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 nn->reconfig_in_progress_update = update; 105 } 106 107 /* Pass 0 as update to run posted reconfigs. */ 108 static void nfp_net_reconfig_start_async(struct nfp_net *nn, u32 update) 109 { 110 update |= nn->reconfig_posted; 111 nn->reconfig_posted = 0; 112 113 nfp_net_reconfig_start(nn, update); 114 115 nn->reconfig_timer_active = true; 116 mod_timer(&nn->reconfig_timer, jiffies + NFP_NET_POLL_TIMEOUT * HZ); 117 } 118 119 static bool nfp_net_reconfig_check_done(struct nfp_net *nn, bool last_check) 120 { 121 u32 reg; 122 123 reg = nn_readl(nn, NFP_NET_CFG_UPDATE); 124 if (reg == 0) 125 return true; 126 if (reg & NFP_NET_CFG_UPDATE_ERR) { 127 nn_err(nn, "Reconfig error (status: 0x%08x update: 0x%08x ctrl: 0x%08x)\n", 128 reg, nn->reconfig_in_progress_update, 129 nn_readl(nn, NFP_NET_CFG_CTRL)); 130 return true; 131 } else if (last_check) { 132 nn_err(nn, "Reconfig timeout (status: 0x%08x update: 0x%08x ctrl: 0x%08x)\n", 133 reg, nn->reconfig_in_progress_update, 134 nn_readl(nn, NFP_NET_CFG_CTRL)); 135 return true; 136 } 137 138 return false; 139 } 140 141 static int nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline) 142 { 143 bool timed_out = false; 144 145 /* Poll update field, waiting for NFP to ack the config */ 146 while (!nfp_net_reconfig_check_done(nn, timed_out)) { 147 msleep(1); 148 timed_out = time_is_before_eq_jiffies(deadline); 149 } 150 151 if (nn_readl(nn, NFP_NET_CFG_UPDATE) & NFP_NET_CFG_UPDATE_ERR) 152 return -EIO; 153 154 return timed_out ? -EIO : 0; 155 } 156 157 static void nfp_net_reconfig_timer(struct timer_list *t) 158 { 159 struct nfp_net *nn = from_timer(nn, t, reconfig_timer); 160 161 spin_lock_bh(&nn->reconfig_lock); 162 163 nn->reconfig_timer_active = false; 164 165 /* If sync caller is present it will take over from us */ 166 if (nn->reconfig_sync_present) 167 goto done; 168 169 /* Read reconfig status and report errors */ 170 nfp_net_reconfig_check_done(nn, true); 171 172 if (nn->reconfig_posted) 173 nfp_net_reconfig_start_async(nn, 0); 174 done: 175 spin_unlock_bh(&nn->reconfig_lock); 176 } 177 178 /** 179 * nfp_net_reconfig_post() - Post async reconfig request 180 * @nn: NFP Net device to reconfigure 181 * @update: The value for the update field in the BAR config 182 * 183 * Record FW reconfiguration request. Reconfiguration will be kicked off 184 * whenever reconfiguration machinery is idle. Multiple requests can be 185 * merged together! 186 */ 187 static void nfp_net_reconfig_post(struct nfp_net *nn, u32 update) 188 { 189 spin_lock_bh(&nn->reconfig_lock); 190 191 /* Sync caller will kick off async reconf when it's done, just post */ 192 if (nn->reconfig_sync_present) { 193 nn->reconfig_posted |= update; 194 goto done; 195 } 196 197 /* Opportunistically check if the previous command is done */ 198 if (!nn->reconfig_timer_active || 199 nfp_net_reconfig_check_done(nn, false)) 200 nfp_net_reconfig_start_async(nn, update); 201 else 202 nn->reconfig_posted |= update; 203 done: 204 spin_unlock_bh(&nn->reconfig_lock); 205 } 206 207 static void nfp_net_reconfig_sync_enter(struct nfp_net *nn) 208 { 209 bool cancelled_timer = false; 210 u32 pre_posted_requests; 211 212 spin_lock_bh(&nn->reconfig_lock); 213 214 nn->reconfig_sync_present = true; 215 216 if (nn->reconfig_timer_active) { 217 nn->reconfig_timer_active = false; 218 cancelled_timer = true; 219 } 220 pre_posted_requests = nn->reconfig_posted; 221 nn->reconfig_posted = 0; 222 223 spin_unlock_bh(&nn->reconfig_lock); 224 225 if (cancelled_timer) { 226 del_timer_sync(&nn->reconfig_timer); 227 nfp_net_reconfig_wait(nn, nn->reconfig_timer.expires); 228 } 229 230 /* Run the posted reconfigs which were issued before we started */ 231 if (pre_posted_requests) { 232 nfp_net_reconfig_start(nn, pre_posted_requests); 233 nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT); 234 } 235 } 236 237 static void nfp_net_reconfig_wait_posted(struct nfp_net *nn) 238 { 239 nfp_net_reconfig_sync_enter(nn); 240 241 spin_lock_bh(&nn->reconfig_lock); 242 nn->reconfig_sync_present = false; 243 spin_unlock_bh(&nn->reconfig_lock); 244 } 245 246 /** 247 * nfp_net_reconfig() - Reconfigure the firmware 248 * @nn: NFP Net device to reconfigure 249 * @update: The value for the update field in the BAR config 250 * 251 * Write the update word to the BAR and ping the reconfig queue. The 252 * poll until the firmware has acknowledged the update by zeroing the 253 * update word. 254 * 255 * Return: Negative errno on error, 0 on success 256 */ 257 int nfp_net_reconfig(struct nfp_net *nn, u32 update) 258 { 259 int ret; 260 261 nfp_net_reconfig_sync_enter(nn); 262 263 nfp_net_reconfig_start(nn, update); 264 ret = nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT); 265 266 spin_lock_bh(&nn->reconfig_lock); 267 268 if (nn->reconfig_posted) 269 nfp_net_reconfig_start_async(nn, 0); 270 271 nn->reconfig_sync_present = false; 272 273 spin_unlock_bh(&nn->reconfig_lock); 274 275 return ret; 276 } 277 278 /** 279 * nfp_net_reconfig_mbox() - Reconfigure the firmware via the mailbox 280 * @nn: NFP Net device to reconfigure 281 * @mbox_cmd: The value for the mailbox command 282 * 283 * Helper function for mailbox updates 284 * 285 * Return: Negative errno on error, 0 on success 286 */ 287 int nfp_net_reconfig_mbox(struct nfp_net *nn, u32 mbox_cmd) 288 { 289 u32 mbox = nn->tlv_caps.mbox_off; 290 int ret; 291 292 if (!nfp_net_has_mbox(&nn->tlv_caps)) { 293 nn_err(nn, "no mailbox present, command: %u\n", mbox_cmd); 294 return -EIO; 295 } 296 297 nn_writeq(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_CMD, mbox_cmd); 298 299 ret = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MBOX); 300 if (ret) { 301 nn_err(nn, "Mailbox update error\n"); 302 return ret; 303 } 304 305 return -nn_readl(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_RET); 306 } 307 308 /* Interrupt configuration and handling 309 */ 310 311 /** 312 * nfp_net_irq_unmask() - Unmask automasked interrupt 313 * @nn: NFP Network structure 314 * @entry_nr: MSI-X table entry 315 * 316 * Clear the ICR for the IRQ entry. 317 */ 318 static void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr) 319 { 320 nn_writeb(nn, NFP_NET_CFG_ICR(entry_nr), NFP_NET_CFG_ICR_UNMASKED); 321 nn_pci_flush(nn); 322 } 323 324 /** 325 * nfp_net_irqs_alloc() - allocates MSI-X irqs 326 * @pdev: PCI device structure 327 * @irq_entries: Array to be initialized and used to hold the irq entries 328 * @min_irqs: Minimal acceptable number of interrupts 329 * @wanted_irqs: Target number of interrupts to allocate 330 * 331 * Return: Number of irqs obtained or 0 on error. 332 */ 333 unsigned int 334 nfp_net_irqs_alloc(struct pci_dev *pdev, struct msix_entry *irq_entries, 335 unsigned int min_irqs, unsigned int wanted_irqs) 336 { 337 unsigned int i; 338 int got_irqs; 339 340 for (i = 0; i < wanted_irqs; i++) 341 irq_entries[i].entry = i; 342 343 got_irqs = pci_enable_msix_range(pdev, irq_entries, 344 min_irqs, wanted_irqs); 345 if (got_irqs < 0) { 346 dev_err(&pdev->dev, "Failed to enable %d-%d MSI-X (err=%d)\n", 347 min_irqs, wanted_irqs, got_irqs); 348 return 0; 349 } 350 351 if (got_irqs < wanted_irqs) 352 dev_warn(&pdev->dev, "Unable to allocate %d IRQs got only %d\n", 353 wanted_irqs, got_irqs); 354 355 return got_irqs; 356 } 357 358 /** 359 * nfp_net_irqs_assign() - Assign interrupts allocated externally to netdev 360 * @nn: NFP Network structure 361 * @irq_entries: Table of allocated interrupts 362 * @n: Size of @irq_entries (number of entries to grab) 363 * 364 * After interrupts are allocated with nfp_net_irqs_alloc() this function 365 * should be called to assign them to a specific netdev (port). 366 */ 367 void 368 nfp_net_irqs_assign(struct nfp_net *nn, struct msix_entry *irq_entries, 369 unsigned int n) 370 { 371 struct nfp_net_dp *dp = &nn->dp; 372 373 nn->max_r_vecs = n - NFP_NET_NON_Q_VECTORS; 374 dp->num_r_vecs = nn->max_r_vecs; 375 376 memcpy(nn->irq_entries, irq_entries, sizeof(*irq_entries) * n); 377 378 if (dp->num_rx_rings > dp->num_r_vecs || 379 dp->num_tx_rings > dp->num_r_vecs) 380 dev_warn(nn->dp.dev, "More rings (%d,%d) than vectors (%d).\n", 381 dp->num_rx_rings, dp->num_tx_rings, 382 dp->num_r_vecs); 383 384 dp->num_rx_rings = min(dp->num_r_vecs, dp->num_rx_rings); 385 dp->num_tx_rings = min(dp->num_r_vecs, dp->num_tx_rings); 386 dp->num_stack_tx_rings = dp->num_tx_rings; 387 } 388 389 /** 390 * nfp_net_irqs_disable() - Disable interrupts 391 * @pdev: PCI device structure 392 * 393 * Undoes what @nfp_net_irqs_alloc() does. 394 */ 395 void nfp_net_irqs_disable(struct pci_dev *pdev) 396 { 397 pci_disable_msix(pdev); 398 } 399 400 /** 401 * nfp_net_irq_rxtx() - Interrupt service routine for RX/TX rings. 402 * @irq: Interrupt 403 * @data: Opaque data structure 404 * 405 * Return: Indicate if the interrupt has been handled. 406 */ 407 static irqreturn_t nfp_net_irq_rxtx(int irq, void *data) 408 { 409 struct nfp_net_r_vector *r_vec = data; 410 411 napi_schedule_irqoff(&r_vec->napi); 412 413 /* The FW auto-masks any interrupt, either via the MASK bit in 414 * the MSI-X table or via the per entry ICR field. So there 415 * is no need to disable interrupts here. 416 */ 417 return IRQ_HANDLED; 418 } 419 420 static irqreturn_t nfp_ctrl_irq_rxtx(int irq, void *data) 421 { 422 struct nfp_net_r_vector *r_vec = data; 423 424 tasklet_schedule(&r_vec->tasklet); 425 426 return IRQ_HANDLED; 427 } 428 429 /** 430 * nfp_net_read_link_status() - Reread link status from control BAR 431 * @nn: NFP Network structure 432 */ 433 static void nfp_net_read_link_status(struct nfp_net *nn) 434 { 435 unsigned long flags; 436 bool link_up; 437 u32 sts; 438 439 spin_lock_irqsave(&nn->link_status_lock, flags); 440 441 sts = nn_readl(nn, NFP_NET_CFG_STS); 442 link_up = !!(sts & NFP_NET_CFG_STS_LINK); 443 444 if (nn->link_up == link_up) 445 goto out; 446 447 nn->link_up = link_up; 448 if (nn->port) 449 set_bit(NFP_PORT_CHANGED, &nn->port->flags); 450 451 if (nn->link_up) { 452 netif_carrier_on(nn->dp.netdev); 453 netdev_info(nn->dp.netdev, "NIC Link is Up\n"); 454 } else { 455 netif_carrier_off(nn->dp.netdev); 456 netdev_info(nn->dp.netdev, "NIC Link is Down\n"); 457 } 458 out: 459 spin_unlock_irqrestore(&nn->link_status_lock, flags); 460 } 461 462 /** 463 * nfp_net_irq_lsc() - Interrupt service routine for link state changes 464 * @irq: Interrupt 465 * @data: Opaque data structure 466 * 467 * Return: Indicate if the interrupt has been handled. 468 */ 469 static irqreturn_t nfp_net_irq_lsc(int irq, void *data) 470 { 471 struct nfp_net *nn = data; 472 struct msix_entry *entry; 473 474 entry = &nn->irq_entries[NFP_NET_IRQ_LSC_IDX]; 475 476 nfp_net_read_link_status(nn); 477 478 nfp_net_irq_unmask(nn, entry->entry); 479 480 return IRQ_HANDLED; 481 } 482 483 /** 484 * nfp_net_irq_exn() - Interrupt service routine for exceptions 485 * @irq: Interrupt 486 * @data: Opaque data structure 487 * 488 * Return: Indicate if the interrupt has been handled. 489 */ 490 static irqreturn_t nfp_net_irq_exn(int irq, void *data) 491 { 492 struct nfp_net *nn = data; 493 494 nn_err(nn, "%s: UNIMPLEMENTED.\n", __func__); 495 /* XXX TO BE IMPLEMENTED */ 496 return IRQ_HANDLED; 497 } 498 499 /** 500 * nfp_net_tx_ring_init() - Fill in the boilerplate for a TX ring 501 * @tx_ring: TX ring structure 502 * @r_vec: IRQ vector servicing this ring 503 * @idx: Ring index 504 * @is_xdp: Is this an XDP TX ring? 505 */ 506 static void 507 nfp_net_tx_ring_init(struct nfp_net_tx_ring *tx_ring, 508 struct nfp_net_r_vector *r_vec, unsigned int idx, 509 bool is_xdp) 510 { 511 struct nfp_net *nn = r_vec->nfp_net; 512 513 tx_ring->idx = idx; 514 tx_ring->r_vec = r_vec; 515 tx_ring->is_xdp = is_xdp; 516 u64_stats_init(&tx_ring->r_vec->tx_sync); 517 518 tx_ring->qcidx = tx_ring->idx * nn->stride_tx; 519 tx_ring->qcp_q = nn->tx_bar + NFP_QCP_QUEUE_OFF(tx_ring->qcidx); 520 } 521 522 /** 523 * nfp_net_rx_ring_init() - Fill in the boilerplate for a RX ring 524 * @rx_ring: RX ring structure 525 * @r_vec: IRQ vector servicing this ring 526 * @idx: Ring index 527 */ 528 static void 529 nfp_net_rx_ring_init(struct nfp_net_rx_ring *rx_ring, 530 struct nfp_net_r_vector *r_vec, unsigned int idx) 531 { 532 struct nfp_net *nn = r_vec->nfp_net; 533 534 rx_ring->idx = idx; 535 rx_ring->r_vec = r_vec; 536 u64_stats_init(&rx_ring->r_vec->rx_sync); 537 538 rx_ring->fl_qcidx = rx_ring->idx * nn->stride_rx; 539 rx_ring->qcp_fl = nn->rx_bar + NFP_QCP_QUEUE_OFF(rx_ring->fl_qcidx); 540 } 541 542 /** 543 * nfp_net_aux_irq_request() - Request an auxiliary interrupt (LSC or EXN) 544 * @nn: NFP Network structure 545 * @ctrl_offset: Control BAR offset where IRQ configuration should be written 546 * @format: printf-style format to construct the interrupt name 547 * @name: Pointer to allocated space for interrupt name 548 * @name_sz: Size of space for interrupt name 549 * @vector_idx: Index of MSI-X vector used for this interrupt 550 * @handler: IRQ handler to register for this interrupt 551 */ 552 static int 553 nfp_net_aux_irq_request(struct nfp_net *nn, u32 ctrl_offset, 554 const char *format, char *name, size_t name_sz, 555 unsigned int vector_idx, irq_handler_t handler) 556 { 557 struct msix_entry *entry; 558 int err; 559 560 entry = &nn->irq_entries[vector_idx]; 561 562 snprintf(name, name_sz, format, nfp_net_name(nn)); 563 err = request_irq(entry->vector, handler, 0, name, nn); 564 if (err) { 565 nn_err(nn, "Failed to request IRQ %d (err=%d).\n", 566 entry->vector, err); 567 return err; 568 } 569 nn_writeb(nn, ctrl_offset, entry->entry); 570 nfp_net_irq_unmask(nn, entry->entry); 571 572 return 0; 573 } 574 575 /** 576 * nfp_net_aux_irq_free() - Free an auxiliary interrupt (LSC or EXN) 577 * @nn: NFP Network structure 578 * @ctrl_offset: Control BAR offset where IRQ configuration should be written 579 * @vector_idx: Index of MSI-X vector used for this interrupt 580 */ 581 static void nfp_net_aux_irq_free(struct nfp_net *nn, u32 ctrl_offset, 582 unsigned int vector_idx) 583 { 584 nn_writeb(nn, ctrl_offset, 0xff); 585 nn_pci_flush(nn); 586 free_irq(nn->irq_entries[vector_idx].vector, nn); 587 } 588 589 /* Transmit 590 * 591 * One queue controller peripheral queue is used for transmit. The 592 * driver en-queues packets for transmit by advancing the write 593 * pointer. The device indicates that packets have transmitted by 594 * advancing the read pointer. The driver maintains a local copy of 595 * the read and write pointer in @struct nfp_net_tx_ring. The driver 596 * keeps @wr_p in sync with the queue controller write pointer and can 597 * determine how many packets have been transmitted by comparing its 598 * copy of the read pointer @rd_p with the read pointer maintained by 599 * the queue controller peripheral. 600 */ 601 602 /** 603 * nfp_net_tx_full() - Check if the TX ring is full 604 * @tx_ring: TX ring to check 605 * @dcnt: Number of descriptors that need to be enqueued (must be >= 1) 606 * 607 * This function checks, based on the *host copy* of read/write 608 * pointer if a given TX ring is full. The real TX queue may have 609 * some newly made available slots. 610 * 611 * Return: True if the ring is full. 612 */ 613 static int nfp_net_tx_full(struct nfp_net_tx_ring *tx_ring, int dcnt) 614 { 615 return (tx_ring->wr_p - tx_ring->rd_p) >= (tx_ring->cnt - dcnt); 616 } 617 618 /* Wrappers for deciding when to stop and restart TX queues */ 619 static int nfp_net_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring) 620 { 621 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4); 622 } 623 624 static int nfp_net_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring) 625 { 626 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1); 627 } 628 629 /** 630 * nfp_net_tx_ring_stop() - stop tx ring 631 * @nd_q: netdev queue 632 * @tx_ring: driver tx queue structure 633 * 634 * Safely stop TX ring. Remember that while we are running .start_xmit() 635 * someone else may be cleaning the TX ring completions so we need to be 636 * extra careful here. 637 */ 638 static void nfp_net_tx_ring_stop(struct netdev_queue *nd_q, 639 struct nfp_net_tx_ring *tx_ring) 640 { 641 netif_tx_stop_queue(nd_q); 642 643 /* We can race with the TX completion out of NAPI so recheck */ 644 smp_mb(); 645 if (unlikely(nfp_net_tx_ring_should_wake(tx_ring))) 646 netif_tx_start_queue(nd_q); 647 } 648 649 /** 650 * nfp_net_tx_tso() - Set up Tx descriptor for LSO 651 * @r_vec: per-ring structure 652 * @txbuf: Pointer to driver soft TX descriptor 653 * @txd: Pointer to HW TX descriptor 654 * @skb: Pointer to SKB 655 * @md_bytes: Prepend length 656 * 657 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs. 658 * Return error on packet header greater than maximum supported LSO header size. 659 */ 660 static void nfp_net_tx_tso(struct nfp_net_r_vector *r_vec, 661 struct nfp_net_tx_buf *txbuf, 662 struct nfp_net_tx_desc *txd, struct sk_buff *skb, 663 u32 md_bytes) 664 { 665 u32 l3_offset, l4_offset, hdrlen; 666 u16 mss; 667 668 if (!skb_is_gso(skb)) 669 return; 670 671 if (!skb->encapsulation) { 672 l3_offset = skb_network_offset(skb); 673 l4_offset = skb_transport_offset(skb); 674 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb); 675 } else { 676 l3_offset = skb_inner_network_offset(skb); 677 l4_offset = skb_inner_transport_offset(skb); 678 hdrlen = skb_inner_transport_header(skb) - skb->data + 679 inner_tcp_hdrlen(skb); 680 } 681 682 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs; 683 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1); 684 685 mss = skb_shinfo(skb)->gso_size & PCIE_DESC_TX_MSS_MASK; 686 txd->l3_offset = l3_offset - md_bytes; 687 txd->l4_offset = l4_offset - md_bytes; 688 txd->lso_hdrlen = hdrlen - md_bytes; 689 txd->mss = cpu_to_le16(mss); 690 txd->flags |= PCIE_DESC_TX_LSO; 691 692 u64_stats_update_begin(&r_vec->tx_sync); 693 r_vec->tx_lso++; 694 u64_stats_update_end(&r_vec->tx_sync); 695 } 696 697 /** 698 * nfp_net_tx_csum() - Set TX CSUM offload flags in TX descriptor 699 * @dp: NFP Net data path struct 700 * @r_vec: per-ring structure 701 * @txbuf: Pointer to driver soft TX descriptor 702 * @txd: Pointer to TX descriptor 703 * @skb: Pointer to SKB 704 * 705 * This function sets the TX checksum flags in the TX descriptor based 706 * on the configuration and the protocol of the packet to be transmitted. 707 */ 708 static void nfp_net_tx_csum(struct nfp_net_dp *dp, 709 struct nfp_net_r_vector *r_vec, 710 struct nfp_net_tx_buf *txbuf, 711 struct nfp_net_tx_desc *txd, struct sk_buff *skb) 712 { 713 struct ipv6hdr *ipv6h; 714 struct iphdr *iph; 715 u8 l4_hdr; 716 717 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM)) 718 return; 719 720 if (skb->ip_summed != CHECKSUM_PARTIAL) 721 return; 722 723 txd->flags |= PCIE_DESC_TX_CSUM; 724 if (skb->encapsulation) 725 txd->flags |= PCIE_DESC_TX_ENCAP; 726 727 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb); 728 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb); 729 730 if (iph->version == 4) { 731 txd->flags |= PCIE_DESC_TX_IP4_CSUM; 732 l4_hdr = iph->protocol; 733 } else if (ipv6h->version == 6) { 734 l4_hdr = ipv6h->nexthdr; 735 } else { 736 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version); 737 return; 738 } 739 740 switch (l4_hdr) { 741 case IPPROTO_TCP: 742 txd->flags |= PCIE_DESC_TX_TCP_CSUM; 743 break; 744 case IPPROTO_UDP: 745 txd->flags |= PCIE_DESC_TX_UDP_CSUM; 746 break; 747 default: 748 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr); 749 return; 750 } 751 752 u64_stats_update_begin(&r_vec->tx_sync); 753 if (skb->encapsulation) 754 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt; 755 else 756 r_vec->hw_csum_tx += txbuf->pkt_cnt; 757 u64_stats_update_end(&r_vec->tx_sync); 758 } 759 760 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring) 761 { 762 wmb(); 763 nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add); 764 tx_ring->wr_ptr_add = 0; 765 } 766 767 static int nfp_net_prep_port_id(struct sk_buff *skb) 768 { 769 struct metadata_dst *md_dst = skb_metadata_dst(skb); 770 unsigned char *data; 771 772 if (likely(!md_dst)) 773 return 0; 774 if (unlikely(md_dst->type != METADATA_HW_PORT_MUX)) 775 return 0; 776 777 if (unlikely(skb_cow_head(skb, 8))) 778 return -ENOMEM; 779 780 data = skb_push(skb, 8); 781 put_unaligned_be32(NFP_NET_META_PORTID, data); 782 put_unaligned_be32(md_dst->u.port_info.port_id, data + 4); 783 784 return 8; 785 } 786 787 /** 788 * nfp_net_tx() - Main transmit entry point 789 * @skb: SKB to transmit 790 * @netdev: netdev structure 791 * 792 * Return: NETDEV_TX_OK on success. 793 */ 794 static int nfp_net_tx(struct sk_buff *skb, struct net_device *netdev) 795 { 796 struct nfp_net *nn = netdev_priv(netdev); 797 const struct skb_frag_struct *frag; 798 int f, nr_frags, wr_idx, md_bytes; 799 struct nfp_net_tx_ring *tx_ring; 800 struct nfp_net_r_vector *r_vec; 801 struct nfp_net_tx_buf *txbuf; 802 struct nfp_net_tx_desc *txd; 803 struct netdev_queue *nd_q; 804 struct nfp_net_dp *dp; 805 dma_addr_t dma_addr; 806 unsigned int fsize; 807 u16 qidx; 808 809 dp = &nn->dp; 810 qidx = skb_get_queue_mapping(skb); 811 tx_ring = &dp->tx_rings[qidx]; 812 r_vec = tx_ring->r_vec; 813 814 nr_frags = skb_shinfo(skb)->nr_frags; 815 816 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) { 817 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n", 818 qidx, tx_ring->wr_p, tx_ring->rd_p); 819 nd_q = netdev_get_tx_queue(dp->netdev, qidx); 820 netif_tx_stop_queue(nd_q); 821 nfp_net_tx_xmit_more_flush(tx_ring); 822 u64_stats_update_begin(&r_vec->tx_sync); 823 r_vec->tx_busy++; 824 u64_stats_update_end(&r_vec->tx_sync); 825 return NETDEV_TX_BUSY; 826 } 827 828 md_bytes = nfp_net_prep_port_id(skb); 829 if (unlikely(md_bytes < 0)) { 830 nfp_net_tx_xmit_more_flush(tx_ring); 831 dev_kfree_skb_any(skb); 832 return NETDEV_TX_OK; 833 } 834 835 /* Start with the head skbuf */ 836 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 837 DMA_TO_DEVICE); 838 if (dma_mapping_error(dp->dev, dma_addr)) 839 goto err_free; 840 841 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 842 843 /* Stash the soft descriptor of the head then initialize it */ 844 txbuf = &tx_ring->txbufs[wr_idx]; 845 txbuf->skb = skb; 846 txbuf->dma_addr = dma_addr; 847 txbuf->fidx = -1; 848 txbuf->pkt_cnt = 1; 849 txbuf->real_len = skb->len; 850 851 /* Build TX descriptor */ 852 txd = &tx_ring->txds[wr_idx]; 853 txd->offset_eop = (nr_frags ? 0 : PCIE_DESC_TX_EOP) | md_bytes; 854 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 855 nfp_desc_set_dma_addr(txd, dma_addr); 856 txd->data_len = cpu_to_le16(skb->len); 857 858 txd->flags = 0; 859 txd->mss = 0; 860 txd->lso_hdrlen = 0; 861 862 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */ 863 nfp_net_tx_tso(r_vec, txbuf, txd, skb, md_bytes); 864 nfp_net_tx_csum(dp, r_vec, txbuf, txd, skb); 865 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) { 866 txd->flags |= PCIE_DESC_TX_VLAN; 867 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb)); 868 } 869 870 /* Gather DMA */ 871 if (nr_frags > 0) { 872 __le64 second_half; 873 874 /* all descs must match except for in addr, length and eop */ 875 second_half = txd->vals8[1]; 876 877 for (f = 0; f < nr_frags; f++) { 878 frag = &skb_shinfo(skb)->frags[f]; 879 fsize = skb_frag_size(frag); 880 881 dma_addr = skb_frag_dma_map(dp->dev, frag, 0, 882 fsize, DMA_TO_DEVICE); 883 if (dma_mapping_error(dp->dev, dma_addr)) 884 goto err_unmap; 885 886 wr_idx = D_IDX(tx_ring, wr_idx + 1); 887 tx_ring->txbufs[wr_idx].skb = skb; 888 tx_ring->txbufs[wr_idx].dma_addr = dma_addr; 889 tx_ring->txbufs[wr_idx].fidx = f; 890 891 txd = &tx_ring->txds[wr_idx]; 892 txd->dma_len = cpu_to_le16(fsize); 893 nfp_desc_set_dma_addr(txd, dma_addr); 894 txd->offset_eop = md_bytes | 895 ((f == nr_frags - 1) ? PCIE_DESC_TX_EOP : 0); 896 txd->vals8[1] = second_half; 897 } 898 899 u64_stats_update_begin(&r_vec->tx_sync); 900 r_vec->tx_gather++; 901 u64_stats_update_end(&r_vec->tx_sync); 902 } 903 904 skb_tx_timestamp(skb); 905 906 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 907 908 tx_ring->wr_p += nr_frags + 1; 909 if (nfp_net_tx_ring_should_stop(tx_ring)) 910 nfp_net_tx_ring_stop(nd_q, tx_ring); 911 912 tx_ring->wr_ptr_add += nr_frags + 1; 913 if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, skb->xmit_more)) 914 nfp_net_tx_xmit_more_flush(tx_ring); 915 916 return NETDEV_TX_OK; 917 918 err_unmap: 919 while (--f >= 0) { 920 frag = &skb_shinfo(skb)->frags[f]; 921 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 922 skb_frag_size(frag), DMA_TO_DEVICE); 923 tx_ring->txbufs[wr_idx].skb = NULL; 924 tx_ring->txbufs[wr_idx].dma_addr = 0; 925 tx_ring->txbufs[wr_idx].fidx = -2; 926 wr_idx = wr_idx - 1; 927 if (wr_idx < 0) 928 wr_idx += tx_ring->cnt; 929 } 930 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr, 931 skb_headlen(skb), DMA_TO_DEVICE); 932 tx_ring->txbufs[wr_idx].skb = NULL; 933 tx_ring->txbufs[wr_idx].dma_addr = 0; 934 tx_ring->txbufs[wr_idx].fidx = -2; 935 err_free: 936 nn_dp_warn(dp, "Failed to map DMA TX buffer\n"); 937 nfp_net_tx_xmit_more_flush(tx_ring); 938 u64_stats_update_begin(&r_vec->tx_sync); 939 r_vec->tx_errors++; 940 u64_stats_update_end(&r_vec->tx_sync); 941 dev_kfree_skb_any(skb); 942 return NETDEV_TX_OK; 943 } 944 945 /** 946 * nfp_net_tx_complete() - Handled completed TX packets 947 * @tx_ring: TX ring structure 948 * @budget: NAPI budget (only used as bool to determine if in NAPI context) 949 */ 950 static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget) 951 { 952 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 953 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 954 struct netdev_queue *nd_q; 955 u32 done_pkts = 0, done_bytes = 0; 956 u32 qcp_rd_p; 957 int todo; 958 959 if (tx_ring->wr_p == tx_ring->rd_p) 960 return; 961 962 /* Work out how many descriptors have been transmitted */ 963 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q); 964 965 if (qcp_rd_p == tx_ring->qcp_rd_p) 966 return; 967 968 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 969 970 while (todo--) { 971 const struct skb_frag_struct *frag; 972 struct nfp_net_tx_buf *tx_buf; 973 struct sk_buff *skb; 974 int fidx, nr_frags; 975 int idx; 976 977 idx = D_IDX(tx_ring, tx_ring->rd_p++); 978 tx_buf = &tx_ring->txbufs[idx]; 979 980 skb = tx_buf->skb; 981 if (!skb) 982 continue; 983 984 nr_frags = skb_shinfo(skb)->nr_frags; 985 fidx = tx_buf->fidx; 986 987 if (fidx == -1) { 988 /* unmap head */ 989 dma_unmap_single(dp->dev, tx_buf->dma_addr, 990 skb_headlen(skb), DMA_TO_DEVICE); 991 992 done_pkts += tx_buf->pkt_cnt; 993 done_bytes += tx_buf->real_len; 994 } else { 995 /* unmap fragment */ 996 frag = &skb_shinfo(skb)->frags[fidx]; 997 dma_unmap_page(dp->dev, tx_buf->dma_addr, 998 skb_frag_size(frag), DMA_TO_DEVICE); 999 } 1000 1001 /* check for last gather fragment */ 1002 if (fidx == nr_frags - 1) 1003 napi_consume_skb(skb, budget); 1004 1005 tx_buf->dma_addr = 0; 1006 tx_buf->skb = NULL; 1007 tx_buf->fidx = -2; 1008 } 1009 1010 tx_ring->qcp_rd_p = qcp_rd_p; 1011 1012 u64_stats_update_begin(&r_vec->tx_sync); 1013 r_vec->tx_bytes += done_bytes; 1014 r_vec->tx_pkts += done_pkts; 1015 u64_stats_update_end(&r_vec->tx_sync); 1016 1017 if (!dp->netdev) 1018 return; 1019 1020 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 1021 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes); 1022 if (nfp_net_tx_ring_should_wake(tx_ring)) { 1023 /* Make sure TX thread will see updated tx_ring->rd_p */ 1024 smp_mb(); 1025 1026 if (unlikely(netif_tx_queue_stopped(nd_q))) 1027 netif_tx_wake_queue(nd_q); 1028 } 1029 1030 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 1031 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 1032 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 1033 } 1034 1035 static bool nfp_net_xdp_complete(struct nfp_net_tx_ring *tx_ring) 1036 { 1037 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 1038 u32 done_pkts = 0, done_bytes = 0; 1039 bool done_all; 1040 int idx, todo; 1041 u32 qcp_rd_p; 1042 1043 /* Work out how many descriptors have been transmitted */ 1044 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q); 1045 1046 if (qcp_rd_p == tx_ring->qcp_rd_p) 1047 return true; 1048 1049 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p); 1050 1051 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE; 1052 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE); 1053 1054 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo); 1055 1056 done_pkts = todo; 1057 while (todo--) { 1058 idx = D_IDX(tx_ring, tx_ring->rd_p); 1059 tx_ring->rd_p++; 1060 1061 done_bytes += tx_ring->txbufs[idx].real_len; 1062 } 1063 1064 u64_stats_update_begin(&r_vec->tx_sync); 1065 r_vec->tx_bytes += done_bytes; 1066 r_vec->tx_pkts += done_pkts; 1067 u64_stats_update_end(&r_vec->tx_sync); 1068 1069 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt, 1070 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n", 1071 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt); 1072 1073 return done_all; 1074 } 1075 1076 /** 1077 * nfp_net_tx_ring_reset() - Free any untransmitted buffers and reset pointers 1078 * @dp: NFP Net data path struct 1079 * @tx_ring: TX ring structure 1080 * 1081 * Assumes that the device is stopped, must be idempotent. 1082 */ 1083 static void 1084 nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring) 1085 { 1086 const struct skb_frag_struct *frag; 1087 struct netdev_queue *nd_q; 1088 1089 while (!tx_ring->is_xdp && tx_ring->rd_p != tx_ring->wr_p) { 1090 struct nfp_net_tx_buf *tx_buf; 1091 struct sk_buff *skb; 1092 int idx, nr_frags; 1093 1094 idx = D_IDX(tx_ring, tx_ring->rd_p); 1095 tx_buf = &tx_ring->txbufs[idx]; 1096 1097 skb = tx_ring->txbufs[idx].skb; 1098 nr_frags = skb_shinfo(skb)->nr_frags; 1099 1100 if (tx_buf->fidx == -1) { 1101 /* unmap head */ 1102 dma_unmap_single(dp->dev, tx_buf->dma_addr, 1103 skb_headlen(skb), DMA_TO_DEVICE); 1104 } else { 1105 /* unmap fragment */ 1106 frag = &skb_shinfo(skb)->frags[tx_buf->fidx]; 1107 dma_unmap_page(dp->dev, tx_buf->dma_addr, 1108 skb_frag_size(frag), DMA_TO_DEVICE); 1109 } 1110 1111 /* check for last gather fragment */ 1112 if (tx_buf->fidx == nr_frags - 1) 1113 dev_kfree_skb_any(skb); 1114 1115 tx_buf->dma_addr = 0; 1116 tx_buf->skb = NULL; 1117 tx_buf->fidx = -2; 1118 1119 tx_ring->qcp_rd_p++; 1120 tx_ring->rd_p++; 1121 } 1122 1123 memset(tx_ring->txds, 0, tx_ring->size); 1124 tx_ring->wr_p = 0; 1125 tx_ring->rd_p = 0; 1126 tx_ring->qcp_rd_p = 0; 1127 tx_ring->wr_ptr_add = 0; 1128 1129 if (tx_ring->is_xdp || !dp->netdev) 1130 return; 1131 1132 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx); 1133 netdev_tx_reset_queue(nd_q); 1134 } 1135 1136 static void nfp_net_tx_timeout(struct net_device *netdev) 1137 { 1138 struct nfp_net *nn = netdev_priv(netdev); 1139 int i; 1140 1141 for (i = 0; i < nn->dp.netdev->real_num_tx_queues; i++) { 1142 if (!netif_tx_queue_stopped(netdev_get_tx_queue(netdev, i))) 1143 continue; 1144 nn_warn(nn, "TX timeout on ring: %d\n", i); 1145 } 1146 nn_warn(nn, "TX watchdog timeout\n"); 1147 } 1148 1149 /* Receive processing 1150 */ 1151 static unsigned int 1152 nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp) 1153 { 1154 unsigned int fl_bufsz; 1155 1156 fl_bufsz = NFP_NET_RX_BUF_HEADROOM; 1157 fl_bufsz += dp->rx_dma_off; 1158 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1159 fl_bufsz += NFP_NET_MAX_PREPEND; 1160 else 1161 fl_bufsz += dp->rx_offset; 1162 fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu; 1163 1164 fl_bufsz = SKB_DATA_ALIGN(fl_bufsz); 1165 fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1166 1167 return fl_bufsz; 1168 } 1169 1170 static void 1171 nfp_net_free_frag(void *frag, bool xdp) 1172 { 1173 if (!xdp) 1174 skb_free_frag(frag); 1175 else 1176 __free_page(virt_to_page(frag)); 1177 } 1178 1179 /** 1180 * nfp_net_rx_alloc_one() - Allocate and map page frag for RX 1181 * @dp: NFP Net data path struct 1182 * @dma_addr: Pointer to storage for DMA address (output param) 1183 * 1184 * This function will allcate a new page frag, map it for DMA. 1185 * 1186 * Return: allocated page frag or NULL on failure. 1187 */ 1188 static void *nfp_net_rx_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 1189 { 1190 void *frag; 1191 1192 if (!dp->xdp_prog) { 1193 frag = netdev_alloc_frag(dp->fl_bufsz); 1194 } else { 1195 struct page *page; 1196 1197 page = alloc_page(GFP_KERNEL); 1198 frag = page ? page_address(page) : NULL; 1199 } 1200 if (!frag) { 1201 nn_dp_warn(dp, "Failed to alloc receive page frag\n"); 1202 return NULL; 1203 } 1204 1205 *dma_addr = nfp_net_dma_map_rx(dp, frag); 1206 if (dma_mapping_error(dp->dev, *dma_addr)) { 1207 nfp_net_free_frag(frag, dp->xdp_prog); 1208 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 1209 return NULL; 1210 } 1211 1212 return frag; 1213 } 1214 1215 static void *nfp_net_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr) 1216 { 1217 void *frag; 1218 1219 if (!dp->xdp_prog) { 1220 frag = napi_alloc_frag(dp->fl_bufsz); 1221 if (unlikely(!frag)) 1222 return NULL; 1223 } else { 1224 struct page *page; 1225 1226 page = dev_alloc_page(); 1227 if (unlikely(!page)) 1228 return NULL; 1229 frag = page_address(page); 1230 } 1231 1232 *dma_addr = nfp_net_dma_map_rx(dp, frag); 1233 if (dma_mapping_error(dp->dev, *dma_addr)) { 1234 nfp_net_free_frag(frag, dp->xdp_prog); 1235 nn_dp_warn(dp, "Failed to map DMA RX buffer\n"); 1236 return NULL; 1237 } 1238 1239 return frag; 1240 } 1241 1242 /** 1243 * nfp_net_rx_give_one() - Put mapped skb on the software and hardware rings 1244 * @dp: NFP Net data path struct 1245 * @rx_ring: RX ring structure 1246 * @frag: page fragment buffer 1247 * @dma_addr: DMA address of skb mapping 1248 */ 1249 static void nfp_net_rx_give_one(const struct nfp_net_dp *dp, 1250 struct nfp_net_rx_ring *rx_ring, 1251 void *frag, dma_addr_t dma_addr) 1252 { 1253 unsigned int wr_idx; 1254 1255 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 1256 1257 nfp_net_dma_sync_dev_rx(dp, dma_addr); 1258 1259 /* Stash SKB and DMA address away */ 1260 rx_ring->rxbufs[wr_idx].frag = frag; 1261 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr; 1262 1263 /* Fill freelist descriptor */ 1264 rx_ring->rxds[wr_idx].fld.reserved = 0; 1265 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0; 1266 nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld, 1267 dma_addr + dp->rx_dma_off); 1268 1269 rx_ring->wr_p++; 1270 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) { 1271 /* Update write pointer of the freelist queue. Make 1272 * sure all writes are flushed before telling the hardware. 1273 */ 1274 wmb(); 1275 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH); 1276 } 1277 } 1278 1279 /** 1280 * nfp_net_rx_ring_reset() - Reflect in SW state of freelist after disable 1281 * @rx_ring: RX ring structure 1282 * 1283 * Assumes that the device is stopped, must be idempotent. 1284 */ 1285 static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring) 1286 { 1287 unsigned int wr_idx, last_idx; 1288 1289 /* wr_p == rd_p means ring was never fed FL bufs. RX rings are always 1290 * kept at cnt - 1 FL bufs. 1291 */ 1292 if (rx_ring->wr_p == 0 && rx_ring->rd_p == 0) 1293 return; 1294 1295 /* Move the empty entry to the end of the list */ 1296 wr_idx = D_IDX(rx_ring, rx_ring->wr_p); 1297 last_idx = rx_ring->cnt - 1; 1298 rx_ring->rxbufs[wr_idx].dma_addr = rx_ring->rxbufs[last_idx].dma_addr; 1299 rx_ring->rxbufs[wr_idx].frag = rx_ring->rxbufs[last_idx].frag; 1300 rx_ring->rxbufs[last_idx].dma_addr = 0; 1301 rx_ring->rxbufs[last_idx].frag = NULL; 1302 1303 memset(rx_ring->rxds, 0, rx_ring->size); 1304 rx_ring->wr_p = 0; 1305 rx_ring->rd_p = 0; 1306 } 1307 1308 /** 1309 * nfp_net_rx_ring_bufs_free() - Free any buffers currently on the RX ring 1310 * @dp: NFP Net data path struct 1311 * @rx_ring: RX ring to remove buffers from 1312 * 1313 * Assumes that the device is stopped and buffers are in [0, ring->cnt - 1) 1314 * entries. After device is disabled nfp_net_rx_ring_reset() must be called 1315 * to restore required ring geometry. 1316 */ 1317 static void 1318 nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp, 1319 struct nfp_net_rx_ring *rx_ring) 1320 { 1321 unsigned int i; 1322 1323 for (i = 0; i < rx_ring->cnt - 1; i++) { 1324 /* NULL skb can only happen when initial filling of the ring 1325 * fails to allocate enough buffers and calls here to free 1326 * already allocated ones. 1327 */ 1328 if (!rx_ring->rxbufs[i].frag) 1329 continue; 1330 1331 nfp_net_dma_unmap_rx(dp, rx_ring->rxbufs[i].dma_addr); 1332 nfp_net_free_frag(rx_ring->rxbufs[i].frag, dp->xdp_prog); 1333 rx_ring->rxbufs[i].dma_addr = 0; 1334 rx_ring->rxbufs[i].frag = NULL; 1335 } 1336 } 1337 1338 /** 1339 * nfp_net_rx_ring_bufs_alloc() - Fill RX ring with buffers (don't give to FW) 1340 * @dp: NFP Net data path struct 1341 * @rx_ring: RX ring to remove buffers from 1342 */ 1343 static int 1344 nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp, 1345 struct nfp_net_rx_ring *rx_ring) 1346 { 1347 struct nfp_net_rx_buf *rxbufs; 1348 unsigned int i; 1349 1350 rxbufs = rx_ring->rxbufs; 1351 1352 for (i = 0; i < rx_ring->cnt - 1; i++) { 1353 rxbufs[i].frag = nfp_net_rx_alloc_one(dp, &rxbufs[i].dma_addr); 1354 if (!rxbufs[i].frag) { 1355 nfp_net_rx_ring_bufs_free(dp, rx_ring); 1356 return -ENOMEM; 1357 } 1358 } 1359 1360 return 0; 1361 } 1362 1363 /** 1364 * nfp_net_rx_ring_fill_freelist() - Give buffers from the ring to FW 1365 * @dp: NFP Net data path struct 1366 * @rx_ring: RX ring to fill 1367 */ 1368 static void 1369 nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp, 1370 struct nfp_net_rx_ring *rx_ring) 1371 { 1372 unsigned int i; 1373 1374 for (i = 0; i < rx_ring->cnt - 1; i++) 1375 nfp_net_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag, 1376 rx_ring->rxbufs[i].dma_addr); 1377 } 1378 1379 /** 1380 * nfp_net_rx_csum_has_errors() - group check if rxd has any csum errors 1381 * @flags: RX descriptor flags field in CPU byte order 1382 */ 1383 static int nfp_net_rx_csum_has_errors(u16 flags) 1384 { 1385 u16 csum_all_checked, csum_all_ok; 1386 1387 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL; 1388 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK; 1389 1390 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT); 1391 } 1392 1393 /** 1394 * nfp_net_rx_csum() - set SKB checksum field based on RX descriptor flags 1395 * @dp: NFP Net data path struct 1396 * @r_vec: per-ring structure 1397 * @rxd: Pointer to RX descriptor 1398 * @meta: Parsed metadata prepend 1399 * @skb: Pointer to SKB 1400 */ 1401 static void nfp_net_rx_csum(struct nfp_net_dp *dp, 1402 struct nfp_net_r_vector *r_vec, 1403 struct nfp_net_rx_desc *rxd, 1404 struct nfp_meta_parsed *meta, struct sk_buff *skb) 1405 { 1406 skb_checksum_none_assert(skb); 1407 1408 if (!(dp->netdev->features & NETIF_F_RXCSUM)) 1409 return; 1410 1411 if (meta->csum_type) { 1412 skb->ip_summed = meta->csum_type; 1413 skb->csum = meta->csum; 1414 u64_stats_update_begin(&r_vec->rx_sync); 1415 r_vec->hw_csum_rx_complete++; 1416 u64_stats_update_end(&r_vec->rx_sync); 1417 return; 1418 } 1419 1420 if (nfp_net_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) { 1421 u64_stats_update_begin(&r_vec->rx_sync); 1422 r_vec->hw_csum_rx_error++; 1423 u64_stats_update_end(&r_vec->rx_sync); 1424 return; 1425 } 1426 1427 /* Assume that the firmware will never report inner CSUM_OK unless outer 1428 * L4 headers were successfully parsed. FW will always report zero UDP 1429 * checksum as CSUM_OK. 1430 */ 1431 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK || 1432 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) { 1433 __skb_incr_checksum_unnecessary(skb); 1434 u64_stats_update_begin(&r_vec->rx_sync); 1435 r_vec->hw_csum_rx_ok++; 1436 u64_stats_update_end(&r_vec->rx_sync); 1437 } 1438 1439 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK || 1440 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) { 1441 __skb_incr_checksum_unnecessary(skb); 1442 u64_stats_update_begin(&r_vec->rx_sync); 1443 r_vec->hw_csum_rx_inner_ok++; 1444 u64_stats_update_end(&r_vec->rx_sync); 1445 } 1446 } 1447 1448 static void 1449 nfp_net_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta, 1450 unsigned int type, __be32 *hash) 1451 { 1452 if (!(netdev->features & NETIF_F_RXHASH)) 1453 return; 1454 1455 switch (type) { 1456 case NFP_NET_RSS_IPV4: 1457 case NFP_NET_RSS_IPV6: 1458 case NFP_NET_RSS_IPV6_EX: 1459 meta->hash_type = PKT_HASH_TYPE_L3; 1460 break; 1461 default: 1462 meta->hash_type = PKT_HASH_TYPE_L4; 1463 break; 1464 } 1465 1466 meta->hash = get_unaligned_be32(hash); 1467 } 1468 1469 static void 1470 nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta, 1471 void *data, struct nfp_net_rx_desc *rxd) 1472 { 1473 struct nfp_net_rx_hash *rx_hash = data; 1474 1475 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS)) 1476 return; 1477 1478 nfp_net_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type), 1479 &rx_hash->hash); 1480 } 1481 1482 static void * 1483 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta, 1484 void *data, int meta_len) 1485 { 1486 u32 meta_info; 1487 1488 meta_info = get_unaligned_be32(data); 1489 data += 4; 1490 1491 while (meta_info) { 1492 switch (meta_info & NFP_NET_META_FIELD_MASK) { 1493 case NFP_NET_META_HASH: 1494 meta_info >>= NFP_NET_META_FIELD_SIZE; 1495 nfp_net_set_hash(netdev, meta, 1496 meta_info & NFP_NET_META_FIELD_MASK, 1497 (__be32 *)data); 1498 data += 4; 1499 break; 1500 case NFP_NET_META_MARK: 1501 meta->mark = get_unaligned_be32(data); 1502 data += 4; 1503 break; 1504 case NFP_NET_META_PORTID: 1505 meta->portid = get_unaligned_be32(data); 1506 data += 4; 1507 break; 1508 case NFP_NET_META_CSUM: 1509 meta->csum_type = CHECKSUM_COMPLETE; 1510 meta->csum = 1511 (__force __wsum)__get_unaligned_cpu32(data); 1512 data += 4; 1513 break; 1514 default: 1515 return NULL; 1516 } 1517 1518 meta_info >>= NFP_NET_META_FIELD_SIZE; 1519 } 1520 1521 return data; 1522 } 1523 1524 static void 1525 nfp_net_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, 1526 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf, 1527 struct sk_buff *skb) 1528 { 1529 u64_stats_update_begin(&r_vec->rx_sync); 1530 r_vec->rx_drops++; 1531 /* If we have both skb and rxbuf the replacement buffer allocation 1532 * must have failed, count this as an alloc failure. 1533 */ 1534 if (skb && rxbuf) 1535 r_vec->rx_replace_buf_alloc_fail++; 1536 u64_stats_update_end(&r_vec->rx_sync); 1537 1538 /* skb is build based on the frag, free_skb() would free the frag 1539 * so to be able to reuse it we need an extra ref. 1540 */ 1541 if (skb && rxbuf && skb->head == rxbuf->frag) 1542 page_ref_inc(virt_to_head_page(rxbuf->frag)); 1543 if (rxbuf) 1544 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr); 1545 if (skb) 1546 dev_kfree_skb_any(skb); 1547 } 1548 1549 static bool 1550 nfp_net_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring, 1551 struct nfp_net_tx_ring *tx_ring, 1552 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off, 1553 unsigned int pkt_len, bool *completed) 1554 { 1555 struct nfp_net_tx_buf *txbuf; 1556 struct nfp_net_tx_desc *txd; 1557 int wr_idx; 1558 1559 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1560 if (!*completed) { 1561 nfp_net_xdp_complete(tx_ring); 1562 *completed = true; 1563 } 1564 1565 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1566 nfp_net_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf, 1567 NULL); 1568 return false; 1569 } 1570 } 1571 1572 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1573 1574 /* Stash the soft descriptor of the head then initialize it */ 1575 txbuf = &tx_ring->txbufs[wr_idx]; 1576 1577 nfp_net_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr); 1578 1579 txbuf->frag = rxbuf->frag; 1580 txbuf->dma_addr = rxbuf->dma_addr; 1581 txbuf->fidx = -1; 1582 txbuf->pkt_cnt = 1; 1583 txbuf->real_len = pkt_len; 1584 1585 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off, 1586 pkt_len, DMA_BIDIRECTIONAL); 1587 1588 /* Build TX descriptor */ 1589 txd = &tx_ring->txds[wr_idx]; 1590 txd->offset_eop = PCIE_DESC_TX_EOP; 1591 txd->dma_len = cpu_to_le16(pkt_len); 1592 nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off); 1593 txd->data_len = cpu_to_le16(pkt_len); 1594 1595 txd->flags = 0; 1596 txd->mss = 0; 1597 txd->lso_hdrlen = 0; 1598 1599 tx_ring->wr_p++; 1600 tx_ring->wr_ptr_add++; 1601 return true; 1602 } 1603 1604 /** 1605 * nfp_net_rx() - receive up to @budget packets on @rx_ring 1606 * @rx_ring: RX ring to receive from 1607 * @budget: NAPI budget 1608 * 1609 * Note, this function is separated out from the napi poll function to 1610 * more cleanly separate packet receive code from other bookkeeping 1611 * functions performed in the napi poll function. 1612 * 1613 * Return: Number of packets received. 1614 */ 1615 static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget) 1616 { 1617 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 1618 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 1619 struct nfp_net_tx_ring *tx_ring; 1620 struct bpf_prog *xdp_prog; 1621 bool xdp_tx_cmpl = false; 1622 unsigned int true_bufsz; 1623 struct sk_buff *skb; 1624 int pkts_polled = 0; 1625 struct xdp_buff xdp; 1626 int idx; 1627 1628 rcu_read_lock(); 1629 xdp_prog = READ_ONCE(dp->xdp_prog); 1630 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz; 1631 xdp.rxq = &rx_ring->xdp_rxq; 1632 tx_ring = r_vec->xdp_ring; 1633 1634 while (pkts_polled < budget) { 1635 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1636 struct nfp_net_rx_buf *rxbuf; 1637 struct nfp_net_rx_desc *rxd; 1638 struct nfp_meta_parsed meta; 1639 struct net_device *netdev; 1640 dma_addr_t new_dma_addr; 1641 u32 meta_len_xdp = 0; 1642 void *new_frag; 1643 1644 idx = D_IDX(rx_ring, rx_ring->rd_p); 1645 1646 rxd = &rx_ring->rxds[idx]; 1647 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 1648 break; 1649 1650 /* Memory barrier to ensure that we won't do other reads 1651 * before the DD bit. 1652 */ 1653 dma_rmb(); 1654 1655 memset(&meta, 0, sizeof(meta)); 1656 1657 rx_ring->rd_p++; 1658 pkts_polled++; 1659 1660 rxbuf = &rx_ring->rxbufs[idx]; 1661 /* < meta_len > 1662 * <-- [rx_offset] --> 1663 * --------------------------------------------------------- 1664 * | [XX] | metadata | packet | XXXX | 1665 * --------------------------------------------------------- 1666 * <---------------- data_len ---------------> 1667 * 1668 * The rx_offset is fixed for all packets, the meta_len can vary 1669 * on a packet by packet basis. If rx_offset is set to zero 1670 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the 1671 * buffer and is immediately followed by the packet (no [XX]). 1672 */ 1673 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 1674 data_len = le16_to_cpu(rxd->rxd.data_len); 1675 pkt_len = data_len - meta_len; 1676 1677 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 1678 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 1679 pkt_off += meta_len; 1680 else 1681 pkt_off += dp->rx_offset; 1682 meta_off = pkt_off - meta_len; 1683 1684 /* Stats update */ 1685 u64_stats_update_begin(&r_vec->rx_sync); 1686 r_vec->rx_pkts++; 1687 r_vec->rx_bytes += pkt_len; 1688 u64_stats_update_end(&r_vec->rx_sync); 1689 1690 if (unlikely(meta_len > NFP_NET_MAX_PREPEND || 1691 (dp->rx_offset && meta_len > dp->rx_offset))) { 1692 nn_dp_warn(dp, "oversized RX packet metadata %u\n", 1693 meta_len); 1694 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1695 continue; 1696 } 1697 1698 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, 1699 data_len); 1700 1701 if (!dp->chained_metadata_format) { 1702 nfp_net_set_hash_desc(dp->netdev, &meta, 1703 rxbuf->frag + meta_off, rxd); 1704 } else if (meta_len) { 1705 void *end; 1706 1707 end = nfp_net_parse_meta(dp->netdev, &meta, 1708 rxbuf->frag + meta_off, 1709 meta_len); 1710 if (unlikely(end != rxbuf->frag + pkt_off)) { 1711 nn_dp_warn(dp, "invalid RX packet metadata\n"); 1712 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, 1713 NULL); 1714 continue; 1715 } 1716 } 1717 1718 if (xdp_prog && !meta.portid) { 1719 void *orig_data = rxbuf->frag + pkt_off; 1720 unsigned int dma_off; 1721 int act; 1722 1723 xdp.data_hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM; 1724 xdp.data = orig_data; 1725 xdp.data_meta = orig_data; 1726 xdp.data_end = orig_data + pkt_len; 1727 1728 act = bpf_prog_run_xdp(xdp_prog, &xdp); 1729 1730 pkt_len = xdp.data_end - xdp.data; 1731 pkt_off += xdp.data - orig_data; 1732 1733 switch (act) { 1734 case XDP_PASS: 1735 meta_len_xdp = xdp.data - xdp.data_meta; 1736 break; 1737 case XDP_TX: 1738 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM; 1739 if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring, 1740 tx_ring, rxbuf, 1741 dma_off, 1742 pkt_len, 1743 &xdp_tx_cmpl))) 1744 trace_xdp_exception(dp->netdev, 1745 xdp_prog, act); 1746 continue; 1747 default: 1748 bpf_warn_invalid_xdp_action(act); 1749 /* fall through */ 1750 case XDP_ABORTED: 1751 trace_xdp_exception(dp->netdev, xdp_prog, act); 1752 /* fall through */ 1753 case XDP_DROP: 1754 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, 1755 rxbuf->dma_addr); 1756 continue; 1757 } 1758 } 1759 1760 if (likely(!meta.portid)) { 1761 netdev = dp->netdev; 1762 } else if (meta.portid == NFP_META_PORT_ID_CTRL) { 1763 struct nfp_net *nn = netdev_priv(dp->netdev); 1764 1765 nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off, 1766 pkt_len); 1767 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, 1768 rxbuf->dma_addr); 1769 continue; 1770 } else { 1771 struct nfp_net *nn; 1772 1773 nn = netdev_priv(dp->netdev); 1774 netdev = nfp_app_repr_get(nn->app, meta.portid); 1775 if (unlikely(!netdev)) { 1776 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, 1777 NULL); 1778 continue; 1779 } 1780 nfp_repr_inc_rx_stats(netdev, pkt_len); 1781 } 1782 1783 skb = build_skb(rxbuf->frag, true_bufsz); 1784 if (unlikely(!skb)) { 1785 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 1786 continue; 1787 } 1788 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr); 1789 if (unlikely(!new_frag)) { 1790 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 1791 continue; 1792 } 1793 1794 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 1795 1796 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 1797 1798 skb_reserve(skb, pkt_off); 1799 skb_put(skb, pkt_len); 1800 1801 skb->mark = meta.mark; 1802 skb_set_hash(skb, meta.hash, meta.hash_type); 1803 1804 skb_record_rx_queue(skb, rx_ring->idx); 1805 skb->protocol = eth_type_trans(skb, netdev); 1806 1807 nfp_net_rx_csum(dp, r_vec, rxd, &meta, skb); 1808 1809 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN) 1810 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), 1811 le16_to_cpu(rxd->rxd.vlan)); 1812 if (meta_len_xdp) 1813 skb_metadata_set(skb, meta_len_xdp); 1814 1815 napi_gro_receive(&rx_ring->r_vec->napi, skb); 1816 } 1817 1818 if (xdp_prog) { 1819 if (tx_ring->wr_ptr_add) 1820 nfp_net_tx_xmit_more_flush(tx_ring); 1821 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) && 1822 !xdp_tx_cmpl) 1823 if (!nfp_net_xdp_complete(tx_ring)) 1824 pkts_polled = budget; 1825 } 1826 rcu_read_unlock(); 1827 1828 return pkts_polled; 1829 } 1830 1831 /** 1832 * nfp_net_poll() - napi poll function 1833 * @napi: NAPI structure 1834 * @budget: NAPI budget 1835 * 1836 * Return: number of packets polled. 1837 */ 1838 static int nfp_net_poll(struct napi_struct *napi, int budget) 1839 { 1840 struct nfp_net_r_vector *r_vec = 1841 container_of(napi, struct nfp_net_r_vector, napi); 1842 unsigned int pkts_polled = 0; 1843 1844 if (r_vec->tx_ring) 1845 nfp_net_tx_complete(r_vec->tx_ring, budget); 1846 if (r_vec->rx_ring) 1847 pkts_polled = nfp_net_rx(r_vec->rx_ring, budget); 1848 1849 if (pkts_polled < budget) 1850 if (napi_complete_done(napi, pkts_polled)) 1851 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 1852 1853 return pkts_polled; 1854 } 1855 1856 /* Control device data path 1857 */ 1858 1859 static bool 1860 nfp_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 1861 struct sk_buff *skb, bool old) 1862 { 1863 unsigned int real_len = skb->len, meta_len = 0; 1864 struct nfp_net_tx_ring *tx_ring; 1865 struct nfp_net_tx_buf *txbuf; 1866 struct nfp_net_tx_desc *txd; 1867 struct nfp_net_dp *dp; 1868 dma_addr_t dma_addr; 1869 int wr_idx; 1870 1871 dp = &r_vec->nfp_net->dp; 1872 tx_ring = r_vec->tx_ring; 1873 1874 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) { 1875 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n"); 1876 goto err_free; 1877 } 1878 1879 if (unlikely(nfp_net_tx_full(tx_ring, 1))) { 1880 u64_stats_update_begin(&r_vec->tx_sync); 1881 r_vec->tx_busy++; 1882 u64_stats_update_end(&r_vec->tx_sync); 1883 if (!old) 1884 __skb_queue_tail(&r_vec->queue, skb); 1885 else 1886 __skb_queue_head(&r_vec->queue, skb); 1887 return true; 1888 } 1889 1890 if (nfp_app_ctrl_has_meta(nn->app)) { 1891 if (unlikely(skb_headroom(skb) < 8)) { 1892 nn_dp_warn(dp, "CTRL TX on skb without headroom\n"); 1893 goto err_free; 1894 } 1895 meta_len = 8; 1896 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4)); 1897 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4)); 1898 } 1899 1900 /* Start with the head skbuf */ 1901 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb), 1902 DMA_TO_DEVICE); 1903 if (dma_mapping_error(dp->dev, dma_addr)) 1904 goto err_dma_warn; 1905 1906 wr_idx = D_IDX(tx_ring, tx_ring->wr_p); 1907 1908 /* Stash the soft descriptor of the head then initialize it */ 1909 txbuf = &tx_ring->txbufs[wr_idx]; 1910 txbuf->skb = skb; 1911 txbuf->dma_addr = dma_addr; 1912 txbuf->fidx = -1; 1913 txbuf->pkt_cnt = 1; 1914 txbuf->real_len = real_len; 1915 1916 /* Build TX descriptor */ 1917 txd = &tx_ring->txds[wr_idx]; 1918 txd->offset_eop = meta_len | PCIE_DESC_TX_EOP; 1919 txd->dma_len = cpu_to_le16(skb_headlen(skb)); 1920 nfp_desc_set_dma_addr(txd, dma_addr); 1921 txd->data_len = cpu_to_le16(skb->len); 1922 1923 txd->flags = 0; 1924 txd->mss = 0; 1925 txd->lso_hdrlen = 0; 1926 1927 tx_ring->wr_p++; 1928 tx_ring->wr_ptr_add++; 1929 nfp_net_tx_xmit_more_flush(tx_ring); 1930 1931 return false; 1932 1933 err_dma_warn: 1934 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n"); 1935 err_free: 1936 u64_stats_update_begin(&r_vec->tx_sync); 1937 r_vec->tx_errors++; 1938 u64_stats_update_end(&r_vec->tx_sync); 1939 dev_kfree_skb_any(skb); 1940 return false; 1941 } 1942 1943 bool __nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb) 1944 { 1945 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0]; 1946 1947 return nfp_ctrl_tx_one(nn, r_vec, skb, false); 1948 } 1949 1950 bool nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb) 1951 { 1952 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0]; 1953 bool ret; 1954 1955 spin_lock_bh(&r_vec->lock); 1956 ret = nfp_ctrl_tx_one(nn, r_vec, skb, false); 1957 spin_unlock_bh(&r_vec->lock); 1958 1959 return ret; 1960 } 1961 1962 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec) 1963 { 1964 struct sk_buff *skb; 1965 1966 while ((skb = __skb_dequeue(&r_vec->queue))) 1967 if (nfp_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true)) 1968 return; 1969 } 1970 1971 static bool 1972 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len) 1973 { 1974 u32 meta_type, meta_tag; 1975 1976 if (!nfp_app_ctrl_has_meta(nn->app)) 1977 return !meta_len; 1978 1979 if (meta_len != 8) 1980 return false; 1981 1982 meta_type = get_unaligned_be32(data); 1983 meta_tag = get_unaligned_be32(data + 4); 1984 1985 return (meta_type == NFP_NET_META_PORTID && 1986 meta_tag == NFP_META_PORT_ID_CTRL); 1987 } 1988 1989 static bool 1990 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp, 1991 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring) 1992 { 1993 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off; 1994 struct nfp_net_rx_buf *rxbuf; 1995 struct nfp_net_rx_desc *rxd; 1996 dma_addr_t new_dma_addr; 1997 struct sk_buff *skb; 1998 void *new_frag; 1999 int idx; 2000 2001 idx = D_IDX(rx_ring, rx_ring->rd_p); 2002 2003 rxd = &rx_ring->rxds[idx]; 2004 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD)) 2005 return false; 2006 2007 /* Memory barrier to ensure that we won't do other reads 2008 * before the DD bit. 2009 */ 2010 dma_rmb(); 2011 2012 rx_ring->rd_p++; 2013 2014 rxbuf = &rx_ring->rxbufs[idx]; 2015 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK; 2016 data_len = le16_to_cpu(rxd->rxd.data_len); 2017 pkt_len = data_len - meta_len; 2018 2019 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off; 2020 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC) 2021 pkt_off += meta_len; 2022 else 2023 pkt_off += dp->rx_offset; 2024 meta_off = pkt_off - meta_len; 2025 2026 /* Stats update */ 2027 u64_stats_update_begin(&r_vec->rx_sync); 2028 r_vec->rx_pkts++; 2029 r_vec->rx_bytes += pkt_len; 2030 u64_stats_update_end(&r_vec->rx_sync); 2031 2032 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len); 2033 2034 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) { 2035 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n", 2036 meta_len); 2037 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 2038 return true; 2039 } 2040 2041 skb = build_skb(rxbuf->frag, dp->fl_bufsz); 2042 if (unlikely(!skb)) { 2043 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL); 2044 return true; 2045 } 2046 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr); 2047 if (unlikely(!new_frag)) { 2048 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb); 2049 return true; 2050 } 2051 2052 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr); 2053 2054 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr); 2055 2056 skb_reserve(skb, pkt_off); 2057 skb_put(skb, pkt_len); 2058 2059 nfp_app_ctrl_rx(nn->app, skb); 2060 2061 return true; 2062 } 2063 2064 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec) 2065 { 2066 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring; 2067 struct nfp_net *nn = r_vec->nfp_net; 2068 struct nfp_net_dp *dp = &nn->dp; 2069 unsigned int budget = 512; 2070 2071 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--) 2072 continue; 2073 2074 return budget; 2075 } 2076 2077 static void nfp_ctrl_poll(unsigned long arg) 2078 { 2079 struct nfp_net_r_vector *r_vec = (void *)arg; 2080 2081 spin_lock(&r_vec->lock); 2082 nfp_net_tx_complete(r_vec->tx_ring, 0); 2083 __nfp_ctrl_tx_queued(r_vec); 2084 spin_unlock(&r_vec->lock); 2085 2086 if (nfp_ctrl_rx(r_vec)) { 2087 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry); 2088 } else { 2089 tasklet_schedule(&r_vec->tasklet); 2090 nn_dp_warn(&r_vec->nfp_net->dp, 2091 "control message budget exceeded!\n"); 2092 } 2093 } 2094 2095 /* Setup and Configuration 2096 */ 2097 2098 /** 2099 * nfp_net_vecs_init() - Assign IRQs and setup rvecs. 2100 * @nn: NFP Network structure 2101 */ 2102 static void nfp_net_vecs_init(struct nfp_net *nn) 2103 { 2104 struct nfp_net_r_vector *r_vec; 2105 int r; 2106 2107 nn->lsc_handler = nfp_net_irq_lsc; 2108 nn->exn_handler = nfp_net_irq_exn; 2109 2110 for (r = 0; r < nn->max_r_vecs; r++) { 2111 struct msix_entry *entry; 2112 2113 entry = &nn->irq_entries[NFP_NET_NON_Q_VECTORS + r]; 2114 2115 r_vec = &nn->r_vecs[r]; 2116 r_vec->nfp_net = nn; 2117 r_vec->irq_entry = entry->entry; 2118 r_vec->irq_vector = entry->vector; 2119 2120 if (nn->dp.netdev) { 2121 r_vec->handler = nfp_net_irq_rxtx; 2122 } else { 2123 r_vec->handler = nfp_ctrl_irq_rxtx; 2124 2125 __skb_queue_head_init(&r_vec->queue); 2126 spin_lock_init(&r_vec->lock); 2127 tasklet_init(&r_vec->tasklet, nfp_ctrl_poll, 2128 (unsigned long)r_vec); 2129 tasklet_disable(&r_vec->tasklet); 2130 } 2131 2132 cpumask_set_cpu(r, &r_vec->affinity_mask); 2133 } 2134 } 2135 2136 /** 2137 * nfp_net_tx_ring_free() - Free resources allocated to a TX ring 2138 * @tx_ring: TX ring to free 2139 */ 2140 static void nfp_net_tx_ring_free(struct nfp_net_tx_ring *tx_ring) 2141 { 2142 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 2143 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 2144 2145 kvfree(tx_ring->txbufs); 2146 2147 if (tx_ring->txds) 2148 dma_free_coherent(dp->dev, tx_ring->size, 2149 tx_ring->txds, tx_ring->dma); 2150 2151 tx_ring->cnt = 0; 2152 tx_ring->txbufs = NULL; 2153 tx_ring->txds = NULL; 2154 tx_ring->dma = 0; 2155 tx_ring->size = 0; 2156 } 2157 2158 /** 2159 * nfp_net_tx_ring_alloc() - Allocate resource for a TX ring 2160 * @dp: NFP Net data path struct 2161 * @tx_ring: TX Ring structure to allocate 2162 * 2163 * Return: 0 on success, negative errno otherwise. 2164 */ 2165 static int 2166 nfp_net_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring) 2167 { 2168 struct nfp_net_r_vector *r_vec = tx_ring->r_vec; 2169 2170 tx_ring->cnt = dp->txd_cnt; 2171 2172 tx_ring->size = array_size(tx_ring->cnt, sizeof(*tx_ring->txds)); 2173 tx_ring->txds = dma_alloc_coherent(dp->dev, tx_ring->size, 2174 &tx_ring->dma, 2175 GFP_KERNEL | __GFP_NOWARN); 2176 if (!tx_ring->txds) { 2177 netdev_warn(dp->netdev, "failed to allocate TX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n", 2178 tx_ring->cnt); 2179 goto err_alloc; 2180 } 2181 2182 tx_ring->txbufs = kvcalloc(tx_ring->cnt, sizeof(*tx_ring->txbufs), 2183 GFP_KERNEL); 2184 if (!tx_ring->txbufs) 2185 goto err_alloc; 2186 2187 if (!tx_ring->is_xdp && dp->netdev) 2188 netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask, 2189 tx_ring->idx); 2190 2191 return 0; 2192 2193 err_alloc: 2194 nfp_net_tx_ring_free(tx_ring); 2195 return -ENOMEM; 2196 } 2197 2198 static void 2199 nfp_net_tx_ring_bufs_free(struct nfp_net_dp *dp, 2200 struct nfp_net_tx_ring *tx_ring) 2201 { 2202 unsigned int i; 2203 2204 if (!tx_ring->is_xdp) 2205 return; 2206 2207 for (i = 0; i < tx_ring->cnt; i++) { 2208 if (!tx_ring->txbufs[i].frag) 2209 return; 2210 2211 nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr); 2212 __free_page(virt_to_page(tx_ring->txbufs[i].frag)); 2213 } 2214 } 2215 2216 static int 2217 nfp_net_tx_ring_bufs_alloc(struct nfp_net_dp *dp, 2218 struct nfp_net_tx_ring *tx_ring) 2219 { 2220 struct nfp_net_tx_buf *txbufs = tx_ring->txbufs; 2221 unsigned int i; 2222 2223 if (!tx_ring->is_xdp) 2224 return 0; 2225 2226 for (i = 0; i < tx_ring->cnt; i++) { 2227 txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr); 2228 if (!txbufs[i].frag) { 2229 nfp_net_tx_ring_bufs_free(dp, tx_ring); 2230 return -ENOMEM; 2231 } 2232 } 2233 2234 return 0; 2235 } 2236 2237 static int nfp_net_tx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp) 2238 { 2239 unsigned int r; 2240 2241 dp->tx_rings = kcalloc(dp->num_tx_rings, sizeof(*dp->tx_rings), 2242 GFP_KERNEL); 2243 if (!dp->tx_rings) 2244 return -ENOMEM; 2245 2246 for (r = 0; r < dp->num_tx_rings; r++) { 2247 int bias = 0; 2248 2249 if (r >= dp->num_stack_tx_rings) 2250 bias = dp->num_stack_tx_rings; 2251 2252 nfp_net_tx_ring_init(&dp->tx_rings[r], &nn->r_vecs[r - bias], 2253 r, bias); 2254 2255 if (nfp_net_tx_ring_alloc(dp, &dp->tx_rings[r])) 2256 goto err_free_prev; 2257 2258 if (nfp_net_tx_ring_bufs_alloc(dp, &dp->tx_rings[r])) 2259 goto err_free_ring; 2260 } 2261 2262 return 0; 2263 2264 err_free_prev: 2265 while (r--) { 2266 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]); 2267 err_free_ring: 2268 nfp_net_tx_ring_free(&dp->tx_rings[r]); 2269 } 2270 kfree(dp->tx_rings); 2271 return -ENOMEM; 2272 } 2273 2274 static void nfp_net_tx_rings_free(struct nfp_net_dp *dp) 2275 { 2276 unsigned int r; 2277 2278 for (r = 0; r < dp->num_tx_rings; r++) { 2279 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]); 2280 nfp_net_tx_ring_free(&dp->tx_rings[r]); 2281 } 2282 2283 kfree(dp->tx_rings); 2284 } 2285 2286 /** 2287 * nfp_net_rx_ring_free() - Free resources allocated to a RX ring 2288 * @rx_ring: RX ring to free 2289 */ 2290 static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring) 2291 { 2292 struct nfp_net_r_vector *r_vec = rx_ring->r_vec; 2293 struct nfp_net_dp *dp = &r_vec->nfp_net->dp; 2294 2295 if (dp->netdev) 2296 xdp_rxq_info_unreg(&rx_ring->xdp_rxq); 2297 kvfree(rx_ring->rxbufs); 2298 2299 if (rx_ring->rxds) 2300 dma_free_coherent(dp->dev, rx_ring->size, 2301 rx_ring->rxds, rx_ring->dma); 2302 2303 rx_ring->cnt = 0; 2304 rx_ring->rxbufs = NULL; 2305 rx_ring->rxds = NULL; 2306 rx_ring->dma = 0; 2307 rx_ring->size = 0; 2308 } 2309 2310 /** 2311 * nfp_net_rx_ring_alloc() - Allocate resource for a RX ring 2312 * @dp: NFP Net data path struct 2313 * @rx_ring: RX ring to allocate 2314 * 2315 * Return: 0 on success, negative errno otherwise. 2316 */ 2317 static int 2318 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring) 2319 { 2320 int err; 2321 2322 if (dp->netdev) { 2323 err = xdp_rxq_info_reg(&rx_ring->xdp_rxq, dp->netdev, 2324 rx_ring->idx); 2325 if (err < 0) 2326 return err; 2327 } 2328 2329 rx_ring->cnt = dp->rxd_cnt; 2330 rx_ring->size = array_size(rx_ring->cnt, sizeof(*rx_ring->rxds)); 2331 rx_ring->rxds = dma_alloc_coherent(dp->dev, rx_ring->size, 2332 &rx_ring->dma, 2333 GFP_KERNEL | __GFP_NOWARN); 2334 if (!rx_ring->rxds) { 2335 netdev_warn(dp->netdev, "failed to allocate RX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n", 2336 rx_ring->cnt); 2337 goto err_alloc; 2338 } 2339 2340 rx_ring->rxbufs = kvcalloc(rx_ring->cnt, sizeof(*rx_ring->rxbufs), 2341 GFP_KERNEL); 2342 if (!rx_ring->rxbufs) 2343 goto err_alloc; 2344 2345 return 0; 2346 2347 err_alloc: 2348 nfp_net_rx_ring_free(rx_ring); 2349 return -ENOMEM; 2350 } 2351 2352 static int nfp_net_rx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp) 2353 { 2354 unsigned int r; 2355 2356 dp->rx_rings = kcalloc(dp->num_rx_rings, sizeof(*dp->rx_rings), 2357 GFP_KERNEL); 2358 if (!dp->rx_rings) 2359 return -ENOMEM; 2360 2361 for (r = 0; r < dp->num_rx_rings; r++) { 2362 nfp_net_rx_ring_init(&dp->rx_rings[r], &nn->r_vecs[r], r); 2363 2364 if (nfp_net_rx_ring_alloc(dp, &dp->rx_rings[r])) 2365 goto err_free_prev; 2366 2367 if (nfp_net_rx_ring_bufs_alloc(dp, &dp->rx_rings[r])) 2368 goto err_free_ring; 2369 } 2370 2371 return 0; 2372 2373 err_free_prev: 2374 while (r--) { 2375 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]); 2376 err_free_ring: 2377 nfp_net_rx_ring_free(&dp->rx_rings[r]); 2378 } 2379 kfree(dp->rx_rings); 2380 return -ENOMEM; 2381 } 2382 2383 static void nfp_net_rx_rings_free(struct nfp_net_dp *dp) 2384 { 2385 unsigned int r; 2386 2387 for (r = 0; r < dp->num_rx_rings; r++) { 2388 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]); 2389 nfp_net_rx_ring_free(&dp->rx_rings[r]); 2390 } 2391 2392 kfree(dp->rx_rings); 2393 } 2394 2395 static void 2396 nfp_net_vector_assign_rings(struct nfp_net_dp *dp, 2397 struct nfp_net_r_vector *r_vec, int idx) 2398 { 2399 r_vec->rx_ring = idx < dp->num_rx_rings ? &dp->rx_rings[idx] : NULL; 2400 r_vec->tx_ring = 2401 idx < dp->num_stack_tx_rings ? &dp->tx_rings[idx] : NULL; 2402 2403 r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ? 2404 &dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL; 2405 } 2406 2407 static int 2408 nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec, 2409 int idx) 2410 { 2411 int err; 2412 2413 /* Setup NAPI */ 2414 if (nn->dp.netdev) 2415 netif_napi_add(nn->dp.netdev, &r_vec->napi, 2416 nfp_net_poll, NAPI_POLL_WEIGHT); 2417 else 2418 tasklet_enable(&r_vec->tasklet); 2419 2420 snprintf(r_vec->name, sizeof(r_vec->name), 2421 "%s-rxtx-%d", nfp_net_name(nn), idx); 2422 err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name, 2423 r_vec); 2424 if (err) { 2425 if (nn->dp.netdev) 2426 netif_napi_del(&r_vec->napi); 2427 else 2428 tasklet_disable(&r_vec->tasklet); 2429 2430 nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector); 2431 return err; 2432 } 2433 disable_irq(r_vec->irq_vector); 2434 2435 irq_set_affinity_hint(r_vec->irq_vector, &r_vec->affinity_mask); 2436 2437 nn_dbg(nn, "RV%02d: irq=%03d/%03d\n", idx, r_vec->irq_vector, 2438 r_vec->irq_entry); 2439 2440 return 0; 2441 } 2442 2443 static void 2444 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec) 2445 { 2446 irq_set_affinity_hint(r_vec->irq_vector, NULL); 2447 if (nn->dp.netdev) 2448 netif_napi_del(&r_vec->napi); 2449 else 2450 tasklet_disable(&r_vec->tasklet); 2451 2452 free_irq(r_vec->irq_vector, r_vec); 2453 } 2454 2455 /** 2456 * nfp_net_rss_write_itbl() - Write RSS indirection table to device 2457 * @nn: NFP Net device to reconfigure 2458 */ 2459 void nfp_net_rss_write_itbl(struct nfp_net *nn) 2460 { 2461 int i; 2462 2463 for (i = 0; i < NFP_NET_CFG_RSS_ITBL_SZ; i += 4) 2464 nn_writel(nn, NFP_NET_CFG_RSS_ITBL + i, 2465 get_unaligned_le32(nn->rss_itbl + i)); 2466 } 2467 2468 /** 2469 * nfp_net_rss_write_key() - Write RSS hash key to device 2470 * @nn: NFP Net device to reconfigure 2471 */ 2472 void nfp_net_rss_write_key(struct nfp_net *nn) 2473 { 2474 int i; 2475 2476 for (i = 0; i < nfp_net_rss_key_sz(nn); i += 4) 2477 nn_writel(nn, NFP_NET_CFG_RSS_KEY + i, 2478 get_unaligned_le32(nn->rss_key + i)); 2479 } 2480 2481 /** 2482 * nfp_net_coalesce_write_cfg() - Write irq coalescence configuration to HW 2483 * @nn: NFP Net device to reconfigure 2484 */ 2485 void nfp_net_coalesce_write_cfg(struct nfp_net *nn) 2486 { 2487 u8 i; 2488 u32 factor; 2489 u32 value; 2490 2491 /* Compute factor used to convert coalesce '_usecs' parameters to 2492 * ME timestamp ticks. There are 16 ME clock cycles for each timestamp 2493 * count. 2494 */ 2495 factor = nn->tlv_caps.me_freq_mhz / 16; 2496 2497 /* copy RX interrupt coalesce parameters */ 2498 value = (nn->rx_coalesce_max_frames << 16) | 2499 (factor * nn->rx_coalesce_usecs); 2500 for (i = 0; i < nn->dp.num_rx_rings; i++) 2501 nn_writel(nn, NFP_NET_CFG_RXR_IRQ_MOD(i), value); 2502 2503 /* copy TX interrupt coalesce parameters */ 2504 value = (nn->tx_coalesce_max_frames << 16) | 2505 (factor * nn->tx_coalesce_usecs); 2506 for (i = 0; i < nn->dp.num_tx_rings; i++) 2507 nn_writel(nn, NFP_NET_CFG_TXR_IRQ_MOD(i), value); 2508 } 2509 2510 /** 2511 * nfp_net_write_mac_addr() - Write mac address to the device control BAR 2512 * @nn: NFP Net device to reconfigure 2513 * @addr: MAC address to write 2514 * 2515 * Writes the MAC address from the netdev to the device control BAR. Does not 2516 * perform the required reconfig. We do a bit of byte swapping dance because 2517 * firmware is LE. 2518 */ 2519 static void nfp_net_write_mac_addr(struct nfp_net *nn, const u8 *addr) 2520 { 2521 nn_writel(nn, NFP_NET_CFG_MACADDR + 0, get_unaligned_be32(addr)); 2522 nn_writew(nn, NFP_NET_CFG_MACADDR + 6, get_unaligned_be16(addr + 4)); 2523 } 2524 2525 static void nfp_net_vec_clear_ring_data(struct nfp_net *nn, unsigned int idx) 2526 { 2527 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), 0); 2528 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), 0); 2529 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), 0); 2530 2531 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), 0); 2532 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), 0); 2533 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), 0); 2534 } 2535 2536 /** 2537 * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP 2538 * @nn: NFP Net device to reconfigure 2539 * 2540 * Warning: must be fully idempotent. 2541 */ 2542 static void nfp_net_clear_config_and_disable(struct nfp_net *nn) 2543 { 2544 u32 new_ctrl, update; 2545 unsigned int r; 2546 int err; 2547 2548 new_ctrl = nn->dp.ctrl; 2549 new_ctrl &= ~NFP_NET_CFG_CTRL_ENABLE; 2550 update = NFP_NET_CFG_UPDATE_GEN; 2551 update |= NFP_NET_CFG_UPDATE_MSIX; 2552 update |= NFP_NET_CFG_UPDATE_RING; 2553 2554 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG) 2555 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG; 2556 2557 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0); 2558 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0); 2559 2560 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2561 err = nfp_net_reconfig(nn, update); 2562 if (err) 2563 nn_err(nn, "Could not disable device: %d\n", err); 2564 2565 for (r = 0; r < nn->dp.num_rx_rings; r++) 2566 nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]); 2567 for (r = 0; r < nn->dp.num_tx_rings; r++) 2568 nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]); 2569 for (r = 0; r < nn->dp.num_r_vecs; r++) 2570 nfp_net_vec_clear_ring_data(nn, r); 2571 2572 nn->dp.ctrl = new_ctrl; 2573 } 2574 2575 static void 2576 nfp_net_rx_ring_hw_cfg_write(struct nfp_net *nn, 2577 struct nfp_net_rx_ring *rx_ring, unsigned int idx) 2578 { 2579 /* Write the DMA address, size and MSI-X info to the device */ 2580 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), rx_ring->dma); 2581 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), ilog2(rx_ring->cnt)); 2582 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), rx_ring->r_vec->irq_entry); 2583 } 2584 2585 static void 2586 nfp_net_tx_ring_hw_cfg_write(struct nfp_net *nn, 2587 struct nfp_net_tx_ring *tx_ring, unsigned int idx) 2588 { 2589 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), tx_ring->dma); 2590 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), ilog2(tx_ring->cnt)); 2591 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), tx_ring->r_vec->irq_entry); 2592 } 2593 2594 /** 2595 * nfp_net_set_config_and_enable() - Write control BAR and enable NFP 2596 * @nn: NFP Net device to reconfigure 2597 */ 2598 static int nfp_net_set_config_and_enable(struct nfp_net *nn) 2599 { 2600 u32 bufsz, new_ctrl, update = 0; 2601 unsigned int r; 2602 int err; 2603 2604 new_ctrl = nn->dp.ctrl; 2605 2606 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_RSS_ANY) { 2607 nfp_net_rss_write_key(nn); 2608 nfp_net_rss_write_itbl(nn); 2609 nn_writel(nn, NFP_NET_CFG_RSS_CTRL, nn->rss_cfg); 2610 update |= NFP_NET_CFG_UPDATE_RSS; 2611 } 2612 2613 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_IRQMOD) { 2614 nfp_net_coalesce_write_cfg(nn); 2615 update |= NFP_NET_CFG_UPDATE_IRQMOD; 2616 } 2617 2618 for (r = 0; r < nn->dp.num_tx_rings; r++) 2619 nfp_net_tx_ring_hw_cfg_write(nn, &nn->dp.tx_rings[r], r); 2620 for (r = 0; r < nn->dp.num_rx_rings; r++) 2621 nfp_net_rx_ring_hw_cfg_write(nn, &nn->dp.rx_rings[r], r); 2622 2623 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, nn->dp.num_tx_rings == 64 ? 2624 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_tx_rings) - 1); 2625 2626 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, nn->dp.num_rx_rings == 64 ? 2627 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_rx_rings) - 1); 2628 2629 if (nn->dp.netdev) 2630 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr); 2631 2632 nn_writel(nn, NFP_NET_CFG_MTU, nn->dp.mtu); 2633 2634 bufsz = nn->dp.fl_bufsz - nn->dp.rx_dma_off - NFP_NET_RX_BUF_NON_DATA; 2635 nn_writel(nn, NFP_NET_CFG_FLBUFSZ, bufsz); 2636 2637 /* Enable device */ 2638 new_ctrl |= NFP_NET_CFG_CTRL_ENABLE; 2639 update |= NFP_NET_CFG_UPDATE_GEN; 2640 update |= NFP_NET_CFG_UPDATE_MSIX; 2641 update |= NFP_NET_CFG_UPDATE_RING; 2642 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG) 2643 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG; 2644 2645 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2646 err = nfp_net_reconfig(nn, update); 2647 if (err) { 2648 nfp_net_clear_config_and_disable(nn); 2649 return err; 2650 } 2651 2652 nn->dp.ctrl = new_ctrl; 2653 2654 for (r = 0; r < nn->dp.num_rx_rings; r++) 2655 nfp_net_rx_ring_fill_freelist(&nn->dp, &nn->dp.rx_rings[r]); 2656 2657 /* Since reconfiguration requests while NFP is down are ignored we 2658 * have to wipe the entire VXLAN configuration and reinitialize it. 2659 */ 2660 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN) { 2661 memset(&nn->vxlan_ports, 0, sizeof(nn->vxlan_ports)); 2662 memset(&nn->vxlan_usecnt, 0, sizeof(nn->vxlan_usecnt)); 2663 udp_tunnel_get_rx_info(nn->dp.netdev); 2664 } 2665 2666 return 0; 2667 } 2668 2669 /** 2670 * nfp_net_close_stack() - Quiesce the stack (part of close) 2671 * @nn: NFP Net device to reconfigure 2672 */ 2673 static void nfp_net_close_stack(struct nfp_net *nn) 2674 { 2675 unsigned int r; 2676 2677 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2678 netif_carrier_off(nn->dp.netdev); 2679 nn->link_up = false; 2680 2681 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2682 disable_irq(nn->r_vecs[r].irq_vector); 2683 napi_disable(&nn->r_vecs[r].napi); 2684 } 2685 2686 netif_tx_disable(nn->dp.netdev); 2687 } 2688 2689 /** 2690 * nfp_net_close_free_all() - Free all runtime resources 2691 * @nn: NFP Net device to reconfigure 2692 */ 2693 static void nfp_net_close_free_all(struct nfp_net *nn) 2694 { 2695 unsigned int r; 2696 2697 nfp_net_tx_rings_free(&nn->dp); 2698 nfp_net_rx_rings_free(&nn->dp); 2699 2700 for (r = 0; r < nn->dp.num_r_vecs; r++) 2701 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 2702 2703 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX); 2704 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX); 2705 } 2706 2707 /** 2708 * nfp_net_netdev_close() - Called when the device is downed 2709 * @netdev: netdev structure 2710 */ 2711 static int nfp_net_netdev_close(struct net_device *netdev) 2712 { 2713 struct nfp_net *nn = netdev_priv(netdev); 2714 2715 /* Step 1: Disable RX and TX rings from the Linux kernel perspective 2716 */ 2717 nfp_net_close_stack(nn); 2718 2719 /* Step 2: Tell NFP 2720 */ 2721 nfp_net_clear_config_and_disable(nn); 2722 nfp_port_configure(netdev, false); 2723 2724 /* Step 3: Free resources 2725 */ 2726 nfp_net_close_free_all(nn); 2727 2728 nn_dbg(nn, "%s down", netdev->name); 2729 return 0; 2730 } 2731 2732 void nfp_ctrl_close(struct nfp_net *nn) 2733 { 2734 int r; 2735 2736 rtnl_lock(); 2737 2738 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2739 disable_irq(nn->r_vecs[r].irq_vector); 2740 tasklet_disable(&nn->r_vecs[r].tasklet); 2741 } 2742 2743 nfp_net_clear_config_and_disable(nn); 2744 2745 nfp_net_close_free_all(nn); 2746 2747 rtnl_unlock(); 2748 } 2749 2750 /** 2751 * nfp_net_open_stack() - Start the device from stack's perspective 2752 * @nn: NFP Net device to reconfigure 2753 */ 2754 static void nfp_net_open_stack(struct nfp_net *nn) 2755 { 2756 unsigned int r; 2757 2758 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2759 napi_enable(&nn->r_vecs[r].napi); 2760 enable_irq(nn->r_vecs[r].irq_vector); 2761 } 2762 2763 netif_tx_wake_all_queues(nn->dp.netdev); 2764 2765 enable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2766 nfp_net_read_link_status(nn); 2767 } 2768 2769 static int nfp_net_open_alloc_all(struct nfp_net *nn) 2770 { 2771 int err, r; 2772 2773 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_EXN, "%s-exn", 2774 nn->exn_name, sizeof(nn->exn_name), 2775 NFP_NET_IRQ_EXN_IDX, nn->exn_handler); 2776 if (err) 2777 return err; 2778 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_LSC, "%s-lsc", 2779 nn->lsc_name, sizeof(nn->lsc_name), 2780 NFP_NET_IRQ_LSC_IDX, nn->lsc_handler); 2781 if (err) 2782 goto err_free_exn; 2783 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector); 2784 2785 for (r = 0; r < nn->dp.num_r_vecs; r++) { 2786 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r); 2787 if (err) 2788 goto err_cleanup_vec_p; 2789 } 2790 2791 err = nfp_net_rx_rings_prepare(nn, &nn->dp); 2792 if (err) 2793 goto err_cleanup_vec; 2794 2795 err = nfp_net_tx_rings_prepare(nn, &nn->dp); 2796 if (err) 2797 goto err_free_rx_rings; 2798 2799 for (r = 0; r < nn->max_r_vecs; r++) 2800 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r); 2801 2802 return 0; 2803 2804 err_free_rx_rings: 2805 nfp_net_rx_rings_free(&nn->dp); 2806 err_cleanup_vec: 2807 r = nn->dp.num_r_vecs; 2808 err_cleanup_vec_p: 2809 while (r--) 2810 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 2811 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX); 2812 err_free_exn: 2813 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX); 2814 return err; 2815 } 2816 2817 static int nfp_net_netdev_open(struct net_device *netdev) 2818 { 2819 struct nfp_net *nn = netdev_priv(netdev); 2820 int err; 2821 2822 /* Step 1: Allocate resources for rings and the like 2823 * - Request interrupts 2824 * - Allocate RX and TX ring resources 2825 * - Setup initial RSS table 2826 */ 2827 err = nfp_net_open_alloc_all(nn); 2828 if (err) 2829 return err; 2830 2831 err = netif_set_real_num_tx_queues(netdev, nn->dp.num_stack_tx_rings); 2832 if (err) 2833 goto err_free_all; 2834 2835 err = netif_set_real_num_rx_queues(netdev, nn->dp.num_rx_rings); 2836 if (err) 2837 goto err_free_all; 2838 2839 /* Step 2: Configure the NFP 2840 * - Ifup the physical interface if it exists 2841 * - Enable rings from 0 to tx_rings/rx_rings - 1. 2842 * - Write MAC address (in case it changed) 2843 * - Set the MTU 2844 * - Set the Freelist buffer size 2845 * - Enable the FW 2846 */ 2847 err = nfp_port_configure(netdev, true); 2848 if (err) 2849 goto err_free_all; 2850 2851 err = nfp_net_set_config_and_enable(nn); 2852 if (err) 2853 goto err_port_disable; 2854 2855 /* Step 3: Enable for kernel 2856 * - put some freelist descriptors on each RX ring 2857 * - enable NAPI on each ring 2858 * - enable all TX queues 2859 * - set link state 2860 */ 2861 nfp_net_open_stack(nn); 2862 2863 return 0; 2864 2865 err_port_disable: 2866 nfp_port_configure(netdev, false); 2867 err_free_all: 2868 nfp_net_close_free_all(nn); 2869 return err; 2870 } 2871 2872 int nfp_ctrl_open(struct nfp_net *nn) 2873 { 2874 int err, r; 2875 2876 /* ring dumping depends on vNICs being opened/closed under rtnl */ 2877 rtnl_lock(); 2878 2879 err = nfp_net_open_alloc_all(nn); 2880 if (err) 2881 goto err_unlock; 2882 2883 err = nfp_net_set_config_and_enable(nn); 2884 if (err) 2885 goto err_free_all; 2886 2887 for (r = 0; r < nn->dp.num_r_vecs; r++) 2888 enable_irq(nn->r_vecs[r].irq_vector); 2889 2890 rtnl_unlock(); 2891 2892 return 0; 2893 2894 err_free_all: 2895 nfp_net_close_free_all(nn); 2896 err_unlock: 2897 rtnl_unlock(); 2898 return err; 2899 } 2900 2901 static void nfp_net_set_rx_mode(struct net_device *netdev) 2902 { 2903 struct nfp_net *nn = netdev_priv(netdev); 2904 u32 new_ctrl; 2905 2906 new_ctrl = nn->dp.ctrl; 2907 2908 if (!netdev_mc_empty(netdev) || netdev->flags & IFF_ALLMULTI) 2909 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_L2MC; 2910 else 2911 new_ctrl &= ~NFP_NET_CFG_CTRL_L2MC; 2912 2913 if (netdev->flags & IFF_PROMISC) { 2914 if (nn->cap & NFP_NET_CFG_CTRL_PROMISC) 2915 new_ctrl |= NFP_NET_CFG_CTRL_PROMISC; 2916 else 2917 nn_warn(nn, "FW does not support promiscuous mode\n"); 2918 } else { 2919 new_ctrl &= ~NFP_NET_CFG_CTRL_PROMISC; 2920 } 2921 2922 if (new_ctrl == nn->dp.ctrl) 2923 return; 2924 2925 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 2926 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_GEN); 2927 2928 nn->dp.ctrl = new_ctrl; 2929 } 2930 2931 static void nfp_net_rss_init_itbl(struct nfp_net *nn) 2932 { 2933 int i; 2934 2935 for (i = 0; i < sizeof(nn->rss_itbl); i++) 2936 nn->rss_itbl[i] = 2937 ethtool_rxfh_indir_default(i, nn->dp.num_rx_rings); 2938 } 2939 2940 static void nfp_net_dp_swap(struct nfp_net *nn, struct nfp_net_dp *dp) 2941 { 2942 struct nfp_net_dp new_dp = *dp; 2943 2944 *dp = nn->dp; 2945 nn->dp = new_dp; 2946 2947 nn->dp.netdev->mtu = new_dp.mtu; 2948 2949 if (!netif_is_rxfh_configured(nn->dp.netdev)) 2950 nfp_net_rss_init_itbl(nn); 2951 } 2952 2953 static int nfp_net_dp_swap_enable(struct nfp_net *nn, struct nfp_net_dp *dp) 2954 { 2955 unsigned int r; 2956 int err; 2957 2958 nfp_net_dp_swap(nn, dp); 2959 2960 for (r = 0; r < nn->max_r_vecs; r++) 2961 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r); 2962 2963 err = netif_set_real_num_rx_queues(nn->dp.netdev, nn->dp.num_rx_rings); 2964 if (err) 2965 return err; 2966 2967 if (nn->dp.netdev->real_num_tx_queues != nn->dp.num_stack_tx_rings) { 2968 err = netif_set_real_num_tx_queues(nn->dp.netdev, 2969 nn->dp.num_stack_tx_rings); 2970 if (err) 2971 return err; 2972 } 2973 2974 return nfp_net_set_config_and_enable(nn); 2975 } 2976 2977 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn) 2978 { 2979 struct nfp_net_dp *new; 2980 2981 new = kmalloc(sizeof(*new), GFP_KERNEL); 2982 if (!new) 2983 return NULL; 2984 2985 *new = nn->dp; 2986 2987 /* Clear things which need to be recomputed */ 2988 new->fl_bufsz = 0; 2989 new->tx_rings = NULL; 2990 new->rx_rings = NULL; 2991 new->num_r_vecs = 0; 2992 new->num_stack_tx_rings = 0; 2993 2994 return new; 2995 } 2996 2997 static int 2998 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp, 2999 struct netlink_ext_ack *extack) 3000 { 3001 /* XDP-enabled tests */ 3002 if (!dp->xdp_prog) 3003 return 0; 3004 if (dp->fl_bufsz > PAGE_SIZE) { 3005 NL_SET_ERR_MSG_MOD(extack, "MTU too large w/ XDP enabled"); 3006 return -EINVAL; 3007 } 3008 if (dp->num_tx_rings > nn->max_tx_rings) { 3009 NL_SET_ERR_MSG_MOD(extack, "Insufficient number of TX rings w/ XDP enabled"); 3010 return -EINVAL; 3011 } 3012 3013 return 0; 3014 } 3015 3016 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp, 3017 struct netlink_ext_ack *extack) 3018 { 3019 int r, err; 3020 3021 dp->fl_bufsz = nfp_net_calc_fl_bufsz(dp); 3022 3023 dp->num_stack_tx_rings = dp->num_tx_rings; 3024 if (dp->xdp_prog) 3025 dp->num_stack_tx_rings -= dp->num_rx_rings; 3026 3027 dp->num_r_vecs = max(dp->num_rx_rings, dp->num_stack_tx_rings); 3028 3029 err = nfp_net_check_config(nn, dp, extack); 3030 if (err) 3031 goto exit_free_dp; 3032 3033 if (!netif_running(dp->netdev)) { 3034 nfp_net_dp_swap(nn, dp); 3035 err = 0; 3036 goto exit_free_dp; 3037 } 3038 3039 /* Prepare new rings */ 3040 for (r = nn->dp.num_r_vecs; r < dp->num_r_vecs; r++) { 3041 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r); 3042 if (err) { 3043 dp->num_r_vecs = r; 3044 goto err_cleanup_vecs; 3045 } 3046 } 3047 3048 err = nfp_net_rx_rings_prepare(nn, dp); 3049 if (err) 3050 goto err_cleanup_vecs; 3051 3052 err = nfp_net_tx_rings_prepare(nn, dp); 3053 if (err) 3054 goto err_free_rx; 3055 3056 /* Stop device, swap in new rings, try to start the firmware */ 3057 nfp_net_close_stack(nn); 3058 nfp_net_clear_config_and_disable(nn); 3059 3060 err = nfp_net_dp_swap_enable(nn, dp); 3061 if (err) { 3062 int err2; 3063 3064 nfp_net_clear_config_and_disable(nn); 3065 3066 /* Try with old configuration and old rings */ 3067 err2 = nfp_net_dp_swap_enable(nn, dp); 3068 if (err2) 3069 nn_err(nn, "Can't restore ring config - FW communication failed (%d,%d)\n", 3070 err, err2); 3071 } 3072 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--) 3073 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 3074 3075 nfp_net_rx_rings_free(dp); 3076 nfp_net_tx_rings_free(dp); 3077 3078 nfp_net_open_stack(nn); 3079 exit_free_dp: 3080 kfree(dp); 3081 3082 return err; 3083 3084 err_free_rx: 3085 nfp_net_rx_rings_free(dp); 3086 err_cleanup_vecs: 3087 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--) 3088 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]); 3089 kfree(dp); 3090 return err; 3091 } 3092 3093 static int nfp_net_change_mtu(struct net_device *netdev, int new_mtu) 3094 { 3095 struct nfp_net *nn = netdev_priv(netdev); 3096 struct nfp_net_dp *dp; 3097 int err; 3098 3099 err = nfp_app_check_mtu(nn->app, netdev, new_mtu); 3100 if (err) 3101 return err; 3102 3103 dp = nfp_net_clone_dp(nn); 3104 if (!dp) 3105 return -ENOMEM; 3106 3107 dp->mtu = new_mtu; 3108 3109 return nfp_net_ring_reconfig(nn, dp, NULL); 3110 } 3111 3112 static int 3113 nfp_net_vlan_rx_add_vid(struct net_device *netdev, __be16 proto, u16 vid) 3114 { 3115 struct nfp_net *nn = netdev_priv(netdev); 3116 3117 /* Priority tagged packets with vlan id 0 are processed by the 3118 * NFP as untagged packets 3119 */ 3120 if (!vid) 3121 return 0; 3122 3123 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid); 3124 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO, 3125 ETH_P_8021Q); 3126 3127 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_ADD); 3128 } 3129 3130 static int 3131 nfp_net_vlan_rx_kill_vid(struct net_device *netdev, __be16 proto, u16 vid) 3132 { 3133 struct nfp_net *nn = netdev_priv(netdev); 3134 3135 /* Priority tagged packets with vlan id 0 are processed by the 3136 * NFP as untagged packets 3137 */ 3138 if (!vid) 3139 return 0; 3140 3141 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid); 3142 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO, 3143 ETH_P_8021Q); 3144 3145 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_KILL); 3146 } 3147 3148 static void nfp_net_stat64(struct net_device *netdev, 3149 struct rtnl_link_stats64 *stats) 3150 { 3151 struct nfp_net *nn = netdev_priv(netdev); 3152 int r; 3153 3154 /* Collect software stats */ 3155 for (r = 0; r < nn->max_r_vecs; r++) { 3156 struct nfp_net_r_vector *r_vec = &nn->r_vecs[r]; 3157 u64 data[3]; 3158 unsigned int start; 3159 3160 do { 3161 start = u64_stats_fetch_begin(&r_vec->rx_sync); 3162 data[0] = r_vec->rx_pkts; 3163 data[1] = r_vec->rx_bytes; 3164 data[2] = r_vec->rx_drops; 3165 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start)); 3166 stats->rx_packets += data[0]; 3167 stats->rx_bytes += data[1]; 3168 stats->rx_dropped += data[2]; 3169 3170 do { 3171 start = u64_stats_fetch_begin(&r_vec->tx_sync); 3172 data[0] = r_vec->tx_pkts; 3173 data[1] = r_vec->tx_bytes; 3174 data[2] = r_vec->tx_errors; 3175 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start)); 3176 stats->tx_packets += data[0]; 3177 stats->tx_bytes += data[1]; 3178 stats->tx_errors += data[2]; 3179 } 3180 3181 /* Add in device stats */ 3182 stats->multicast += nn_readq(nn, NFP_NET_CFG_STATS_RX_MC_FRAMES); 3183 stats->rx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_RX_DISCARDS); 3184 stats->rx_errors += nn_readq(nn, NFP_NET_CFG_STATS_RX_ERRORS); 3185 3186 stats->tx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_TX_DISCARDS); 3187 stats->tx_errors += nn_readq(nn, NFP_NET_CFG_STATS_TX_ERRORS); 3188 } 3189 3190 static int nfp_net_set_features(struct net_device *netdev, 3191 netdev_features_t features) 3192 { 3193 netdev_features_t changed = netdev->features ^ features; 3194 struct nfp_net *nn = netdev_priv(netdev); 3195 u32 new_ctrl; 3196 int err; 3197 3198 /* Assume this is not called with features we have not advertised */ 3199 3200 new_ctrl = nn->dp.ctrl; 3201 3202 if (changed & NETIF_F_RXCSUM) { 3203 if (features & NETIF_F_RXCSUM) 3204 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY; 3205 else 3206 new_ctrl &= ~NFP_NET_CFG_CTRL_RXCSUM_ANY; 3207 } 3208 3209 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3210 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) 3211 new_ctrl |= NFP_NET_CFG_CTRL_TXCSUM; 3212 else 3213 new_ctrl &= ~NFP_NET_CFG_CTRL_TXCSUM; 3214 } 3215 3216 if (changed & (NETIF_F_TSO | NETIF_F_TSO6)) { 3217 if (features & (NETIF_F_TSO | NETIF_F_TSO6)) 3218 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?: 3219 NFP_NET_CFG_CTRL_LSO; 3220 else 3221 new_ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY; 3222 } 3223 3224 if (changed & NETIF_F_HW_VLAN_CTAG_RX) { 3225 if (features & NETIF_F_HW_VLAN_CTAG_RX) 3226 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN; 3227 else 3228 new_ctrl &= ~NFP_NET_CFG_CTRL_RXVLAN; 3229 } 3230 3231 if (changed & NETIF_F_HW_VLAN_CTAG_TX) { 3232 if (features & NETIF_F_HW_VLAN_CTAG_TX) 3233 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN; 3234 else 3235 new_ctrl &= ~NFP_NET_CFG_CTRL_TXVLAN; 3236 } 3237 3238 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) { 3239 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) 3240 new_ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER; 3241 else 3242 new_ctrl &= ~NFP_NET_CFG_CTRL_CTAG_FILTER; 3243 } 3244 3245 if (changed & NETIF_F_SG) { 3246 if (features & NETIF_F_SG) 3247 new_ctrl |= NFP_NET_CFG_CTRL_GATHER; 3248 else 3249 new_ctrl &= ~NFP_NET_CFG_CTRL_GATHER; 3250 } 3251 3252 err = nfp_port_set_features(netdev, features); 3253 if (err) 3254 return err; 3255 3256 nn_dbg(nn, "Feature change 0x%llx -> 0x%llx (changed=0x%llx)\n", 3257 netdev->features, features, changed); 3258 3259 if (new_ctrl == nn->dp.ctrl) 3260 return 0; 3261 3262 nn_dbg(nn, "NIC ctrl: 0x%x -> 0x%x\n", nn->dp.ctrl, new_ctrl); 3263 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl); 3264 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN); 3265 if (err) 3266 return err; 3267 3268 nn->dp.ctrl = new_ctrl; 3269 3270 return 0; 3271 } 3272 3273 static netdev_features_t 3274 nfp_net_features_check(struct sk_buff *skb, struct net_device *dev, 3275 netdev_features_t features) 3276 { 3277 u8 l4_hdr; 3278 3279 /* We can't do TSO over double tagged packets (802.1AD) */ 3280 features &= vlan_features_check(skb, features); 3281 3282 if (!skb->encapsulation) 3283 return features; 3284 3285 /* Ensure that inner L4 header offset fits into TX descriptor field */ 3286 if (skb_is_gso(skb)) { 3287 u32 hdrlen; 3288 3289 hdrlen = skb_inner_transport_header(skb) - skb->data + 3290 inner_tcp_hdrlen(skb); 3291 3292 /* Assume worst case scenario of having longest possible 3293 * metadata prepend - 8B 3294 */ 3295 if (unlikely(hdrlen > NFP_NET_LSO_MAX_HDR_SZ - 8)) 3296 features &= ~NETIF_F_GSO_MASK; 3297 } 3298 3299 /* VXLAN/GRE check */ 3300 switch (vlan_get_protocol(skb)) { 3301 case htons(ETH_P_IP): 3302 l4_hdr = ip_hdr(skb)->protocol; 3303 break; 3304 case htons(ETH_P_IPV6): 3305 l4_hdr = ipv6_hdr(skb)->nexthdr; 3306 break; 3307 default: 3308 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3309 } 3310 3311 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER || 3312 skb->inner_protocol != htons(ETH_P_TEB) || 3313 (l4_hdr != IPPROTO_UDP && l4_hdr != IPPROTO_GRE) || 3314 (l4_hdr == IPPROTO_UDP && 3315 (skb_inner_mac_header(skb) - skb_transport_header(skb) != 3316 sizeof(struct udphdr) + sizeof(struct vxlanhdr)))) 3317 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3318 3319 return features; 3320 } 3321 3322 static int 3323 nfp_net_get_phys_port_name(struct net_device *netdev, char *name, size_t len) 3324 { 3325 struct nfp_net *nn = netdev_priv(netdev); 3326 int n; 3327 3328 if (nn->port) 3329 return nfp_port_get_phys_port_name(netdev, name, len); 3330 3331 if (nn->dp.is_vf || nn->vnic_no_name) 3332 return -EOPNOTSUPP; 3333 3334 n = snprintf(name, len, "n%d", nn->id); 3335 if (n >= len) 3336 return -EINVAL; 3337 3338 return 0; 3339 } 3340 3341 /** 3342 * nfp_net_set_vxlan_port() - set vxlan port in SW and reconfigure HW 3343 * @nn: NFP Net device to reconfigure 3344 * @idx: Index into the port table where new port should be written 3345 * @port: UDP port to configure (pass zero to remove VXLAN port) 3346 */ 3347 static void nfp_net_set_vxlan_port(struct nfp_net *nn, int idx, __be16 port) 3348 { 3349 int i; 3350 3351 nn->vxlan_ports[idx] = port; 3352 3353 if (!(nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN)) 3354 return; 3355 3356 BUILD_BUG_ON(NFP_NET_N_VXLAN_PORTS & 1); 3357 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i += 2) 3358 nn_writel(nn, NFP_NET_CFG_VXLAN_PORT + i * sizeof(port), 3359 be16_to_cpu(nn->vxlan_ports[i + 1]) << 16 | 3360 be16_to_cpu(nn->vxlan_ports[i])); 3361 3362 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_VXLAN); 3363 } 3364 3365 /** 3366 * nfp_net_find_vxlan_idx() - find table entry of the port or a free one 3367 * @nn: NFP Network structure 3368 * @port: UDP port to look for 3369 * 3370 * Return: if the port is already in the table -- it's position; 3371 * if the port is not in the table -- free position to use; 3372 * if the table is full -- -ENOSPC. 3373 */ 3374 static int nfp_net_find_vxlan_idx(struct nfp_net *nn, __be16 port) 3375 { 3376 int i, free_idx = -ENOSPC; 3377 3378 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i++) { 3379 if (nn->vxlan_ports[i] == port) 3380 return i; 3381 if (!nn->vxlan_usecnt[i]) 3382 free_idx = i; 3383 } 3384 3385 return free_idx; 3386 } 3387 3388 static void nfp_net_add_vxlan_port(struct net_device *netdev, 3389 struct udp_tunnel_info *ti) 3390 { 3391 struct nfp_net *nn = netdev_priv(netdev); 3392 int idx; 3393 3394 if (ti->type != UDP_TUNNEL_TYPE_VXLAN) 3395 return; 3396 3397 idx = nfp_net_find_vxlan_idx(nn, ti->port); 3398 if (idx == -ENOSPC) 3399 return; 3400 3401 if (!nn->vxlan_usecnt[idx]++) 3402 nfp_net_set_vxlan_port(nn, idx, ti->port); 3403 } 3404 3405 static void nfp_net_del_vxlan_port(struct net_device *netdev, 3406 struct udp_tunnel_info *ti) 3407 { 3408 struct nfp_net *nn = netdev_priv(netdev); 3409 int idx; 3410 3411 if (ti->type != UDP_TUNNEL_TYPE_VXLAN) 3412 return; 3413 3414 idx = nfp_net_find_vxlan_idx(nn, ti->port); 3415 if (idx == -ENOSPC || !nn->vxlan_usecnt[idx]) 3416 return; 3417 3418 if (!--nn->vxlan_usecnt[idx]) 3419 nfp_net_set_vxlan_port(nn, idx, 0); 3420 } 3421 3422 static int nfp_net_xdp_setup_drv(struct nfp_net *nn, struct netdev_bpf *bpf) 3423 { 3424 struct bpf_prog *prog = bpf->prog; 3425 struct nfp_net_dp *dp; 3426 int err; 3427 3428 if (!xdp_attachment_flags_ok(&nn->xdp, bpf)) 3429 return -EBUSY; 3430 3431 if (!prog == !nn->dp.xdp_prog) { 3432 WRITE_ONCE(nn->dp.xdp_prog, prog); 3433 xdp_attachment_setup(&nn->xdp, bpf); 3434 return 0; 3435 } 3436 3437 dp = nfp_net_clone_dp(nn); 3438 if (!dp) 3439 return -ENOMEM; 3440 3441 dp->xdp_prog = prog; 3442 dp->num_tx_rings += prog ? nn->dp.num_rx_rings : -nn->dp.num_rx_rings; 3443 dp->rx_dma_dir = prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE; 3444 dp->rx_dma_off = prog ? XDP_PACKET_HEADROOM - nn->dp.rx_offset : 0; 3445 3446 /* We need RX reconfig to remap the buffers (BIDIR vs FROM_DEV) */ 3447 err = nfp_net_ring_reconfig(nn, dp, bpf->extack); 3448 if (err) 3449 return err; 3450 3451 xdp_attachment_setup(&nn->xdp, bpf); 3452 return 0; 3453 } 3454 3455 static int nfp_net_xdp_setup_hw(struct nfp_net *nn, struct netdev_bpf *bpf) 3456 { 3457 int err; 3458 3459 if (!xdp_attachment_flags_ok(&nn->xdp_hw, bpf)) 3460 return -EBUSY; 3461 3462 err = nfp_app_xdp_offload(nn->app, nn, bpf->prog, bpf->extack); 3463 if (err) 3464 return err; 3465 3466 xdp_attachment_setup(&nn->xdp_hw, bpf); 3467 return 0; 3468 } 3469 3470 static int nfp_net_xdp(struct net_device *netdev, struct netdev_bpf *xdp) 3471 { 3472 struct nfp_net *nn = netdev_priv(netdev); 3473 3474 switch (xdp->command) { 3475 case XDP_SETUP_PROG: 3476 return nfp_net_xdp_setup_drv(nn, xdp); 3477 case XDP_SETUP_PROG_HW: 3478 return nfp_net_xdp_setup_hw(nn, xdp); 3479 case XDP_QUERY_PROG: 3480 return xdp_attachment_query(&nn->xdp, xdp); 3481 case XDP_QUERY_PROG_HW: 3482 return xdp_attachment_query(&nn->xdp_hw, xdp); 3483 default: 3484 return nfp_app_bpf(nn->app, nn, xdp); 3485 } 3486 } 3487 3488 static int nfp_net_set_mac_address(struct net_device *netdev, void *addr) 3489 { 3490 struct nfp_net *nn = netdev_priv(netdev); 3491 struct sockaddr *saddr = addr; 3492 int err; 3493 3494 err = eth_prepare_mac_addr_change(netdev, addr); 3495 if (err) 3496 return err; 3497 3498 nfp_net_write_mac_addr(nn, saddr->sa_data); 3499 3500 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MACADDR); 3501 if (err) 3502 return err; 3503 3504 eth_commit_mac_addr_change(netdev, addr); 3505 3506 return 0; 3507 } 3508 3509 const struct net_device_ops nfp_net_netdev_ops = { 3510 .ndo_init = nfp_app_ndo_init, 3511 .ndo_uninit = nfp_app_ndo_uninit, 3512 .ndo_open = nfp_net_netdev_open, 3513 .ndo_stop = nfp_net_netdev_close, 3514 .ndo_start_xmit = nfp_net_tx, 3515 .ndo_get_stats64 = nfp_net_stat64, 3516 .ndo_vlan_rx_add_vid = nfp_net_vlan_rx_add_vid, 3517 .ndo_vlan_rx_kill_vid = nfp_net_vlan_rx_kill_vid, 3518 .ndo_set_vf_mac = nfp_app_set_vf_mac, 3519 .ndo_set_vf_vlan = nfp_app_set_vf_vlan, 3520 .ndo_set_vf_spoofchk = nfp_app_set_vf_spoofchk, 3521 .ndo_get_vf_config = nfp_app_get_vf_config, 3522 .ndo_set_vf_link_state = nfp_app_set_vf_link_state, 3523 .ndo_setup_tc = nfp_port_setup_tc, 3524 .ndo_tx_timeout = nfp_net_tx_timeout, 3525 .ndo_set_rx_mode = nfp_net_set_rx_mode, 3526 .ndo_change_mtu = nfp_net_change_mtu, 3527 .ndo_set_mac_address = nfp_net_set_mac_address, 3528 .ndo_set_features = nfp_net_set_features, 3529 .ndo_features_check = nfp_net_features_check, 3530 .ndo_get_phys_port_name = nfp_net_get_phys_port_name, 3531 .ndo_udp_tunnel_add = nfp_net_add_vxlan_port, 3532 .ndo_udp_tunnel_del = nfp_net_del_vxlan_port, 3533 .ndo_bpf = nfp_net_xdp, 3534 }; 3535 3536 /** 3537 * nfp_net_info() - Print general info about the NIC 3538 * @nn: NFP Net device to reconfigure 3539 */ 3540 void nfp_net_info(struct nfp_net *nn) 3541 { 3542 nn_info(nn, "Netronome NFP-6xxx %sNetdev: TxQs=%d/%d RxQs=%d/%d\n", 3543 nn->dp.is_vf ? "VF " : "", 3544 nn->dp.num_tx_rings, nn->max_tx_rings, 3545 nn->dp.num_rx_rings, nn->max_rx_rings); 3546 nn_info(nn, "VER: %d.%d.%d.%d, Maximum supported MTU: %d\n", 3547 nn->fw_ver.resv, nn->fw_ver.class, 3548 nn->fw_ver.major, nn->fw_ver.minor, 3549 nn->max_mtu); 3550 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", 3551 nn->cap, 3552 nn->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "", 3553 nn->cap & NFP_NET_CFG_CTRL_L2BC ? "L2BCFILT " : "", 3554 nn->cap & NFP_NET_CFG_CTRL_L2MC ? "L2MCFILT " : "", 3555 nn->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "", 3556 nn->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "", 3557 nn->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "", 3558 nn->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "", 3559 nn->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "", 3560 nn->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "", 3561 nn->cap & NFP_NET_CFG_CTRL_LSO ? "TSO1 " : "", 3562 nn->cap & NFP_NET_CFG_CTRL_LSO2 ? "TSO2 " : "", 3563 nn->cap & NFP_NET_CFG_CTRL_RSS ? "RSS1 " : "", 3564 nn->cap & NFP_NET_CFG_CTRL_RSS2 ? "RSS2 " : "", 3565 nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER ? "CTAG_FILTER " : "", 3566 nn->cap & NFP_NET_CFG_CTRL_L2SWITCH ? "L2SWITCH " : "", 3567 nn->cap & NFP_NET_CFG_CTRL_MSIXAUTO ? "AUTOMASK " : "", 3568 nn->cap & NFP_NET_CFG_CTRL_IRQMOD ? "IRQMOD " : "", 3569 nn->cap & NFP_NET_CFG_CTRL_VXLAN ? "VXLAN " : "", 3570 nn->cap & NFP_NET_CFG_CTRL_NVGRE ? "NVGRE " : "", 3571 nn->cap & NFP_NET_CFG_CTRL_CSUM_COMPLETE ? 3572 "RXCSUM_COMPLETE " : "", 3573 nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR ? "LIVE_ADDR " : "", 3574 nfp_app_extra_cap(nn->app, nn)); 3575 } 3576 3577 /** 3578 * nfp_net_alloc() - Allocate netdev and related structure 3579 * @pdev: PCI device 3580 * @ctrl_bar: PCI IOMEM with vNIC config memory 3581 * @needs_netdev: Whether to allocate a netdev for this vNIC 3582 * @max_tx_rings: Maximum number of TX rings supported by device 3583 * @max_rx_rings: Maximum number of RX rings supported by device 3584 * 3585 * This function allocates a netdev device and fills in the initial 3586 * part of the @struct nfp_net structure. In case of control device 3587 * nfp_net structure is allocated without the netdev. 3588 * 3589 * Return: NFP Net device structure, or ERR_PTR on error. 3590 */ 3591 struct nfp_net * 3592 nfp_net_alloc(struct pci_dev *pdev, void __iomem *ctrl_bar, bool needs_netdev, 3593 unsigned int max_tx_rings, unsigned int max_rx_rings) 3594 { 3595 struct nfp_net *nn; 3596 int err; 3597 3598 if (needs_netdev) { 3599 struct net_device *netdev; 3600 3601 netdev = alloc_etherdev_mqs(sizeof(struct nfp_net), 3602 max_tx_rings, max_rx_rings); 3603 if (!netdev) 3604 return ERR_PTR(-ENOMEM); 3605 3606 SET_NETDEV_DEV(netdev, &pdev->dev); 3607 nn = netdev_priv(netdev); 3608 nn->dp.netdev = netdev; 3609 } else { 3610 nn = vzalloc(sizeof(*nn)); 3611 if (!nn) 3612 return ERR_PTR(-ENOMEM); 3613 } 3614 3615 nn->dp.dev = &pdev->dev; 3616 nn->dp.ctrl_bar = ctrl_bar; 3617 nn->pdev = pdev; 3618 3619 nn->max_tx_rings = max_tx_rings; 3620 nn->max_rx_rings = max_rx_rings; 3621 3622 nn->dp.num_tx_rings = min_t(unsigned int, 3623 max_tx_rings, num_online_cpus()); 3624 nn->dp.num_rx_rings = min_t(unsigned int, max_rx_rings, 3625 netif_get_num_default_rss_queues()); 3626 3627 nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings); 3628 nn->dp.num_r_vecs = min_t(unsigned int, 3629 nn->dp.num_r_vecs, num_online_cpus()); 3630 3631 nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT; 3632 nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT; 3633 3634 spin_lock_init(&nn->reconfig_lock); 3635 spin_lock_init(&nn->link_status_lock); 3636 3637 timer_setup(&nn->reconfig_timer, nfp_net_reconfig_timer, 0); 3638 3639 err = nfp_net_tlv_caps_parse(&nn->pdev->dev, nn->dp.ctrl_bar, 3640 &nn->tlv_caps); 3641 if (err) 3642 goto err_free_nn; 3643 3644 return nn; 3645 3646 err_free_nn: 3647 if (nn->dp.netdev) 3648 free_netdev(nn->dp.netdev); 3649 else 3650 vfree(nn); 3651 return ERR_PTR(err); 3652 } 3653 3654 /** 3655 * nfp_net_free() - Undo what @nfp_net_alloc() did 3656 * @nn: NFP Net device to reconfigure 3657 */ 3658 void nfp_net_free(struct nfp_net *nn) 3659 { 3660 WARN_ON(timer_pending(&nn->reconfig_timer) || nn->reconfig_posted); 3661 if (nn->dp.netdev) 3662 free_netdev(nn->dp.netdev); 3663 else 3664 vfree(nn); 3665 } 3666 3667 /** 3668 * nfp_net_rss_key_sz() - Get current size of the RSS key 3669 * @nn: NFP Net device instance 3670 * 3671 * Return: size of the RSS key for currently selected hash function. 3672 */ 3673 unsigned int nfp_net_rss_key_sz(struct nfp_net *nn) 3674 { 3675 switch (nn->rss_hfunc) { 3676 case ETH_RSS_HASH_TOP: 3677 return NFP_NET_CFG_RSS_KEY_SZ; 3678 case ETH_RSS_HASH_XOR: 3679 return 0; 3680 case ETH_RSS_HASH_CRC32: 3681 return 4; 3682 } 3683 3684 nn_warn(nn, "Unknown hash function: %u\n", nn->rss_hfunc); 3685 return 0; 3686 } 3687 3688 /** 3689 * nfp_net_rss_init() - Set the initial RSS parameters 3690 * @nn: NFP Net device to reconfigure 3691 */ 3692 static void nfp_net_rss_init(struct nfp_net *nn) 3693 { 3694 unsigned long func_bit, rss_cap_hfunc; 3695 u32 reg; 3696 3697 /* Read the RSS function capability and select first supported func */ 3698 reg = nn_readl(nn, NFP_NET_CFG_RSS_CAP); 3699 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, reg); 3700 if (!rss_cap_hfunc) 3701 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, 3702 NFP_NET_CFG_RSS_TOEPLITZ); 3703 3704 func_bit = find_first_bit(&rss_cap_hfunc, NFP_NET_CFG_RSS_HFUNCS); 3705 if (func_bit == NFP_NET_CFG_RSS_HFUNCS) { 3706 dev_warn(nn->dp.dev, 3707 "Bad RSS config, defaulting to Toeplitz hash\n"); 3708 func_bit = ETH_RSS_HASH_TOP_BIT; 3709 } 3710 nn->rss_hfunc = 1 << func_bit; 3711 3712 netdev_rss_key_fill(nn->rss_key, nfp_net_rss_key_sz(nn)); 3713 3714 nfp_net_rss_init_itbl(nn); 3715 3716 /* Enable IPv4/IPv6 TCP by default */ 3717 nn->rss_cfg = NFP_NET_CFG_RSS_IPV4_TCP | 3718 NFP_NET_CFG_RSS_IPV6_TCP | 3719 FIELD_PREP(NFP_NET_CFG_RSS_HFUNC, nn->rss_hfunc) | 3720 NFP_NET_CFG_RSS_MASK; 3721 } 3722 3723 /** 3724 * nfp_net_irqmod_init() - Set the initial IRQ moderation parameters 3725 * @nn: NFP Net device to reconfigure 3726 */ 3727 static void nfp_net_irqmod_init(struct nfp_net *nn) 3728 { 3729 nn->rx_coalesce_usecs = 50; 3730 nn->rx_coalesce_max_frames = 64; 3731 nn->tx_coalesce_usecs = 50; 3732 nn->tx_coalesce_max_frames = 64; 3733 } 3734 3735 static void nfp_net_netdev_init(struct nfp_net *nn) 3736 { 3737 struct net_device *netdev = nn->dp.netdev; 3738 3739 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr); 3740 3741 netdev->mtu = nn->dp.mtu; 3742 3743 /* Advertise/enable offloads based on capabilities 3744 * 3745 * Note: netdev->features show the currently enabled features 3746 * and netdev->hw_features advertises which features are 3747 * supported. By default we enable most features. 3748 */ 3749 if (nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR) 3750 netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE; 3751 3752 netdev->hw_features = NETIF_F_HIGHDMA; 3753 if (nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY) { 3754 netdev->hw_features |= NETIF_F_RXCSUM; 3755 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY; 3756 } 3757 if (nn->cap & NFP_NET_CFG_CTRL_TXCSUM) { 3758 netdev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 3759 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXCSUM; 3760 } 3761 if (nn->cap & NFP_NET_CFG_CTRL_GATHER) { 3762 netdev->hw_features |= NETIF_F_SG; 3763 nn->dp.ctrl |= NFP_NET_CFG_CTRL_GATHER; 3764 } 3765 if ((nn->cap & NFP_NET_CFG_CTRL_LSO && nn->fw_ver.major > 2) || 3766 nn->cap & NFP_NET_CFG_CTRL_LSO2) { 3767 netdev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 3768 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?: 3769 NFP_NET_CFG_CTRL_LSO; 3770 } 3771 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) 3772 netdev->hw_features |= NETIF_F_RXHASH; 3773 if (nn->cap & NFP_NET_CFG_CTRL_VXLAN) { 3774 if (nn->cap & NFP_NET_CFG_CTRL_LSO) 3775 netdev->hw_features |= NETIF_F_GSO_UDP_TUNNEL; 3776 nn->dp.ctrl |= NFP_NET_CFG_CTRL_VXLAN; 3777 } 3778 if (nn->cap & NFP_NET_CFG_CTRL_NVGRE) { 3779 if (nn->cap & NFP_NET_CFG_CTRL_LSO) 3780 netdev->hw_features |= NETIF_F_GSO_GRE; 3781 nn->dp.ctrl |= NFP_NET_CFG_CTRL_NVGRE; 3782 } 3783 if (nn->cap & (NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE)) 3784 netdev->hw_enc_features = netdev->hw_features; 3785 3786 netdev->vlan_features = netdev->hw_features; 3787 3788 if (nn->cap & NFP_NET_CFG_CTRL_RXVLAN) { 3789 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX; 3790 nn->dp.ctrl |= NFP_NET_CFG_CTRL_RXVLAN; 3791 } 3792 if (nn->cap & NFP_NET_CFG_CTRL_TXVLAN) { 3793 if (nn->cap & NFP_NET_CFG_CTRL_LSO2) { 3794 nn_warn(nn, "Device advertises both TSO2 and TXVLAN. Refusing to enable TXVLAN.\n"); 3795 } else { 3796 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX; 3797 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXVLAN; 3798 } 3799 } 3800 if (nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER) { 3801 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER; 3802 nn->dp.ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER; 3803 } 3804 3805 netdev->features = netdev->hw_features; 3806 3807 if (nfp_app_has_tc(nn->app) && nn->port) 3808 netdev->hw_features |= NETIF_F_HW_TC; 3809 3810 /* Advertise but disable TSO by default. */ 3811 netdev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6); 3812 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY; 3813 3814 /* Finalise the netdev setup */ 3815 netdev->netdev_ops = &nfp_net_netdev_ops; 3816 netdev->watchdog_timeo = msecs_to_jiffies(5 * 1000); 3817 3818 SWITCHDEV_SET_OPS(netdev, &nfp_port_switchdev_ops); 3819 3820 /* MTU range: 68 - hw-specific max */ 3821 netdev->min_mtu = ETH_MIN_MTU; 3822 netdev->max_mtu = nn->max_mtu; 3823 3824 netdev->gso_max_segs = NFP_NET_LSO_MAX_SEGS; 3825 3826 netif_carrier_off(netdev); 3827 3828 nfp_net_set_ethtool_ops(netdev); 3829 } 3830 3831 static int nfp_net_read_caps(struct nfp_net *nn) 3832 { 3833 /* Get some of the read-only fields from the BAR */ 3834 nn->cap = nn_readl(nn, NFP_NET_CFG_CAP); 3835 nn->max_mtu = nn_readl(nn, NFP_NET_CFG_MAX_MTU); 3836 3837 /* ABI 4.x and ctrl vNIC always use chained metadata, in other cases 3838 * we allow use of non-chained metadata if RSS(v1) is the only 3839 * advertised capability requiring metadata. 3840 */ 3841 nn->dp.chained_metadata_format = nn->fw_ver.major == 4 || 3842 !nn->dp.netdev || 3843 !(nn->cap & NFP_NET_CFG_CTRL_RSS) || 3844 nn->cap & NFP_NET_CFG_CTRL_CHAIN_META; 3845 /* RSS(v1) uses non-chained metadata format, except in ABI 4.x where 3846 * it has the same meaning as RSSv2. 3847 */ 3848 if (nn->dp.chained_metadata_format && nn->fw_ver.major != 4) 3849 nn->cap &= ~NFP_NET_CFG_CTRL_RSS; 3850 3851 /* Determine RX packet/metadata boundary offset */ 3852 if (nn->fw_ver.major >= 2) { 3853 u32 reg; 3854 3855 reg = nn_readl(nn, NFP_NET_CFG_RX_OFFSET); 3856 if (reg > NFP_NET_MAX_PREPEND) { 3857 nn_err(nn, "Invalid rx offset: %d\n", reg); 3858 return -EINVAL; 3859 } 3860 nn->dp.rx_offset = reg; 3861 } else { 3862 nn->dp.rx_offset = NFP_NET_RX_OFFSET; 3863 } 3864 3865 /* For control vNICs mask out the capabilities app doesn't want. */ 3866 if (!nn->dp.netdev) 3867 nn->cap &= nn->app->type->ctrl_cap_mask; 3868 3869 return 0; 3870 } 3871 3872 /** 3873 * nfp_net_init() - Initialise/finalise the nfp_net structure 3874 * @nn: NFP Net device structure 3875 * 3876 * Return: 0 on success or negative errno on error. 3877 */ 3878 int nfp_net_init(struct nfp_net *nn) 3879 { 3880 int err; 3881 3882 nn->dp.rx_dma_dir = DMA_FROM_DEVICE; 3883 3884 err = nfp_net_read_caps(nn); 3885 if (err) 3886 return err; 3887 3888 /* Set default MTU and Freelist buffer size */ 3889 if (!nfp_net_is_data_vnic(nn) && nn->app->ctrl_mtu) { 3890 if (nn->app->ctrl_mtu <= nn->max_mtu) { 3891 nn->dp.mtu = nn->app->ctrl_mtu; 3892 } else { 3893 if (nn->app->ctrl_mtu != NFP_APP_CTRL_MTU_MAX) 3894 nn_warn(nn, "app requested MTU above max supported %u > %u\n", 3895 nn->app->ctrl_mtu, nn->max_mtu); 3896 nn->dp.mtu = nn->max_mtu; 3897 } 3898 } else if (nn->max_mtu < NFP_NET_DEFAULT_MTU) { 3899 nn->dp.mtu = nn->max_mtu; 3900 } else { 3901 nn->dp.mtu = NFP_NET_DEFAULT_MTU; 3902 } 3903 nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp); 3904 3905 if (nfp_app_ctrl_uses_data_vnics(nn->app)) 3906 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_CMSG_DATA; 3907 3908 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) { 3909 nfp_net_rss_init(nn); 3910 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?: 3911 NFP_NET_CFG_CTRL_RSS; 3912 } 3913 3914 /* Allow L2 Broadcast and Multicast through by default, if supported */ 3915 if (nn->cap & NFP_NET_CFG_CTRL_L2BC) 3916 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2BC; 3917 3918 /* Allow IRQ moderation, if supported */ 3919 if (nn->cap & NFP_NET_CFG_CTRL_IRQMOD) { 3920 nfp_net_irqmod_init(nn); 3921 nn->dp.ctrl |= NFP_NET_CFG_CTRL_IRQMOD; 3922 } 3923 3924 if (nn->dp.netdev) 3925 nfp_net_netdev_init(nn); 3926 3927 /* Stash the re-configuration queue away. First odd queue in TX Bar */ 3928 nn->qcp_cfg = nn->tx_bar + NFP_QCP_QUEUE_ADDR_SZ; 3929 3930 /* Make sure the FW knows the netdev is supposed to be disabled here */ 3931 nn_writel(nn, NFP_NET_CFG_CTRL, 0); 3932 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0); 3933 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0); 3934 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_RING | 3935 NFP_NET_CFG_UPDATE_GEN); 3936 if (err) 3937 return err; 3938 3939 nfp_net_vecs_init(nn); 3940 3941 if (!nn->dp.netdev) 3942 return 0; 3943 return register_netdev(nn->dp.netdev); 3944 } 3945 3946 /** 3947 * nfp_net_clean() - Undo what nfp_net_init() did. 3948 * @nn: NFP Net device structure 3949 */ 3950 void nfp_net_clean(struct nfp_net *nn) 3951 { 3952 if (!nn->dp.netdev) 3953 return; 3954 3955 unregister_netdev(nn->dp.netdev); 3956 nfp_net_reconfig_wait_posted(nn); 3957 } 3958