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