1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
2 /* Copyright (C) 2015-2018 Netronome Systems, Inc. */
6 * Netronome network device driver: Common functions between PF and VF
7 * Authors: Jakub Kicinski <jakub.kicinski@netronome.com>
8 * Jason McMullan <jason.mcmullan@netronome.com>
9 * Rolf Neugebauer <rolf.neugebauer@netronome.com>
10 * Brad Petrus <brad.petrus@netronome.com>
11 * Chris Telfer <chris.telfer@netronome.com>
14 #include <linux/bitfield.h>
15 #include <linux/bpf.h>
16 #include <linux/bpf_trace.h>
17 #include <linux/module.h>
18 #include <linux/kernel.h>
19 #include <linux/init.h>
21 #include <linux/netdevice.h>
22 #include <linux/etherdevice.h>
23 #include <linux/interrupt.h>
25 #include <linux/ipv6.h>
27 #include <linux/overflow.h>
28 #include <linux/page_ref.h>
29 #include <linux/pci.h>
30 #include <linux/pci_regs.h>
31 #include <linux/msi.h>
32 #include <linux/ethtool.h>
33 #include <linux/log2.h>
34 #include <linux/if_vlan.h>
35 #include <linux/random.h>
36 #include <linux/vmalloc.h>
37 #include <linux/ktime.h>
39 #include <net/switchdev.h>
40 #include <net/vxlan.h>
42 #include "nfpcore/nfp_nsp.h"
44 #include "nfp_net_ctrl.h"
46 #include "nfp_net_sriov.h"
50 * nfp_net_get_fw_version() - Read and parse the FW version
51 * @fw_ver: Output fw_version structure to read to
52 * @ctrl_bar: Mapped address of the control BAR
54 void nfp_net_get_fw_version(struct nfp_net_fw_version *fw_ver,
55 void __iomem *ctrl_bar)
59 reg = readl(ctrl_bar + NFP_NET_CFG_VERSION);
60 put_unaligned_le32(reg, fw_ver);
63 static dma_addr_t nfp_net_dma_map_rx(struct nfp_net_dp *dp, void *frag)
65 return dma_map_single_attrs(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM,
66 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
67 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
71 nfp_net_dma_sync_dev_rx(const struct nfp_net_dp *dp, dma_addr_t dma_addr)
73 dma_sync_single_for_device(dp->dev, dma_addr,
74 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
78 static void nfp_net_dma_unmap_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr)
80 dma_unmap_single_attrs(dp->dev, dma_addr,
81 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
82 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
85 static void nfp_net_dma_sync_cpu_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr,
88 dma_sync_single_for_cpu(dp->dev, dma_addr - NFP_NET_RX_BUF_HEADROOM,
94 * Firmware reconfig may take a while so we have two versions of it -
95 * synchronous and asynchronous (posted). All synchronous callers are holding
96 * RTNL so we don't have to worry about serializing them.
98 static void nfp_net_reconfig_start(struct nfp_net *nn, u32 update)
100 nn_writel(nn, NFP_NET_CFG_UPDATE, update);
101 /* ensure update is written before pinging HW */
103 nfp_qcp_wr_ptr_add(nn->qcp_cfg, 1);
104 nn->reconfig_in_progress_update = update;
107 /* Pass 0 as update to run posted reconfigs. */
108 static void nfp_net_reconfig_start_async(struct nfp_net *nn, u32 update)
110 update |= nn->reconfig_posted;
111 nn->reconfig_posted = 0;
113 nfp_net_reconfig_start(nn, update);
115 nn->reconfig_timer_active = true;
116 mod_timer(&nn->reconfig_timer, jiffies + NFP_NET_POLL_TIMEOUT * HZ);
119 static bool nfp_net_reconfig_check_done(struct nfp_net *nn, bool last_check)
123 reg = nn_readl(nn, NFP_NET_CFG_UPDATE);
126 if (reg & NFP_NET_CFG_UPDATE_ERR) {
127 nn_err(nn, "Reconfig error (status: 0x%08x update: 0x%08x ctrl: 0x%08x)\n",
128 reg, nn->reconfig_in_progress_update,
129 nn_readl(nn, NFP_NET_CFG_CTRL));
131 } else if (last_check) {
132 nn_err(nn, "Reconfig timeout (status: 0x%08x update: 0x%08x ctrl: 0x%08x)\n",
133 reg, nn->reconfig_in_progress_update,
134 nn_readl(nn, NFP_NET_CFG_CTRL));
141 static int nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline)
143 bool timed_out = false;
145 /* Poll update field, waiting for NFP to ack the config */
146 while (!nfp_net_reconfig_check_done(nn, timed_out)) {
148 timed_out = time_is_before_eq_jiffies(deadline);
151 if (nn_readl(nn, NFP_NET_CFG_UPDATE) & NFP_NET_CFG_UPDATE_ERR)
154 return timed_out ? -EIO : 0;
157 static void nfp_net_reconfig_timer(struct timer_list *t)
159 struct nfp_net *nn = from_timer(nn, t, reconfig_timer);
161 spin_lock_bh(&nn->reconfig_lock);
163 nn->reconfig_timer_active = false;
165 /* If sync caller is present it will take over from us */
166 if (nn->reconfig_sync_present)
169 /* Read reconfig status and report errors */
170 nfp_net_reconfig_check_done(nn, true);
172 if (nn->reconfig_posted)
173 nfp_net_reconfig_start_async(nn, 0);
175 spin_unlock_bh(&nn->reconfig_lock);
179 * nfp_net_reconfig_post() - Post async reconfig request
180 * @nn: NFP Net device to reconfigure
181 * @update: The value for the update field in the BAR config
183 * Record FW reconfiguration request. Reconfiguration will be kicked off
184 * whenever reconfiguration machinery is idle. Multiple requests can be
187 static void nfp_net_reconfig_post(struct nfp_net *nn, u32 update)
189 spin_lock_bh(&nn->reconfig_lock);
191 /* Sync caller will kick off async reconf when it's done, just post */
192 if (nn->reconfig_sync_present) {
193 nn->reconfig_posted |= update;
197 /* Opportunistically check if the previous command is done */
198 if (!nn->reconfig_timer_active ||
199 nfp_net_reconfig_check_done(nn, false))
200 nfp_net_reconfig_start_async(nn, update);
202 nn->reconfig_posted |= update;
204 spin_unlock_bh(&nn->reconfig_lock);
207 static void nfp_net_reconfig_sync_enter(struct nfp_net *nn)
209 bool cancelled_timer = false;
210 u32 pre_posted_requests;
212 spin_lock_bh(&nn->reconfig_lock);
214 nn->reconfig_sync_present = true;
216 if (nn->reconfig_timer_active) {
217 nn->reconfig_timer_active = false;
218 cancelled_timer = true;
220 pre_posted_requests = nn->reconfig_posted;
221 nn->reconfig_posted = 0;
223 spin_unlock_bh(&nn->reconfig_lock);
225 if (cancelled_timer) {
226 del_timer_sync(&nn->reconfig_timer);
227 nfp_net_reconfig_wait(nn, nn->reconfig_timer.expires);
230 /* Run the posted reconfigs which were issued before we started */
231 if (pre_posted_requests) {
232 nfp_net_reconfig_start(nn, pre_posted_requests);
233 nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT);
237 static void nfp_net_reconfig_wait_posted(struct nfp_net *nn)
239 nfp_net_reconfig_sync_enter(nn);
241 spin_lock_bh(&nn->reconfig_lock);
242 nn->reconfig_sync_present = false;
243 spin_unlock_bh(&nn->reconfig_lock);
247 * nfp_net_reconfig() - Reconfigure the firmware
248 * @nn: NFP Net device to reconfigure
249 * @update: The value for the update field in the BAR config
251 * Write the update word to the BAR and ping the reconfig queue. The
252 * poll until the firmware has acknowledged the update by zeroing the
255 * Return: Negative errno on error, 0 on success
257 int nfp_net_reconfig(struct nfp_net *nn, u32 update)
261 nfp_net_reconfig_sync_enter(nn);
263 nfp_net_reconfig_start(nn, update);
264 ret = nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT);
266 spin_lock_bh(&nn->reconfig_lock);
268 if (nn->reconfig_posted)
269 nfp_net_reconfig_start_async(nn, 0);
271 nn->reconfig_sync_present = false;
273 spin_unlock_bh(&nn->reconfig_lock);
279 * nfp_net_reconfig_mbox() - Reconfigure the firmware via the mailbox
280 * @nn: NFP Net device to reconfigure
281 * @mbox_cmd: The value for the mailbox command
283 * Helper function for mailbox updates
285 * Return: Negative errno on error, 0 on success
287 int nfp_net_reconfig_mbox(struct nfp_net *nn, u32 mbox_cmd)
289 u32 mbox = nn->tlv_caps.mbox_off;
292 if (!nfp_net_has_mbox(&nn->tlv_caps)) {
293 nn_err(nn, "no mailbox present, command: %u\n", mbox_cmd);
297 nn_writeq(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_CMD, mbox_cmd);
299 ret = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MBOX);
301 nn_err(nn, "Mailbox update error\n");
305 return -nn_readl(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_RET);
308 /* Interrupt configuration and handling
312 * nfp_net_irq_unmask() - Unmask automasked interrupt
313 * @nn: NFP Network structure
314 * @entry_nr: MSI-X table entry
316 * Clear the ICR for the IRQ entry.
318 static void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr)
320 nn_writeb(nn, NFP_NET_CFG_ICR(entry_nr), NFP_NET_CFG_ICR_UNMASKED);
325 * nfp_net_irqs_alloc() - allocates MSI-X irqs
326 * @pdev: PCI device structure
327 * @irq_entries: Array to be initialized and used to hold the irq entries
328 * @min_irqs: Minimal acceptable number of interrupts
329 * @wanted_irqs: Target number of interrupts to allocate
331 * Return: Number of irqs obtained or 0 on error.
334 nfp_net_irqs_alloc(struct pci_dev *pdev, struct msix_entry *irq_entries,
335 unsigned int min_irqs, unsigned int wanted_irqs)
340 for (i = 0; i < wanted_irqs; i++)
341 irq_entries[i].entry = i;
343 got_irqs = pci_enable_msix_range(pdev, irq_entries,
344 min_irqs, wanted_irqs);
346 dev_err(&pdev->dev, "Failed to enable %d-%d MSI-X (err=%d)\n",
347 min_irqs, wanted_irqs, got_irqs);
351 if (got_irqs < wanted_irqs)
352 dev_warn(&pdev->dev, "Unable to allocate %d IRQs got only %d\n",
353 wanted_irqs, got_irqs);
359 * nfp_net_irqs_assign() - Assign interrupts allocated externally to netdev
360 * @nn: NFP Network structure
361 * @irq_entries: Table of allocated interrupts
362 * @n: Size of @irq_entries (number of entries to grab)
364 * After interrupts are allocated with nfp_net_irqs_alloc() this function
365 * should be called to assign them to a specific netdev (port).
368 nfp_net_irqs_assign(struct nfp_net *nn, struct msix_entry *irq_entries,
371 struct nfp_net_dp *dp = &nn->dp;
373 nn->max_r_vecs = n - NFP_NET_NON_Q_VECTORS;
374 dp->num_r_vecs = nn->max_r_vecs;
376 memcpy(nn->irq_entries, irq_entries, sizeof(*irq_entries) * n);
378 if (dp->num_rx_rings > dp->num_r_vecs ||
379 dp->num_tx_rings > dp->num_r_vecs)
380 dev_warn(nn->dp.dev, "More rings (%d,%d) than vectors (%d).\n",
381 dp->num_rx_rings, dp->num_tx_rings,
384 dp->num_rx_rings = min(dp->num_r_vecs, dp->num_rx_rings);
385 dp->num_tx_rings = min(dp->num_r_vecs, dp->num_tx_rings);
386 dp->num_stack_tx_rings = dp->num_tx_rings;
390 * nfp_net_irqs_disable() - Disable interrupts
391 * @pdev: PCI device structure
393 * Undoes what @nfp_net_irqs_alloc() does.
395 void nfp_net_irqs_disable(struct pci_dev *pdev)
397 pci_disable_msix(pdev);
401 * nfp_net_irq_rxtx() - Interrupt service routine for RX/TX rings.
403 * @data: Opaque data structure
405 * Return: Indicate if the interrupt has been handled.
407 static irqreturn_t nfp_net_irq_rxtx(int irq, void *data)
409 struct nfp_net_r_vector *r_vec = data;
411 napi_schedule_irqoff(&r_vec->napi);
413 /* The FW auto-masks any interrupt, either via the MASK bit in
414 * the MSI-X table or via the per entry ICR field. So there
415 * is no need to disable interrupts here.
420 static irqreturn_t nfp_ctrl_irq_rxtx(int irq, void *data)
422 struct nfp_net_r_vector *r_vec = data;
424 tasklet_schedule(&r_vec->tasklet);
430 * nfp_net_read_link_status() - Reread link status from control BAR
431 * @nn: NFP Network structure
433 static void nfp_net_read_link_status(struct nfp_net *nn)
439 spin_lock_irqsave(&nn->link_status_lock, flags);
441 sts = nn_readl(nn, NFP_NET_CFG_STS);
442 link_up = !!(sts & NFP_NET_CFG_STS_LINK);
444 if (nn->link_up == link_up)
447 nn->link_up = link_up;
449 set_bit(NFP_PORT_CHANGED, &nn->port->flags);
452 netif_carrier_on(nn->dp.netdev);
453 netdev_info(nn->dp.netdev, "NIC Link is Up\n");
455 netif_carrier_off(nn->dp.netdev);
456 netdev_info(nn->dp.netdev, "NIC Link is Down\n");
459 spin_unlock_irqrestore(&nn->link_status_lock, flags);
463 * nfp_net_irq_lsc() - Interrupt service routine for link state changes
465 * @data: Opaque data structure
467 * Return: Indicate if the interrupt has been handled.
469 static irqreturn_t nfp_net_irq_lsc(int irq, void *data)
471 struct nfp_net *nn = data;
472 struct msix_entry *entry;
474 entry = &nn->irq_entries[NFP_NET_IRQ_LSC_IDX];
476 nfp_net_read_link_status(nn);
478 nfp_net_irq_unmask(nn, entry->entry);
484 * nfp_net_irq_exn() - Interrupt service routine for exceptions
486 * @data: Opaque data structure
488 * Return: Indicate if the interrupt has been handled.
490 static irqreturn_t nfp_net_irq_exn(int irq, void *data)
492 struct nfp_net *nn = data;
494 nn_err(nn, "%s: UNIMPLEMENTED.\n", __func__);
495 /* XXX TO BE IMPLEMENTED */
500 * nfp_net_tx_ring_init() - Fill in the boilerplate for a TX ring
501 * @tx_ring: TX ring structure
502 * @r_vec: IRQ vector servicing this ring
504 * @is_xdp: Is this an XDP TX ring?
507 nfp_net_tx_ring_init(struct nfp_net_tx_ring *tx_ring,
508 struct nfp_net_r_vector *r_vec, unsigned int idx,
511 struct nfp_net *nn = r_vec->nfp_net;
514 tx_ring->r_vec = r_vec;
515 tx_ring->is_xdp = is_xdp;
516 u64_stats_init(&tx_ring->r_vec->tx_sync);
518 tx_ring->qcidx = tx_ring->idx * nn->stride_tx;
519 tx_ring->qcp_q = nn->tx_bar + NFP_QCP_QUEUE_OFF(tx_ring->qcidx);
523 * nfp_net_rx_ring_init() - Fill in the boilerplate for a RX ring
524 * @rx_ring: RX ring structure
525 * @r_vec: IRQ vector servicing this ring
529 nfp_net_rx_ring_init(struct nfp_net_rx_ring *rx_ring,
530 struct nfp_net_r_vector *r_vec, unsigned int idx)
532 struct nfp_net *nn = r_vec->nfp_net;
535 rx_ring->r_vec = r_vec;
536 u64_stats_init(&rx_ring->r_vec->rx_sync);
538 rx_ring->fl_qcidx = rx_ring->idx * nn->stride_rx;
539 rx_ring->qcp_fl = nn->rx_bar + NFP_QCP_QUEUE_OFF(rx_ring->fl_qcidx);
543 * nfp_net_aux_irq_request() - Request an auxiliary interrupt (LSC or EXN)
544 * @nn: NFP Network structure
545 * @ctrl_offset: Control BAR offset where IRQ configuration should be written
546 * @format: printf-style format to construct the interrupt name
547 * @name: Pointer to allocated space for interrupt name
548 * @name_sz: Size of space for interrupt name
549 * @vector_idx: Index of MSI-X vector used for this interrupt
550 * @handler: IRQ handler to register for this interrupt
553 nfp_net_aux_irq_request(struct nfp_net *nn, u32 ctrl_offset,
554 const char *format, char *name, size_t name_sz,
555 unsigned int vector_idx, irq_handler_t handler)
557 struct msix_entry *entry;
560 entry = &nn->irq_entries[vector_idx];
562 snprintf(name, name_sz, format, nfp_net_name(nn));
563 err = request_irq(entry->vector, handler, 0, name, nn);
565 nn_err(nn, "Failed to request IRQ %d (err=%d).\n",
569 nn_writeb(nn, ctrl_offset, entry->entry);
570 nfp_net_irq_unmask(nn, entry->entry);
576 * nfp_net_aux_irq_free() - Free an auxiliary interrupt (LSC or EXN)
577 * @nn: NFP Network structure
578 * @ctrl_offset: Control BAR offset where IRQ configuration should be written
579 * @vector_idx: Index of MSI-X vector used for this interrupt
581 static void nfp_net_aux_irq_free(struct nfp_net *nn, u32 ctrl_offset,
582 unsigned int vector_idx)
584 nn_writeb(nn, ctrl_offset, 0xff);
586 free_irq(nn->irq_entries[vector_idx].vector, nn);
591 * One queue controller peripheral queue is used for transmit. The
592 * driver en-queues packets for transmit by advancing the write
593 * pointer. The device indicates that packets have transmitted by
594 * advancing the read pointer. The driver maintains a local copy of
595 * the read and write pointer in @struct nfp_net_tx_ring. The driver
596 * keeps @wr_p in sync with the queue controller write pointer and can
597 * determine how many packets have been transmitted by comparing its
598 * copy of the read pointer @rd_p with the read pointer maintained by
599 * the queue controller peripheral.
603 * nfp_net_tx_full() - Check if the TX ring is full
604 * @tx_ring: TX ring to check
605 * @dcnt: Number of descriptors that need to be enqueued (must be >= 1)
607 * This function checks, based on the *host copy* of read/write
608 * pointer if a given TX ring is full. The real TX queue may have
609 * some newly made available slots.
611 * Return: True if the ring is full.
613 static int nfp_net_tx_full(struct nfp_net_tx_ring *tx_ring, int dcnt)
615 return (tx_ring->wr_p - tx_ring->rd_p) >= (tx_ring->cnt - dcnt);
618 /* Wrappers for deciding when to stop and restart TX queues */
619 static int nfp_net_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
621 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
624 static int nfp_net_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
626 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
630 * nfp_net_tx_ring_stop() - stop tx ring
631 * @nd_q: netdev queue
632 * @tx_ring: driver tx queue structure
634 * Safely stop TX ring. Remember that while we are running .start_xmit()
635 * someone else may be cleaning the TX ring completions so we need to be
636 * extra careful here.
638 static void nfp_net_tx_ring_stop(struct netdev_queue *nd_q,
639 struct nfp_net_tx_ring *tx_ring)
641 netif_tx_stop_queue(nd_q);
643 /* We can race with the TX completion out of NAPI so recheck */
645 if (unlikely(nfp_net_tx_ring_should_wake(tx_ring)))
646 netif_tx_start_queue(nd_q);
650 * nfp_net_tx_tso() - Set up Tx descriptor for LSO
651 * @r_vec: per-ring structure
652 * @txbuf: Pointer to driver soft TX descriptor
653 * @txd: Pointer to HW TX descriptor
654 * @skb: Pointer to SKB
655 * @md_bytes: Prepend length
657 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
658 * Return error on packet header greater than maximum supported LSO header size.
660 static void nfp_net_tx_tso(struct nfp_net_r_vector *r_vec,
661 struct nfp_net_tx_buf *txbuf,
662 struct nfp_net_tx_desc *txd, struct sk_buff *skb,
665 u32 l3_offset, l4_offset, hdrlen;
668 if (!skb_is_gso(skb))
671 if (!skb->encapsulation) {
672 l3_offset = skb_network_offset(skb);
673 l4_offset = skb_transport_offset(skb);
674 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
676 l3_offset = skb_inner_network_offset(skb);
677 l4_offset = skb_inner_transport_offset(skb);
678 hdrlen = skb_inner_transport_header(skb) - skb->data +
679 inner_tcp_hdrlen(skb);
682 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
683 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
685 mss = skb_shinfo(skb)->gso_size & PCIE_DESC_TX_MSS_MASK;
686 txd->l3_offset = l3_offset - md_bytes;
687 txd->l4_offset = l4_offset - md_bytes;
688 txd->lso_hdrlen = hdrlen - md_bytes;
689 txd->mss = cpu_to_le16(mss);
690 txd->flags |= PCIE_DESC_TX_LSO;
692 u64_stats_update_begin(&r_vec->tx_sync);
694 u64_stats_update_end(&r_vec->tx_sync);
698 * nfp_net_tx_csum() - Set TX CSUM offload flags in TX descriptor
699 * @dp: NFP Net data path struct
700 * @r_vec: per-ring structure
701 * @txbuf: Pointer to driver soft TX descriptor
702 * @txd: Pointer to TX descriptor
703 * @skb: Pointer to SKB
705 * This function sets the TX checksum flags in the TX descriptor based
706 * on the configuration and the protocol of the packet to be transmitted.
708 static void nfp_net_tx_csum(struct nfp_net_dp *dp,
709 struct nfp_net_r_vector *r_vec,
710 struct nfp_net_tx_buf *txbuf,
711 struct nfp_net_tx_desc *txd, struct sk_buff *skb)
713 struct ipv6hdr *ipv6h;
717 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
720 if (skb->ip_summed != CHECKSUM_PARTIAL)
723 txd->flags |= PCIE_DESC_TX_CSUM;
724 if (skb->encapsulation)
725 txd->flags |= PCIE_DESC_TX_ENCAP;
727 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
728 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
730 if (iph->version == 4) {
731 txd->flags |= PCIE_DESC_TX_IP4_CSUM;
732 l4_hdr = iph->protocol;
733 } else if (ipv6h->version == 6) {
734 l4_hdr = ipv6h->nexthdr;
736 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
742 txd->flags |= PCIE_DESC_TX_TCP_CSUM;
745 txd->flags |= PCIE_DESC_TX_UDP_CSUM;
748 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
752 u64_stats_update_begin(&r_vec->tx_sync);
753 if (skb->encapsulation)
754 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
756 r_vec->hw_csum_tx += txbuf->pkt_cnt;
757 u64_stats_update_end(&r_vec->tx_sync);
760 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring)
763 nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add);
764 tx_ring->wr_ptr_add = 0;
767 static int nfp_net_prep_port_id(struct sk_buff *skb)
769 struct metadata_dst *md_dst = skb_metadata_dst(skb);
774 if (unlikely(md_dst->type != METADATA_HW_PORT_MUX))
777 if (unlikely(skb_cow_head(skb, 8)))
780 data = skb_push(skb, 8);
781 put_unaligned_be32(NFP_NET_META_PORTID, data);
782 put_unaligned_be32(md_dst->u.port_info.port_id, data + 4);
788 * nfp_net_tx() - Main transmit entry point
789 * @skb: SKB to transmit
790 * @netdev: netdev structure
792 * Return: NETDEV_TX_OK on success.
794 static int nfp_net_tx(struct sk_buff *skb, struct net_device *netdev)
796 struct nfp_net *nn = netdev_priv(netdev);
797 const struct skb_frag_struct *frag;
798 int f, nr_frags, wr_idx, md_bytes;
799 struct nfp_net_tx_ring *tx_ring;
800 struct nfp_net_r_vector *r_vec;
801 struct nfp_net_tx_buf *txbuf;
802 struct nfp_net_tx_desc *txd;
803 struct netdev_queue *nd_q;
804 struct nfp_net_dp *dp;
810 qidx = skb_get_queue_mapping(skb);
811 tx_ring = &dp->tx_rings[qidx];
812 r_vec = tx_ring->r_vec;
814 nr_frags = skb_shinfo(skb)->nr_frags;
816 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
817 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
818 qidx, tx_ring->wr_p, tx_ring->rd_p);
819 nd_q = netdev_get_tx_queue(dp->netdev, qidx);
820 netif_tx_stop_queue(nd_q);
821 nfp_net_tx_xmit_more_flush(tx_ring);
822 u64_stats_update_begin(&r_vec->tx_sync);
824 u64_stats_update_end(&r_vec->tx_sync);
825 return NETDEV_TX_BUSY;
828 md_bytes = nfp_net_prep_port_id(skb);
829 if (unlikely(md_bytes < 0)) {
830 nfp_net_tx_xmit_more_flush(tx_ring);
831 dev_kfree_skb_any(skb);
835 /* Start with the head skbuf */
836 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
838 if (dma_mapping_error(dp->dev, dma_addr))
841 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
843 /* Stash the soft descriptor of the head then initialize it */
844 txbuf = &tx_ring->txbufs[wr_idx];
846 txbuf->dma_addr = dma_addr;
849 txbuf->real_len = skb->len;
851 /* Build TX descriptor */
852 txd = &tx_ring->txds[wr_idx];
853 txd->offset_eop = (nr_frags ? 0 : PCIE_DESC_TX_EOP) | md_bytes;
854 txd->dma_len = cpu_to_le16(skb_headlen(skb));
855 nfp_desc_set_dma_addr(txd, dma_addr);
856 txd->data_len = cpu_to_le16(skb->len);
862 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
863 nfp_net_tx_tso(r_vec, txbuf, txd, skb, md_bytes);
864 nfp_net_tx_csum(dp, r_vec, txbuf, txd, skb);
865 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
866 txd->flags |= PCIE_DESC_TX_VLAN;
867 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
874 /* all descs must match except for in addr, length and eop */
875 second_half = txd->vals8[1];
877 for (f = 0; f < nr_frags; f++) {
878 frag = &skb_shinfo(skb)->frags[f];
879 fsize = skb_frag_size(frag);
881 dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
882 fsize, DMA_TO_DEVICE);
883 if (dma_mapping_error(dp->dev, dma_addr))
886 wr_idx = D_IDX(tx_ring, wr_idx + 1);
887 tx_ring->txbufs[wr_idx].skb = skb;
888 tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
889 tx_ring->txbufs[wr_idx].fidx = f;
891 txd = &tx_ring->txds[wr_idx];
892 txd->dma_len = cpu_to_le16(fsize);
893 nfp_desc_set_dma_addr(txd, dma_addr);
894 txd->offset_eop = md_bytes |
895 ((f == nr_frags - 1) ? PCIE_DESC_TX_EOP : 0);
896 txd->vals8[1] = second_half;
899 u64_stats_update_begin(&r_vec->tx_sync);
901 u64_stats_update_end(&r_vec->tx_sync);
904 skb_tx_timestamp(skb);
906 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
908 tx_ring->wr_p += nr_frags + 1;
909 if (nfp_net_tx_ring_should_stop(tx_ring))
910 nfp_net_tx_ring_stop(nd_q, tx_ring);
912 tx_ring->wr_ptr_add += nr_frags + 1;
913 if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, skb->xmit_more))
914 nfp_net_tx_xmit_more_flush(tx_ring);
920 frag = &skb_shinfo(skb)->frags[f];
921 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
922 skb_frag_size(frag), DMA_TO_DEVICE);
923 tx_ring->txbufs[wr_idx].skb = NULL;
924 tx_ring->txbufs[wr_idx].dma_addr = 0;
925 tx_ring->txbufs[wr_idx].fidx = -2;
928 wr_idx += tx_ring->cnt;
930 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
931 skb_headlen(skb), DMA_TO_DEVICE);
932 tx_ring->txbufs[wr_idx].skb = NULL;
933 tx_ring->txbufs[wr_idx].dma_addr = 0;
934 tx_ring->txbufs[wr_idx].fidx = -2;
936 nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
937 nfp_net_tx_xmit_more_flush(tx_ring);
938 u64_stats_update_begin(&r_vec->tx_sync);
940 u64_stats_update_end(&r_vec->tx_sync);
941 dev_kfree_skb_any(skb);
946 * nfp_net_tx_complete() - Handled completed TX packets
947 * @tx_ring: TX ring structure
948 * @budget: NAPI budget (only used as bool to determine if in NAPI context)
950 static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
952 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
953 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
954 struct netdev_queue *nd_q;
955 u32 done_pkts = 0, done_bytes = 0;
959 if (tx_ring->wr_p == tx_ring->rd_p)
962 /* Work out how many descriptors have been transmitted */
963 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
965 if (qcp_rd_p == tx_ring->qcp_rd_p)
968 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
971 const struct skb_frag_struct *frag;
972 struct nfp_net_tx_buf *tx_buf;
977 idx = D_IDX(tx_ring, tx_ring->rd_p++);
978 tx_buf = &tx_ring->txbufs[idx];
984 nr_frags = skb_shinfo(skb)->nr_frags;
989 dma_unmap_single(dp->dev, tx_buf->dma_addr,
990 skb_headlen(skb), DMA_TO_DEVICE);
992 done_pkts += tx_buf->pkt_cnt;
993 done_bytes += tx_buf->real_len;
996 frag = &skb_shinfo(skb)->frags[fidx];
997 dma_unmap_page(dp->dev, tx_buf->dma_addr,
998 skb_frag_size(frag), DMA_TO_DEVICE);
1001 /* check for last gather fragment */
1002 if (fidx == nr_frags - 1)
1003 napi_consume_skb(skb, budget);
1005 tx_buf->dma_addr = 0;
1010 tx_ring->qcp_rd_p = qcp_rd_p;
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);
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 */
1026 if (unlikely(netif_tx_queue_stopped(nd_q)))
1027 netif_tx_wake_queue(nd_q);
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);
1035 static bool nfp_net_xdp_complete(struct nfp_net_tx_ring *tx_ring)
1037 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
1038 u32 done_pkts = 0, done_bytes = 0;
1043 /* Work out how many descriptors have been transmitted */
1044 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
1046 if (qcp_rd_p == tx_ring->qcp_rd_p)
1049 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
1051 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
1052 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
1054 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
1058 idx = D_IDX(tx_ring, tx_ring->rd_p);
1061 done_bytes += tx_ring->txbufs[idx].real_len;
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);
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);
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
1081 * Assumes that the device is stopped, must be idempotent.
1084 nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
1086 const struct skb_frag_struct *frag;
1087 struct netdev_queue *nd_q;
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;
1094 idx = D_IDX(tx_ring, tx_ring->rd_p);
1095 tx_buf = &tx_ring->txbufs[idx];
1097 skb = tx_ring->txbufs[idx].skb;
1098 nr_frags = skb_shinfo(skb)->nr_frags;
1100 if (tx_buf->fidx == -1) {
1102 dma_unmap_single(dp->dev, tx_buf->dma_addr,
1103 skb_headlen(skb), DMA_TO_DEVICE);
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);
1111 /* check for last gather fragment */
1112 if (tx_buf->fidx == nr_frags - 1)
1113 dev_kfree_skb_any(skb);
1115 tx_buf->dma_addr = 0;
1119 tx_ring->qcp_rd_p++;
1123 memset(tx_ring->txds, 0, tx_ring->size);
1126 tx_ring->qcp_rd_p = 0;
1127 tx_ring->wr_ptr_add = 0;
1129 if (tx_ring->is_xdp || !dp->netdev)
1132 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
1133 netdev_tx_reset_queue(nd_q);
1136 static void nfp_net_tx_timeout(struct net_device *netdev)
1138 struct nfp_net *nn = netdev_priv(netdev);
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)))
1144 nn_warn(nn, "TX timeout on ring: %d\n", i);
1146 nn_warn(nn, "TX watchdog timeout\n");
1149 /* Receive processing
1152 nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp)
1154 unsigned int fl_bufsz;
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;
1161 fl_bufsz += dp->rx_offset;
1162 fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu;
1164 fl_bufsz = SKB_DATA_ALIGN(fl_bufsz);
1165 fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1171 nfp_net_free_frag(void *frag, bool xdp)
1174 skb_free_frag(frag);
1176 __free_page(virt_to_page(frag));
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)
1184 * This function will allcate a new page frag, map it for DMA.
1186 * Return: allocated page frag or NULL on failure.
1188 static void *nfp_net_rx_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
1192 if (!dp->xdp_prog) {
1193 frag = netdev_alloc_frag(dp->fl_bufsz);
1197 page = alloc_page(GFP_KERNEL);
1198 frag = page ? page_address(page) : NULL;
1201 nn_dp_warn(dp, "Failed to alloc receive page frag\n");
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");
1215 static void *nfp_net_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
1219 if (!dp->xdp_prog) {
1220 frag = napi_alloc_frag(dp->fl_bufsz);
1221 if (unlikely(!frag))
1226 page = dev_alloc_page();
1227 if (unlikely(!page))
1229 frag = page_address(page);
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");
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
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)
1253 unsigned int wr_idx;
1255 wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
1257 nfp_net_dma_sync_dev_rx(dp, dma_addr);
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;
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);
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.
1275 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
1280 * nfp_net_rx_ring_reset() - Reflect in SW state of freelist after disable
1281 * @rx_ring: RX ring structure
1283 * Assumes that the device is stopped, must be idempotent.
1285 static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring)
1287 unsigned int wr_idx, last_idx;
1289 /* wr_p == rd_p means ring was never fed FL bufs. RX rings are always
1290 * kept at cnt - 1 FL bufs.
1292 if (rx_ring->wr_p == 0 && rx_ring->rd_p == 0)
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;
1303 memset(rx_ring->rxds, 0, rx_ring->size);
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
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.
1318 nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp,
1319 struct nfp_net_rx_ring *rx_ring)
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.
1328 if (!rx_ring->rxbufs[i].frag)
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;
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
1344 nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp,
1345 struct nfp_net_rx_ring *rx_ring)
1347 struct nfp_net_rx_buf *rxbufs;
1350 rxbufs = rx_ring->rxbufs;
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);
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
1369 nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp,
1370 struct nfp_net_rx_ring *rx_ring)
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);
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
1383 static int nfp_net_rx_csum_has_errors(u16 flags)
1385 u16 csum_all_checked, csum_all_ok;
1387 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
1388 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
1390 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
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
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)
1406 skb_checksum_none_assert(skb);
1408 if (!(dp->netdev->features & NETIF_F_RXCSUM))
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);
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);
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.
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);
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);
1449 nfp_net_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
1450 unsigned int type, __be32 *hash)
1452 if (!(netdev->features & NETIF_F_RXHASH))
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;
1462 meta->hash_type = PKT_HASH_TYPE_L4;
1466 meta->hash = get_unaligned_be32(hash);
1470 nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
1471 void *data, struct nfp_net_rx_desc *rxd)
1473 struct nfp_net_rx_hash *rx_hash = data;
1475 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
1478 nfp_net_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
1483 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
1484 void *data, int meta_len)
1488 meta_info = get_unaligned_be32(data);
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,
1500 case NFP_NET_META_MARK:
1501 meta->mark = get_unaligned_be32(data);
1504 case NFP_NET_META_PORTID:
1505 meta->portid = get_unaligned_be32(data);
1508 case NFP_NET_META_CSUM:
1509 meta->csum_type = CHECKSUM_COMPLETE;
1511 (__force __wsum)__get_unaligned_cpu32(data);
1518 meta_info >>= NFP_NET_META_FIELD_SIZE;
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)
1529 u64_stats_update_begin(&r_vec->rx_sync);
1531 /* If we have both skb and rxbuf the replacement buffer allocation
1532 * must have failed, count this as an alloc failure.
1535 r_vec->rx_replace_buf_alloc_fail++;
1536 u64_stats_update_end(&r_vec->rx_sync);
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.
1541 if (skb && rxbuf && skb->head == rxbuf->frag)
1542 page_ref_inc(virt_to_head_page(rxbuf->frag));
1544 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
1546 dev_kfree_skb_any(skb);
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)
1555 struct nfp_net_tx_buf *txbuf;
1556 struct nfp_net_tx_desc *txd;
1559 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1561 nfp_net_xdp_complete(tx_ring);
1565 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1566 nfp_net_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
1572 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1574 /* Stash the soft descriptor of the head then initialize it */
1575 txbuf = &tx_ring->txbufs[wr_idx];
1577 nfp_net_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
1579 txbuf->frag = rxbuf->frag;
1580 txbuf->dma_addr = rxbuf->dma_addr;
1583 txbuf->real_len = pkt_len;
1585 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
1586 pkt_len, DMA_BIDIRECTIONAL);
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);
1597 txd->lso_hdrlen = 0;
1600 tx_ring->wr_ptr_add++;
1605 * nfp_net_rx() - receive up to @budget packets on @rx_ring
1606 * @rx_ring: RX ring to receive from
1607 * @budget: NAPI budget
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.
1613 * Return: Number of packets received.
1615 static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
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;
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;
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;
1644 idx = D_IDX(rx_ring, rx_ring->rd_p);
1646 rxd = &rx_ring->rxds[idx];
1647 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1650 /* Memory barrier to ensure that we won't do other reads
1651 * before the DD bit.
1655 memset(&meta, 0, sizeof(meta));
1660 rxbuf = &rx_ring->rxbufs[idx];
1662 * <-- [rx_offset] -->
1663 * ---------------------------------------------------------
1664 * | [XX] | metadata | packet | XXXX |
1665 * ---------------------------------------------------------
1666 * <---------------- data_len --------------->
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]).
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;
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;
1681 pkt_off += dp->rx_offset;
1682 meta_off = pkt_off - meta_len;
1685 u64_stats_update_begin(&r_vec->rx_sync);
1687 r_vec->rx_bytes += pkt_len;
1688 u64_stats_update_end(&r_vec->rx_sync);
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",
1694 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1698 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
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) {
1707 end = nfp_net_parse_meta(dp->netdev, &meta,
1708 rxbuf->frag + meta_off,
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,
1718 if (xdp_prog && !meta.portid) {
1719 void *orig_data = rxbuf->frag + pkt_off;
1720 unsigned int dma_off;
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;
1728 act = bpf_prog_run_xdp(xdp_prog, &xdp);
1730 pkt_len = xdp.data_end - xdp.data;
1731 pkt_off += xdp.data - orig_data;
1735 meta_len_xdp = xdp.data - xdp.data_meta;
1738 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
1739 if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring,
1744 trace_xdp_exception(dp->netdev,
1748 bpf_warn_invalid_xdp_action(act);
1751 trace_xdp_exception(dp->netdev, xdp_prog, act);
1754 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag,
1760 if (likely(!meta.portid)) {
1761 netdev = dp->netdev;
1762 } else if (meta.portid == NFP_META_PORT_ID_CTRL) {
1763 struct nfp_net *nn = netdev_priv(dp->netdev);
1765 nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
1767 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag,
1773 nn = netdev_priv(dp->netdev);
1774 netdev = nfp_app_repr_get(nn->app, meta.portid);
1775 if (unlikely(!netdev)) {
1776 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf,
1780 nfp_repr_inc_rx_stats(netdev, pkt_len);
1783 skb = build_skb(rxbuf->frag, true_bufsz);
1784 if (unlikely(!skb)) {
1785 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1788 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr);
1789 if (unlikely(!new_frag)) {
1790 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1794 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1796 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1798 skb_reserve(skb, pkt_off);
1799 skb_put(skb, pkt_len);
1801 skb->mark = meta.mark;
1802 skb_set_hash(skb, meta.hash, meta.hash_type);
1804 skb_record_rx_queue(skb, rx_ring->idx);
1805 skb->protocol = eth_type_trans(skb, netdev);
1807 nfp_net_rx_csum(dp, r_vec, rxd, &meta, skb);
1809 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
1810 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
1811 le16_to_cpu(rxd->rxd.vlan));
1813 skb_metadata_set(skb, meta_len_xdp);
1815 napi_gro_receive(&rx_ring->r_vec->napi, skb);
1819 if (tx_ring->wr_ptr_add)
1820 nfp_net_tx_xmit_more_flush(tx_ring);
1821 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1823 if (!nfp_net_xdp_complete(tx_ring))
1824 pkts_polled = budget;
1832 * nfp_net_poll() - napi poll function
1833 * @napi: NAPI structure
1834 * @budget: NAPI budget
1836 * Return: number of packets polled.
1838 static int nfp_net_poll(struct napi_struct *napi, int budget)
1840 struct nfp_net_r_vector *r_vec =
1841 container_of(napi, struct nfp_net_r_vector, napi);
1842 unsigned int pkts_polled = 0;
1845 nfp_net_tx_complete(r_vec->tx_ring, budget);
1847 pkts_polled = nfp_net_rx(r_vec->rx_ring, budget);
1849 if (pkts_polled < budget)
1850 if (napi_complete_done(napi, pkts_polled))
1851 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1856 /* Control device data path
1860 nfp_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1861 struct sk_buff *skb, bool old)
1863 unsigned int real_len = skb->len, meta_len = 0;
1864 struct nfp_net_tx_ring *tx_ring;
1865 struct nfp_net_tx_buf *txbuf;
1866 struct nfp_net_tx_desc *txd;
1867 struct nfp_net_dp *dp;
1868 dma_addr_t dma_addr;
1871 dp = &r_vec->nfp_net->dp;
1872 tx_ring = r_vec->tx_ring;
1874 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1875 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1879 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1880 u64_stats_update_begin(&r_vec->tx_sync);
1882 u64_stats_update_end(&r_vec->tx_sync);
1884 __skb_queue_tail(&r_vec->queue, skb);
1886 __skb_queue_head(&r_vec->queue, skb);
1890 if (nfp_app_ctrl_has_meta(nn->app)) {
1891 if (unlikely(skb_headroom(skb) < 8)) {
1892 nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1896 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1897 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
1900 /* Start with the head skbuf */
1901 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1903 if (dma_mapping_error(dp->dev, dma_addr))
1906 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1908 /* Stash the soft descriptor of the head then initialize it */
1909 txbuf = &tx_ring->txbufs[wr_idx];
1911 txbuf->dma_addr = dma_addr;
1914 txbuf->real_len = real_len;
1916 /* Build TX descriptor */
1917 txd = &tx_ring->txds[wr_idx];
1918 txd->offset_eop = meta_len | PCIE_DESC_TX_EOP;
1919 txd->dma_len = cpu_to_le16(skb_headlen(skb));
1920 nfp_desc_set_dma_addr(txd, dma_addr);
1921 txd->data_len = cpu_to_le16(skb->len);
1925 txd->lso_hdrlen = 0;
1928 tx_ring->wr_ptr_add++;
1929 nfp_net_tx_xmit_more_flush(tx_ring);
1934 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
1936 u64_stats_update_begin(&r_vec->tx_sync);
1938 u64_stats_update_end(&r_vec->tx_sync);
1939 dev_kfree_skb_any(skb);
1943 bool __nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb)
1945 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0];
1947 return nfp_ctrl_tx_one(nn, r_vec, skb, false);
1950 bool nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb)
1952 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0];
1955 spin_lock_bh(&r_vec->lock);
1956 ret = nfp_ctrl_tx_one(nn, r_vec, skb, false);
1957 spin_unlock_bh(&r_vec->lock);
1962 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1964 struct sk_buff *skb;
1966 while ((skb = __skb_dequeue(&r_vec->queue)))
1967 if (nfp_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1972 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1974 u32 meta_type, meta_tag;
1976 if (!nfp_app_ctrl_has_meta(nn->app))
1982 meta_type = get_unaligned_be32(data);
1983 meta_tag = get_unaligned_be32(data + 4);
1985 return (meta_type == NFP_NET_META_PORTID &&
1986 meta_tag == NFP_META_PORT_ID_CTRL);
1990 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1991 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1993 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1994 struct nfp_net_rx_buf *rxbuf;
1995 struct nfp_net_rx_desc *rxd;
1996 dma_addr_t new_dma_addr;
1997 struct sk_buff *skb;
2001 idx = D_IDX(rx_ring, rx_ring->rd_p);
2003 rxd = &rx_ring->rxds[idx];
2004 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
2007 /* Memory barrier to ensure that we won't do other reads
2008 * before the DD bit.
2014 rxbuf = &rx_ring->rxbufs[idx];
2015 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
2016 data_len = le16_to_cpu(rxd->rxd.data_len);
2017 pkt_len = data_len - meta_len;
2019 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
2020 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
2021 pkt_off += meta_len;
2023 pkt_off += dp->rx_offset;
2024 meta_off = pkt_off - meta_len;
2027 u64_stats_update_begin(&r_vec->rx_sync);
2029 r_vec->rx_bytes += pkt_len;
2030 u64_stats_update_end(&r_vec->rx_sync);
2032 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len);
2034 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
2035 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
2037 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
2041 skb = build_skb(rxbuf->frag, dp->fl_bufsz);
2042 if (unlikely(!skb)) {
2043 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
2046 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr);
2047 if (unlikely(!new_frag)) {
2048 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
2052 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
2054 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
2056 skb_reserve(skb, pkt_off);
2057 skb_put(skb, pkt_len);
2059 nfp_app_ctrl_rx(nn->app, skb);
2064 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
2066 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
2067 struct nfp_net *nn = r_vec->nfp_net;
2068 struct nfp_net_dp *dp = &nn->dp;
2069 unsigned int budget = 512;
2071 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
2077 static void nfp_ctrl_poll(unsigned long arg)
2079 struct nfp_net_r_vector *r_vec = (void *)arg;
2081 spin_lock(&r_vec->lock);
2082 nfp_net_tx_complete(r_vec->tx_ring, 0);
2083 __nfp_ctrl_tx_queued(r_vec);
2084 spin_unlock(&r_vec->lock);
2086 if (nfp_ctrl_rx(r_vec)) {
2087 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
2089 tasklet_schedule(&r_vec->tasklet);
2090 nn_dp_warn(&r_vec->nfp_net->dp,
2091 "control message budget exceeded!\n");
2095 /* Setup and Configuration
2099 * nfp_net_vecs_init() - Assign IRQs and setup rvecs.
2100 * @nn: NFP Network structure
2102 static void nfp_net_vecs_init(struct nfp_net *nn)
2104 struct nfp_net_r_vector *r_vec;
2107 nn->lsc_handler = nfp_net_irq_lsc;
2108 nn->exn_handler = nfp_net_irq_exn;
2110 for (r = 0; r < nn->max_r_vecs; r++) {
2111 struct msix_entry *entry;
2113 entry = &nn->irq_entries[NFP_NET_NON_Q_VECTORS + r];
2115 r_vec = &nn->r_vecs[r];
2116 r_vec->nfp_net = nn;
2117 r_vec->irq_entry = entry->entry;
2118 r_vec->irq_vector = entry->vector;
2120 if (nn->dp.netdev) {
2121 r_vec->handler = nfp_net_irq_rxtx;
2123 r_vec->handler = nfp_ctrl_irq_rxtx;
2125 __skb_queue_head_init(&r_vec->queue);
2126 spin_lock_init(&r_vec->lock);
2127 tasklet_init(&r_vec->tasklet, nfp_ctrl_poll,
2128 (unsigned long)r_vec);
2129 tasklet_disable(&r_vec->tasklet);
2132 cpumask_set_cpu(r, &r_vec->affinity_mask);
2137 * nfp_net_tx_ring_free() - Free resources allocated to a TX ring
2138 * @tx_ring: TX ring to free
2140 static void nfp_net_tx_ring_free(struct nfp_net_tx_ring *tx_ring)
2142 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
2143 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
2145 kvfree(tx_ring->txbufs);
2148 dma_free_coherent(dp->dev, tx_ring->size,
2149 tx_ring->txds, tx_ring->dma);
2152 tx_ring->txbufs = NULL;
2153 tx_ring->txds = NULL;
2159 * nfp_net_tx_ring_alloc() - Allocate resource for a TX ring
2160 * @dp: NFP Net data path struct
2161 * @tx_ring: TX Ring structure to allocate
2163 * Return: 0 on success, negative errno otherwise.
2166 nfp_net_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
2168 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
2170 tx_ring->cnt = dp->txd_cnt;
2172 tx_ring->size = array_size(tx_ring->cnt, sizeof(*tx_ring->txds));
2173 tx_ring->txds = dma_alloc_coherent(dp->dev, tx_ring->size,
2175 GFP_KERNEL | __GFP_NOWARN);
2176 if (!tx_ring->txds) {
2177 netdev_warn(dp->netdev, "failed to allocate TX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n",
2182 tx_ring->txbufs = kvcalloc(tx_ring->cnt, sizeof(*tx_ring->txbufs),
2184 if (!tx_ring->txbufs)
2187 if (!tx_ring->is_xdp && dp->netdev)
2188 netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask,
2194 nfp_net_tx_ring_free(tx_ring);
2199 nfp_net_tx_ring_bufs_free(struct nfp_net_dp *dp,
2200 struct nfp_net_tx_ring *tx_ring)
2204 if (!tx_ring->is_xdp)
2207 for (i = 0; i < tx_ring->cnt; i++) {
2208 if (!tx_ring->txbufs[i].frag)
2211 nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr);
2212 __free_page(virt_to_page(tx_ring->txbufs[i].frag));
2217 nfp_net_tx_ring_bufs_alloc(struct nfp_net_dp *dp,
2218 struct nfp_net_tx_ring *tx_ring)
2220 struct nfp_net_tx_buf *txbufs = tx_ring->txbufs;
2223 if (!tx_ring->is_xdp)
2226 for (i = 0; i < tx_ring->cnt; i++) {
2227 txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr);
2228 if (!txbufs[i].frag) {
2229 nfp_net_tx_ring_bufs_free(dp, tx_ring);
2237 static int nfp_net_tx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp)
2241 dp->tx_rings = kcalloc(dp->num_tx_rings, sizeof(*dp->tx_rings),
2246 for (r = 0; r < dp->num_tx_rings; r++) {
2249 if (r >= dp->num_stack_tx_rings)
2250 bias = dp->num_stack_tx_rings;
2252 nfp_net_tx_ring_init(&dp->tx_rings[r], &nn->r_vecs[r - bias],
2255 if (nfp_net_tx_ring_alloc(dp, &dp->tx_rings[r]))
2258 if (nfp_net_tx_ring_bufs_alloc(dp, &dp->tx_rings[r]))
2266 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]);
2268 nfp_net_tx_ring_free(&dp->tx_rings[r]);
2270 kfree(dp->tx_rings);
2274 static void nfp_net_tx_rings_free(struct nfp_net_dp *dp)
2278 for (r = 0; r < dp->num_tx_rings; r++) {
2279 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]);
2280 nfp_net_tx_ring_free(&dp->tx_rings[r]);
2283 kfree(dp->tx_rings);
2287 * nfp_net_rx_ring_free() - Free resources allocated to a RX ring
2288 * @rx_ring: RX ring to free
2290 static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring)
2292 struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
2293 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
2296 xdp_rxq_info_unreg(&rx_ring->xdp_rxq);
2297 kvfree(rx_ring->rxbufs);
2300 dma_free_coherent(dp->dev, rx_ring->size,
2301 rx_ring->rxds, rx_ring->dma);
2304 rx_ring->rxbufs = NULL;
2305 rx_ring->rxds = NULL;
2311 * nfp_net_rx_ring_alloc() - Allocate resource for a RX ring
2312 * @dp: NFP Net data path struct
2313 * @rx_ring: RX ring to allocate
2315 * Return: 0 on success, negative errno otherwise.
2318 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring)
2323 err = xdp_rxq_info_reg(&rx_ring->xdp_rxq, dp->netdev,
2329 rx_ring->cnt = dp->rxd_cnt;
2330 rx_ring->size = array_size(rx_ring->cnt, sizeof(*rx_ring->rxds));
2331 rx_ring->rxds = dma_alloc_coherent(dp->dev, rx_ring->size,
2333 GFP_KERNEL | __GFP_NOWARN);
2334 if (!rx_ring->rxds) {
2335 netdev_warn(dp->netdev, "failed to allocate RX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n",
2340 rx_ring->rxbufs = kvcalloc(rx_ring->cnt, sizeof(*rx_ring->rxbufs),
2342 if (!rx_ring->rxbufs)
2348 nfp_net_rx_ring_free(rx_ring);
2352 static int nfp_net_rx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp)
2356 dp->rx_rings = kcalloc(dp->num_rx_rings, sizeof(*dp->rx_rings),
2361 for (r = 0; r < dp->num_rx_rings; r++) {
2362 nfp_net_rx_ring_init(&dp->rx_rings[r], &nn->r_vecs[r], r);
2364 if (nfp_net_rx_ring_alloc(dp, &dp->rx_rings[r]))
2367 if (nfp_net_rx_ring_bufs_alloc(dp, &dp->rx_rings[r]))
2375 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]);
2377 nfp_net_rx_ring_free(&dp->rx_rings[r]);
2379 kfree(dp->rx_rings);
2383 static void nfp_net_rx_rings_free(struct nfp_net_dp *dp)
2387 for (r = 0; r < dp->num_rx_rings; r++) {
2388 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]);
2389 nfp_net_rx_ring_free(&dp->rx_rings[r]);
2392 kfree(dp->rx_rings);
2396 nfp_net_vector_assign_rings(struct nfp_net_dp *dp,
2397 struct nfp_net_r_vector *r_vec, int idx)
2399 r_vec->rx_ring = idx < dp->num_rx_rings ? &dp->rx_rings[idx] : NULL;
2401 idx < dp->num_stack_tx_rings ? &dp->tx_rings[idx] : NULL;
2403 r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ?
2404 &dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL;
2408 nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
2415 netif_napi_add(nn->dp.netdev, &r_vec->napi,
2416 nfp_net_poll, NAPI_POLL_WEIGHT);
2418 tasklet_enable(&r_vec->tasklet);
2420 snprintf(r_vec->name, sizeof(r_vec->name),
2421 "%s-rxtx-%d", nfp_net_name(nn), idx);
2422 err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name,
2426 netif_napi_del(&r_vec->napi);
2428 tasklet_disable(&r_vec->tasklet);
2430 nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector);
2433 disable_irq(r_vec->irq_vector);
2435 irq_set_affinity_hint(r_vec->irq_vector, &r_vec->affinity_mask);
2437 nn_dbg(nn, "RV%02d: irq=%03d/%03d\n", idx, r_vec->irq_vector,
2444 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec)
2446 irq_set_affinity_hint(r_vec->irq_vector, NULL);
2448 netif_napi_del(&r_vec->napi);
2450 tasklet_disable(&r_vec->tasklet);
2452 free_irq(r_vec->irq_vector, r_vec);
2456 * nfp_net_rss_write_itbl() - Write RSS indirection table to device
2457 * @nn: NFP Net device to reconfigure
2459 void nfp_net_rss_write_itbl(struct nfp_net *nn)
2463 for (i = 0; i < NFP_NET_CFG_RSS_ITBL_SZ; i += 4)
2464 nn_writel(nn, NFP_NET_CFG_RSS_ITBL + i,
2465 get_unaligned_le32(nn->rss_itbl + i));
2469 * nfp_net_rss_write_key() - Write RSS hash key to device
2470 * @nn: NFP Net device to reconfigure
2472 void nfp_net_rss_write_key(struct nfp_net *nn)
2476 for (i = 0; i < nfp_net_rss_key_sz(nn); i += 4)
2477 nn_writel(nn, NFP_NET_CFG_RSS_KEY + i,
2478 get_unaligned_le32(nn->rss_key + i));
2482 * nfp_net_coalesce_write_cfg() - Write irq coalescence configuration to HW
2483 * @nn: NFP Net device to reconfigure
2485 void nfp_net_coalesce_write_cfg(struct nfp_net *nn)
2491 /* Compute factor used to convert coalesce '_usecs' parameters to
2492 * ME timestamp ticks. There are 16 ME clock cycles for each timestamp
2495 factor = nn->tlv_caps.me_freq_mhz / 16;
2497 /* copy RX interrupt coalesce parameters */
2498 value = (nn->rx_coalesce_max_frames << 16) |
2499 (factor * nn->rx_coalesce_usecs);
2500 for (i = 0; i < nn->dp.num_rx_rings; i++)
2501 nn_writel(nn, NFP_NET_CFG_RXR_IRQ_MOD(i), value);
2503 /* copy TX interrupt coalesce parameters */
2504 value = (nn->tx_coalesce_max_frames << 16) |
2505 (factor * nn->tx_coalesce_usecs);
2506 for (i = 0; i < nn->dp.num_tx_rings; i++)
2507 nn_writel(nn, NFP_NET_CFG_TXR_IRQ_MOD(i), value);
2511 * nfp_net_write_mac_addr() - Write mac address to the device control BAR
2512 * @nn: NFP Net device to reconfigure
2513 * @addr: MAC address to write
2515 * Writes the MAC address from the netdev to the device control BAR. Does not
2516 * perform the required reconfig. We do a bit of byte swapping dance because
2519 static void nfp_net_write_mac_addr(struct nfp_net *nn, const u8 *addr)
2521 nn_writel(nn, NFP_NET_CFG_MACADDR + 0, get_unaligned_be32(addr));
2522 nn_writew(nn, NFP_NET_CFG_MACADDR + 6, get_unaligned_be16(addr + 4));
2525 static void nfp_net_vec_clear_ring_data(struct nfp_net *nn, unsigned int idx)
2527 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), 0);
2528 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), 0);
2529 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), 0);
2531 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), 0);
2532 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), 0);
2533 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), 0);
2537 * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP
2538 * @nn: NFP Net device to reconfigure
2540 * Warning: must be fully idempotent.
2542 static void nfp_net_clear_config_and_disable(struct nfp_net *nn)
2544 u32 new_ctrl, update;
2548 new_ctrl = nn->dp.ctrl;
2549 new_ctrl &= ~NFP_NET_CFG_CTRL_ENABLE;
2550 update = NFP_NET_CFG_UPDATE_GEN;
2551 update |= NFP_NET_CFG_UPDATE_MSIX;
2552 update |= NFP_NET_CFG_UPDATE_RING;
2554 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG)
2555 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG;
2557 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0);
2558 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0);
2560 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2561 err = nfp_net_reconfig(nn, update);
2563 nn_err(nn, "Could not disable device: %d\n", err);
2565 for (r = 0; r < nn->dp.num_rx_rings; r++)
2566 nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]);
2567 for (r = 0; r < nn->dp.num_tx_rings; r++)
2568 nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]);
2569 for (r = 0; r < nn->dp.num_r_vecs; r++)
2570 nfp_net_vec_clear_ring_data(nn, r);
2572 nn->dp.ctrl = new_ctrl;
2576 nfp_net_rx_ring_hw_cfg_write(struct nfp_net *nn,
2577 struct nfp_net_rx_ring *rx_ring, unsigned int idx)
2579 /* Write the DMA address, size and MSI-X info to the device */
2580 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), rx_ring->dma);
2581 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), ilog2(rx_ring->cnt));
2582 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), rx_ring->r_vec->irq_entry);
2586 nfp_net_tx_ring_hw_cfg_write(struct nfp_net *nn,
2587 struct nfp_net_tx_ring *tx_ring, unsigned int idx)
2589 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), tx_ring->dma);
2590 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), ilog2(tx_ring->cnt));
2591 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), tx_ring->r_vec->irq_entry);
2595 * nfp_net_set_config_and_enable() - Write control BAR and enable NFP
2596 * @nn: NFP Net device to reconfigure
2598 static int nfp_net_set_config_and_enable(struct nfp_net *nn)
2600 u32 bufsz, new_ctrl, update = 0;
2604 new_ctrl = nn->dp.ctrl;
2606 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_RSS_ANY) {
2607 nfp_net_rss_write_key(nn);
2608 nfp_net_rss_write_itbl(nn);
2609 nn_writel(nn, NFP_NET_CFG_RSS_CTRL, nn->rss_cfg);
2610 update |= NFP_NET_CFG_UPDATE_RSS;
2613 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_IRQMOD) {
2614 nfp_net_coalesce_write_cfg(nn);
2615 update |= NFP_NET_CFG_UPDATE_IRQMOD;
2618 for (r = 0; r < nn->dp.num_tx_rings; r++)
2619 nfp_net_tx_ring_hw_cfg_write(nn, &nn->dp.tx_rings[r], r);
2620 for (r = 0; r < nn->dp.num_rx_rings; r++)
2621 nfp_net_rx_ring_hw_cfg_write(nn, &nn->dp.rx_rings[r], r);
2623 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, nn->dp.num_tx_rings == 64 ?
2624 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_tx_rings) - 1);
2626 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, nn->dp.num_rx_rings == 64 ?
2627 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_rx_rings) - 1);
2630 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr);
2632 nn_writel(nn, NFP_NET_CFG_MTU, nn->dp.mtu);
2634 bufsz = nn->dp.fl_bufsz - nn->dp.rx_dma_off - NFP_NET_RX_BUF_NON_DATA;
2635 nn_writel(nn, NFP_NET_CFG_FLBUFSZ, bufsz);
2638 new_ctrl |= NFP_NET_CFG_CTRL_ENABLE;
2639 update |= NFP_NET_CFG_UPDATE_GEN;
2640 update |= NFP_NET_CFG_UPDATE_MSIX;
2641 update |= NFP_NET_CFG_UPDATE_RING;
2642 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG)
2643 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG;
2645 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2646 err = nfp_net_reconfig(nn, update);
2648 nfp_net_clear_config_and_disable(nn);
2652 nn->dp.ctrl = new_ctrl;
2654 for (r = 0; r < nn->dp.num_rx_rings; r++)
2655 nfp_net_rx_ring_fill_freelist(&nn->dp, &nn->dp.rx_rings[r]);
2657 /* Since reconfiguration requests while NFP is down are ignored we
2658 * have to wipe the entire VXLAN configuration and reinitialize it.
2660 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN) {
2661 memset(&nn->vxlan_ports, 0, sizeof(nn->vxlan_ports));
2662 memset(&nn->vxlan_usecnt, 0, sizeof(nn->vxlan_usecnt));
2663 udp_tunnel_get_rx_info(nn->dp.netdev);
2670 * nfp_net_close_stack() - Quiesce the stack (part of close)
2671 * @nn: NFP Net device to reconfigure
2673 static void nfp_net_close_stack(struct nfp_net *nn)
2677 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2678 netif_carrier_off(nn->dp.netdev);
2679 nn->link_up = false;
2681 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2682 disable_irq(nn->r_vecs[r].irq_vector);
2683 napi_disable(&nn->r_vecs[r].napi);
2686 netif_tx_disable(nn->dp.netdev);
2690 * nfp_net_close_free_all() - Free all runtime resources
2691 * @nn: NFP Net device to reconfigure
2693 static void nfp_net_close_free_all(struct nfp_net *nn)
2697 nfp_net_tx_rings_free(&nn->dp);
2698 nfp_net_rx_rings_free(&nn->dp);
2700 for (r = 0; r < nn->dp.num_r_vecs; r++)
2701 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
2703 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2704 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX);
2708 * nfp_net_netdev_close() - Called when the device is downed
2709 * @netdev: netdev structure
2711 static int nfp_net_netdev_close(struct net_device *netdev)
2713 struct nfp_net *nn = netdev_priv(netdev);
2715 /* Step 1: Disable RX and TX rings from the Linux kernel perspective
2717 nfp_net_close_stack(nn);
2721 nfp_net_clear_config_and_disable(nn);
2722 nfp_port_configure(netdev, false);
2724 /* Step 3: Free resources
2726 nfp_net_close_free_all(nn);
2728 nn_dbg(nn, "%s down", netdev->name);
2732 void nfp_ctrl_close(struct nfp_net *nn)
2738 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2739 disable_irq(nn->r_vecs[r].irq_vector);
2740 tasklet_disable(&nn->r_vecs[r].tasklet);
2743 nfp_net_clear_config_and_disable(nn);
2745 nfp_net_close_free_all(nn);
2751 * nfp_net_open_stack() - Start the device from stack's perspective
2752 * @nn: NFP Net device to reconfigure
2754 static void nfp_net_open_stack(struct nfp_net *nn)
2758 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2759 napi_enable(&nn->r_vecs[r].napi);
2760 enable_irq(nn->r_vecs[r].irq_vector);
2763 netif_tx_wake_all_queues(nn->dp.netdev);
2765 enable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2766 nfp_net_read_link_status(nn);
2769 static int nfp_net_open_alloc_all(struct nfp_net *nn)
2773 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_EXN, "%s-exn",
2774 nn->exn_name, sizeof(nn->exn_name),
2775 NFP_NET_IRQ_EXN_IDX, nn->exn_handler);
2778 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_LSC, "%s-lsc",
2779 nn->lsc_name, sizeof(nn->lsc_name),
2780 NFP_NET_IRQ_LSC_IDX, nn->lsc_handler);
2783 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2785 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2786 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r);
2788 goto err_cleanup_vec_p;
2791 err = nfp_net_rx_rings_prepare(nn, &nn->dp);
2793 goto err_cleanup_vec;
2795 err = nfp_net_tx_rings_prepare(nn, &nn->dp);
2797 goto err_free_rx_rings;
2799 for (r = 0; r < nn->max_r_vecs; r++)
2800 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r);
2805 nfp_net_rx_rings_free(&nn->dp);
2807 r = nn->dp.num_r_vecs;
2810 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
2811 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2813 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX);
2817 static int nfp_net_netdev_open(struct net_device *netdev)
2819 struct nfp_net *nn = netdev_priv(netdev);
2822 /* Step 1: Allocate resources for rings and the like
2823 * - Request interrupts
2824 * - Allocate RX and TX ring resources
2825 * - Setup initial RSS table
2827 err = nfp_net_open_alloc_all(nn);
2831 err = netif_set_real_num_tx_queues(netdev, nn->dp.num_stack_tx_rings);
2835 err = netif_set_real_num_rx_queues(netdev, nn->dp.num_rx_rings);
2839 /* Step 2: Configure the NFP
2840 * - Ifup the physical interface if it exists
2841 * - Enable rings from 0 to tx_rings/rx_rings - 1.
2842 * - Write MAC address (in case it changed)
2844 * - Set the Freelist buffer size
2847 err = nfp_port_configure(netdev, true);
2851 err = nfp_net_set_config_and_enable(nn);
2853 goto err_port_disable;
2855 /* Step 3: Enable for kernel
2856 * - put some freelist descriptors on each RX ring
2857 * - enable NAPI on each ring
2858 * - enable all TX queues
2861 nfp_net_open_stack(nn);
2866 nfp_port_configure(netdev, false);
2868 nfp_net_close_free_all(nn);
2872 int nfp_ctrl_open(struct nfp_net *nn)
2876 /* ring dumping depends on vNICs being opened/closed under rtnl */
2879 err = nfp_net_open_alloc_all(nn);
2883 err = nfp_net_set_config_and_enable(nn);
2887 for (r = 0; r < nn->dp.num_r_vecs; r++)
2888 enable_irq(nn->r_vecs[r].irq_vector);
2895 nfp_net_close_free_all(nn);
2901 static void nfp_net_set_rx_mode(struct net_device *netdev)
2903 struct nfp_net *nn = netdev_priv(netdev);
2906 new_ctrl = nn->dp.ctrl;
2908 if (!netdev_mc_empty(netdev) || netdev->flags & IFF_ALLMULTI)
2909 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_L2MC;
2911 new_ctrl &= ~NFP_NET_CFG_CTRL_L2MC;
2913 if (netdev->flags & IFF_PROMISC) {
2914 if (nn->cap & NFP_NET_CFG_CTRL_PROMISC)
2915 new_ctrl |= NFP_NET_CFG_CTRL_PROMISC;
2917 nn_warn(nn, "FW does not support promiscuous mode\n");
2919 new_ctrl &= ~NFP_NET_CFG_CTRL_PROMISC;
2922 if (new_ctrl == nn->dp.ctrl)
2925 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2926 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_GEN);
2928 nn->dp.ctrl = new_ctrl;
2931 static void nfp_net_rss_init_itbl(struct nfp_net *nn)
2935 for (i = 0; i < sizeof(nn->rss_itbl); i++)
2937 ethtool_rxfh_indir_default(i, nn->dp.num_rx_rings);
2940 static void nfp_net_dp_swap(struct nfp_net *nn, struct nfp_net_dp *dp)
2942 struct nfp_net_dp new_dp = *dp;
2947 nn->dp.netdev->mtu = new_dp.mtu;
2949 if (!netif_is_rxfh_configured(nn->dp.netdev))
2950 nfp_net_rss_init_itbl(nn);
2953 static int nfp_net_dp_swap_enable(struct nfp_net *nn, struct nfp_net_dp *dp)
2958 nfp_net_dp_swap(nn, dp);
2960 for (r = 0; r < nn->max_r_vecs; r++)
2961 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r);
2963 err = netif_set_real_num_rx_queues(nn->dp.netdev, nn->dp.num_rx_rings);
2967 if (nn->dp.netdev->real_num_tx_queues != nn->dp.num_stack_tx_rings) {
2968 err = netif_set_real_num_tx_queues(nn->dp.netdev,
2969 nn->dp.num_stack_tx_rings);
2974 return nfp_net_set_config_and_enable(nn);
2977 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn)
2979 struct nfp_net_dp *new;
2981 new = kmalloc(sizeof(*new), GFP_KERNEL);
2987 /* Clear things which need to be recomputed */
2989 new->tx_rings = NULL;
2990 new->rx_rings = NULL;
2991 new->num_r_vecs = 0;
2992 new->num_stack_tx_rings = 0;
2998 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp,
2999 struct netlink_ext_ack *extack)
3001 /* XDP-enabled tests */
3004 if (dp->fl_bufsz > PAGE_SIZE) {
3005 NL_SET_ERR_MSG_MOD(extack, "MTU too large w/ XDP enabled");
3008 if (dp->num_tx_rings > nn->max_tx_rings) {
3009 NL_SET_ERR_MSG_MOD(extack, "Insufficient number of TX rings w/ XDP enabled");
3016 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp,
3017 struct netlink_ext_ack *extack)
3021 dp->fl_bufsz = nfp_net_calc_fl_bufsz(dp);
3023 dp->num_stack_tx_rings = dp->num_tx_rings;
3025 dp->num_stack_tx_rings -= dp->num_rx_rings;
3027 dp->num_r_vecs = max(dp->num_rx_rings, dp->num_stack_tx_rings);
3029 err = nfp_net_check_config(nn, dp, extack);
3033 if (!netif_running(dp->netdev)) {
3034 nfp_net_dp_swap(nn, dp);
3039 /* Prepare new rings */
3040 for (r = nn->dp.num_r_vecs; r < dp->num_r_vecs; r++) {
3041 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r);
3044 goto err_cleanup_vecs;
3048 err = nfp_net_rx_rings_prepare(nn, dp);
3050 goto err_cleanup_vecs;
3052 err = nfp_net_tx_rings_prepare(nn, dp);
3056 /* Stop device, swap in new rings, try to start the firmware */
3057 nfp_net_close_stack(nn);
3058 nfp_net_clear_config_and_disable(nn);
3060 err = nfp_net_dp_swap_enable(nn, dp);
3064 nfp_net_clear_config_and_disable(nn);
3066 /* Try with old configuration and old rings */
3067 err2 = nfp_net_dp_swap_enable(nn, dp);
3069 nn_err(nn, "Can't restore ring config - FW communication failed (%d,%d)\n",
3072 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
3073 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3075 nfp_net_rx_rings_free(dp);
3076 nfp_net_tx_rings_free(dp);
3078 nfp_net_open_stack(nn);
3085 nfp_net_rx_rings_free(dp);
3087 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
3088 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3093 static int nfp_net_change_mtu(struct net_device *netdev, int new_mtu)
3095 struct nfp_net *nn = netdev_priv(netdev);
3096 struct nfp_net_dp *dp;
3099 err = nfp_app_check_mtu(nn->app, netdev, new_mtu);
3103 dp = nfp_net_clone_dp(nn);
3109 return nfp_net_ring_reconfig(nn, dp, NULL);
3113 nfp_net_vlan_rx_add_vid(struct net_device *netdev, __be16 proto, u16 vid)
3115 struct nfp_net *nn = netdev_priv(netdev);
3117 /* Priority tagged packets with vlan id 0 are processed by the
3118 * NFP as untagged packets
3123 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid);
3124 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO,
3127 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_ADD);
3131 nfp_net_vlan_rx_kill_vid(struct net_device *netdev, __be16 proto, u16 vid)
3133 struct nfp_net *nn = netdev_priv(netdev);
3135 /* Priority tagged packets with vlan id 0 are processed by the
3136 * NFP as untagged packets
3141 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid);
3142 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO,
3145 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_KILL);
3148 static void nfp_net_stat64(struct net_device *netdev,
3149 struct rtnl_link_stats64 *stats)
3151 struct nfp_net *nn = netdev_priv(netdev);
3154 /* Collect software stats */
3155 for (r = 0; r < nn->max_r_vecs; r++) {
3156 struct nfp_net_r_vector *r_vec = &nn->r_vecs[r];
3161 start = u64_stats_fetch_begin(&r_vec->rx_sync);
3162 data[0] = r_vec->rx_pkts;
3163 data[1] = r_vec->rx_bytes;
3164 data[2] = r_vec->rx_drops;
3165 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
3166 stats->rx_packets += data[0];
3167 stats->rx_bytes += data[1];
3168 stats->rx_dropped += data[2];
3171 start = u64_stats_fetch_begin(&r_vec->tx_sync);
3172 data[0] = r_vec->tx_pkts;
3173 data[1] = r_vec->tx_bytes;
3174 data[2] = r_vec->tx_errors;
3175 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
3176 stats->tx_packets += data[0];
3177 stats->tx_bytes += data[1];
3178 stats->tx_errors += data[2];
3181 /* Add in device stats */
3182 stats->multicast += nn_readq(nn, NFP_NET_CFG_STATS_RX_MC_FRAMES);
3183 stats->rx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_RX_DISCARDS);
3184 stats->rx_errors += nn_readq(nn, NFP_NET_CFG_STATS_RX_ERRORS);
3186 stats->tx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_TX_DISCARDS);
3187 stats->tx_errors += nn_readq(nn, NFP_NET_CFG_STATS_TX_ERRORS);
3190 static int nfp_net_set_features(struct net_device *netdev,
3191 netdev_features_t features)
3193 netdev_features_t changed = netdev->features ^ features;
3194 struct nfp_net *nn = netdev_priv(netdev);
3198 /* Assume this is not called with features we have not advertised */
3200 new_ctrl = nn->dp.ctrl;
3202 if (changed & NETIF_F_RXCSUM) {
3203 if (features & NETIF_F_RXCSUM)
3204 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY;
3206 new_ctrl &= ~NFP_NET_CFG_CTRL_RXCSUM_ANY;
3209 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3210 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM))
3211 new_ctrl |= NFP_NET_CFG_CTRL_TXCSUM;
3213 new_ctrl &= ~NFP_NET_CFG_CTRL_TXCSUM;
3216 if (changed & (NETIF_F_TSO | NETIF_F_TSO6)) {
3217 if (features & (NETIF_F_TSO | NETIF_F_TSO6))
3218 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?:
3219 NFP_NET_CFG_CTRL_LSO;
3221 new_ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY;
3224 if (changed & NETIF_F_HW_VLAN_CTAG_RX) {
3225 if (features & NETIF_F_HW_VLAN_CTAG_RX)
3226 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
3228 new_ctrl &= ~NFP_NET_CFG_CTRL_RXVLAN;
3231 if (changed & NETIF_F_HW_VLAN_CTAG_TX) {
3232 if (features & NETIF_F_HW_VLAN_CTAG_TX)
3233 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
3235 new_ctrl &= ~NFP_NET_CFG_CTRL_TXVLAN;
3238 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) {
3239 if (features & NETIF_F_HW_VLAN_CTAG_FILTER)
3240 new_ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER;
3242 new_ctrl &= ~NFP_NET_CFG_CTRL_CTAG_FILTER;
3245 if (changed & NETIF_F_SG) {
3246 if (features & NETIF_F_SG)
3247 new_ctrl |= NFP_NET_CFG_CTRL_GATHER;
3249 new_ctrl &= ~NFP_NET_CFG_CTRL_GATHER;
3252 err = nfp_port_set_features(netdev, features);
3256 nn_dbg(nn, "Feature change 0x%llx -> 0x%llx (changed=0x%llx)\n",
3257 netdev->features, features, changed);
3259 if (new_ctrl == nn->dp.ctrl)
3262 nn_dbg(nn, "NIC ctrl: 0x%x -> 0x%x\n", nn->dp.ctrl, new_ctrl);
3263 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
3264 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN);
3268 nn->dp.ctrl = new_ctrl;
3273 static netdev_features_t
3274 nfp_net_features_check(struct sk_buff *skb, struct net_device *dev,
3275 netdev_features_t features)
3279 /* We can't do TSO over double tagged packets (802.1AD) */
3280 features &= vlan_features_check(skb, features);
3282 if (!skb->encapsulation)
3285 /* Ensure that inner L4 header offset fits into TX descriptor field */
3286 if (skb_is_gso(skb)) {
3289 hdrlen = skb_inner_transport_header(skb) - skb->data +
3290 inner_tcp_hdrlen(skb);
3292 /* Assume worst case scenario of having longest possible
3293 * metadata prepend - 8B
3295 if (unlikely(hdrlen > NFP_NET_LSO_MAX_HDR_SZ - 8))
3296 features &= ~NETIF_F_GSO_MASK;
3299 /* VXLAN/GRE check */
3300 switch (vlan_get_protocol(skb)) {
3301 case htons(ETH_P_IP):
3302 l4_hdr = ip_hdr(skb)->protocol;
3304 case htons(ETH_P_IPV6):
3305 l4_hdr = ipv6_hdr(skb)->nexthdr;
3308 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3311 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
3312 skb->inner_protocol != htons(ETH_P_TEB) ||
3313 (l4_hdr != IPPROTO_UDP && l4_hdr != IPPROTO_GRE) ||
3314 (l4_hdr == IPPROTO_UDP &&
3315 (skb_inner_mac_header(skb) - skb_transport_header(skb) !=
3316 sizeof(struct udphdr) + sizeof(struct vxlanhdr))))
3317 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3323 nfp_net_get_phys_port_name(struct net_device *netdev, char *name, size_t len)
3325 struct nfp_net *nn = netdev_priv(netdev);
3329 return nfp_port_get_phys_port_name(netdev, name, len);
3331 if (nn->dp.is_vf || nn->vnic_no_name)
3334 n = snprintf(name, len, "n%d", nn->id);
3342 * nfp_net_set_vxlan_port() - set vxlan port in SW and reconfigure HW
3343 * @nn: NFP Net device to reconfigure
3344 * @idx: Index into the port table where new port should be written
3345 * @port: UDP port to configure (pass zero to remove VXLAN port)
3347 static void nfp_net_set_vxlan_port(struct nfp_net *nn, int idx, __be16 port)
3351 nn->vxlan_ports[idx] = port;
3353 if (!(nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN))
3356 BUILD_BUG_ON(NFP_NET_N_VXLAN_PORTS & 1);
3357 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i += 2)
3358 nn_writel(nn, NFP_NET_CFG_VXLAN_PORT + i * sizeof(port),
3359 be16_to_cpu(nn->vxlan_ports[i + 1]) << 16 |
3360 be16_to_cpu(nn->vxlan_ports[i]));
3362 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_VXLAN);
3366 * nfp_net_find_vxlan_idx() - find table entry of the port or a free one
3367 * @nn: NFP Network structure
3368 * @port: UDP port to look for
3370 * Return: if the port is already in the table -- it's position;
3371 * if the port is not in the table -- free position to use;
3372 * if the table is full -- -ENOSPC.
3374 static int nfp_net_find_vxlan_idx(struct nfp_net *nn, __be16 port)
3376 int i, free_idx = -ENOSPC;
3378 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i++) {
3379 if (nn->vxlan_ports[i] == port)
3381 if (!nn->vxlan_usecnt[i])
3388 static void nfp_net_add_vxlan_port(struct net_device *netdev,
3389 struct udp_tunnel_info *ti)
3391 struct nfp_net *nn = netdev_priv(netdev);
3394 if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3397 idx = nfp_net_find_vxlan_idx(nn, ti->port);
3401 if (!nn->vxlan_usecnt[idx]++)
3402 nfp_net_set_vxlan_port(nn, idx, ti->port);
3405 static void nfp_net_del_vxlan_port(struct net_device *netdev,
3406 struct udp_tunnel_info *ti)
3408 struct nfp_net *nn = netdev_priv(netdev);
3411 if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3414 idx = nfp_net_find_vxlan_idx(nn, ti->port);
3415 if (idx == -ENOSPC || !nn->vxlan_usecnt[idx])
3418 if (!--nn->vxlan_usecnt[idx])
3419 nfp_net_set_vxlan_port(nn, idx, 0);
3422 static int nfp_net_xdp_setup_drv(struct nfp_net *nn, struct netdev_bpf *bpf)
3424 struct bpf_prog *prog = bpf->prog;
3425 struct nfp_net_dp *dp;
3428 if (!xdp_attachment_flags_ok(&nn->xdp, bpf))
3431 if (!prog == !nn->dp.xdp_prog) {
3432 WRITE_ONCE(nn->dp.xdp_prog, prog);
3433 xdp_attachment_setup(&nn->xdp, bpf);
3437 dp = nfp_net_clone_dp(nn);
3441 dp->xdp_prog = prog;
3442 dp->num_tx_rings += prog ? nn->dp.num_rx_rings : -nn->dp.num_rx_rings;
3443 dp->rx_dma_dir = prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
3444 dp->rx_dma_off = prog ? XDP_PACKET_HEADROOM - nn->dp.rx_offset : 0;
3446 /* We need RX reconfig to remap the buffers (BIDIR vs FROM_DEV) */
3447 err = nfp_net_ring_reconfig(nn, dp, bpf->extack);
3451 xdp_attachment_setup(&nn->xdp, bpf);
3455 static int nfp_net_xdp_setup_hw(struct nfp_net *nn, struct netdev_bpf *bpf)
3459 if (!xdp_attachment_flags_ok(&nn->xdp_hw, bpf))
3462 err = nfp_app_xdp_offload(nn->app, nn, bpf->prog, bpf->extack);
3466 xdp_attachment_setup(&nn->xdp_hw, bpf);
3470 static int nfp_net_xdp(struct net_device *netdev, struct netdev_bpf *xdp)
3472 struct nfp_net *nn = netdev_priv(netdev);
3474 switch (xdp->command) {
3475 case XDP_SETUP_PROG:
3476 return nfp_net_xdp_setup_drv(nn, xdp);
3477 case XDP_SETUP_PROG_HW:
3478 return nfp_net_xdp_setup_hw(nn, xdp);
3479 case XDP_QUERY_PROG:
3480 return xdp_attachment_query(&nn->xdp, xdp);
3481 case XDP_QUERY_PROG_HW:
3482 return xdp_attachment_query(&nn->xdp_hw, xdp);
3484 return nfp_app_bpf(nn->app, nn, xdp);
3488 static int nfp_net_set_mac_address(struct net_device *netdev, void *addr)
3490 struct nfp_net *nn = netdev_priv(netdev);
3491 struct sockaddr *saddr = addr;
3494 err = eth_prepare_mac_addr_change(netdev, addr);
3498 nfp_net_write_mac_addr(nn, saddr->sa_data);
3500 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MACADDR);
3504 eth_commit_mac_addr_change(netdev, addr);
3509 const struct net_device_ops nfp_net_netdev_ops = {
3510 .ndo_init = nfp_app_ndo_init,
3511 .ndo_uninit = nfp_app_ndo_uninit,
3512 .ndo_open = nfp_net_netdev_open,
3513 .ndo_stop = nfp_net_netdev_close,
3514 .ndo_start_xmit = nfp_net_tx,
3515 .ndo_get_stats64 = nfp_net_stat64,
3516 .ndo_vlan_rx_add_vid = nfp_net_vlan_rx_add_vid,
3517 .ndo_vlan_rx_kill_vid = nfp_net_vlan_rx_kill_vid,
3518 .ndo_set_vf_mac = nfp_app_set_vf_mac,
3519 .ndo_set_vf_vlan = nfp_app_set_vf_vlan,
3520 .ndo_set_vf_spoofchk = nfp_app_set_vf_spoofchk,
3521 .ndo_get_vf_config = nfp_app_get_vf_config,
3522 .ndo_set_vf_link_state = nfp_app_set_vf_link_state,
3523 .ndo_setup_tc = nfp_port_setup_tc,
3524 .ndo_tx_timeout = nfp_net_tx_timeout,
3525 .ndo_set_rx_mode = nfp_net_set_rx_mode,
3526 .ndo_change_mtu = nfp_net_change_mtu,
3527 .ndo_set_mac_address = nfp_net_set_mac_address,
3528 .ndo_set_features = nfp_net_set_features,
3529 .ndo_features_check = nfp_net_features_check,
3530 .ndo_get_phys_port_name = nfp_net_get_phys_port_name,
3531 .ndo_udp_tunnel_add = nfp_net_add_vxlan_port,
3532 .ndo_udp_tunnel_del = nfp_net_del_vxlan_port,
3533 .ndo_bpf = nfp_net_xdp,
3537 * nfp_net_info() - Print general info about the NIC
3538 * @nn: NFP Net device to reconfigure
3540 void nfp_net_info(struct nfp_net *nn)
3542 nn_info(nn, "Netronome NFP-6xxx %sNetdev: TxQs=%d/%d RxQs=%d/%d\n",
3543 nn->dp.is_vf ? "VF " : "",
3544 nn->dp.num_tx_rings, nn->max_tx_rings,
3545 nn->dp.num_rx_rings, nn->max_rx_rings);
3546 nn_info(nn, "VER: %d.%d.%d.%d, Maximum supported MTU: %d\n",
3547 nn->fw_ver.resv, nn->fw_ver.class,
3548 nn->fw_ver.major, nn->fw_ver.minor,
3550 nn_info(nn, "CAP: %#x %s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n",
3552 nn->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "",
3553 nn->cap & NFP_NET_CFG_CTRL_L2BC ? "L2BCFILT " : "",
3554 nn->cap & NFP_NET_CFG_CTRL_L2MC ? "L2MCFILT " : "",
3555 nn->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "",
3556 nn->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "",
3557 nn->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "",
3558 nn->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "",
3559 nn->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "",
3560 nn->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "",
3561 nn->cap & NFP_NET_CFG_CTRL_LSO ? "TSO1 " : "",
3562 nn->cap & NFP_NET_CFG_CTRL_LSO2 ? "TSO2 " : "",
3563 nn->cap & NFP_NET_CFG_CTRL_RSS ? "RSS1 " : "",
3564 nn->cap & NFP_NET_CFG_CTRL_RSS2 ? "RSS2 " : "",
3565 nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER ? "CTAG_FILTER " : "",
3566 nn->cap & NFP_NET_CFG_CTRL_L2SWITCH ? "L2SWITCH " : "",
3567 nn->cap & NFP_NET_CFG_CTRL_MSIXAUTO ? "AUTOMASK " : "",
3568 nn->cap & NFP_NET_CFG_CTRL_IRQMOD ? "IRQMOD " : "",
3569 nn->cap & NFP_NET_CFG_CTRL_VXLAN ? "VXLAN " : "",
3570 nn->cap & NFP_NET_CFG_CTRL_NVGRE ? "NVGRE " : "",
3571 nn->cap & NFP_NET_CFG_CTRL_CSUM_COMPLETE ?
3572 "RXCSUM_COMPLETE " : "",
3573 nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR ? "LIVE_ADDR " : "",
3574 nfp_app_extra_cap(nn->app, nn));
3578 * nfp_net_alloc() - Allocate netdev and related structure
3580 * @ctrl_bar: PCI IOMEM with vNIC config memory
3581 * @needs_netdev: Whether to allocate a netdev for this vNIC
3582 * @max_tx_rings: Maximum number of TX rings supported by device
3583 * @max_rx_rings: Maximum number of RX rings supported by device
3585 * This function allocates a netdev device and fills in the initial
3586 * part of the @struct nfp_net structure. In case of control device
3587 * nfp_net structure is allocated without the netdev.
3589 * Return: NFP Net device structure, or ERR_PTR on error.
3592 nfp_net_alloc(struct pci_dev *pdev, void __iomem *ctrl_bar, bool needs_netdev,
3593 unsigned int max_tx_rings, unsigned int max_rx_rings)
3599 struct net_device *netdev;
3601 netdev = alloc_etherdev_mqs(sizeof(struct nfp_net),
3602 max_tx_rings, max_rx_rings);
3604 return ERR_PTR(-ENOMEM);
3606 SET_NETDEV_DEV(netdev, &pdev->dev);
3607 nn = netdev_priv(netdev);
3608 nn->dp.netdev = netdev;
3610 nn = vzalloc(sizeof(*nn));
3612 return ERR_PTR(-ENOMEM);
3615 nn->dp.dev = &pdev->dev;
3616 nn->dp.ctrl_bar = ctrl_bar;
3619 nn->max_tx_rings = max_tx_rings;
3620 nn->max_rx_rings = max_rx_rings;
3622 nn->dp.num_tx_rings = min_t(unsigned int,
3623 max_tx_rings, num_online_cpus());
3624 nn->dp.num_rx_rings = min_t(unsigned int, max_rx_rings,
3625 netif_get_num_default_rss_queues());
3627 nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings);
3628 nn->dp.num_r_vecs = min_t(unsigned int,
3629 nn->dp.num_r_vecs, num_online_cpus());
3631 nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT;
3632 nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT;
3634 spin_lock_init(&nn->reconfig_lock);
3635 spin_lock_init(&nn->link_status_lock);
3637 timer_setup(&nn->reconfig_timer, nfp_net_reconfig_timer, 0);
3639 err = nfp_net_tlv_caps_parse(&nn->pdev->dev, nn->dp.ctrl_bar,
3648 free_netdev(nn->dp.netdev);
3651 return ERR_PTR(err);
3655 * nfp_net_free() - Undo what @nfp_net_alloc() did
3656 * @nn: NFP Net device to reconfigure
3658 void nfp_net_free(struct nfp_net *nn)
3660 WARN_ON(timer_pending(&nn->reconfig_timer) || nn->reconfig_posted);
3662 free_netdev(nn->dp.netdev);
3668 * nfp_net_rss_key_sz() - Get current size of the RSS key
3669 * @nn: NFP Net device instance
3671 * Return: size of the RSS key for currently selected hash function.
3673 unsigned int nfp_net_rss_key_sz(struct nfp_net *nn)
3675 switch (nn->rss_hfunc) {
3676 case ETH_RSS_HASH_TOP:
3677 return NFP_NET_CFG_RSS_KEY_SZ;
3678 case ETH_RSS_HASH_XOR:
3680 case ETH_RSS_HASH_CRC32:
3684 nn_warn(nn, "Unknown hash function: %u\n", nn->rss_hfunc);
3689 * nfp_net_rss_init() - Set the initial RSS parameters
3690 * @nn: NFP Net device to reconfigure
3692 static void nfp_net_rss_init(struct nfp_net *nn)
3694 unsigned long func_bit, rss_cap_hfunc;
3697 /* Read the RSS function capability and select first supported func */
3698 reg = nn_readl(nn, NFP_NET_CFG_RSS_CAP);
3699 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, reg);
3701 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC,
3702 NFP_NET_CFG_RSS_TOEPLITZ);
3704 func_bit = find_first_bit(&rss_cap_hfunc, NFP_NET_CFG_RSS_HFUNCS);
3705 if (func_bit == NFP_NET_CFG_RSS_HFUNCS) {
3706 dev_warn(nn->dp.dev,
3707 "Bad RSS config, defaulting to Toeplitz hash\n");
3708 func_bit = ETH_RSS_HASH_TOP_BIT;
3710 nn->rss_hfunc = 1 << func_bit;
3712 netdev_rss_key_fill(nn->rss_key, nfp_net_rss_key_sz(nn));
3714 nfp_net_rss_init_itbl(nn);
3716 /* Enable IPv4/IPv6 TCP by default */
3717 nn->rss_cfg = NFP_NET_CFG_RSS_IPV4_TCP |
3718 NFP_NET_CFG_RSS_IPV6_TCP |
3719 FIELD_PREP(NFP_NET_CFG_RSS_HFUNC, nn->rss_hfunc) |
3720 NFP_NET_CFG_RSS_MASK;
3724 * nfp_net_irqmod_init() - Set the initial IRQ moderation parameters
3725 * @nn: NFP Net device to reconfigure
3727 static void nfp_net_irqmod_init(struct nfp_net *nn)
3729 nn->rx_coalesce_usecs = 50;
3730 nn->rx_coalesce_max_frames = 64;
3731 nn->tx_coalesce_usecs = 50;
3732 nn->tx_coalesce_max_frames = 64;
3735 static void nfp_net_netdev_init(struct nfp_net *nn)
3737 struct net_device *netdev = nn->dp.netdev;
3739 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr);
3741 netdev->mtu = nn->dp.mtu;
3743 /* Advertise/enable offloads based on capabilities
3745 * Note: netdev->features show the currently enabled features
3746 * and netdev->hw_features advertises which features are
3747 * supported. By default we enable most features.
3749 if (nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR)
3750 netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
3752 netdev->hw_features = NETIF_F_HIGHDMA;
3753 if (nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY) {
3754 netdev->hw_features |= NETIF_F_RXCSUM;
3755 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY;
3757 if (nn->cap & NFP_NET_CFG_CTRL_TXCSUM) {
3758 netdev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
3759 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXCSUM;
3761 if (nn->cap & NFP_NET_CFG_CTRL_GATHER) {
3762 netdev->hw_features |= NETIF_F_SG;
3763 nn->dp.ctrl |= NFP_NET_CFG_CTRL_GATHER;
3765 if ((nn->cap & NFP_NET_CFG_CTRL_LSO && nn->fw_ver.major > 2) ||
3766 nn->cap & NFP_NET_CFG_CTRL_LSO2) {
3767 netdev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
3768 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?:
3769 NFP_NET_CFG_CTRL_LSO;
3771 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY)
3772 netdev->hw_features |= NETIF_F_RXHASH;
3773 if (nn->cap & NFP_NET_CFG_CTRL_VXLAN) {
3774 if (nn->cap & NFP_NET_CFG_CTRL_LSO)
3775 netdev->hw_features |= NETIF_F_GSO_UDP_TUNNEL;
3776 nn->dp.ctrl |= NFP_NET_CFG_CTRL_VXLAN;
3778 if (nn->cap & NFP_NET_CFG_CTRL_NVGRE) {
3779 if (nn->cap & NFP_NET_CFG_CTRL_LSO)
3780 netdev->hw_features |= NETIF_F_GSO_GRE;
3781 nn->dp.ctrl |= NFP_NET_CFG_CTRL_NVGRE;
3783 if (nn->cap & (NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE))
3784 netdev->hw_enc_features = netdev->hw_features;
3786 netdev->vlan_features = netdev->hw_features;
3788 if (nn->cap & NFP_NET_CFG_CTRL_RXVLAN) {
3789 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
3790 nn->dp.ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
3792 if (nn->cap & NFP_NET_CFG_CTRL_TXVLAN) {
3793 if (nn->cap & NFP_NET_CFG_CTRL_LSO2) {
3794 nn_warn(nn, "Device advertises both TSO2 and TXVLAN. Refusing to enable TXVLAN.\n");
3796 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX;
3797 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
3800 if (nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER) {
3801 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3802 nn->dp.ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER;
3805 netdev->features = netdev->hw_features;
3807 if (nfp_app_has_tc(nn->app) && nn->port)
3808 netdev->hw_features |= NETIF_F_HW_TC;
3810 /* Advertise but disable TSO by default. */
3811 netdev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6);
3812 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY;
3814 /* Finalise the netdev setup */
3815 netdev->netdev_ops = &nfp_net_netdev_ops;
3816 netdev->watchdog_timeo = msecs_to_jiffies(5 * 1000);
3818 SWITCHDEV_SET_OPS(netdev, &nfp_port_switchdev_ops);
3820 /* MTU range: 68 - hw-specific max */
3821 netdev->min_mtu = ETH_MIN_MTU;
3822 netdev->max_mtu = nn->max_mtu;
3824 netdev->gso_max_segs = NFP_NET_LSO_MAX_SEGS;
3826 netif_carrier_off(netdev);
3828 nfp_net_set_ethtool_ops(netdev);
3831 static int nfp_net_read_caps(struct nfp_net *nn)
3833 /* Get some of the read-only fields from the BAR */
3834 nn->cap = nn_readl(nn, NFP_NET_CFG_CAP);
3835 nn->max_mtu = nn_readl(nn, NFP_NET_CFG_MAX_MTU);
3837 /* ABI 4.x and ctrl vNIC always use chained metadata, in other cases
3838 * we allow use of non-chained metadata if RSS(v1) is the only
3839 * advertised capability requiring metadata.
3841 nn->dp.chained_metadata_format = nn->fw_ver.major == 4 ||
3843 !(nn->cap & NFP_NET_CFG_CTRL_RSS) ||
3844 nn->cap & NFP_NET_CFG_CTRL_CHAIN_META;
3845 /* RSS(v1) uses non-chained metadata format, except in ABI 4.x where
3846 * it has the same meaning as RSSv2.
3848 if (nn->dp.chained_metadata_format && nn->fw_ver.major != 4)
3849 nn->cap &= ~NFP_NET_CFG_CTRL_RSS;
3851 /* Determine RX packet/metadata boundary offset */
3852 if (nn->fw_ver.major >= 2) {
3855 reg = nn_readl(nn, NFP_NET_CFG_RX_OFFSET);
3856 if (reg > NFP_NET_MAX_PREPEND) {
3857 nn_err(nn, "Invalid rx offset: %d\n", reg);
3860 nn->dp.rx_offset = reg;
3862 nn->dp.rx_offset = NFP_NET_RX_OFFSET;
3865 /* For control vNICs mask out the capabilities app doesn't want. */
3867 nn->cap &= nn->app->type->ctrl_cap_mask;
3873 * nfp_net_init() - Initialise/finalise the nfp_net structure
3874 * @nn: NFP Net device structure
3876 * Return: 0 on success or negative errno on error.
3878 int nfp_net_init(struct nfp_net *nn)
3882 nn->dp.rx_dma_dir = DMA_FROM_DEVICE;
3884 err = nfp_net_read_caps(nn);
3888 /* Set default MTU and Freelist buffer size */
3889 if (!nfp_net_is_data_vnic(nn) && nn->app->ctrl_mtu) {
3890 if (nn->app->ctrl_mtu <= nn->max_mtu) {
3891 nn->dp.mtu = nn->app->ctrl_mtu;
3893 if (nn->app->ctrl_mtu != NFP_APP_CTRL_MTU_MAX)
3894 nn_warn(nn, "app requested MTU above max supported %u > %u\n",
3895 nn->app->ctrl_mtu, nn->max_mtu);
3896 nn->dp.mtu = nn->max_mtu;
3898 } else if (nn->max_mtu < NFP_NET_DEFAULT_MTU) {
3899 nn->dp.mtu = nn->max_mtu;
3901 nn->dp.mtu = NFP_NET_DEFAULT_MTU;
3903 nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp);
3905 if (nfp_app_ctrl_uses_data_vnics(nn->app))
3906 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_CMSG_DATA;
3908 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) {
3909 nfp_net_rss_init(nn);
3910 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?:
3911 NFP_NET_CFG_CTRL_RSS;
3914 /* Allow L2 Broadcast and Multicast through by default, if supported */
3915 if (nn->cap & NFP_NET_CFG_CTRL_L2BC)
3916 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2BC;
3918 /* Allow IRQ moderation, if supported */
3919 if (nn->cap & NFP_NET_CFG_CTRL_IRQMOD) {
3920 nfp_net_irqmod_init(nn);
3921 nn->dp.ctrl |= NFP_NET_CFG_CTRL_IRQMOD;
3925 nfp_net_netdev_init(nn);
3927 /* Stash the re-configuration queue away. First odd queue in TX Bar */
3928 nn->qcp_cfg = nn->tx_bar + NFP_QCP_QUEUE_ADDR_SZ;
3930 /* Make sure the FW knows the netdev is supposed to be disabled here */
3931 nn_writel(nn, NFP_NET_CFG_CTRL, 0);
3932 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0);
3933 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0);
3934 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_RING |
3935 NFP_NET_CFG_UPDATE_GEN);
3939 nfp_net_vecs_init(nn);
3943 return register_netdev(nn->dp.netdev);
3947 * nfp_net_clean() - Undo what nfp_net_init() did.
3948 * @nn: NFP Net device structure
3950 void nfp_net_clean(struct nfp_net *nn)
3955 unregister_netdev(nn->dp.netdev);
3956 nfp_net_reconfig_wait_posted(nn);