260 261 spin_lock_bh(&nn->reconfig_lock); 262 263 if (nn->reconfig_posted) 264 nfp_net_reconfig_start_async(nn, 0); 265 266 nn->reconfig_sync_present = false; 267 268 spin_unlock_bh(&nn->reconfig_lock); 269 270 return ret; 271 } 272 273 /** 274 * nfp_net_reconfig_mbox() - Reconfigure the firmware via the mailbox 275 * @nn: NFP Net device to reconfigure 276 * @mbox_cmd: The value for the mailbox command 277 * 278 * Helper function for mailbox updates 279 * 280 * Return: Negative errno on error, 0 on success 281 */ 282 int nfp_net_reconfig_mbox(struct nfp_net *nn, u32 mbox_cmd) 283 { 284 u32 mbox = nn->tlv_caps.mbox_off; 285 int ret; 286 287 if (!nfp_net_has_mbox(&nn->tlv_caps)) { 288 nn_err(nn, "no mailbox present, command: %u\n", mbox_cmd); 289 return -EIO; 290 } 291 292 nn_writeq(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_CMD, mbox_cmd); 293 294 ret = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MBOX); 295 if (ret) { 296 nn_err(nn, "Mailbox update error\n"); 297 return ret; 298 } 299 300 return -nn_readl(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_RET); 301 } 302 303 /* Interrupt configuration and handling 304 */ 305 306 /** 307 * nfp_net_irq_unmask() - Unmask automasked interrupt 308 * @nn: NFP Network structure 309 * @entry_nr: MSI-X table entry 310 * 311 * Clear the ICR for the IRQ entry. 312 */ 313 static void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr) 314 { 315 nn_writeb(nn, NFP_NET_CFG_ICR(entry_nr), NFP_NET_CFG_ICR_UNMASKED); 316 nn_pci_flush(nn); 317 } 318 319 /** 320 * nfp_net_irqs_alloc() - allocates MSI-X irqs 321 * @pdev: PCI device structure 322 * @irq_entries: Array to be initialized and used to hold the irq entries 323 * @min_irqs: Minimal acceptable number of interrupts 324 * @wanted_irqs: Target number of interrupts to allocate 325 * 326 * Return: Number of irqs obtained or 0 on error. 327 */ 328 unsigned int 329 nfp_net_irqs_alloc(struct pci_dev *pdev, struct msix_entry *irq_entries, 330 unsigned int min_irqs, unsigned int wanted_irqs) 331 { 332 unsigned int i; 333 int got_irqs; 334 335 for (i = 0; i < wanted_irqs; i++) 336 irq_entries[i].entry = i; 337 338 got_irqs = pci_enable_msix_range(pdev, irq_entries, 339 min_irqs, wanted_irqs); 340 if (got_irqs < 0) { 341 dev_err(&pdev->dev, "Failed to enable %d-%d MSI-X (err=%d)\n", 342 min_irqs, wanted_irqs, got_irqs); 343 return 0; 344 } 345 346 if (got_irqs < wanted_irqs) 347 dev_warn(&pdev->dev, "Unable to allocate %d IRQs got only %d\n", 348 wanted_irqs, got_irqs); 349 350 return got_irqs; 351 } 352 353 /** 354 * nfp_net_irqs_assign() - Assign interrupts allocated externally to netdev 355 * @nn: NFP Network structure 356 * @irq_entries: Table of allocated interrupts 357 * @n: Size of @irq_entries (number of entries to grab) 358 * 359 * After interrupts are allocated with nfp_net_irqs_alloc() this function 360 * should be called to assign them to a specific netdev (port). 361 */ 362 void 363 nfp_net_irqs_assign(struct nfp_net *nn, struct msix_entry *irq_entries, 364 unsigned int n) 365 { 366 struct nfp_net_dp *dp = &nn->dp; 367 368 nn->max_r_vecs = n - NFP_NET_NON_Q_VECTORS; 369 dp->num_r_vecs = nn->max_r_vecs; 370 371 memcpy(nn->irq_entries, irq_entries, sizeof(*irq_entries) * n); 372 373 if (dp->num_rx_rings > dp->num_r_vecs || 374 dp->num_tx_rings > dp->num_r_vecs) 375 dev_warn(nn->dp.dev, "More rings (%d,%d) than vectors (%d).\n", 376 dp->num_rx_rings, dp->num_tx_rings, 377 dp->num_r_vecs); 378 379 dp->num_rx_rings = min(dp->num_r_vecs, dp->num_rx_rings); 380 dp->num_tx_rings = min(dp->num_r_vecs, dp->num_tx_rings); 381 dp->num_stack_tx_rings = dp->num_tx_rings; 382 } 383 384 /** 385 * nfp_net_irqs_disable() - Disable interrupts 386 * @pdev: PCI device structure 387 * 388 * Undoes what @nfp_net_irqs_alloc() does. 389 */ 390 void nfp_net_irqs_disable(struct pci_dev *pdev) 391 { 392 pci_disable_msix(pdev); 393 } 394 395 /** 396 * nfp_net_irq_rxtx() - Interrupt service routine for RX/TX rings. 397 * @irq: Interrupt 398 * @data: Opaque data structure 399 * 400 * Return: Indicate if the interrupt has been handled. 401 */ 402 static irqreturn_t nfp_net_irq_rxtx(int irq, void *data) 403 { 404 struct nfp_net_r_vector *r_vec = data; 405 406 napi_schedule_irqoff(&r_vec->napi); 407 408 /* The FW auto-masks any interrupt, either via the MASK bit in 409 * the MSI-X table or via the per entry ICR field. So there 410 * is no need to disable interrupts here. 411 */ 412 return IRQ_HANDLED; 413 } 414 415 static irqreturn_t nfp_ctrl_irq_rxtx(int irq, void *data) 416 { 417 struct nfp_net_r_vector *r_vec = data; 418 419 tasklet_schedule(&r_vec->tasklet); 420 421 return IRQ_HANDLED; 422 } 423 424 /** 425 * nfp_net_read_link_status() - Reread link status from control BAR 426 * @nn: NFP Network structure 427 */ 428 static void nfp_net_read_link_status(struct nfp_net *nn) 429 { 430 unsigned long flags; 431 bool link_up; 432 u32 sts; 433 434 spin_lock_irqsave(&nn->link_status_lock, flags); 435 436 sts = nn_readl(nn, NFP_NET_CFG_STS); 437 link_up = !!(sts & NFP_NET_CFG_STS_LINK); 438 439 if (nn->link_up == link_up) 440 goto out; 441 442 nn->link_up = link_up; 443 if (nn->port) 444 set_bit(NFP_PORT_CHANGED, &nn->port->flags); 445 446 if (nn->link_up) { 447 netif_carrier_on(nn->dp.netdev); 448 netdev_info(nn->dp.netdev, "NIC Link is Up\n"); 449 } else { 450 netif_carrier_off(nn->dp.netdev); 451 netdev_info(nn->dp.netdev, "NIC Link is Down\n"); 452 } 453 out: 454 spin_unlock_irqrestore(&nn->link_status_lock, flags); 455 } 456 457 /** 458 * nfp_net_irq_lsc() - Interrupt service routine for link state changes 459 * @irq: Interrupt 460 * @data: Opaque data structure 461 * 462 * Return: Indicate if the interrupt has been handled. 463 */ 464 static irqreturn_t nfp_net_irq_lsc(int irq, void *data) 465 { 466 struct nfp_net *nn = data; 467 struct msix_entry *entry; 468 469 entry = &nn->irq_entries[NFP_NET_IRQ_LSC_IDX]; 470 471 nfp_net_read_link_status(nn); 472 473 nfp_net_irq_unmask(nn, entry->entry); 474 475 return IRQ_HANDLED; 476 } 477 478 /** 479 * nfp_net_irq_exn() - Interrupt service routine for exceptions 480 * @irq: Interrupt 481 * @data: Opaque data structure 482 * 483 * Return: Indicate if the interrupt has been handled. 484 */ 485 static irqreturn_t nfp_net_irq_exn(int irq, void *data) 486 { 487 struct nfp_net *nn = data; 488 489 nn_err(nn, "%s: UNIMPLEMENTED.\n", __func__); 490 /* XXX TO BE IMPLEMENTED */ 491 return IRQ_HANDLED; 492 } 493 494 /** 495 * nfp_net_tx_ring_init() - Fill in the boilerplate for a TX ring 496 * @tx_ring: TX ring structure 497 * @r_vec: IRQ vector servicing this ring 498 * @idx: Ring index 499 * @is_xdp: Is this an XDP TX ring? 500 */ 501 static void 502 nfp_net_tx_ring_init(struct nfp_net_tx_ring *tx_ring, 503 struct nfp_net_r_vector *r_vec, unsigned int idx, 504 bool is_xdp) 505 { 506 struct nfp_net *nn = r_vec->nfp_net; 507 508 tx_ring->idx = idx; 509 tx_ring->r_vec = r_vec; 510 tx_ring->is_xdp = is_xdp; 511 u64_stats_init(&tx_ring->r_vec->tx_sync); 512 513 tx_ring->qcidx = tx_ring->idx * nn->stride_tx; 514 tx_ring->qcp_q = nn->tx_bar + NFP_QCP_QUEUE_OFF(tx_ring->qcidx); 515 } 516 517 /** 518 * nfp_net_rx_ring_init() - Fill in the boilerplate for a RX ring 519 * @rx_ring: RX ring structure 520 * @r_vec: IRQ vector servicing this ring 521 * @idx: Ring index 522 */ 523 static void 524 nfp_net_rx_ring_init(struct nfp_net_rx_ring *rx_ring, 525 struct nfp_net_r_vector *r_vec, unsigned int idx) 526 { 527 struct nfp_net *nn = r_vec->nfp_net; 528 529 rx_ring->idx = idx; 530 rx_ring->r_vec = r_vec; 531 u64_stats_init(&rx_ring->r_vec->rx_sync); 532 533 rx_ring->fl_qcidx = rx_ring->idx * nn->stride_rx; 534 rx_ring->qcp_fl = nn->rx_bar + NFP_QCP_QUEUE_OFF(rx_ring->fl_qcidx); 535 } 536 537 /** 538 * nfp_net_aux_irq_request() - Request an auxiliary interrupt (LSC or EXN) 539 * @nn: NFP Network structure 540 * @ctrl_offset: Control BAR offset where IRQ configuration should be written 541 * @format: printf-style format to construct the interrupt name 542 * @name: Pointer to allocated space for interrupt name 543 * @name_sz: Size of space for interrupt name 544 * @vector_idx: Index of MSI-X vector used for this interrupt 545 * @handler: IRQ handler to register for this interrupt 546 */ 547 static int 548 nfp_net_aux_irq_request(struct nfp_net *nn, u32 ctrl_offset, 549 const char *format, char *name, size_t name_sz, 550 unsigned int vector_idx, irq_handler_t handler) 551 { 552 struct msix_entry *entry; 553 int err; 554 555 entry = &nn->irq_entries[vector_idx]; 556 557 snprintf(name, name_sz, format, nfp_net_name(nn)); 558 err = request_irq(entry->vector, handler, 0, name, nn); 559 if (err) { 560 nn_err(nn, "Failed to request IRQ %d (err=%d).\n", 561 entry->vector, err); 562 return err; 563 } 564 nn_writeb(nn, ctrl_offset, entry->entry); 565 nfp_net_irq_unmask(nn, entry->entry); 566 567 return 0; 568 } 569 570 /** 571 * nfp_net_aux_irq_free() - Free an auxiliary interrupt (LSC or EXN) 572 * @nn: NFP Network structure 573 * @ctrl_offset: Control BAR offset where IRQ configuration should be written 574 * @vector_idx: Index of MSI-X vector used for this interrupt 575 */ 576 static void nfp_net_aux_irq_free(struct nfp_net *nn, u32 ctrl_offset, 577 unsigned int vector_idx) 578 { 579 nn_writeb(nn, ctrl_offset, 0xff); 580 nn_pci_flush(nn); 581 free_irq(nn->irq_entries[vector_idx].vector, nn); 582 } 583 584 /* Transmit 585 * 586 * One queue controller peripheral queue is used for transmit. The 587 * driver en-queues packets for transmit by advancing the write 588 * pointer. The device indicates that packets have transmitted by 589 * advancing the read pointer. The driver maintains a local copy of 590 * the read and write pointer in @struct nfp_net_tx_ring. The driver 591 * keeps @wr_p in sync with the queue controller write pointer and can 592 * determine how many packets have been transmitted by comparing its 593 * copy of the read pointer @rd_p with the read pointer maintained by 594 * the queue controller peripheral. 595 */ 596 597 /** 598 * nfp_net_tx_full() - Check if the TX ring is full 599 * @tx_ring: TX ring to check 600 * @dcnt: Number of descriptors that need to be enqueued (must be >= 1) 601 * 602 * This function checks, based on the *host copy* of read/write 603 * pointer if a given TX ring is full. The real TX queue may have 604 * some newly made available slots. 605 * 606 * Return: True if the ring is full. 607 */ 608 static int nfp_net_tx_full(struct nfp_net_tx_ring *tx_ring, int dcnt) 609 { 610 return (tx_ring->wr_p - tx_ring->rd_p) >= (tx_ring->cnt - dcnt); 611 } 612 613 /* Wrappers for deciding when to stop and restart TX queues */ 614 static int nfp_net_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring) 615 { 616 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4); 617 } 618 619 static int nfp_net_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring) 620 { 621 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1); 622 } 623 624 /** 625 * nfp_net_tx_ring_stop() - stop tx ring 626 * @nd_q: netdev queue 627 * @tx_ring: driver tx queue structure 628 * 629 * Safely stop TX ring. Remember that while we are running .start_xmit() 630 * someone else may be cleaning the TX ring completions so we need to be 631 * extra careful here. 632 */ 633 static void nfp_net_tx_ring_stop(struct netdev_queue *nd_q, 634 struct nfp_net_tx_ring *tx_ring) 635 { 636 netif_tx_stop_queue(nd_q); 637 638 /* We can race with the TX completion out of NAPI so recheck */ 639 smp_mb(); 640 if (unlikely(nfp_net_tx_ring_should_wake(tx_ring))) 641 netif_tx_start_queue(nd_q); 642 } 643 644 /** 645 * nfp_net_tx_tso() - Set up Tx descriptor for LSO 646 * @r_vec: per-ring structure 647 * @txbuf: Pointer to driver soft TX descriptor 648 * @txd: Pointer to HW TX descriptor 649 * @skb: Pointer to SKB 650 * 651 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs. 652 * Return error on packet header greater than maximum supported LSO header size. 653 */ 654 static void nfp_net_tx_tso(struct nfp_net_r_vector *r_vec, 655 struct nfp_net_tx_buf *txbuf, 656 struct nfp_net_tx_desc *txd, struct sk_buff *skb) 657 { 658 u32 hdrlen; 659 u16 mss; 660 661 if (!skb_is_gso(skb)) 662 return; 663 664 if (!skb->encapsulation) { 665 txd->l3_offset = skb_network_offset(skb); 666 txd->l4_offset = skb_transport_offset(skb); 667 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb); 668 } else { 669 txd->l3_offset = skb_inner_network_offset(skb); 670 txd->l4_offset = skb_inner_transport_offset(skb); 671 hdrlen = skb_inner_transport_header(skb) - skb->data + 672 inner_tcp_hdrlen(skb); 673 } 674 675 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs; 676 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1); 677 678 mss = skb_shinfo(skb)->gso_size & PCIE_DESC_TX_MSS_MASK; 679 txd->lso_hdrlen = hdrlen; 680 txd->mss = cpu_to_le16(mss); 681 txd->flags |= PCIE_DESC_TX_LSO; 682 683 u64_stats_update_begin(&r_vec->tx_sync); 684 r_vec->tx_lso++; 685 u64_stats_update_end(&r_vec->tx_sync); 686 } 687 688 /** 689 * nfp_net_tx_csum() - Set TX CSUM offload flags in TX descriptor 690 * @dp: NFP Net data path struct 691 * @r_vec: per-ring structure 692 * @txbuf: Pointer to driver soft TX descriptor 693 * @txd: Pointer to TX descriptor 694 * @skb: Pointer to SKB 695 * 696 * This function sets the TX checksum flags in the TX descriptor based 697 * on the configuration and the protocol of the packet to be transmitted. 698 */ 699 static void nfp_net_tx_csum(struct nfp_net_dp *dp, 700 struct nfp_net_r_vector *r_vec, 701 struct nfp_net_tx_buf *txbuf, 702 struct nfp_net_tx_desc *txd, struct sk_buff *skb) 703 { 704 struct ipv6hdr *ipv6h; 705 struct iphdr *iph; 706 u8 l4_hdr; 707 708 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM)) 709 return; 710 711 if (skb->ip_summed != CHECKSUM_PARTIAL) 712 return; 713 714 txd->flags |= PCIE_DESC_TX_CSUM; 715 if (skb->encapsulation) 716 txd->flags |= PCIE_DESC_TX_ENCAP; 717 718 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb); 719 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb); 720 721 if (iph->version == 4) { 722 txd->flags |= PCIE_DESC_TX_IP4_CSUM; 723 l4_hdr = iph->protocol; 724 } else if (ipv6h->version == 6) { 725 l4_hdr = ipv6h->nexthdr; 726 } else { 727 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version); 728 return; 729 } 730 731 switch (l4_hdr) { 732 case IPPROTO_TCP: 733 txd->flags |= PCIE_DESC_TX_TCP_CSUM; 734 break; 735 case IPPROTO_UDP: 736 txd->flags |= PCIE_DESC_TX_UDP_CSUM; 737 break; 738 default: 739 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr); 740 return; 741 } 742 743 u64_stats_update_begin(&r_vec->tx_sync); 744 if (skb->encapsulation) 745 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt; 746 else 747 r_vec->hw_csum_tx += txbuf->pkt_cnt; 748 u64_stats_update_end(&r_vec->tx_sync); 749 } 750 751 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring) 752 { 753 wmb(); 754 nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add); 755 tx_ring->wr_ptr_add = 0; 756 } 757 758 static int nfp_net_prep_port_id(struct sk_buff *skb) 759 { 760 struct metadata_dst *md_dst = skb_metadata_dst(skb); 761 unsigned char *data; 762 763 if (likely(!md_dst)) 764 return 0; 765 if (unlikely(md_dst->type != METADATA_HW_PORT_MUX)) 766 return 0; 767 768 if (unlikely(skb_cow_head(skb, 8))) 769 return -ENOMEM; 770 771 data = skb_push(skb, 8); 772 put_unaligned_be32(NFP_NET_META_PORTID, data); 773 put_unaligned_be32(md_dst->u.port_info.port_id, data + 4); 774 775 return 8; 776 } 777 778 /** 779 * nfp_net_tx() - Main transmit entry point 780 * @skb: SKB to transmit 781 * @netdev: netdev structure 782 * 783 * Return: NETDEV_TX_OK on success. 784 */ 785 static int nfp_net_tx(struct sk_buff *skb, struct net_device *netdev) 786 { 787 struct nfp_net *nn = netdev_priv(netdev); 788 const struct skb_frag_struct *frag; 789 struct nfp_net_tx_desc *txd, txdg; 790 int f, nr_frags, wr_idx, md_bytes; 791 struct nfp_net_tx_ring *tx_ring; 792 struct nfp_net_r_vector *r_vec; 793 struct nfp_net_tx_buf *txbuf; 794 struct netdev_queue *nd_q; 795 struct nfp_net_dp *dp; 796 dma_addr_t dma_addr; 797 unsigned int fsize; 798 u16 qidx; 799 800 dp = &nn->dp; 801 qidx = skb_get_queue_mapping(skb); 802 tx_ring = &dp->tx_rings[qidx]; 803 r_vec = tx_ring->r_vec; 804 nd_q = netdev_get_tx_queue(dp->netdev, qidx); 805 806 nr_frags = skb_shinfo(skb)->nr_frags; 807 808 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) { 809 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n", 810 qidx, tx_ring->wr_p, tx_ring->rd_p); 811 netif_tx_stop_queue(nd_q); 812 nfp_net_tx_xmit_more_flush(tx_ring); 813 u64_stats_update_begin(&r_vec->tx_sync); 814 r_vec->tx_busy++; 815 u64_stats_update_end(&r_vec->tx_sync); 816 return NETDEV_TX_BUSY; 817 } 818 819 md_bytes = nfp_net_prep_port_id(skb); 820 if (unlikely(md_bytes < 0)) { 821 nfp_net_tx_xmit_more_flush(tx_ring); 822 dev_kfree_skb_any(skb); 823 return NETDEV_TX_OK; 824 } 825 826 /* Start with the head skbuf */ 827 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 828 DMA_TO_DEVICE); 829 if (dma_mapping_error(dp->dev, dma_addr)) 830 goto err_free; 831 832 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 833 834 /* Stash the soft descriptor of the head then initialize it */ 835 txbuf = &tx_ring->txbufs[wr_idx]; 836 txbuf->skb = skb; 837 txbuf->dma_addr = dma_addr; 838 txbuf->fidx = -1; 839 txbuf->pkt_cnt = 1; 840 txbuf->real_len = skb->len; 841 842 /* Build TX descriptor */ 843 txd = &tx_ring->txds[wr_idx]; 844 txd->offset_eop = (nr_frags ? 0 : PCIE_DESC_TX_EOP) | md_bytes; 845 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 846 nfp_desc_set_dma_addr(txd, dma_addr); 847 txd->data_len = cpu_to_le16(skb->len); 848 849 txd->flags = 0; 850 txd->mss = 0; 851 txd->lso_hdrlen = 0; 852 853 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */ 854 nfp_net_tx_tso(r_vec, txbuf, txd, skb); 855 nfp_net_tx_csum(dp, r_vec, txbuf, txd, skb); 856 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) { 857 txd->flags |= PCIE_DESC_TX_VLAN; 858 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 859 } 860 861 /* Gather DMA */ 862 if (nr_frags > 0) { 863 /* all descs must match except for in addr, length and eop */ 864 txdg = *txd; 865 866 for (f = 0; f < nr_frags; f++) { 867 frag = &skb_shinfo(skb)->frags[f]; 868 fsize = skb_frag_size(frag); 869 870 dma_addr = skb_frag_dma_map(dp->dev, frag, 0, 871 fsize, DMA_TO_DEVICE); 872 if (dma_mapping_error(dp->dev, dma_addr)) 873 goto err_unmap; 874 875 wr_idx = D_IDX(tx_ring, wr_idx + 1); 876 tx_ring->txbufs[wr_idx].skb = skb; 877 tx_ring->txbufs[wr_idx].dma_addr = dma_addr; 878 tx_ring->txbufs[wr_idx].fidx = f; 879 880 txd = &tx_ring->txds[wr_idx]; 881 *txd = txdg; 882 txd->dma_len = cpu_to_le16(fsize); 883 nfp_desc_set_dma_addr(txd, dma_addr); 884 txd->offset_eop |= 885 (f == nr_frags - 1) ? PCIE_DESC_TX_EOP : 0; 886 } 887 888 u64_stats_update_begin(&r_vec->tx_sync); 889 r_vec->tx_gather++; 890 u64_stats_update_end(&r_vec->tx_sync); 891 } 892 893 skb_tx_timestamp(skb); 894 895 tx_ring->wr_p += nr_frags + 1; 896 if (nfp_net_tx_ring_should_stop(tx_ring)) 897 nfp_net_tx_ring_stop(nd_q, tx_ring); 898 899 tx_ring->wr_ptr_add += nr_frags + 1; 900 if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, skb->xmit_more)) 901 nfp_net_tx_xmit_more_flush(tx_ring); 902 903 return NETDEV_TX_OK; 904 905 err_unmap: 906 while (--f >= 0) { 907 frag = &skb_shinfo(skb)->frags[f]; 908 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 909 skb_frag_size(frag), DMA_TO_DEVICE); 910 tx_ring->txbufs[wr_idx].skb = NULL; 911 tx_ring->txbufs[wr_idx].dma_addr = 0; 912 tx_ring->txbufs[wr_idx].fidx = -2; 913 wr_idx = wr_idx - 1; 914 if (wr_idx < 0) 915 wr_idx += tx_ring->cnt; 916 } 917 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 918 skb_headlen(skb), DMA_TO_DEVICE); 919 tx_ring->txbufs[wr_idx].skb = NULL; 920 tx_ring->txbufs[wr_idx].dma_addr = 0; 921 tx_ring->txbufs[wr_idx].fidx = -2; 922 err_free: 923 nn_dp_warn(dp, "Failed to map DMA TX buffer\n"); 924 nfp_net_tx_xmit_more_flush(tx_ring); 925 u64_stats_update_begin(&r_vec->tx_sync); 926 r_vec->tx_errors++; 927 u64_stats_update_end(&r_vec->tx_sync); 928 dev_kfree_skb_any(skb); 929 return NETDEV_TX_OK; 930 } 931 932 /** 933 * nfp_net_tx_complete() - Handled completed TX packets 934 * @tx_ring: TX ring structure 935 * @budget: NAPI budget (only used as bool to determine if in NAPI context) 936 */ 937 static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget) 938 { 939 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 940 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 941 const struct skb_frag_struct *frag; 942 struct netdev_queue *nd_q; 943 u32 done_pkts = 0, done_bytes = 0; 944 struct sk_buff *skb; 945 int todo, nr_frags; 946 u32 qcp_rd_p; 947 int fidx; 948 int idx; 949 950 if (tx_ring->wr_p == tx_ring->rd_p) 951 return; 952 953 /* Work out how many descriptors have been transmitted */ 954 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q); 955 956 if (qcp_rd_p == tx_ring->qcp_rd_p) 957 return; 958 959 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 960 961 while (todo--) { 962 idx = D_IDX(tx_ring, tx_ring->rd_p++); 963 964 skb = tx_ring->txbufs[idx].skb; 965 if (!skb) 966 continue; 967 968 nr_frags = skb_shinfo(skb)->nr_frags; 969 fidx = tx_ring->txbufs[idx].fidx; 970 971 if (fidx == -1) { 972 /* unmap head */ 973 dma_unmap_single(dp->dev, tx_ring->txbufs[idx].dma_addr, 974 skb_headlen(skb), DMA_TO_DEVICE); 975 976 done_pkts += tx_ring->txbufs[idx].pkt_cnt; 977 done_bytes += tx_ring->txbufs[idx].real_len; 978 } else { 979 /* unmap fragment */ 980 frag = &skb_shinfo(skb)->frags[fidx]; 981 dma_unmap_page(dp->dev, tx_ring->txbufs[idx].dma_addr, 982 skb_frag_size(frag), DMA_TO_DEVICE); 983 } 984 985 /* check for last gather fragment */ 986 if (fidx == nr_frags - 1) 987 napi_consume_skb(skb, budget); 988 989 tx_ring->txbufs[idx].dma_addr = 0; 990 tx_ring->txbufs[idx].skb = NULL; 991 tx_ring->txbufs[idx].fidx = -2; 992 } 993 994 tx_ring->qcp_rd_p = qcp_rd_p; 995 996 u64_stats_update_begin(&r_vec->tx_sync); 997 r_vec->tx_bytes += done_bytes; 998 r_vec->tx_pkts += done_pkts; 999 u64_stats_update_end(&r_vec->tx_sync); 1000 1001 if (!dp->netdev) 1002 return; 1003 1004 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 1005 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes); 1006 if (nfp_net_tx_ring_should_wake(tx_ring)) { 1007 /* Make sure TX thread will see updated tx_ring->rd_p */ 1008 smp_mb(); 1009 1010 if (unlikely(netif_tx_queue_stopped(nd_q))) 1011 netif_tx_wake_queue(nd_q); 1012 } 1013 1014 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 1015 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 1016 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 1017 } 1018 1019 static bool nfp_net_xdp_complete(struct nfp_net_tx_ring *tx_ring) 1020 { 1021 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 1022 u32 done_pkts = 0, done_bytes = 0; 1023 bool done_all; 1024 int idx, todo; 1025 u32 qcp_rd_p; 1026 1027 /* Work out how many descriptors have been transmitted */ 1028 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q); 1029 1030 if (qcp_rd_p == tx_ring->qcp_rd_p) 1031 return true; 1032 1033 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 1034 1035 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE; 1036 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE); 1037 1038 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo); 1039 1040 done_pkts = todo; 1041 while (todo--) { 1042 idx = D_IDX(tx_ring, tx_ring->rd_p); 1043 tx_ring->rd_p++; 1044 1045 done_bytes += tx_ring->txbufs[idx].real_len; 1046 } 1047 1048 u64_stats_update_begin(&r_vec->tx_sync); 1049 r_vec->tx_bytes += done_bytes; 1050 r_vec->tx_pkts += done_pkts; 1051 u64_stats_update_end(&r_vec->tx_sync); 1052 1053 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 1054 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 1055 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 1056 1057 return done_all; 1058 } 1059 1060 /** 1061 * nfp_net_tx_ring_reset() - Free any untransmitted buffers and reset pointers 1062 * @dp: NFP Net data path struct 1063 * @tx_ring: TX ring structure 1064 * 1065 * Assumes that the device is stopped, must be idempotent. 1066 */ 1067 static void 1068 nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring) 1069 { 1070 const struct skb_frag_struct *frag; 1071 struct netdev_queue *nd_q; 1072 1073 while (!tx_ring->is_xdp && tx_ring->rd_p != tx_ring->wr_p) { 1074 struct nfp_net_tx_buf *tx_buf; 1075 struct sk_buff *skb; 1076 int idx, nr_frags; 1077 1078 idx = D_IDX(tx_ring, tx_ring->rd_p); 1079 tx_buf = &tx_ring->txbufs[idx]; 1080 1081 skb = tx_ring->txbufs[idx].skb; 1082 nr_frags = skb_shinfo(skb)->nr_frags; 1083 1084 if (tx_buf->fidx == -1) { 1085 /* unmap head */ 1086 dma_unmap_single(dp->dev, tx_buf->dma_addr, 1087 skb_headlen(skb), DMA_TO_DEVICE); 1088 } else { 1089 /* unmap fragment */ 1090 frag = &skb_shinfo(skb)->frags[tx_buf->fidx]; 1091 dma_unmap_page(dp->dev, tx_buf->dma_addr, 1092 skb_frag_size(frag), DMA_TO_DEVICE); 1093 } 1094 1095 /* check for last gather fragment */ 1096 if (tx_buf->fidx == nr_frags - 1) 1097 dev_kfree_skb_any(skb); 1098 1099 tx_buf->dma_addr = 0; 1100 tx_buf->skb = NULL; 1101 tx_buf->fidx = -2; 1102 1103 tx_ring->qcp_rd_p++; 1104 tx_ring->rd_p++; 1105 } 1106 1107 memset(tx_ring->txds, 0, tx_ring->size); 1108 tx_ring->wr_p = 0; 1109 tx_ring->rd_p = 0; 1110 tx_ring->qcp_rd_p = 0; 1111 tx_ring->wr_ptr_add = 0; 1112 1113 if (tx_ring->is_xdp || !dp->netdev) 1114 return; 1115 1116 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 1117 netdev_tx_reset_queue(nd_q); 1118 } 1119 1120 static void nfp_net_tx_timeout(struct net_device *netdev) 1121 { 1122 struct nfp_net *nn = netdev_priv(netdev); 1123 int i; 1124 1125 for (i = 0; i < nn->dp.netdev->real_num_tx_queues; i++) { 1126 if (!netif_tx_queue_stopped(netdev_get_tx_queue(netdev, i))) 1127 continue; 1128 nn_warn(nn, "TX timeout on ring: %d\n", i); 1129 } 1130 nn_warn(nn, "TX watchdog timeout\n"); 1131 } 1132 1133 /* Receive processing 1134 */ 1135 static unsigned int 1136 nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp) 1137 { 1138 unsigned int fl_bufsz; 1139 1140 fl_bufsz = NFP_NET_RX_BUF_HEADROOM; 1141 fl_bufsz += dp->rx_dma_off; 1142 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1143 fl_bufsz += NFP_NET_MAX_PREPEND; 1144 else 1145 fl_bufsz += dp->rx_offset; 1146 fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu; 1147 1148 fl_bufsz = SKB_DATA_ALIGN(fl_bufsz); 1149 fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1150 1151 return fl_bufsz; 1152 } 1153 1154 static void 1155 nfp_net_free_frag(void *frag, bool xdp) 1156 { 1157 if (!xdp) 1158 skb_free_frag(frag); 1159 else 1160 __free_page(virt_to_page(frag)); 1161 } 1162 1163 /** 1164 * nfp_net_rx_alloc_one() - Allocate and map page frag for RX 1165 * @dp: NFP Net data path struct 1166 * @dma_addr: Pointer to storage for DMA address (output param) 1167 * 1168 * This function will allcate a new page frag, map it for DMA. 1169 * 1170 * Return: allocated page frag or NULL on failure. 1171 */ 1172 static void *nfp_net_rx_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 1173 { 1174 void *frag; 1175 1176 if (!dp->xdp_prog) { 1177 frag = netdev_alloc_frag(dp->fl_bufsz); 1178 } else { 1179 struct page *page; 1180 1181 page = alloc_page(GFP_KERNEL); 1182 frag = page ? page_address(page) : NULL; 1183 } 1184 if (!frag) { 1185 nn_dp_warn(dp, "Failed to alloc receive page frag\n"); 1186 return NULL; 1187 } 1188 1189 *dma_addr = nfp_net_dma_map_rx(dp, frag); 1190 if (dma_mapping_error(dp->dev, *dma_addr)) { 1191 nfp_net_free_frag(frag, dp->xdp_prog); 1192 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 1193 return NULL; 1194 } 1195 1196 return frag; 1197 } 1198 1199 static void *nfp_net_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 1200 { 1201 void *frag; 1202 1203 if (!dp->xdp_prog) { 1204 frag = napi_alloc_frag(dp->fl_bufsz); 1205 if (unlikely(!frag)) 1206 return NULL; 1207 } else { 1208 struct page *page; 1209 1210 page = dev_alloc_page(); 1211 if (unlikely(!page)) 1212 return NULL; 1213 frag = page_address(page); 1214 } 1215 1216 *dma_addr = nfp_net_dma_map_rx(dp, frag); 1217 if (dma_mapping_error(dp->dev, *dma_addr)) { 1218 nfp_net_free_frag(frag, dp->xdp_prog); 1219 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 1220 return NULL; 1221 } 1222 1223 return frag; 1224 } 1225 1226 /** 1227 * nfp_net_rx_give_one() - Put mapped skb on the software and hardware rings 1228 * @dp: NFP Net data path struct 1229 * @rx_ring: RX ring structure 1230 * @frag: page fragment buffer 1231 * @dma_addr: DMA address of skb mapping 1232 */ 1233 static void nfp_net_rx_give_one(const struct nfp_net_dp *dp, 1234 struct nfp_net_rx_ring *rx_ring, 1235 void *frag, dma_addr_t dma_addr) 1236 { 1237 unsigned int wr_idx; 1238 1239 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 1240 1241 nfp_net_dma_sync_dev_rx(dp, dma_addr); 1242 1243 /* Stash SKB and DMA address away */ 1244 rx_ring->rxbufs[wr_idx].frag = frag; 1245 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr; 1246 1247 /* Fill freelist descriptor */ 1248 rx_ring->rxds[wr_idx].fld.reserved = 0; 1249 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0; 1250 nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld, 1251 dma_addr + dp->rx_dma_off); 1252 1253 rx_ring->wr_p++; 1254 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) { 1255 /* Update write pointer of the freelist queue. Make 1256 * sure all writes are flushed before telling the hardware. 1257 */ 1258 wmb(); 1259 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH); 1260 } 1261 } 1262 1263 /** 1264 * nfp_net_rx_ring_reset() - Reflect in SW state of freelist after disable 1265 * @rx_ring: RX ring structure 1266 * 1267 * Assumes that the device is stopped, must be idempotent. 1268 */ 1269 static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring) 1270 { 1271 unsigned int wr_idx, last_idx; 1272 1273 /* wr_p == rd_p means ring was never fed FL bufs. RX rings are always 1274 * kept at cnt - 1 FL bufs. 1275 */ 1276 if (rx_ring->wr_p == 0 && rx_ring->rd_p == 0) 1277 return; 1278 1279 /* Move the empty entry to the end of the list */ 1280 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 1281 last_idx = rx_ring->cnt - 1; 1282 rx_ring->rxbufs[wr_idx].dma_addr = rx_ring->rxbufs[last_idx].dma_addr; 1283 rx_ring->rxbufs[wr_idx].frag = rx_ring->rxbufs[last_idx].frag; 1284 rx_ring->rxbufs[last_idx].dma_addr = 0; 1285 rx_ring->rxbufs[last_idx].frag = NULL; 1286 1287 memset(rx_ring->rxds, 0, rx_ring->size); 1288 rx_ring->wr_p = 0; 1289 rx_ring->rd_p = 0; 1290 } 1291 1292 /** 1293 * nfp_net_rx_ring_bufs_free() - Free any buffers currently on the RX ring 1294 * @dp: NFP Net data path struct 1295 * @rx_ring: RX ring to remove buffers from 1296 * 1297 * Assumes that the device is stopped and buffers are in [0, ring->cnt - 1) 1298 * entries. After device is disabled nfp_net_rx_ring_reset() must be called 1299 * to restore required ring geometry. 1300 */ 1301 static void 1302 nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp, 1303 struct nfp_net_rx_ring *rx_ring) 1304 { 1305 unsigned int i; 1306 1307 for (i = 0; i < rx_ring->cnt - 1; i++) { 1308 /* NULL skb can only happen when initial filling of the ring 1309 * fails to allocate enough buffers and calls here to free 1310 * already allocated ones. 1311 */ 1312 if (!rx_ring->rxbufs[i].frag) 1313 continue; 1314 1315 nfp_net_dma_unmap_rx(dp, rx_ring->rxbufs[i].dma_addr); 1316 nfp_net_free_frag(rx_ring->rxbufs[i].frag, dp->xdp_prog); 1317 rx_ring->rxbufs[i].dma_addr = 0; 1318 rx_ring->rxbufs[i].frag = NULL; 1319 } 1320 } 1321 1322 /** 1323 * nfp_net_rx_ring_bufs_alloc() - Fill RX ring with buffers (don't give to FW) 1324 * @dp: NFP Net data path struct 1325 * @rx_ring: RX ring to remove buffers from 1326 */ 1327 static int 1328 nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp, 1329 struct nfp_net_rx_ring *rx_ring) 1330 { 1331 struct nfp_net_rx_buf *rxbufs; 1332 unsigned int i; 1333 1334 rxbufs = rx_ring->rxbufs; 1335 1336 for (i = 0; i < rx_ring->cnt - 1; i++) { 1337 rxbufs[i].frag = nfp_net_rx_alloc_one(dp, &rxbufs[i].dma_addr); 1338 if (!rxbufs[i].frag) { 1339 nfp_net_rx_ring_bufs_free(dp, rx_ring); 1340 return -ENOMEM; 1341 } 1342 } 1343 1344 return 0; 1345 } 1346 1347 /** 1348 * nfp_net_rx_ring_fill_freelist() - Give buffers from the ring to FW 1349 * @dp: NFP Net data path struct 1350 * @rx_ring: RX ring to fill 1351 */ 1352 static void 1353 nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp, 1354 struct nfp_net_rx_ring *rx_ring) 1355 { 1356 unsigned int i; 1357 1358 for (i = 0; i < rx_ring->cnt - 1; i++) 1359 nfp_net_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag, 1360 rx_ring->rxbufs[i].dma_addr); 1361 } 1362 1363 /** 1364 * nfp_net_rx_csum_has_errors() - group check if rxd has any csum errors 1365 * @flags: RX descriptor flags field in CPU byte order 1366 */ 1367 static int nfp_net_rx_csum_has_errors(u16 flags) 1368 { 1369 u16 csum_all_checked, csum_all_ok; 1370 1371 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL; 1372 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK; 1373 1374 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT); 1375 } 1376 1377 /** 1378 * nfp_net_rx_csum() - set SKB checksum field based on RX descriptor flags 1379 * @dp: NFP Net data path struct 1380 * @r_vec: per-ring structure 1381 * @rxd: Pointer to RX descriptor 1382 * @meta: Parsed metadata prepend 1383 * @skb: Pointer to SKB 1384 */ 1385 static void nfp_net_rx_csum(struct nfp_net_dp *dp, 1386 struct nfp_net_r_vector *r_vec, 1387 struct nfp_net_rx_desc *rxd, 1388 struct nfp_meta_parsed *meta, struct sk_buff *skb) 1389 { 1390 skb_checksum_none_assert(skb); 1391 1392 if (!(dp->netdev->features & NETIF_F_RXCSUM)) 1393 return; 1394 1395 if (meta->csum_type) { 1396 skb->ip_summed = meta->csum_type; 1397 skb->csum = meta->csum; 1398 u64_stats_update_begin(&r_vec->rx_sync); 1399 r_vec->hw_csum_rx_complete++; 1400 u64_stats_update_end(&r_vec->rx_sync); 1401 return; 1402 } 1403 1404 if (nfp_net_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) { 1405 u64_stats_update_begin(&r_vec->rx_sync); 1406 r_vec->hw_csum_rx_error++; 1407 u64_stats_update_end(&r_vec->rx_sync); 1408 return; 1409 } 1410 1411 /* Assume that the firmware will never report inner CSUM_OK unless outer 1412 * L4 headers were successfully parsed. FW will always report zero UDP 1413 * checksum as CSUM_OK. 1414 */ 1415 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK || 1416 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) { 1417 __skb_incr_checksum_unnecessary(skb); 1418 u64_stats_update_begin(&r_vec->rx_sync); 1419 r_vec->hw_csum_rx_ok++; 1420 u64_stats_update_end(&r_vec->rx_sync); 1421 } 1422 1423 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK || 1424 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) { 1425 __skb_incr_checksum_unnecessary(skb); 1426 u64_stats_update_begin(&r_vec->rx_sync); 1427 r_vec->hw_csum_rx_inner_ok++; 1428 u64_stats_update_end(&r_vec->rx_sync); 1429 } 1430 } 1431 1432 static void 1433 nfp_net_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta, 1434 unsigned int type, __be32 *hash) 1435 { 1436 if (!(netdev->features & NETIF_F_RXHASH)) 1437 return; 1438 1439 switch (type) { 1440 case NFP_NET_RSS_IPV4: 1441 case NFP_NET_RSS_IPV6: 1442 case NFP_NET_RSS_IPV6_EX: 1443 meta->hash_type = PKT_HASH_TYPE_L3; 1444 break; 1445 default: 1446 meta->hash_type = PKT_HASH_TYPE_L4; 1447 break; 1448 } 1449 1450 meta->hash = get_unaligned_be32(hash); 1451 } 1452 1453 static void 1454 nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta, 1455 void *data, struct nfp_net_rx_desc *rxd) 1456 { 1457 struct nfp_net_rx_hash *rx_hash = data; 1458 1459 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS)) 1460 return; 1461 1462 nfp_net_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type), 1463 &rx_hash->hash); 1464 } 1465 1466 static void * 1467 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta, 1468 void *data, int meta_len) 1469 { 1470 u32 meta_info; 1471 1472 meta_info = get_unaligned_be32(data); 1473 data += 4; 1474 1475 while (meta_info) { 1476 switch (meta_info & NFP_NET_META_FIELD_MASK) { 1477 case NFP_NET_META_HASH: 1478 meta_info >>= NFP_NET_META_FIELD_SIZE; 1479 nfp_net_set_hash(netdev, meta, 1480 meta_info & NFP_NET_META_FIELD_MASK, 1481 (__be32 *)data); 1482 data += 4; 1483 break; 1484 case NFP_NET_META_MARK: 1485 meta->mark = get_unaligned_be32(data); 1486 data += 4; 1487 break; 1488 case NFP_NET_META_PORTID: 1489 meta->portid = get_unaligned_be32(data); 1490 data += 4; 1491 break; 1492 case NFP_NET_META_CSUM: 1493 meta->csum_type = CHECKSUM_COMPLETE; 1494 meta->csum = 1495 (__force __wsum)__get_unaligned_cpu32(data); 1496 data += 4; 1497 break; 1498 default: 1499 return NULL; 1500 } 1501 1502 meta_info >>= NFP_NET_META_FIELD_SIZE; 1503 } 1504 1505 return data; 1506 } 1507 1508 static void 1509 nfp_net_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, 1510 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf, 1511 struct sk_buff *skb) 1512 { 1513 u64_stats_update_begin(&r_vec->rx_sync); 1514 r_vec->rx_drops++; 1515 /* If we have both skb and rxbuf the replacement buffer allocation 1516 * must have failed, count this as an alloc failure. 1517 */ 1518 if (skb && rxbuf) 1519 r_vec->rx_replace_buf_alloc_fail++; 1520 u64_stats_update_end(&r_vec->rx_sync); 1521 1522 /* skb is build based on the frag, free_skb() would free the frag 1523 * so to be able to reuse it we need an extra ref. 1524 */ 1525 if (skb && rxbuf && skb->head == rxbuf->frag) 1526 page_ref_inc(virt_to_head_page(rxbuf->frag)); 1527 if (rxbuf) 1528 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr); 1529 if (skb) 1530 dev_kfree_skb_any(skb); 1531 } 1532 1533 static bool 1534 nfp_net_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring, 1535 struct nfp_net_tx_ring *tx_ring, 1536 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off, 1537 unsigned int pkt_len, bool *completed) 1538 { 1539 struct nfp_net_tx_buf *txbuf; 1540 struct nfp_net_tx_desc *txd; 1541 int wr_idx; 1542 1543 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1544 if (!*completed) { 1545 nfp_net_xdp_complete(tx_ring); 1546 *completed = true; 1547 } 1548 1549 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1550 nfp_net_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf, 1551 NULL); 1552 return false; 1553 } 1554 } 1555 1556 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1557 1558 /* Stash the soft descriptor of the head then initialize it */ 1559 txbuf = &tx_ring->txbufs[wr_idx]; 1560 1561 nfp_net_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr); 1562 1563 txbuf->frag = rxbuf->frag; 1564 txbuf->dma_addr = rxbuf->dma_addr; 1565 txbuf->fidx = -1; 1566 txbuf->pkt_cnt = 1; 1567 txbuf->real_len = pkt_len; 1568 1569 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off, 1570 pkt_len, DMA_BIDIRECTIONAL); 1571 1572 /* Build TX descriptor */ 1573 txd = &tx_ring->txds[wr_idx]; 1574 txd->offset_eop = PCIE_DESC_TX_EOP; 1575 txd->dma_len = cpu_to_le16(pkt_len); 1576 nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off); 1577 txd->data_len = cpu_to_le16(pkt_len); 1578 1579 txd->flags = 0; 1580 txd->mss = 0; 1581 txd->lso_hdrlen = 0; 1582 1583 tx_ring->wr_p++; 1584 tx_ring->wr_ptr_add++; 1585 return true; 1586 } 1587 1588 /** 1589 * nfp_net_rx() - receive up to @budget packets on @rx_ring 1590 * @rx_ring: RX ring to receive from 1591 * @budget: NAPI budget 1592 * 1593 * Note, this function is separated out from the napi poll function to 1594 * more cleanly separate packet receive code from other bookkeeping 1595 * functions performed in the napi poll function. 1596 * 1597 * Return: Number of packets received. 1598 */ 1599 static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget) 1600 { 1601 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 1602 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 1603 struct nfp_net_tx_ring *tx_ring; 1604 struct bpf_prog *xdp_prog; 1605 bool xdp_tx_cmpl = false; 1606 unsigned int true_bufsz; 1607 struct sk_buff *skb; 1608 int pkts_polled = 0; 1609 struct xdp_buff xdp; 1610 int idx; 1611 1612 rcu_read_lock(); 1613 xdp_prog = READ_ONCE(dp->xdp_prog); 1614 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz; 1615 xdp.rxq = &rx_ring->xdp_rxq; 1616 tx_ring = r_vec->xdp_ring; 1617 1618 while (pkts_polled < budget) { 1619 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1620 struct nfp_net_rx_buf *rxbuf; 1621 struct nfp_net_rx_desc *rxd; 1622 struct nfp_meta_parsed meta; 1623 struct net_device *netdev; 1624 dma_addr_t new_dma_addr; 1625 u32 meta_len_xdp = 0; 1626 void *new_frag; 1627 1628 idx = D_IDX(rx_ring, rx_ring->rd_p); 1629 1630 rxd = &rx_ring->rxds[idx]; 1631 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 1632 break; 1633 1634 /* Memory barrier to ensure that we won't do other reads 1635 * before the DD bit. 1636 */ 1637 dma_rmb(); 1638 1639 memset(&meta, 0, sizeof(meta)); 1640 1641 rx_ring->rd_p++; 1642 pkts_polled++; 1643 1644 rxbuf = &rx_ring->rxbufs[idx]; 1645 /* < meta_len > 1646 * <-- [rx_offset] --> 1647 * --------------------------------------------------------- 1648 * | [XX] | metadata | packet | XXXX | 1649 * --------------------------------------------------------- 1650 * <---------------- data_len ---------------> 1651 * 1652 * The rx_offset is fixed for all packets, the meta_len can vary 1653 * on a packet by packet basis. If rx_offset is set to zero 1654 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the 1655 * buffer and is immediately followed by the packet (no [XX]). 1656 */ 1657 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 1658 data_len = le16_to_cpu(rxd->rxd.data_len); 1659 pkt_len = data_len - meta_len; 1660 1661 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 1662 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1663 pkt_off += meta_len; 1664 else 1665 pkt_off += dp->rx_offset; 1666 meta_off = pkt_off - meta_len; 1667 1668 /* Stats update */ 1669 u64_stats_update_begin(&r_vec->rx_sync); 1670 r_vec->rx_pkts++; 1671 r_vec->rx_bytes += pkt_len; 1672 u64_stats_update_end(&r_vec->rx_sync); 1673 1674 if (unlikely(meta_len > NFP_NET_MAX_PREPEND || 1675 (dp->rx_offset && meta_len > dp->rx_offset))) { 1676 nn_dp_warn(dp, "oversized RX packet metadata %u\n", 1677 meta_len); 1678 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1679 continue; 1680 } 1681 1682 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, 1683 data_len); 1684 1685 if (!dp->chained_metadata_format) { 1686 nfp_net_set_hash_desc(dp->netdev, &meta, 1687 rxbuf->frag + meta_off, rxd); 1688 } else if (meta_len) { 1689 void *end; 1690 1691 end = nfp_net_parse_meta(dp->netdev, &meta, 1692 rxbuf->frag + meta_off, 1693 meta_len); 1694 if (unlikely(end != rxbuf->frag + pkt_off)) { 1695 nn_dp_warn(dp, "invalid RX packet metadata\n"); 1696 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, 1697 NULL); 1698 continue; 1699 } 1700 } 1701 1702 if (xdp_prog && !meta.portid) { 1703 void *orig_data = rxbuf->frag + pkt_off; 1704 unsigned int dma_off; 1705 int act; 1706 1707 xdp.data_hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM; 1708 xdp.data = orig_data; 1709 xdp.data_meta = orig_data; 1710 xdp.data_end = orig_data + pkt_len; 1711 1712 act = bpf_prog_run_xdp(xdp_prog, &xdp); 1713 1714 pkt_len = xdp.data_end - xdp.data; 1715 pkt_off += xdp.data - orig_data; 1716 1717 switch (act) { 1718 case XDP_PASS: 1719 meta_len_xdp = xdp.data - xdp.data_meta; 1720 break; 1721 case XDP_TX: 1722 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM; 1723 if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring, 1724 tx_ring, rxbuf, 1725 dma_off, 1726 pkt_len, 1727 &xdp_tx_cmpl))) 1728 trace_xdp_exception(dp->netdev, 1729 xdp_prog, act); 1730 continue; 1731 default: 1732 bpf_warn_invalid_xdp_action(act); 1733 /* fall through */ 1734 case XDP_ABORTED: 1735 trace_xdp_exception(dp->netdev, xdp_prog, act); 1736 /* fall through */ 1737 case XDP_DROP: 1738 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, 1739 rxbuf->dma_addr); 1740 continue; 1741 } 1742 } 1743 1744 if (likely(!meta.portid)) { 1745 netdev = dp->netdev; 1746 } else if (meta.portid == NFP_META_PORT_ID_CTRL) { 1747 struct nfp_net *nn = netdev_priv(dp->netdev); 1748 1749 nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off, 1750 pkt_len); 1751 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, 1752 rxbuf->dma_addr); 1753 continue; 1754 } else { 1755 struct nfp_net *nn; 1756 1757 nn = netdev_priv(dp->netdev); 1758 netdev = nfp_app_repr_get(nn->app, meta.portid); 1759 if (unlikely(!netdev)) { 1760 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, 1761 NULL); 1762 continue; 1763 } 1764 nfp_repr_inc_rx_stats(netdev, pkt_len); 1765 } 1766 1767 skb = build_skb(rxbuf->frag, true_bufsz); 1768 if (unlikely(!skb)) { 1769 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1770 continue; 1771 } 1772 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr); 1773 if (unlikely(!new_frag)) { 1774 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 1775 continue; 1776 } 1777 1778 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 1779 1780 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 1781 1782 skb_reserve(skb, pkt_off); 1783 skb_put(skb, pkt_len); 1784 1785 skb->mark = meta.mark; 1786 skb_set_hash(skb, meta.hash, meta.hash_type); 1787 1788 skb_record_rx_queue(skb, rx_ring->idx); 1789 skb->protocol = eth_type_trans(skb, netdev); 1790 1791 nfp_net_rx_csum(dp, r_vec, rxd, &meta, skb); 1792 1793 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN) 1794 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), 1795 le16_to_cpu(rxd->rxd.vlan)); 1796 if (meta_len_xdp) 1797 skb_metadata_set(skb, meta_len_xdp); 1798 1799 napi_gro_receive(&rx_ring->r_vec->napi, skb); 1800 } 1801 1802 if (xdp_prog) { 1803 if (tx_ring->wr_ptr_add) 1804 nfp_net_tx_xmit_more_flush(tx_ring); 1805 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) && 1806 !xdp_tx_cmpl) 1807 if (!nfp_net_xdp_complete(tx_ring)) 1808 pkts_polled = budget; 1809 } 1810 rcu_read_unlock(); 1811 1812 return pkts_polled; 1813 } 1814 1815 /** 1816 * nfp_net_poll() - napi poll function 1817 * @napi: NAPI structure 1818 * @budget: NAPI budget 1819 * 1820 * Return: number of packets polled. 1821 */ 1822 static int nfp_net_poll(struct napi_struct *napi, int budget) 1823 { 1824 struct nfp_net_r_vector *r_vec = 1825 container_of(napi, struct nfp_net_r_vector, napi); 1826 unsigned int pkts_polled = 0; 1827 1828 if (r_vec->tx_ring) 1829 nfp_net_tx_complete(r_vec->tx_ring, budget); 1830 if (r_vec->rx_ring) 1831 pkts_polled = nfp_net_rx(r_vec->rx_ring, budget); 1832 1833 if (pkts_polled < budget) 1834 if (napi_complete_done(napi, pkts_polled)) 1835 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 1836 1837 return pkts_polled; 1838 } 1839 1840 /* Control device data path 1841 */ 1842 1843 static bool 1844 nfp_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 1845 struct sk_buff *skb, bool old) 1846 { 1847 unsigned int real_len = skb->len, meta_len = 0; 1848 struct nfp_net_tx_ring *tx_ring; 1849 struct nfp_net_tx_buf *txbuf; 1850 struct nfp_net_tx_desc *txd; 1851 struct nfp_net_dp *dp; 1852 dma_addr_t dma_addr; 1853 int wr_idx; 1854 1855 dp = &r_vec->nfp_net->dp; 1856 tx_ring = r_vec->tx_ring; 1857 1858 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) { 1859 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n"); 1860 goto err_free; 1861 } 1862 1863 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1864 u64_stats_update_begin(&r_vec->tx_sync); 1865 r_vec->tx_busy++; 1866 u64_stats_update_end(&r_vec->tx_sync); 1867 if (!old) 1868 __skb_queue_tail(&r_vec->queue, skb); 1869 else 1870 __skb_queue_head(&r_vec->queue, skb); 1871 return true; 1872 } 1873 1874 if (nfp_app_ctrl_has_meta(nn->app)) { 1875 if (unlikely(skb_headroom(skb) < 8)) { 1876 nn_dp_warn(dp, "CTRL TX on skb without headroom\n"); 1877 goto err_free; 1878 } 1879 meta_len = 8; 1880 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4)); 1881 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4)); 1882 } 1883 1884 /* Start with the head skbuf */ 1885 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 1886 DMA_TO_DEVICE); 1887 if (dma_mapping_error(dp->dev, dma_addr)) 1888 goto err_dma_warn; 1889 1890 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1891 1892 /* Stash the soft descriptor of the head then initialize it */ 1893 txbuf = &tx_ring->txbufs[wr_idx]; 1894 txbuf->skb = skb; 1895 txbuf->dma_addr = dma_addr; 1896 txbuf->fidx = -1; 1897 txbuf->pkt_cnt = 1; 1898 txbuf->real_len = real_len; 1899 1900 /* Build TX descriptor */ 1901 txd = &tx_ring->txds[wr_idx]; 1902 txd->offset_eop = meta_len | PCIE_DESC_TX_EOP; 1903 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 1904 nfp_desc_set_dma_addr(txd, dma_addr); 1905 txd->data_len = cpu_to_le16(skb->len); 1906 1907 txd->flags = 0; 1908 txd->mss = 0; 1909 txd->lso_hdrlen = 0; 1910 1911 tx_ring->wr_p++; 1912 tx_ring->wr_ptr_add++; 1913 nfp_net_tx_xmit_more_flush(tx_ring); 1914 1915 return false; 1916 1917 err_dma_warn: 1918 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n"); 1919 err_free: 1920 u64_stats_update_begin(&r_vec->tx_sync); 1921 r_vec->tx_errors++; 1922 u64_stats_update_end(&r_vec->tx_sync); 1923 dev_kfree_skb_any(skb); 1924 return false; 1925 } 1926 1927 bool __nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb) 1928 { 1929 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0]; 1930 1931 return nfp_ctrl_tx_one(nn, r_vec, skb, false); 1932 } 1933 1934 bool nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb) 1935 { 1936 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0]; 1937 bool ret; 1938 1939 spin_lock_bh(&r_vec->lock); 1940 ret = nfp_ctrl_tx_one(nn, r_vec, skb, false); 1941 spin_unlock_bh(&r_vec->lock); 1942 1943 return ret; 1944 } 1945 1946 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec) 1947 { 1948 struct sk_buff *skb; 1949 1950 while ((skb = __skb_dequeue(&r_vec->queue))) 1951 if (nfp_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true)) 1952 return; 1953 } 1954 1955 static bool 1956 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len) 1957 { 1958 u32 meta_type, meta_tag; 1959 1960 if (!nfp_app_ctrl_has_meta(nn->app)) 1961 return !meta_len; 1962 1963 if (meta_len != 8) 1964 return false; 1965 1966 meta_type = get_unaligned_be32(data); 1967 meta_tag = get_unaligned_be32(data + 4); 1968 1969 return (meta_type == NFP_NET_META_PORTID && 1970 meta_tag == NFP_META_PORT_ID_CTRL); 1971 } 1972 1973 static bool 1974 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp, 1975 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring) 1976 { 1977 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1978 struct nfp_net_rx_buf *rxbuf; 1979 struct nfp_net_rx_desc *rxd; 1980 dma_addr_t new_dma_addr; 1981 struct sk_buff *skb; 1982 void *new_frag; 1983 int idx; 1984 1985 idx = D_IDX(rx_ring, rx_ring->rd_p); 1986 1987 rxd = &rx_ring->rxds[idx]; 1988 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 1989 return false; 1990 1991 /* Memory barrier to ensure that we won't do other reads 1992 * before the DD bit. 1993 */ 1994 dma_rmb(); 1995 1996 rx_ring->rd_p++; 1997 1998 rxbuf = &rx_ring->rxbufs[idx]; 1999 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 2000 data_len = le16_to_cpu(rxd->rxd.data_len); 2001 pkt_len = data_len - meta_len; 2002 2003 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 2004 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 2005 pkt_off += meta_len; 2006 else 2007 pkt_off += dp->rx_offset; 2008 meta_off = pkt_off - meta_len; 2009 2010 /* Stats update */ 2011 u64_stats_update_begin(&r_vec->rx_sync); 2012 r_vec->rx_pkts++; 2013 r_vec->rx_bytes += pkt_len; 2014 u64_stats_update_end(&r_vec->rx_sync); 2015 2016 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len); 2017 2018 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) { 2019 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n", 2020 meta_len); 2021 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 2022 return true; 2023 } 2024 2025 skb = build_skb(rxbuf->frag, dp->fl_bufsz); 2026 if (unlikely(!skb)) { 2027 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 2028 return true; 2029 } 2030 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr); 2031 if (unlikely(!new_frag)) { 2032 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 2033 return true; 2034 } 2035 2036 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 2037 2038 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 2039 2040 skb_reserve(skb, pkt_off); 2041 skb_put(skb, pkt_len); 2042 2043 nfp_app_ctrl_rx(nn->app, skb); 2044 2045 return true; 2046 } 2047 2048 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec) 2049 { 2050 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring; 2051 struct nfp_net *nn = r_vec->nfp_net; 2052 struct nfp_net_dp *dp = &nn->dp; 2053 unsigned int budget = 512; 2054 2055 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--) 2056 continue; 2057 2058 return budget; 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(&r_vec->lock); 2066 nfp_net_tx_complete(r_vec->tx_ring, 0); 2067 __nfp_ctrl_tx_queued(r_vec); 2068 spin_unlock(&r_vec->lock); 2069 2070 if (nfp_ctrl_rx(r_vec)) { 2071 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 2072 } else { 2073 tasklet_schedule(&r_vec->tasklet); 2074 nn_dp_warn(&r_vec->nfp_net->dp, 2075 "control message budget exceeded!\n"); 2076 } 2077 } 2078 2079 /* Setup and Configuration 2080 */ 2081 2082 /** 2083 * nfp_net_vecs_init() - Assign IRQs and setup rvecs. 2084 * @nn: NFP Network structure 2085 */ 2086 static void nfp_net_vecs_init(struct nfp_net *nn) 2087 { 2088 struct nfp_net_r_vector *r_vec; 2089 int r; 2090 2091 nn->lsc_handler = nfp_net_irq_lsc; 2092 nn->exn_handler = nfp_net_irq_exn; 2093 2094 for (r = 0; r < nn->max_r_vecs; r++) { 2095 struct msix_entry *entry; 2096 2097 entry = &nn->irq_entries[NFP_NET_NON_Q_VECTORS + r]; 2098 2099 r_vec = &nn->r_vecs[r]; 2100 r_vec->nfp_net = nn; 2101 r_vec->irq_entry = entry->entry; 2102 r_vec->irq_vector = entry->vector; 2103 2104 if (nn->dp.netdev) { 2105 r_vec->handler = nfp_net_irq_rxtx; 2106 } else { 2107 r_vec->handler = nfp_ctrl_irq_rxtx; 2108 2109 __skb_queue_head_init(&r_vec->queue); 2110 spin_lock_init(&r_vec->lock); 2111 tasklet_init(&r_vec->tasklet, nfp_ctrl_poll, 2112 (unsigned long)r_vec); 2113 tasklet_disable(&r_vec->tasklet); 2114 } 2115 2116 cpumask_set_cpu(r, &r_vec->affinity_mask); 2117 } 2118 } 2119 2120 /** 2121 * nfp_net_tx_ring_free() - Free resources allocated to a TX ring 2122 * @tx_ring: TX ring to free 2123 */ 2124 static void nfp_net_tx_ring_free(struct nfp_net_tx_ring *tx_ring) 2125 { 2126 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 2127 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 2128 2129 kvfree(tx_ring->txbufs); 2130 2131 if (tx_ring->txds) 2132 dma_free_coherent(dp->dev, tx_ring->size, 2133 tx_ring->txds, tx_ring->dma); 2134 2135 tx_ring->cnt = 0; 2136 tx_ring->txbufs = NULL; 2137 tx_ring->txds = NULL; 2138 tx_ring->dma = 0; 2139 tx_ring->size = 0; 2140 } 2141 2142 /** 2143 * nfp_net_tx_ring_alloc() - Allocate resource for a TX ring 2144 * @dp: NFP Net data path struct 2145 * @tx_ring: TX Ring structure to allocate 2146 * 2147 * Return: 0 on success, negative errno otherwise. 2148 */ 2149 static int 2150 nfp_net_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring) 2151 { 2152 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 2153 2154 tx_ring->cnt = dp->txd_cnt; 2155 2156 tx_ring->size = array_size(tx_ring->cnt, sizeof(*tx_ring->txds)); 2157 tx_ring->txds = dma_zalloc_coherent(dp->dev, tx_ring->size, 2158 &tx_ring->dma, 2159 GFP_KERNEL | __GFP_NOWARN); 2160 if (!tx_ring->txds) { 2161 netdev_warn(dp->netdev, "failed to allocate TX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n", 2162 tx_ring->cnt); 2163 goto err_alloc; 2164 } 2165 2166 tx_ring->txbufs = kvcalloc(tx_ring->cnt, sizeof(*tx_ring->txbufs), 2167 GFP_KERNEL); 2168 if (!tx_ring->txbufs) 2169 goto err_alloc; 2170 2171 if (!tx_ring->is_xdp && dp->netdev) 2172 netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask, 2173 tx_ring->idx); 2174 2175 return 0; 2176 2177 err_alloc: 2178 nfp_net_tx_ring_free(tx_ring); 2179 return -ENOMEM; 2180 } 2181 2182 static void 2183 nfp_net_tx_ring_bufs_free(struct nfp_net_dp *dp, 2184 struct nfp_net_tx_ring *tx_ring) 2185 { 2186 unsigned int i; 2187 2188 if (!tx_ring->is_xdp) 2189 return; 2190 2191 for (i = 0; i < tx_ring->cnt; i++) { 2192 if (!tx_ring->txbufs[i].frag) 2193 return; 2194 2195 nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr); 2196 __free_page(virt_to_page(tx_ring->txbufs[i].frag)); 2197 } 2198 } 2199 2200 static int 2201 nfp_net_tx_ring_bufs_alloc(struct nfp_net_dp *dp, 2202 struct nfp_net_tx_ring *tx_ring) 2203 { 2204 struct nfp_net_tx_buf *txbufs = tx_ring->txbufs; 2205 unsigned int i; 2206 2207 if (!tx_ring->is_xdp) 2208 return 0; 2209 2210 for (i = 0; i < tx_ring->cnt; i++) { 2211 txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr); 2212 if (!txbufs[i].frag) { 2213 nfp_net_tx_ring_bufs_free(dp, tx_ring); 2214 return -ENOMEM; 2215 } 2216 } 2217 2218 return 0; 2219 } 2220 2221 static int nfp_net_tx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp) 2222 { 2223 unsigned int r; 2224 2225 dp->tx_rings = kcalloc(dp->num_tx_rings, sizeof(*dp->tx_rings), 2226 GFP_KERNEL); 2227 if (!dp->tx_rings) 2228 return -ENOMEM; 2229 2230 for (r = 0; r < dp->num_tx_rings; r++) { 2231 int bias = 0; 2232 2233 if (r >= dp->num_stack_tx_rings) 2234 bias = dp->num_stack_tx_rings; 2235 2236 nfp_net_tx_ring_init(&dp->tx_rings[r], &nn->r_vecs[r - bias], 2237 r, bias); 2238 2239 if (nfp_net_tx_ring_alloc(dp, &dp->tx_rings[r])) 2240 goto err_free_prev; 2241 2242 if (nfp_net_tx_ring_bufs_alloc(dp, &dp->tx_rings[r])) 2243 goto err_free_ring; 2244 } 2245 2246 return 0; 2247 2248 err_free_prev: 2249 while (r--) { 2250 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]); 2251 err_free_ring: 2252 nfp_net_tx_ring_free(&dp->tx_rings[r]); 2253 } 2254 kfree(dp->tx_rings); 2255 return -ENOMEM; 2256 } 2257 2258 static void nfp_net_tx_rings_free(struct nfp_net_dp *dp) 2259 { 2260 unsigned int r; 2261 2262 for (r = 0; r < dp->num_tx_rings; r++) { 2263 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]); 2264 nfp_net_tx_ring_free(&dp->tx_rings[r]); 2265 } 2266 2267 kfree(dp->tx_rings); 2268 } 2269 2270 /** 2271 * nfp_net_rx_ring_free() - Free resources allocated to a RX ring 2272 * @rx_ring: RX ring to free 2273 */ 2274 static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring) 2275 { 2276 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 2277 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 2278 2279 if (dp->netdev) 2280 xdp_rxq_info_unreg(&rx_ring->xdp_rxq); 2281 kvfree(rx_ring->rxbufs); 2282 2283 if (rx_ring->rxds) 2284 dma_free_coherent(dp->dev, rx_ring->size, 2285 rx_ring->rxds, rx_ring->dma); 2286 2287 rx_ring->cnt = 0; 2288 rx_ring->rxbufs = NULL; 2289 rx_ring->rxds = NULL; 2290 rx_ring->dma = 0; 2291 rx_ring->size = 0; 2292 } 2293 2294 /** 2295 * nfp_net_rx_ring_alloc() - Allocate resource for a RX ring 2296 * @dp: NFP Net data path struct 2297 * @rx_ring: RX ring to allocate 2298 * 2299 * Return: 0 on success, negative errno otherwise. 2300 */ 2301 static int 2302 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring) 2303 { 2304 int err; 2305 2306 if (dp->netdev) { 2307 err = xdp_rxq_info_reg(&rx_ring->xdp_rxq, dp->netdev, 2308 rx_ring->idx); 2309 if (err < 0) 2310 return err; 2311 } 2312 2313 rx_ring->cnt = dp->rxd_cnt; 2314 rx_ring->size = array_size(rx_ring->cnt, sizeof(*rx_ring->rxds)); 2315 rx_ring->rxds = dma_zalloc_coherent(dp->dev, rx_ring->size, 2316 &rx_ring->dma, 2317 GFP_KERNEL | __GFP_NOWARN); 2318 if (!rx_ring->rxds) { 2319 netdev_warn(dp->netdev, "failed to allocate RX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n", 2320 rx_ring->cnt); 2321 goto err_alloc; 2322 } 2323 2324 rx_ring->rxbufs = kvcalloc(rx_ring->cnt, sizeof(*rx_ring->rxbufs), 2325 GFP_KERNEL); 2326 if (!rx_ring->rxbufs) 2327 goto err_alloc; 2328 2329 return 0; 2330 2331 err_alloc: 2332 nfp_net_rx_ring_free(rx_ring); 2333 return -ENOMEM; 2334 } 2335 2336 static int nfp_net_rx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp) 2337 { 2338 unsigned int r; 2339 2340 dp->rx_rings = kcalloc(dp->num_rx_rings, sizeof(*dp->rx_rings), 2341 GFP_KERNEL); 2342 if (!dp->rx_rings) 2343 return -ENOMEM; 2344 2345 for (r = 0; r < dp->num_rx_rings; r++) { 2346 nfp_net_rx_ring_init(&dp->rx_rings[r], &nn->r_vecs[r], r); 2347 2348 if (nfp_net_rx_ring_alloc(dp, &dp->rx_rings[r])) 2349 goto err_free_prev; 2350 2351 if (nfp_net_rx_ring_bufs_alloc(dp, &dp->rx_rings[r])) 2352 goto err_free_ring; 2353 } 2354 2355 return 0; 2356 2357 err_free_prev: 2358 while (r--) { 2359 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]); 2360 err_free_ring: 2361 nfp_net_rx_ring_free(&dp->rx_rings[r]); 2362 } 2363 kfree(dp->rx_rings); 2364 return -ENOMEM; 2365 } 2366 2367 static void nfp_net_rx_rings_free(struct nfp_net_dp *dp) 2368 { 2369 unsigned int r; 2370 2371 for (r = 0; r < dp->num_rx_rings; r++) { 2372 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]); 2373 nfp_net_rx_ring_free(&dp->rx_rings[r]); 2374 } 2375 2376 kfree(dp->rx_rings); 2377 } 2378 2379 static void 2380 nfp_net_vector_assign_rings(struct nfp_net_dp *dp, 2381 struct nfp_net_r_vector *r_vec, int idx) 2382 { 2383 r_vec->rx_ring = idx < dp->num_rx_rings ? &dp->rx_rings[idx] : NULL; 2384 r_vec->tx_ring = 2385 idx < dp->num_stack_tx_rings ? &dp->tx_rings[idx] : NULL; 2386 2387 r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ? 2388 &dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL; 2389 } 2390 2391 static int 2392 nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 2393 int idx) 2394 { 2395 int err; 2396 2397 /* Setup NAPI */ 2398 if (nn->dp.netdev) 2399 netif_napi_add(nn->dp.netdev, &r_vec->napi, 2400 nfp_net_poll, NAPI_POLL_WEIGHT); 2401 else 2402 tasklet_enable(&r_vec->tasklet); 2403 2404 snprintf(r_vec->name, sizeof(r_vec->name), 2405 "%s-rxtx-%d", nfp_net_name(nn), idx); 2406 err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name, 2407 r_vec); 2408 if (err) { 2409 if (nn->dp.netdev) 2410 netif_napi_del(&r_vec->napi); 2411 else 2412 tasklet_disable(&r_vec->tasklet); 2413 2414 nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector); 2415 return err; 2416 } 2417 disable_irq(r_vec->irq_vector); 2418 2419 irq_set_affinity_hint(r_vec->irq_vector, &r_vec->affinity_mask); 2420 2421 nn_dbg(nn, "RV%02d: irq=%03d/%03d\n", idx, r_vec->irq_vector, 2422 r_vec->irq_entry); 2423 2424 return 0; 2425 } 2426 2427 static void 2428 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec) 2429 { 2430 irq_set_affinity_hint(r_vec->irq_vector, NULL); 2431 if (nn->dp.netdev) 2432 netif_napi_del(&r_vec->napi); 2433 else 2434 tasklet_disable(&r_vec->tasklet); 2435 2436 free_irq(r_vec->irq_vector, r_vec); 2437 } 2438 2439 /** 2440 * nfp_net_rss_write_itbl() - Write RSS indirection table to device 2441 * @nn: NFP Net device to reconfigure 2442 */ 2443 void nfp_net_rss_write_itbl(struct nfp_net *nn) 2444 { 2445 int i; 2446 2447 for (i = 0; i < NFP_NET_CFG_RSS_ITBL_SZ; i += 4) 2448 nn_writel(nn, NFP_NET_CFG_RSS_ITBL + i, 2449 get_unaligned_le32(nn->rss_itbl + i)); 2450 } 2451 2452 /** 2453 * nfp_net_rss_write_key() - Write RSS hash key to device 2454 * @nn: NFP Net device to reconfigure 2455 */ 2456 void nfp_net_rss_write_key(struct nfp_net *nn) 2457 { 2458 int i; 2459 2460 for (i = 0; i < nfp_net_rss_key_sz(nn); i += 4) 2461 nn_writel(nn, NFP_NET_CFG_RSS_KEY + i, 2462 get_unaligned_le32(nn->rss_key + i)); 2463 } 2464 2465 /** 2466 * nfp_net_coalesce_write_cfg() - Write irq coalescence configuration to HW 2467 * @nn: NFP Net device to reconfigure 2468 */ 2469 void nfp_net_coalesce_write_cfg(struct nfp_net *nn) 2470 { 2471 u8 i; 2472 u32 factor; 2473 u32 value; 2474 2475 /* Compute factor used to convert coalesce '_usecs' parameters to 2476 * ME timestamp ticks. There are 16 ME clock cycles for each timestamp 2477 * count. 2478 */ 2479 factor = nn->tlv_caps.me_freq_mhz / 16; 2480 2481 /* copy RX interrupt coalesce parameters */ 2482 value = (nn->rx_coalesce_max_frames << 16) | 2483 (factor * nn->rx_coalesce_usecs); 2484 for (i = 0; i < nn->dp.num_rx_rings; i++) 2485 nn_writel(nn, NFP_NET_CFG_RXR_IRQ_MOD(i), value); 2486 2487 /* copy TX interrupt coalesce parameters */ 2488 value = (nn->tx_coalesce_max_frames << 16) | 2489 (factor * nn->tx_coalesce_usecs); 2490 for (i = 0; i < nn->dp.num_tx_rings; i++) 2491 nn_writel(nn, NFP_NET_CFG_TXR_IRQ_MOD(i), value); 2492 } 2493 2494 /** 2495 * nfp_net_write_mac_addr() - Write mac address to the device control BAR 2496 * @nn: NFP Net device to reconfigure 2497 * @addr: MAC address to write 2498 * 2499 * Writes the MAC address from the netdev to the device control BAR. Does not 2500 * perform the required reconfig. We do a bit of byte swapping dance because 2501 * firmware is LE. 2502 */ 2503 static void nfp_net_write_mac_addr(struct nfp_net *nn, const u8 *addr) 2504 { 2505 nn_writel(nn, NFP_NET_CFG_MACADDR + 0, get_unaligned_be32(addr)); 2506 nn_writew(nn, NFP_NET_CFG_MACADDR + 6, get_unaligned_be16(addr + 4)); 2507 } 2508 2509 static void nfp_net_vec_clear_ring_data(struct nfp_net *nn, unsigned int idx) 2510 { 2511 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), 0); 2512 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), 0); 2513 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), 0); 2514 2515 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), 0); 2516 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), 0); 2517 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), 0); 2518 } 2519 2520 /** 2521 * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP 2522 * @nn: NFP Net device to reconfigure 2523 * 2524 * Warning: must be fully idempotent. 2525 */ 2526 static void nfp_net_clear_config_and_disable(struct nfp_net *nn) 2527 { 2528 u32 new_ctrl, update; 2529 unsigned int r; 2530 int err; 2531 2532 new_ctrl = nn->dp.ctrl; 2533 new_ctrl &= ~NFP_NET_CFG_CTRL_ENABLE; 2534 update = NFP_NET_CFG_UPDATE_GEN; 2535 update |= NFP_NET_CFG_UPDATE_MSIX; 2536 update |= NFP_NET_CFG_UPDATE_RING; 2537 2538 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG) 2539 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG; 2540 2541 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0); 2542 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0); 2543 2544 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2545 err = nfp_net_reconfig(nn, update); 2546 if (err) 2547 nn_err(nn, "Could not disable device: %d\n", err); 2548 2549 for (r = 0; r < nn->dp.num_rx_rings; r++) 2550 nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]); 2551 for (r = 0; r < nn->dp.num_tx_rings; r++) 2552 nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]); 2553 for (r = 0; r < nn->dp.num_r_vecs; r++) 2554 nfp_net_vec_clear_ring_data(nn, r); 2555 2556 nn->dp.ctrl = new_ctrl; 2557 } 2558 2559 static void 2560 nfp_net_rx_ring_hw_cfg_write(struct nfp_net *nn, 2561 struct nfp_net_rx_ring *rx_ring, unsigned int idx) 2562 { 2563 /* Write the DMA address, size and MSI-X info to the device */ 2564 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), rx_ring->dma); 2565 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), ilog2(rx_ring->cnt)); 2566 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), rx_ring->r_vec->irq_entry); 2567 } 2568 2569 static void 2570 nfp_net_tx_ring_hw_cfg_write(struct nfp_net *nn, 2571 struct nfp_net_tx_ring *tx_ring, unsigned int idx) 2572 { 2573 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), tx_ring->dma); 2574 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), ilog2(tx_ring->cnt)); 2575 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), tx_ring->r_vec->irq_entry); 2576 } 2577 2578 /** 2579 * nfp_net_set_config_and_enable() - Write control BAR and enable NFP 2580 * @nn: NFP Net device to reconfigure 2581 */ 2582 static int nfp_net_set_config_and_enable(struct nfp_net *nn) 2583 { 2584 u32 bufsz, new_ctrl, update = 0; 2585 unsigned int r; 2586 int err; 2587 2588 new_ctrl = nn->dp.ctrl; 2589 2590 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_RSS_ANY) { 2591 nfp_net_rss_write_key(nn); 2592 nfp_net_rss_write_itbl(nn); 2593 nn_writel(nn, NFP_NET_CFG_RSS_CTRL, nn->rss_cfg); 2594 update |= NFP_NET_CFG_UPDATE_RSS; 2595 } 2596 2597 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_IRQMOD) { 2598 nfp_net_coalesce_write_cfg(nn); 2599 update |= NFP_NET_CFG_UPDATE_IRQMOD; 2600 } 2601 2602 for (r = 0; r < nn->dp.num_tx_rings; r++) 2603 nfp_net_tx_ring_hw_cfg_write(nn, &nn->dp.tx_rings[r], r); 2604 for (r = 0; r < nn->dp.num_rx_rings; r++) 2605 nfp_net_rx_ring_hw_cfg_write(nn, &nn->dp.rx_rings[r], r); 2606 2607 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, nn->dp.num_tx_rings == 64 ? 2608 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_tx_rings) - 1); 2609 2610 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, nn->dp.num_rx_rings == 64 ? 2611 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_rx_rings) - 1); 2612 2613 if (nn->dp.netdev) 2614 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr); 2615 2616 nn_writel(nn, NFP_NET_CFG_MTU, nn->dp.mtu); 2617 2618 bufsz = nn->dp.fl_bufsz - nn->dp.rx_dma_off - NFP_NET_RX_BUF_NON_DATA; 2619 nn_writel(nn, NFP_NET_CFG_FLBUFSZ, bufsz); 2620 2621 /* Enable device */ 2622 new_ctrl |= NFP_NET_CFG_CTRL_ENABLE; 2623 update |= NFP_NET_CFG_UPDATE_GEN; 2624 update |= NFP_NET_CFG_UPDATE_MSIX; 2625 update |= NFP_NET_CFG_UPDATE_RING; 2626 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG) 2627 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG; 2628 2629 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2630 err = nfp_net_reconfig(nn, update); 2631 if (err) { 2632 nfp_net_clear_config_and_disable(nn); 2633 return err; 2634 } 2635 2636 nn->dp.ctrl = new_ctrl; 2637 2638 for (r = 0; r < nn->dp.num_rx_rings; r++) 2639 nfp_net_rx_ring_fill_freelist(&nn->dp, &nn->dp.rx_rings[r]); 2640 2641 /* Since reconfiguration requests while NFP is down are ignored we 2642 * have to wipe the entire VXLAN configuration and reinitialize it. 2643 */ 2644 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN) { 2645 memset(&nn->vxlan_ports, 0, sizeof(nn->vxlan_ports)); 2646 memset(&nn->vxlan_usecnt, 0, sizeof(nn->vxlan_usecnt)); 2647 udp_tunnel_get_rx_info(nn->dp.netdev); 2648 } 2649 2650 return 0; 2651 } 2652 2653 /** 2654 * nfp_net_close_stack() - Quiesce the stack (part of close) 2655 * @nn: NFP Net device to reconfigure 2656 */ 2657 static void nfp_net_close_stack(struct nfp_net *nn) 2658 { 2659 unsigned int r; 2660 2661 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2662 netif_carrier_off(nn->dp.netdev); 2663 nn->link_up = false; 2664 2665 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2666 disable_irq(nn->r_vecs[r].irq_vector); 2667 napi_disable(&nn->r_vecs[r].napi); 2668 } 2669 2670 netif_tx_disable(nn->dp.netdev); 2671 } 2672 2673 /** 2674 * nfp_net_close_free_all() - Free all runtime resources 2675 * @nn: NFP Net device to reconfigure 2676 */ 2677 static void nfp_net_close_free_all(struct nfp_net *nn) 2678 { 2679 unsigned int r; 2680 2681 nfp_net_tx_rings_free(&nn->dp); 2682 nfp_net_rx_rings_free(&nn->dp); 2683 2684 for (r = 0; r < nn->dp.num_r_vecs; r++) 2685 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 2686 2687 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX); 2688 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX); 2689 } 2690 2691 /** 2692 * nfp_net_netdev_close() - Called when the device is downed 2693 * @netdev: netdev structure 2694 */ 2695 static int nfp_net_netdev_close(struct net_device *netdev) 2696 { 2697 struct nfp_net *nn = netdev_priv(netdev); 2698 2699 /* Step 1: Disable RX and TX rings from the Linux kernel perspective 2700 */ 2701 nfp_net_close_stack(nn); 2702 2703 /* Step 2: Tell NFP 2704 */ 2705 nfp_net_clear_config_and_disable(nn); 2706 nfp_port_configure(netdev, false); 2707 2708 /* Step 3: Free resources 2709 */ 2710 nfp_net_close_free_all(nn); 2711 2712 nn_dbg(nn, "%s down", netdev->name); 2713 return 0; 2714 } 2715 2716 void nfp_ctrl_close(struct nfp_net *nn) 2717 { 2718 int r; 2719 2720 rtnl_lock(); 2721 2722 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2723 disable_irq(nn->r_vecs[r].irq_vector); 2724 tasklet_disable(&nn->r_vecs[r].tasklet); 2725 } 2726 2727 nfp_net_clear_config_and_disable(nn); 2728 2729 nfp_net_close_free_all(nn); 2730 2731 rtnl_unlock(); 2732 } 2733 2734 /** 2735 * nfp_net_open_stack() - Start the device from stack's perspective 2736 * @nn: NFP Net device to reconfigure 2737 */ 2738 static void nfp_net_open_stack(struct nfp_net *nn) 2739 { 2740 unsigned int r; 2741 2742 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2743 napi_enable(&nn->r_vecs[r].napi); 2744 enable_irq(nn->r_vecs[r].irq_vector); 2745 } 2746 2747 netif_tx_wake_all_queues(nn->dp.netdev); 2748 2749 enable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2750 nfp_net_read_link_status(nn); 2751 } 2752 2753 static int nfp_net_open_alloc_all(struct nfp_net *nn) 2754 { 2755 int err, r; 2756 2757 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_EXN, "%s-exn", 2758 nn->exn_name, sizeof(nn->exn_name), 2759 NFP_NET_IRQ_EXN_IDX, nn->exn_handler); 2760 if (err) 2761 return err; 2762 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_LSC, "%s-lsc", 2763 nn->lsc_name, sizeof(nn->lsc_name), 2764 NFP_NET_IRQ_LSC_IDX, nn->lsc_handler); 2765 if (err) 2766 goto err_free_exn; 2767 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2768 2769 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2770 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r); 2771 if (err) 2772 goto err_cleanup_vec_p; 2773 } 2774 2775 err = nfp_net_rx_rings_prepare(nn, &nn->dp); 2776 if (err) 2777 goto err_cleanup_vec; 2778 2779 err = nfp_net_tx_rings_prepare(nn, &nn->dp); 2780 if (err) 2781 goto err_free_rx_rings; 2782 2783 for (r = 0; r < nn->max_r_vecs; r++) 2784 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r); 2785 2786 return 0; 2787 2788 err_free_rx_rings: 2789 nfp_net_rx_rings_free(&nn->dp); 2790 err_cleanup_vec: 2791 r = nn->dp.num_r_vecs; 2792 err_cleanup_vec_p: 2793 while (r--) 2794 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 2795 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX); 2796 err_free_exn: 2797 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX); 2798 return err; 2799 } 2800 2801 static int nfp_net_netdev_open(struct net_device *netdev) 2802 { 2803 struct nfp_net *nn = netdev_priv(netdev); 2804 int err; 2805 2806 /* Step 1: Allocate resources for rings and the like 2807 * - Request interrupts 2808 * - Allocate RX and TX ring resources 2809 * - Setup initial RSS table 2810 */ 2811 err = nfp_net_open_alloc_all(nn); 2812 if (err) 2813 return err; 2814 2815 err = netif_set_real_num_tx_queues(netdev, nn->dp.num_stack_tx_rings); 2816 if (err) 2817 goto err_free_all; 2818 2819 err = netif_set_real_num_rx_queues(netdev, nn->dp.num_rx_rings); 2820 if (err) 2821 goto err_free_all; 2822 2823 /* Step 2: Configure the NFP 2824 * - Ifup the physical interface if it exists 2825 * - Enable rings from 0 to tx_rings/rx_rings - 1. 2826 * - Write MAC address (in case it changed) 2827 * - Set the MTU 2828 * - Set the Freelist buffer size 2829 * - Enable the FW 2830 */ 2831 err = nfp_port_configure(netdev, true); 2832 if (err) 2833 goto err_free_all; 2834 2835 err = nfp_net_set_config_and_enable(nn); 2836 if (err) 2837 goto err_port_disable; 2838 2839 /* Step 3: Enable for kernel 2840 * - put some freelist descriptors on each RX ring 2841 * - enable NAPI on each ring 2842 * - enable all TX queues 2843 * - set link state 2844 */ 2845 nfp_net_open_stack(nn); 2846 2847 return 0; 2848 2849 err_port_disable: 2850 nfp_port_configure(netdev, false); 2851 err_free_all: 2852 nfp_net_close_free_all(nn); 2853 return err; 2854 } 2855 2856 int nfp_ctrl_open(struct nfp_net *nn) 2857 { 2858 int err, r; 2859 2860 /* ring dumping depends on vNICs being opened/closed under rtnl */ 2861 rtnl_lock(); 2862 2863 err = nfp_net_open_alloc_all(nn); 2864 if (err) 2865 goto err_unlock; 2866 2867 err = nfp_net_set_config_and_enable(nn); 2868 if (err) 2869 goto err_free_all; 2870 2871 for (r = 0; r < nn->dp.num_r_vecs; r++) 2872 enable_irq(nn->r_vecs[r].irq_vector); 2873 2874 rtnl_unlock(); 2875 2876 return 0; 2877 2878 err_free_all: 2879 nfp_net_close_free_all(nn); 2880 err_unlock: 2881 rtnl_unlock(); 2882 return err; 2883 } 2884 2885 static void nfp_net_set_rx_mode(struct net_device *netdev) 2886 { 2887 struct nfp_net *nn = netdev_priv(netdev); 2888 u32 new_ctrl; 2889 2890 new_ctrl = nn->dp.ctrl; 2891 2892 if (!netdev_mc_empty(netdev) || netdev->flags & IFF_ALLMULTI) 2893 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_L2MC; 2894 else 2895 new_ctrl &= ~NFP_NET_CFG_CTRL_L2MC; 2896 2897 if (netdev->flags & IFF_PROMISC) { 2898 if (nn->cap & NFP_NET_CFG_CTRL_PROMISC) 2899 new_ctrl |= NFP_NET_CFG_CTRL_PROMISC; 2900 else 2901 nn_warn(nn, "FW does not support promiscuous mode\n"); 2902 } else { 2903 new_ctrl &= ~NFP_NET_CFG_CTRL_PROMISC; 2904 } 2905 2906 if (new_ctrl == nn->dp.ctrl) 2907 return; 2908 2909 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2910 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_GEN); 2911 2912 nn->dp.ctrl = new_ctrl; 2913 } 2914 2915 static void nfp_net_rss_init_itbl(struct nfp_net *nn) 2916 { 2917 int i; 2918 2919 for (i = 0; i < sizeof(nn->rss_itbl); i++) 2920 nn->rss_itbl[i] = 2921 ethtool_rxfh_indir_default(i, nn->dp.num_rx_rings); 2922 } 2923 2924 static void nfp_net_dp_swap(struct nfp_net *nn, struct nfp_net_dp *dp) 2925 { 2926 struct nfp_net_dp new_dp = *dp; 2927 2928 *dp = nn->dp; 2929 nn->dp = new_dp; 2930 2931 nn->dp.netdev->mtu = new_dp.mtu; 2932 2933 if (!netif_is_rxfh_configured(nn->dp.netdev)) 2934 nfp_net_rss_init_itbl(nn); 2935 } 2936 2937 static int nfp_net_dp_swap_enable(struct nfp_net *nn, struct nfp_net_dp *dp) 2938 { 2939 unsigned int r; 2940 int err; 2941 2942 nfp_net_dp_swap(nn, dp); 2943 2944 for (r = 0; r < nn->max_r_vecs; r++) 2945 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r); 2946 2947 err = netif_set_real_num_rx_queues(nn->dp.netdev, nn->dp.num_rx_rings); 2948 if (err) 2949 return err; 2950 2951 if (nn->dp.netdev->real_num_tx_queues != nn->dp.num_stack_tx_rings) { 2952 err = netif_set_real_num_tx_queues(nn->dp.netdev, 2953 nn->dp.num_stack_tx_rings); 2954 if (err) 2955 return err; 2956 } 2957 2958 return nfp_net_set_config_and_enable(nn); 2959 } 2960 2961 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn) 2962 { 2963 struct nfp_net_dp *new; 2964 2965 new = kmalloc(sizeof(*new), GFP_KERNEL); 2966 if (!new) 2967 return NULL; 2968 2969 *new = nn->dp; 2970 2971 /* Clear things which need to be recomputed */ 2972 new->fl_bufsz = 0; 2973 new->tx_rings = NULL; 2974 new->rx_rings = NULL; 2975 new->num_r_vecs = 0; 2976 new->num_stack_tx_rings = 0; 2977 2978 return new; 2979 } 2980 2981 static int 2982 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp, 2983 struct netlink_ext_ack *extack) 2984 { 2985 /* XDP-enabled tests */ 2986 if (!dp->xdp_prog) 2987 return 0; 2988 if (dp->fl_bufsz > PAGE_SIZE) { 2989 NL_SET_ERR_MSG_MOD(extack, "MTU too large w/ XDP enabled"); 2990 return -EINVAL; 2991 } 2992 if (dp->num_tx_rings > nn->max_tx_rings) { 2993 NL_SET_ERR_MSG_MOD(extack, "Insufficient number of TX rings w/ XDP enabled"); 2994 return -EINVAL; 2995 } 2996 2997 return 0; 2998 } 2999 3000 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp, 3001 struct netlink_ext_ack *extack) 3002 { 3003 int r, err; 3004 3005 dp->fl_bufsz = nfp_net_calc_fl_bufsz(dp); 3006 3007 dp->num_stack_tx_rings = dp->num_tx_rings; 3008 if (dp->xdp_prog) 3009 dp->num_stack_tx_rings -= dp->num_rx_rings; 3010 3011 dp->num_r_vecs = max(dp->num_rx_rings, dp->num_stack_tx_rings); 3012 3013 err = nfp_net_check_config(nn, dp, extack); 3014 if (err) 3015 goto exit_free_dp; 3016 3017 if (!netif_running(dp->netdev)) { 3018 nfp_net_dp_swap(nn, dp); 3019 err = 0; 3020 goto exit_free_dp; 3021 } 3022 3023 /* Prepare new rings */ 3024 for (r = nn->dp.num_r_vecs; r < dp->num_r_vecs; r++) { 3025 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r); 3026 if (err) { 3027 dp->num_r_vecs = r; 3028 goto err_cleanup_vecs; 3029 } 3030 } 3031 3032 err = nfp_net_rx_rings_prepare(nn, dp); 3033 if (err) 3034 goto err_cleanup_vecs; 3035 3036 err = nfp_net_tx_rings_prepare(nn, dp); 3037 if (err) 3038 goto err_free_rx; 3039 3040 /* Stop device, swap in new rings, try to start the firmware */ 3041 nfp_net_close_stack(nn); 3042 nfp_net_clear_config_and_disable(nn); 3043 3044 err = nfp_net_dp_swap_enable(nn, dp); 3045 if (err) { 3046 int err2; 3047 3048 nfp_net_clear_config_and_disable(nn); 3049 3050 /* Try with old configuration and old rings */ 3051 err2 = nfp_net_dp_swap_enable(nn, dp); 3052 if (err2) 3053 nn_err(nn, "Can't restore ring config - FW communication failed (%d,%d)\n", 3054 err, err2); 3055 } 3056 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--) 3057 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 3058 3059 nfp_net_rx_rings_free(dp); 3060 nfp_net_tx_rings_free(dp); 3061 3062 nfp_net_open_stack(nn); 3063 exit_free_dp: 3064 kfree(dp); 3065 3066 return err; 3067 3068 err_free_rx: 3069 nfp_net_rx_rings_free(dp); 3070 err_cleanup_vecs: 3071 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--) 3072 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 3073 kfree(dp); 3074 return err; 3075 } 3076 3077 static int nfp_net_change_mtu(struct net_device *netdev, int new_mtu) 3078 { 3079 struct nfp_net *nn = netdev_priv(netdev); 3080 struct nfp_net_dp *dp; 3081 int err; 3082 3083 err = nfp_app_check_mtu(nn->app, netdev, new_mtu); 3084 if (err) 3085 return err; 3086 3087 dp = nfp_net_clone_dp(nn); 3088 if (!dp) 3089 return -ENOMEM; 3090 3091 dp->mtu = new_mtu; 3092 3093 return nfp_net_ring_reconfig(nn, dp, NULL); 3094 } 3095 3096 static int 3097 nfp_net_vlan_rx_add_vid(struct net_device *netdev, __be16 proto, u16 vid) 3098 { 3099 struct nfp_net *nn = netdev_priv(netdev); 3100 3101 /* Priority tagged packets with vlan id 0 are processed by the 3102 * NFP as untagged packets 3103 */ 3104 if (!vid) 3105 return 0; 3106 3107 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid); 3108 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO, 3109 ETH_P_8021Q); 3110 3111 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_ADD); 3112 } 3113 3114 static int 3115 nfp_net_vlan_rx_kill_vid(struct net_device *netdev, __be16 proto, u16 vid) 3116 { 3117 struct nfp_net *nn = netdev_priv(netdev); 3118 3119 /* Priority tagged packets with vlan id 0 are processed by the 3120 * NFP as untagged packets 3121 */ 3122 if (!vid) 3123 return 0; 3124 3125 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid); 3126 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO, 3127 ETH_P_8021Q); 3128 3129 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_KILL); 3130 } 3131 3132 static void nfp_net_stat64(struct net_device *netdev, 3133 struct rtnl_link_stats64 *stats) 3134 { 3135 struct nfp_net *nn = netdev_priv(netdev); 3136 int r; 3137 3138 /* Collect software stats */ 3139 for (r = 0; r < nn->max_r_vecs; r++) { 3140 struct nfp_net_r_vector *r_vec = &nn->r_vecs[r]; 3141 u64 data[3]; 3142 unsigned int start; 3143 3144 do { 3145 start = u64_stats_fetch_begin(&r_vec->rx_sync); 3146 data[0] = r_vec->rx_pkts; 3147 data[1] = r_vec->rx_bytes; 3148 data[2] = r_vec->rx_drops; 3149 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start)); 3150 stats->rx_packets += data[0]; 3151 stats->rx_bytes += data[1]; 3152 stats->rx_dropped += data[2]; 3153 3154 do { 3155 start = u64_stats_fetch_begin(&r_vec->tx_sync); 3156 data[0] = r_vec->tx_pkts; 3157 data[1] = r_vec->tx_bytes; 3158 data[2] = r_vec->tx_errors; 3159 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start)); 3160 stats->tx_packets += data[0]; 3161 stats->tx_bytes += data[1]; 3162 stats->tx_errors += data[2]; 3163 } 3164 3165 /* Add in device stats */ 3166 stats->multicast += nn_readq(nn, NFP_NET_CFG_STATS_RX_MC_FRAMES); 3167 stats->rx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_RX_DISCARDS); 3168 stats->rx_errors += nn_readq(nn, NFP_NET_CFG_STATS_RX_ERRORS); 3169 3170 stats->tx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_TX_DISCARDS); 3171 stats->tx_errors += nn_readq(nn, NFP_NET_CFG_STATS_TX_ERRORS); 3172 } 3173 3174 static int nfp_net_set_features(struct net_device *netdev, 3175 netdev_features_t features) 3176 { 3177 netdev_features_t changed = netdev->features ^ features; 3178 struct nfp_net *nn = netdev_priv(netdev); 3179 u32 new_ctrl; 3180 int err; 3181 3182 /* Assume this is not called with features we have not advertised */ 3183 3184 new_ctrl = nn->dp.ctrl; 3185 3186 if (changed & NETIF_F_RXCSUM) { 3187 if (features & NETIF_F_RXCSUM) 3188 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY; 3189 else 3190 new_ctrl &= ~NFP_NET_CFG_CTRL_RXCSUM_ANY; 3191 } 3192 3193 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3194 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) 3195 new_ctrl |= NFP_NET_CFG_CTRL_TXCSUM; 3196 else 3197 new_ctrl &= ~NFP_NET_CFG_CTRL_TXCSUM; 3198 } 3199 3200 if (changed & (NETIF_F_TSO | NETIF_F_TSO6)) { 3201 if (features & (NETIF_F_TSO | NETIF_F_TSO6)) 3202 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?: 3203 NFP_NET_CFG_CTRL_LSO; 3204 else 3205 new_ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY; 3206 } 3207 3208 if (changed & NETIF_F_HW_VLAN_CTAG_RX) { 3209 if (features & NETIF_F_HW_VLAN_CTAG_RX) 3210 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN; 3211 else 3212 new_ctrl &= ~NFP_NET_CFG_CTRL_RXVLAN; 3213 } 3214 3215 if (changed & NETIF_F_HW_VLAN_CTAG_TX) { 3216 if (features & NETIF_F_HW_VLAN_CTAG_TX) 3217 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN; 3218 else 3219 new_ctrl &= ~NFP_NET_CFG_CTRL_TXVLAN; 3220 } 3221 3222 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) { 3223 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) 3224 new_ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER; 3225 else 3226 new_ctrl &= ~NFP_NET_CFG_CTRL_CTAG_FILTER; 3227 } 3228 3229 if (changed & NETIF_F_SG) { 3230 if (features & NETIF_F_SG) 3231 new_ctrl |= NFP_NET_CFG_CTRL_GATHER; 3232 else 3233 new_ctrl &= ~NFP_NET_CFG_CTRL_GATHER; 3234 } 3235 3236 err = nfp_port_set_features(netdev, features); 3237 if (err) 3238 return err; 3239 3240 nn_dbg(nn, "Feature change 0x%llx -> 0x%llx (changed=0x%llx)\n", 3241 netdev->features, features, changed); 3242 3243 if (new_ctrl == nn->dp.ctrl) 3244 return 0; 3245 3246 nn_dbg(nn, "NIC ctrl: 0x%x -> 0x%x\n", nn->dp.ctrl, new_ctrl); 3247 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 3248 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN); 3249 if (err) 3250 return err; 3251 3252 nn->dp.ctrl = new_ctrl; 3253 3254 return 0; 3255 } 3256 3257 static netdev_features_t 3258 nfp_net_features_check(struct sk_buff *skb, struct net_device *dev, 3259 netdev_features_t features) 3260 { 3261 u8 l4_hdr; 3262 3263 /* We can't do TSO over double tagged packets (802.1AD) */ 3264 features &= vlan_features_check(skb, features); 3265 3266 if (!skb->encapsulation) 3267 return features; 3268 3269 /* Ensure that inner L4 header offset fits into TX descriptor field */ 3270 if (skb_is_gso(skb)) { 3271 u32 hdrlen; 3272 3273 hdrlen = skb_inner_transport_header(skb) - skb->data + 3274 inner_tcp_hdrlen(skb); 3275 3276 if (unlikely(hdrlen > NFP_NET_LSO_MAX_HDR_SZ)) 3277 features &= ~NETIF_F_GSO_MASK; 3278 } 3279 3280 /* VXLAN/GRE check */ 3281 switch (vlan_get_protocol(skb)) { 3282 case htons(ETH_P_IP): 3283 l4_hdr = ip_hdr(skb)->protocol; 3284 break; 3285 case htons(ETH_P_IPV6): 3286 l4_hdr = ipv6_hdr(skb)->nexthdr; 3287 break; 3288 default: 3289 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3290 } 3291 3292 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER || 3293 skb->inner_protocol != htons(ETH_P_TEB) || 3294 (l4_hdr != IPPROTO_UDP && l4_hdr != IPPROTO_GRE) || 3295 (l4_hdr == IPPROTO_UDP && 3296 (skb_inner_mac_header(skb) - skb_transport_header(skb) != 3297 sizeof(struct udphdr) + sizeof(struct vxlanhdr)))) 3298 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3299 3300 return features; 3301 } 3302 3303 static int 3304 nfp_net_get_phys_port_name(struct net_device *netdev, char *name, size_t len) 3305 { 3306 struct nfp_net *nn = netdev_priv(netdev); 3307 int n; 3308 3309 if (nn->port) 3310 return nfp_port_get_phys_port_name(netdev, name, len); 3311 3312 if (nn->dp.is_vf || nn->vnic_no_name) 3313 return -EOPNOTSUPP; 3314 3315 n = snprintf(name, len, "n%d", nn->id); 3316 if (n >= len) 3317 return -EINVAL; 3318 3319 return 0; 3320 } 3321 3322 /** 3323 * nfp_net_set_vxlan_port() - set vxlan port in SW and reconfigure HW 3324 * @nn: NFP Net device to reconfigure 3325 * @idx: Index into the port table where new port should be written 3326 * @port: UDP port to configure (pass zero to remove VXLAN port) 3327 */ 3328 static void nfp_net_set_vxlan_port(struct nfp_net *nn, int idx, __be16 port) 3329 { 3330 int i; 3331 3332 nn->vxlan_ports[idx] = port; 3333 3334 if (!(nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN)) 3335 return; 3336 3337 BUILD_BUG_ON(NFP_NET_N_VXLAN_PORTS & 1); 3338 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i += 2) 3339 nn_writel(nn, NFP_NET_CFG_VXLAN_PORT + i * sizeof(port), 3340 be16_to_cpu(nn->vxlan_ports[i + 1]) << 16 | 3341 be16_to_cpu(nn->vxlan_ports[i])); 3342 3343 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_VXLAN); 3344 } 3345 3346 /** 3347 * nfp_net_find_vxlan_idx() - find table entry of the port or a free one 3348 * @nn: NFP Network structure 3349 * @port: UDP port to look for 3350 * 3351 * Return: if the port is already in the table -- it's position; 3352 * if the port is not in the table -- free position to use; 3353 * if the table is full -- -ENOSPC. 3354 */ 3355 static int nfp_net_find_vxlan_idx(struct nfp_net *nn, __be16 port) 3356 { 3357 int i, free_idx = -ENOSPC; 3358 3359 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i++) { 3360 if (nn->vxlan_ports[i] == port) 3361 return i; 3362 if (!nn->vxlan_usecnt[i]) 3363 free_idx = i; 3364 } 3365 3366 return free_idx; 3367 } 3368 3369 static void nfp_net_add_vxlan_port(struct net_device *netdev, 3370 struct udp_tunnel_info *ti) 3371 { 3372 struct nfp_net *nn = netdev_priv(netdev); 3373 int idx; 3374 3375 if (ti->type != UDP_TUNNEL_TYPE_VXLAN) 3376 return; 3377 3378 idx = nfp_net_find_vxlan_idx(nn, ti->port); 3379 if (idx == -ENOSPC) 3380 return; 3381 3382 if (!nn->vxlan_usecnt[idx]++) 3383 nfp_net_set_vxlan_port(nn, idx, ti->port); 3384 } 3385 3386 static void nfp_net_del_vxlan_port(struct net_device *netdev, 3387 struct udp_tunnel_info *ti) 3388 { 3389 struct nfp_net *nn = netdev_priv(netdev); 3390 int idx; 3391 3392 if (ti->type != UDP_TUNNEL_TYPE_VXLAN) 3393 return; 3394 3395 idx = nfp_net_find_vxlan_idx(nn, ti->port); 3396 if (idx == -ENOSPC || !nn->vxlan_usecnt[idx]) 3397 return; 3398 3399 if (!--nn->vxlan_usecnt[idx]) 3400 nfp_net_set_vxlan_port(nn, idx, 0); 3401 } 3402 3403 static int nfp_net_xdp_setup_drv(struct nfp_net *nn, struct netdev_bpf *bpf) 3404 { 3405 struct bpf_prog *prog = bpf->prog; 3406 struct nfp_net_dp *dp; 3407 int err; 3408 3409 if (!xdp_attachment_flags_ok(&nn->xdp, bpf)) 3410 return -EBUSY; 3411 3412 if (!prog == !nn->dp.xdp_prog) { 3413 WRITE_ONCE(nn->dp.xdp_prog, prog); 3414 xdp_attachment_setup(&nn->xdp, bpf); 3415 return 0; 3416 } 3417 3418 dp = nfp_net_clone_dp(nn); 3419 if (!dp) 3420 return -ENOMEM; 3421 3422 dp->xdp_prog = prog; 3423 dp->num_tx_rings += prog ? nn->dp.num_rx_rings : -nn->dp.num_rx_rings; 3424 dp->rx_dma_dir = prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE; 3425 dp->rx_dma_off = prog ? XDP_PACKET_HEADROOM - nn->dp.rx_offset : 0; 3426 3427 /* We need RX reconfig to remap the buffers (BIDIR vs FROM_DEV) */ 3428 err = nfp_net_ring_reconfig(nn, dp, bpf->extack); 3429 if (err) 3430 return err; 3431 3432 xdp_attachment_setup(&nn->xdp, bpf); 3433 return 0; 3434 } 3435 3436 static int nfp_net_xdp_setup_hw(struct nfp_net *nn, struct netdev_bpf *bpf) 3437 { 3438 int err; 3439 3440 if (!xdp_attachment_flags_ok(&nn->xdp_hw, bpf)) 3441 return -EBUSY; 3442 3443 err = nfp_app_xdp_offload(nn->app, nn, bpf->prog, bpf->extack); 3444 if (err) 3445 return err; 3446 3447 xdp_attachment_setup(&nn->xdp_hw, bpf); 3448 return 0; 3449 } 3450 3451 static int nfp_net_xdp(struct net_device *netdev, struct netdev_bpf *xdp) 3452 { 3453 struct nfp_net *nn = netdev_priv(netdev); 3454 3455 switch (xdp->command) { 3456 case XDP_SETUP_PROG: 3457 return nfp_net_xdp_setup_drv(nn, xdp); 3458 case XDP_SETUP_PROG_HW: 3459 return nfp_net_xdp_setup_hw(nn, xdp); 3460 case XDP_QUERY_PROG: 3461 return xdp_attachment_query(&nn->xdp, xdp); 3462 case XDP_QUERY_PROG_HW: 3463 return xdp_attachment_query(&nn->xdp_hw, xdp); 3464 default: 3465 return nfp_app_bpf(nn->app, nn, xdp); 3466 } 3467 } 3468 3469 static int nfp_net_set_mac_address(struct net_device *netdev, void *addr) 3470 { 3471 struct nfp_net *nn = netdev_priv(netdev); 3472 struct sockaddr *saddr = addr; 3473 int err; 3474 3475 err = eth_prepare_mac_addr_change(netdev, addr); 3476 if (err) 3477 return err; 3478 3479 nfp_net_write_mac_addr(nn, saddr->sa_data); 3480 3481 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MACADDR); 3482 if (err) 3483 return err; 3484 3485 eth_commit_mac_addr_change(netdev, addr); 3486 3487 return 0; 3488 } 3489 3490 const struct net_device_ops nfp_net_netdev_ops = { 3491 .ndo_init = nfp_app_ndo_init, 3492 .ndo_uninit = nfp_app_ndo_uninit, 3493 .ndo_open = nfp_net_netdev_open, 3494 .ndo_stop = nfp_net_netdev_close, 3495 .ndo_start_xmit = nfp_net_tx, 3496 .ndo_get_stats64 = nfp_net_stat64, 3497 .ndo_vlan_rx_add_vid = nfp_net_vlan_rx_add_vid, 3498 .ndo_vlan_rx_kill_vid = nfp_net_vlan_rx_kill_vid, 3499 .ndo_set_vf_mac = nfp_app_set_vf_mac, 3500 .ndo_set_vf_vlan = nfp_app_set_vf_vlan, 3501 .ndo_set_vf_spoofchk = nfp_app_set_vf_spoofchk, 3502 .ndo_get_vf_config = nfp_app_get_vf_config, 3503 .ndo_set_vf_link_state = nfp_app_set_vf_link_state, 3504 .ndo_setup_tc = nfp_port_setup_tc, 3505 .ndo_tx_timeout = nfp_net_tx_timeout, 3506 .ndo_set_rx_mode = nfp_net_set_rx_mode, 3507 .ndo_change_mtu = nfp_net_change_mtu, 3508 .ndo_set_mac_address = nfp_net_set_mac_address, 3509 .ndo_set_features = nfp_net_set_features, 3510 .ndo_features_check = nfp_net_features_check, 3511 .ndo_get_phys_port_name = nfp_net_get_phys_port_name, 3512 .ndo_udp_tunnel_add = nfp_net_add_vxlan_port, 3513 .ndo_udp_tunnel_del = nfp_net_del_vxlan_port, 3514 .ndo_bpf = nfp_net_xdp, 3515 }; 3516 3517 /** 3518 * nfp_net_info() - Print general info about the NIC 3519 * @nn: NFP Net device to reconfigure 3520 */ 3521 void nfp_net_info(struct nfp_net *nn) 3522 { 3523 nn_info(nn, "Netronome NFP-6xxx %sNetdev: TxQs=%d/%d RxQs=%d/%d\n", 3524 nn->dp.is_vf ? "VF " : "", 3525 nn->dp.num_tx_rings, nn->max_tx_rings, 3526 nn->dp.num_rx_rings, nn->max_rx_rings); 3527 nn_info(nn, "VER: %d.%d.%d.%d, Maximum supported MTU: %d\n", 3528 nn->fw_ver.resv, nn->fw_ver.class, 3529 nn->fw_ver.major, nn->fw_ver.minor, 3530 nn->max_mtu); 3531 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", 3532 nn->cap, 3533 nn->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "", 3534 nn->cap & NFP_NET_CFG_CTRL_L2BC ? "L2BCFILT " : "", 3535 nn->cap & NFP_NET_CFG_CTRL_L2MC ? "L2MCFILT " : "", 3536 nn->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "", 3537 nn->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "", 3538 nn->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "", 3539 nn->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "", 3540 nn->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "", 3541 nn->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "", 3542 nn->cap & NFP_NET_CFG_CTRL_LSO ? "TSO1 " : "", 3543 nn->cap & NFP_NET_CFG_CTRL_LSO2 ? "TSO2 " : "", 3544 nn->cap & NFP_NET_CFG_CTRL_RSS ? "RSS1 " : "", 3545 nn->cap & NFP_NET_CFG_CTRL_RSS2 ? "RSS2 " : "", 3546 nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER ? "CTAG_FILTER " : "", 3547 nn->cap & NFP_NET_CFG_CTRL_L2SWITCH ? "L2SWITCH " : "", 3548 nn->cap & NFP_NET_CFG_CTRL_MSIXAUTO ? "AUTOMASK " : "", 3549 nn->cap & NFP_NET_CFG_CTRL_IRQMOD ? "IRQMOD " : "", 3550 nn->cap & NFP_NET_CFG_CTRL_VXLAN ? "VXLAN " : "", 3551 nn->cap & NFP_NET_CFG_CTRL_NVGRE ? "NVGRE " : "", 3552 nn->cap & NFP_NET_CFG_CTRL_CSUM_COMPLETE ? 3553 "RXCSUM_COMPLETE " : "", 3554 nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR ? "LIVE_ADDR " : "", 3555 nfp_app_extra_cap(nn->app, nn)); 3556 } 3557 3558 /** 3559 * nfp_net_alloc() - Allocate netdev and related structure 3560 * @pdev: PCI device 3561 * @ctrl_bar: PCI IOMEM with vNIC config memory 3562 * @needs_netdev: Whether to allocate a netdev for this vNIC 3563 * @max_tx_rings: Maximum number of TX rings supported by device 3564 * @max_rx_rings: Maximum number of RX rings supported by device 3565 * 3566 * This function allocates a netdev device and fills in the initial 3567 * part of the @struct nfp_net structure. In case of control device 3568 * nfp_net structure is allocated without the netdev. 3569 * 3570 * Return: NFP Net device structure, or ERR_PTR on error. 3571 */ 3572 struct nfp_net * 3573 nfp_net_alloc(struct pci_dev *pdev, void __iomem *ctrl_bar, bool needs_netdev, 3574 unsigned int max_tx_rings, unsigned int max_rx_rings) 3575 { 3576 struct nfp_net *nn; 3577 int err; 3578 3579 if (needs_netdev) { 3580 struct net_device *netdev; 3581 3582 netdev = alloc_etherdev_mqs(sizeof(struct nfp_net), 3583 max_tx_rings, max_rx_rings); 3584 if (!netdev) 3585 return ERR_PTR(-ENOMEM); 3586 3587 SET_NETDEV_DEV(netdev, &pdev->dev); 3588 nn = netdev_priv(netdev); 3589 nn->dp.netdev = netdev; 3590 } else { 3591 nn = vzalloc(sizeof(*nn)); 3592 if (!nn) 3593 return ERR_PTR(-ENOMEM); 3594 } 3595 3596 nn->dp.dev = &pdev->dev; 3597 nn->dp.ctrl_bar = ctrl_bar; 3598 nn->pdev = pdev; 3599 3600 nn->max_tx_rings = max_tx_rings; 3601 nn->max_rx_rings = max_rx_rings; 3602 3603 nn->dp.num_tx_rings = min_t(unsigned int, 3604 max_tx_rings, num_online_cpus()); 3605 nn->dp.num_rx_rings = min_t(unsigned int, max_rx_rings, 3606 netif_get_num_default_rss_queues()); 3607 3608 nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings); 3609 nn->dp.num_r_vecs = min_t(unsigned int, 3610 nn->dp.num_r_vecs, num_online_cpus()); 3611 3612 nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT; 3613 nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT; 3614 3615 spin_lock_init(&nn->reconfig_lock); 3616 spin_lock_init(&nn->link_status_lock); 3617 3618 timer_setup(&nn->reconfig_timer, nfp_net_reconfig_timer, 0); 3619 3620 err = nfp_net_tlv_caps_parse(&nn->pdev->dev, nn->dp.ctrl_bar, 3621 &nn->tlv_caps); 3622 if (err) 3623 goto err_free_nn; 3624 3625 return nn; 3626 3627 err_free_nn: 3628 if (nn->dp.netdev) 3629 free_netdev(nn->dp.netdev); 3630 else 3631 vfree(nn); 3632 return ERR_PTR(err); 3633 } 3634 3635 /** 3636 * nfp_net_free() - Undo what @nfp_net_alloc() did 3637 * @nn: NFP Net device to reconfigure 3638 */ 3639 void nfp_net_free(struct nfp_net *nn) 3640 { 3641 WARN_ON(timer_pending(&nn->reconfig_timer) || nn->reconfig_posted); 3642 if (nn->dp.netdev) 3643 free_netdev(nn->dp.netdev); 3644 else 3645 vfree(nn); 3646 } 3647 3648 /** 3649 * nfp_net_rss_key_sz() - Get current size of the RSS key 3650 * @nn: NFP Net device instance 3651 * 3652 * Return: size of the RSS key for currently selected hash function. 3653 */ 3654 unsigned int nfp_net_rss_key_sz(struct nfp_net *nn) 3655 { 3656 switch (nn->rss_hfunc) { 3657 case ETH_RSS_HASH_TOP: 3658 return NFP_NET_CFG_RSS_KEY_SZ; 3659 case ETH_RSS_HASH_XOR: 3660 return 0; 3661 case ETH_RSS_HASH_CRC32: 3662 return 4; 3663 } 3664 3665 nn_warn(nn, "Unknown hash function: %u\n", nn->rss_hfunc); 3666 return 0; 3667 } 3668 3669 /** 3670 * nfp_net_rss_init() - Set the initial RSS parameters 3671 * @nn: NFP Net device to reconfigure 3672 */ 3673 static void nfp_net_rss_init(struct nfp_net *nn) 3674 { 3675 unsigned long func_bit, rss_cap_hfunc; 3676 u32 reg; 3677 3678 /* Read the RSS function capability and select first supported func */ 3679 reg = nn_readl(nn, NFP_NET_CFG_RSS_CAP); 3680 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, reg); 3681 if (!rss_cap_hfunc) 3682 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, 3683 NFP_NET_CFG_RSS_TOEPLITZ); 3684 3685 func_bit = find_first_bit(&rss_cap_hfunc, NFP_NET_CFG_RSS_HFUNCS); 3686 if (func_bit == NFP_NET_CFG_RSS_HFUNCS) { 3687 dev_warn(nn->dp.dev, 3688 "Bad RSS config, defaulting to Toeplitz hash\n"); 3689 func_bit = ETH_RSS_HASH_TOP_BIT; 3690 } 3691 nn->rss_hfunc = 1 << func_bit; 3692 3693 netdev_rss_key_fill(nn->rss_key, nfp_net_rss_key_sz(nn)); 3694 3695 nfp_net_rss_init_itbl(nn); 3696 3697 /* Enable IPv4/IPv6 TCP by default */ 3698 nn->rss_cfg = NFP_NET_CFG_RSS_IPV4_TCP | 3699 NFP_NET_CFG_RSS_IPV6_TCP | 3700 FIELD_PREP(NFP_NET_CFG_RSS_HFUNC, nn->rss_hfunc) | 3701 NFP_NET_CFG_RSS_MASK; 3702 } 3703 3704 /** 3705 * nfp_net_irqmod_init() - Set the initial IRQ moderation parameters 3706 * @nn: NFP Net device to reconfigure 3707 */ 3708 static void nfp_net_irqmod_init(struct nfp_net *nn) 3709 { 3710 nn->rx_coalesce_usecs = 50; 3711 nn->rx_coalesce_max_frames = 64; 3712 nn->tx_coalesce_usecs = 50; 3713 nn->tx_coalesce_max_frames = 64; 3714 } 3715 3716 static void nfp_net_netdev_init(struct nfp_net *nn) 3717 { 3718 struct net_device *netdev = nn->dp.netdev; 3719 3720 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr); 3721 3722 netdev->mtu = nn->dp.mtu; 3723 3724 /* Advertise/enable offloads based on capabilities 3725 * 3726 * Note: netdev->features show the currently enabled features 3727 * and netdev->hw_features advertises which features are 3728 * supported. By default we enable most features. 3729 */ 3730 if (nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR) 3731 netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 3732 3733 netdev->hw_features = NETIF_F_HIGHDMA; 3734 if (nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY) { 3735 netdev->hw_features |= NETIF_F_RXCSUM; 3736 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY; 3737 } 3738 if (nn->cap & NFP_NET_CFG_CTRL_TXCSUM) { 3739 netdev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 3740 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXCSUM; 3741 } 3742 if (nn->cap & NFP_NET_CFG_CTRL_GATHER) { 3743 netdev->hw_features |= NETIF_F_SG; 3744 nn->dp.ctrl |= NFP_NET_CFG_CTRL_GATHER; 3745 } 3746 if ((nn->cap & NFP_NET_CFG_CTRL_LSO && nn->fw_ver.major > 2) || 3747 nn->cap & NFP_NET_CFG_CTRL_LSO2) { 3748 netdev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 3749 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?: 3750 NFP_NET_CFG_CTRL_LSO; 3751 } 3752 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) 3753 netdev->hw_features |= NETIF_F_RXHASH; 3754 if (nn->cap & NFP_NET_CFG_CTRL_VXLAN) { 3755 if (nn->cap & NFP_NET_CFG_CTRL_LSO) 3756 netdev->hw_features |= NETIF_F_GSO_UDP_TUNNEL; 3757 nn->dp.ctrl |= NFP_NET_CFG_CTRL_VXLAN; 3758 } 3759 if (nn->cap & NFP_NET_CFG_CTRL_NVGRE) { 3760 if (nn->cap & NFP_NET_CFG_CTRL_LSO) 3761 netdev->hw_features |= NETIF_F_GSO_GRE; 3762 nn->dp.ctrl |= NFP_NET_CFG_CTRL_NVGRE; 3763 } 3764 if (nn->cap & (NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE)) 3765 netdev->hw_enc_features = netdev->hw_features; 3766 3767 netdev->vlan_features = netdev->hw_features; 3768 3769 if (nn->cap & NFP_NET_CFG_CTRL_RXVLAN) { 3770 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX; 3771 nn->dp.ctrl |= NFP_NET_CFG_CTRL_RXVLAN; 3772 } 3773 if (nn->cap & NFP_NET_CFG_CTRL_TXVLAN) { 3774 if (nn->cap & NFP_NET_CFG_CTRL_LSO2) { 3775 nn_warn(nn, "Device advertises both TSO2 and TXVLAN. Refusing to enable TXVLAN.\n"); 3776 } else { 3777 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX; 3778 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXVLAN; 3779 } 3780 } 3781 if (nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER) { 3782 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER; 3783 nn->dp.ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER; 3784 } 3785 3786 netdev->features = netdev->hw_features; 3787 3788 if (nfp_app_has_tc(nn->app) && nn->port) 3789 netdev->hw_features |= NETIF_F_HW_TC; 3790 3791 /* Advertise but disable TSO by default. */ 3792 netdev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6); 3793 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY; 3794 3795 /* Finalise the netdev setup */ 3796 netdev->netdev_ops = &nfp_net_netdev_ops; 3797 netdev->watchdog_timeo = msecs_to_jiffies(5 * 1000); 3798 3799 SWITCHDEV_SET_OPS(netdev, &nfp_port_switchdev_ops); 3800 3801 /* MTU range: 68 - hw-specific max */ 3802 netdev->min_mtu = ETH_MIN_MTU; 3803 netdev->max_mtu = nn->max_mtu; 3804 3805 netdev->gso_max_segs = NFP_NET_LSO_MAX_SEGS; 3806 3807 netif_carrier_off(netdev); 3808 3809 nfp_net_set_ethtool_ops(netdev); 3810 } 3811 3812 static int nfp_net_read_caps(struct nfp_net *nn) 3813 { 3814 /* Get some of the read-only fields from the BAR */ 3815 nn->cap = nn_readl(nn, NFP_NET_CFG_CAP); 3816 nn->max_mtu = nn_readl(nn, NFP_NET_CFG_MAX_MTU); 3817 3818 /* ABI 4.x and ctrl vNIC always use chained metadata, in other cases 3819 * we allow use of non-chained metadata if RSS(v1) is the only 3820 * advertised capability requiring metadata. 3821 */ 3822 nn->dp.chained_metadata_format = nn->fw_ver.major == 4 || 3823 !nn->dp.netdev || 3824 !(nn->cap & NFP_NET_CFG_CTRL_RSS) || 3825 nn->cap & NFP_NET_CFG_CTRL_CHAIN_META; 3826 /* RSS(v1) uses non-chained metadata format, except in ABI 4.x where 3827 * it has the same meaning as RSSv2. 3828 */ 3829 if (nn->dp.chained_metadata_format && nn->fw_ver.major != 4) 3830 nn->cap &= ~NFP_NET_CFG_CTRL_RSS; 3831 3832 /* Determine RX packet/metadata boundary offset */ 3833 if (nn->fw_ver.major >= 2) { 3834 u32 reg; 3835 3836 reg = nn_readl(nn, NFP_NET_CFG_RX_OFFSET); 3837 if (reg > NFP_NET_MAX_PREPEND) { 3838 nn_err(nn, "Invalid rx offset: %d\n", reg); 3839 return -EINVAL; 3840 } 3841 nn->dp.rx_offset = reg; 3842 } else { 3843 nn->dp.rx_offset = NFP_NET_RX_OFFSET; 3844 } 3845 3846 /* For control vNICs mask out the capabilities app doesn't want. */ 3847 if (!nn->dp.netdev) 3848 nn->cap &= nn->app->type->ctrl_cap_mask; 3849 3850 return 0; 3851 } 3852 3853 /** 3854 * nfp_net_init() - Initialise/finalise the nfp_net structure 3855 * @nn: NFP Net device structure 3856 * 3857 * Return: 0 on success or negative errno on error. 3858 */ 3859 int nfp_net_init(struct nfp_net *nn) 3860 { 3861 int err; 3862 3863 nn->dp.rx_dma_dir = DMA_FROM_DEVICE; 3864 3865 err = nfp_net_read_caps(nn); 3866 if (err) 3867 return err; 3868 3869 /* Set default MTU and Freelist buffer size */ 3870 if (!nfp_net_is_data_vnic(nn) && nn->app->ctrl_mtu) { 3871 if (nn->app->ctrl_mtu <= nn->max_mtu) { 3872 nn->dp.mtu = nn->app->ctrl_mtu; 3873 } else { 3874 if (nn->app->ctrl_mtu != NFP_APP_CTRL_MTU_MAX) 3875 nn_warn(nn, "app requested MTU above max supported %u > %u\n", 3876 nn->app->ctrl_mtu, nn->max_mtu); 3877 nn->dp.mtu = nn->max_mtu; 3878 } 3879 } else if (nn->max_mtu < NFP_NET_DEFAULT_MTU) { 3880 nn->dp.mtu = nn->max_mtu; 3881 } else { 3882 nn->dp.mtu = NFP_NET_DEFAULT_MTU; 3883 } 3884 nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp); 3885 3886 if (nfp_app_ctrl_uses_data_vnics(nn->app)) 3887 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_CMSG_DATA; 3888 3889 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) { 3890 nfp_net_rss_init(nn); 3891 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?: 3892 NFP_NET_CFG_CTRL_RSS; 3893 } 3894 3895 /* Allow L2 Broadcast and Multicast through by default, if supported */ 3896 if (nn->cap & NFP_NET_CFG_CTRL_L2BC) 3897 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2BC; 3898 3899 /* Allow IRQ moderation, if supported */ 3900 if (nn->cap & NFP_NET_CFG_CTRL_IRQMOD) { 3901 nfp_net_irqmod_init(nn); 3902 nn->dp.ctrl |= NFP_NET_CFG_CTRL_IRQMOD; 3903 } 3904 3905 if (nn->dp.netdev) 3906 nfp_net_netdev_init(nn); 3907 3908 /* Stash the re-configuration queue away. First odd queue in TX Bar */ 3909 nn->qcp_cfg = nn->tx_bar + NFP_QCP_QUEUE_ADDR_SZ; 3910 3911 /* Make sure the FW knows the netdev is supposed to be disabled here */ 3912 nn_writel(nn, NFP_NET_CFG_CTRL, 0); 3913 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0); 3914 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0); 3915 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_RING | 3916 NFP_NET_CFG_UPDATE_GEN); 3917 if (err) 3918 return err; 3919 3920 nfp_net_vecs_init(nn); 3921 3922 if (!nn->dp.netdev) 3923 return 0; 3924 return register_netdev(nn->dp.netdev); 3925 } 3926 3927 /** 3928 * nfp_net_clean() - Undo what nfp_net_init() did. 3929 * @nn: NFP Net device structure 3930 */ 3931 void nfp_net_clean(struct nfp_net *nn) 3932 { 3933 if (!nn->dp.netdev) 3934 return; 3935 3936 unregister_netdev(nn->dp.netdev); 3937 nfp_net_reconfig_wait_posted(nn); 3938 } 3939