2 * Copyright (C) 2015-2017 Netronome Systems, Inc.
4 * This software is dual licensed under the GNU General License Version 2,
5 * June 1991 as shown in the file COPYING in the top-level directory of this
6 * source tree or the BSD 2-Clause License provided below. You have the
7 * option to license this software under the complete terms of either license.
9 * The BSD 2-Clause License:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * 1. Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * 2. Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 * Netronome network device driver: Common functions between PF and VF
37 * Authors: Jakub Kicinski <jakub.kicinski@netronome.com>
38 * Jason McMullan <jason.mcmullan@netronome.com>
39 * Rolf Neugebauer <rolf.neugebauer@netronome.com>
40 * Brad Petrus <brad.petrus@netronome.com>
41 * Chris Telfer <chris.telfer@netronome.com>
44 #include <linux/bitfield.h>
45 #include <linux/bpf.h>
46 #include <linux/bpf_trace.h>
47 #include <linux/module.h>
48 #include <linux/kernel.h>
49 #include <linux/init.h>
51 #include <linux/netdevice.h>
52 #include <linux/etherdevice.h>
53 #include <linux/interrupt.h>
55 #include <linux/ipv6.h>
56 #include <linux/page_ref.h>
57 #include <linux/pci.h>
58 #include <linux/pci_regs.h>
59 #include <linux/msi.h>
60 #include <linux/ethtool.h>
61 #include <linux/log2.h>
62 #include <linux/if_vlan.h>
63 #include <linux/random.h>
64 #include <linux/vmalloc.h>
65 #include <linux/ktime.h>
67 #include <net/switchdev.h>
68 #include <net/vxlan.h>
70 #include "nfpcore/nfp_nsp.h"
72 #include "nfp_net_ctrl.h"
77 * nfp_net_get_fw_version() - Read and parse the FW version
78 * @fw_ver: Output fw_version structure to read to
79 * @ctrl_bar: Mapped address of the control BAR
81 void nfp_net_get_fw_version(struct nfp_net_fw_version *fw_ver,
82 void __iomem *ctrl_bar)
86 reg = readl(ctrl_bar + NFP_NET_CFG_VERSION);
87 put_unaligned_le32(reg, fw_ver);
90 static dma_addr_t nfp_net_dma_map_rx(struct nfp_net_dp *dp, void *frag)
92 return dma_map_single_attrs(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM,
93 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
94 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
98 nfp_net_dma_sync_dev_rx(const struct nfp_net_dp *dp, dma_addr_t dma_addr)
100 dma_sync_single_for_device(dp->dev, dma_addr,
101 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
105 static void nfp_net_dma_unmap_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr)
107 dma_unmap_single_attrs(dp->dev, dma_addr,
108 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
109 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
112 static void nfp_net_dma_sync_cpu_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr,
115 dma_sync_single_for_cpu(dp->dev, dma_addr - NFP_NET_RX_BUF_HEADROOM,
116 len, dp->rx_dma_dir);
121 * Firmware reconfig may take a while so we have two versions of it -
122 * synchronous and asynchronous (posted). All synchronous callers are holding
123 * RTNL so we don't have to worry about serializing them.
125 static void nfp_net_reconfig_start(struct nfp_net *nn, u32 update)
127 nn_writel(nn, NFP_NET_CFG_UPDATE, update);
128 /* ensure update is written before pinging HW */
130 nfp_qcp_wr_ptr_add(nn->qcp_cfg, 1);
133 /* Pass 0 as update to run posted reconfigs. */
134 static void nfp_net_reconfig_start_async(struct nfp_net *nn, u32 update)
136 update |= nn->reconfig_posted;
137 nn->reconfig_posted = 0;
139 nfp_net_reconfig_start(nn, update);
141 nn->reconfig_timer_active = true;
142 mod_timer(&nn->reconfig_timer, jiffies + NFP_NET_POLL_TIMEOUT * HZ);
145 static bool nfp_net_reconfig_check_done(struct nfp_net *nn, bool last_check)
149 reg = nn_readl(nn, NFP_NET_CFG_UPDATE);
152 if (reg & NFP_NET_CFG_UPDATE_ERR) {
153 nn_err(nn, "Reconfig error: 0x%08x\n", reg);
155 } else if (last_check) {
156 nn_err(nn, "Reconfig timeout: 0x%08x\n", reg);
163 static int nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline)
165 bool timed_out = false;
167 /* Poll update field, waiting for NFP to ack the config */
168 while (!nfp_net_reconfig_check_done(nn, timed_out)) {
170 timed_out = time_is_before_eq_jiffies(deadline);
173 if (nn_readl(nn, NFP_NET_CFG_UPDATE) & NFP_NET_CFG_UPDATE_ERR)
176 return timed_out ? -EIO : 0;
179 static void nfp_net_reconfig_timer(unsigned long data)
181 struct nfp_net *nn = (void *)data;
183 spin_lock_bh(&nn->reconfig_lock);
185 nn->reconfig_timer_active = false;
187 /* If sync caller is present it will take over from us */
188 if (nn->reconfig_sync_present)
191 /* Read reconfig status and report errors */
192 nfp_net_reconfig_check_done(nn, true);
194 if (nn->reconfig_posted)
195 nfp_net_reconfig_start_async(nn, 0);
197 spin_unlock_bh(&nn->reconfig_lock);
201 * nfp_net_reconfig_post() - Post async reconfig request
202 * @nn: NFP Net device to reconfigure
203 * @update: The value for the update field in the BAR config
205 * Record FW reconfiguration request. Reconfiguration will be kicked off
206 * whenever reconfiguration machinery is idle. Multiple requests can be
209 static void nfp_net_reconfig_post(struct nfp_net *nn, u32 update)
211 spin_lock_bh(&nn->reconfig_lock);
213 /* Sync caller will kick off async reconf when it's done, just post */
214 if (nn->reconfig_sync_present) {
215 nn->reconfig_posted |= update;
219 /* Opportunistically check if the previous command is done */
220 if (!nn->reconfig_timer_active ||
221 nfp_net_reconfig_check_done(nn, false))
222 nfp_net_reconfig_start_async(nn, update);
224 nn->reconfig_posted |= update;
226 spin_unlock_bh(&nn->reconfig_lock);
230 * nfp_net_reconfig() - Reconfigure the firmware
231 * @nn: NFP Net device to reconfigure
232 * @update: The value for the update field in the BAR config
234 * Write the update word to the BAR and ping the reconfig queue. The
235 * poll until the firmware has acknowledged the update by zeroing the
238 * Return: Negative errno on error, 0 on success
240 int nfp_net_reconfig(struct nfp_net *nn, u32 update)
242 bool cancelled_timer = false;
243 u32 pre_posted_requests;
246 spin_lock_bh(&nn->reconfig_lock);
248 nn->reconfig_sync_present = true;
250 if (nn->reconfig_timer_active) {
251 del_timer(&nn->reconfig_timer);
252 nn->reconfig_timer_active = false;
253 cancelled_timer = true;
255 pre_posted_requests = nn->reconfig_posted;
256 nn->reconfig_posted = 0;
258 spin_unlock_bh(&nn->reconfig_lock);
261 nfp_net_reconfig_wait(nn, nn->reconfig_timer.expires);
263 /* Run the posted reconfigs which were issued before we started */
264 if (pre_posted_requests) {
265 nfp_net_reconfig_start(nn, pre_posted_requests);
266 nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT);
269 nfp_net_reconfig_start(nn, update);
270 ret = nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT);
272 spin_lock_bh(&nn->reconfig_lock);
274 if (nn->reconfig_posted)
275 nfp_net_reconfig_start_async(nn, 0);
277 nn->reconfig_sync_present = false;
279 spin_unlock_bh(&nn->reconfig_lock);
285 * nfp_net_reconfig_mbox() - Reconfigure the firmware via the mailbox
286 * @nn: NFP Net device to reconfigure
287 * @mbox_cmd: The value for the mailbox command
289 * Helper function for mailbox updates
291 * Return: Negative errno on error, 0 on success
293 static int nfp_net_reconfig_mbox(struct nfp_net *nn, u32 mbox_cmd)
297 nn_writeq(nn, NFP_NET_CFG_MBOX_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, NFP_NET_CFG_MBOX_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;
517 tx_ring->qcidx = tx_ring->idx * nn->stride_tx;
518 tx_ring->qcp_q = nn->tx_bar + NFP_QCP_QUEUE_OFF(tx_ring->qcidx);
522 * nfp_net_rx_ring_init() - Fill in the boilerplate for a RX ring
523 * @rx_ring: RX ring structure
524 * @r_vec: IRQ vector servicing this ring
528 nfp_net_rx_ring_init(struct nfp_net_rx_ring *rx_ring,
529 struct nfp_net_r_vector *r_vec, unsigned int idx)
531 struct nfp_net *nn = r_vec->nfp_net;
534 rx_ring->r_vec = r_vec;
536 rx_ring->fl_qcidx = rx_ring->idx * nn->stride_rx;
537 rx_ring->qcp_fl = nn->rx_bar + NFP_QCP_QUEUE_OFF(rx_ring->fl_qcidx);
541 * nfp_net_aux_irq_request() - Request an auxiliary interrupt (LSC or EXN)
542 * @nn: NFP Network structure
543 * @ctrl_offset: Control BAR offset where IRQ configuration should be written
544 * @format: printf-style format to construct the interrupt name
545 * @name: Pointer to allocated space for interrupt name
546 * @name_sz: Size of space for interrupt name
547 * @vector_idx: Index of MSI-X vector used for this interrupt
548 * @handler: IRQ handler to register for this interrupt
551 nfp_net_aux_irq_request(struct nfp_net *nn, u32 ctrl_offset,
552 const char *format, char *name, size_t name_sz,
553 unsigned int vector_idx, irq_handler_t handler)
555 struct msix_entry *entry;
558 entry = &nn->irq_entries[vector_idx];
560 snprintf(name, name_sz, format, nfp_net_name(nn));
561 err = request_irq(entry->vector, handler, 0, name, nn);
563 nn_err(nn, "Failed to request IRQ %d (err=%d).\n",
567 nn_writeb(nn, ctrl_offset, entry->entry);
573 * nfp_net_aux_irq_free() - Free an auxiliary interrupt (LSC or EXN)
574 * @nn: NFP Network structure
575 * @ctrl_offset: Control BAR offset where IRQ configuration should be written
576 * @vector_idx: Index of MSI-X vector used for this interrupt
578 static void nfp_net_aux_irq_free(struct nfp_net *nn, u32 ctrl_offset,
579 unsigned int vector_idx)
581 nn_writeb(nn, ctrl_offset, 0xff);
582 free_irq(nn->irq_entries[vector_idx].vector, nn);
587 * One queue controller peripheral queue is used for transmit. The
588 * driver en-queues packets for transmit by advancing the write
589 * pointer. The device indicates that packets have transmitted by
590 * advancing the read pointer. The driver maintains a local copy of
591 * the read and write pointer in @struct nfp_net_tx_ring. The driver
592 * keeps @wr_p in sync with the queue controller write pointer and can
593 * determine how many packets have been transmitted by comparing its
594 * copy of the read pointer @rd_p with the read pointer maintained by
595 * the queue controller peripheral.
599 * nfp_net_tx_full() - Check if the TX ring is full
600 * @tx_ring: TX ring to check
601 * @dcnt: Number of descriptors that need to be enqueued (must be >= 1)
603 * This function checks, based on the *host copy* of read/write
604 * pointer if a given TX ring is full. The real TX queue may have
605 * some newly made available slots.
607 * Return: True if the ring is full.
609 static int nfp_net_tx_full(struct nfp_net_tx_ring *tx_ring, int dcnt)
611 return (tx_ring->wr_p - tx_ring->rd_p) >= (tx_ring->cnt - dcnt);
614 /* Wrappers for deciding when to stop and restart TX queues */
615 static int nfp_net_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
617 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
620 static int nfp_net_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
622 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
626 * nfp_net_tx_ring_stop() - stop tx ring
627 * @nd_q: netdev queue
628 * @tx_ring: driver tx queue structure
630 * Safely stop TX ring. Remember that while we are running .start_xmit()
631 * someone else may be cleaning the TX ring completions so we need to be
632 * extra careful here.
634 static void nfp_net_tx_ring_stop(struct netdev_queue *nd_q,
635 struct nfp_net_tx_ring *tx_ring)
637 netif_tx_stop_queue(nd_q);
639 /* We can race with the TX completion out of NAPI so recheck */
641 if (unlikely(nfp_net_tx_ring_should_wake(tx_ring)))
642 netif_tx_start_queue(nd_q);
646 * nfp_net_tx_tso() - Set up Tx descriptor for LSO
647 * @r_vec: per-ring structure
648 * @txbuf: Pointer to driver soft TX descriptor
649 * @txd: Pointer to HW TX descriptor
650 * @skb: Pointer to SKB
652 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
653 * Return error on packet header greater than maximum supported LSO header size.
655 static void nfp_net_tx_tso(struct nfp_net_r_vector *r_vec,
656 struct nfp_net_tx_buf *txbuf,
657 struct nfp_net_tx_desc *txd, struct sk_buff *skb)
662 if (!skb_is_gso(skb))
665 if (!skb->encapsulation) {
666 txd->l3_offset = skb_network_offset(skb);
667 txd->l4_offset = skb_transport_offset(skb);
668 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
670 txd->l3_offset = skb_inner_network_offset(skb);
671 txd->l4_offset = skb_inner_transport_offset(skb);
672 hdrlen = skb_inner_transport_header(skb) - skb->data +
673 inner_tcp_hdrlen(skb);
676 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
677 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
679 mss = skb_shinfo(skb)->gso_size & PCIE_DESC_TX_MSS_MASK;
680 txd->lso_hdrlen = hdrlen;
681 txd->mss = cpu_to_le16(mss);
682 txd->flags |= PCIE_DESC_TX_LSO;
684 u64_stats_update_begin(&r_vec->tx_sync);
686 u64_stats_update_end(&r_vec->tx_sync);
690 * nfp_net_tx_csum() - Set TX CSUM offload flags in TX descriptor
691 * @dp: NFP Net data path struct
692 * @r_vec: per-ring structure
693 * @txbuf: Pointer to driver soft TX descriptor
694 * @txd: Pointer to TX descriptor
695 * @skb: Pointer to SKB
697 * This function sets the TX checksum flags in the TX descriptor based
698 * on the configuration and the protocol of the packet to be transmitted.
700 static void nfp_net_tx_csum(struct nfp_net_dp *dp,
701 struct nfp_net_r_vector *r_vec,
702 struct nfp_net_tx_buf *txbuf,
703 struct nfp_net_tx_desc *txd, struct sk_buff *skb)
705 struct ipv6hdr *ipv6h;
709 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
712 if (skb->ip_summed != CHECKSUM_PARTIAL)
715 txd->flags |= PCIE_DESC_TX_CSUM;
716 if (skb->encapsulation)
717 txd->flags |= PCIE_DESC_TX_ENCAP;
719 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
720 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
722 if (iph->version == 4) {
723 txd->flags |= PCIE_DESC_TX_IP4_CSUM;
724 l4_hdr = iph->protocol;
725 } else if (ipv6h->version == 6) {
726 l4_hdr = ipv6h->nexthdr;
728 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
734 txd->flags |= PCIE_DESC_TX_TCP_CSUM;
737 txd->flags |= PCIE_DESC_TX_UDP_CSUM;
740 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
744 u64_stats_update_begin(&r_vec->tx_sync);
745 if (skb->encapsulation)
746 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
748 r_vec->hw_csum_tx += txbuf->pkt_cnt;
749 u64_stats_update_end(&r_vec->tx_sync);
752 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring)
755 nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add);
756 tx_ring->wr_ptr_add = 0;
759 static int nfp_net_prep_port_id(struct sk_buff *skb)
761 struct metadata_dst *md_dst = skb_metadata_dst(skb);
766 if (unlikely(md_dst->type != METADATA_HW_PORT_MUX))
769 if (unlikely(skb_cow_head(skb, 8)))
772 data = skb_push(skb, 8);
773 put_unaligned_be32(NFP_NET_META_PORTID, data);
774 put_unaligned_be32(md_dst->u.port_info.port_id, data + 4);
780 * nfp_net_tx() - Main transmit entry point
781 * @skb: SKB to transmit
782 * @netdev: netdev structure
784 * Return: NETDEV_TX_OK on success.
786 static int nfp_net_tx(struct sk_buff *skb, struct net_device *netdev)
788 struct nfp_net *nn = netdev_priv(netdev);
789 const struct skb_frag_struct *frag;
790 struct nfp_net_tx_desc *txd, txdg;
791 int f, nr_frags, wr_idx, md_bytes;
792 struct nfp_net_tx_ring *tx_ring;
793 struct nfp_net_r_vector *r_vec;
794 struct nfp_net_tx_buf *txbuf;
795 struct netdev_queue *nd_q;
796 struct nfp_net_dp *dp;
802 qidx = skb_get_queue_mapping(skb);
803 tx_ring = &dp->tx_rings[qidx];
804 r_vec = tx_ring->r_vec;
805 nd_q = netdev_get_tx_queue(dp->netdev, qidx);
807 nr_frags = skb_shinfo(skb)->nr_frags;
809 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
810 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
811 qidx, tx_ring->wr_p, tx_ring->rd_p);
812 netif_tx_stop_queue(nd_q);
813 nfp_net_tx_xmit_more_flush(tx_ring);
814 u64_stats_update_begin(&r_vec->tx_sync);
816 u64_stats_update_end(&r_vec->tx_sync);
817 return NETDEV_TX_BUSY;
820 md_bytes = nfp_net_prep_port_id(skb);
821 if (unlikely(md_bytes < 0)) {
822 nfp_net_tx_xmit_more_flush(tx_ring);
823 dev_kfree_skb_any(skb);
827 /* Start with the head skbuf */
828 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
830 if (dma_mapping_error(dp->dev, dma_addr))
833 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
835 /* Stash the soft descriptor of the head then initialize it */
836 txbuf = &tx_ring->txbufs[wr_idx];
838 txbuf->dma_addr = dma_addr;
841 txbuf->real_len = skb->len;
843 /* Build TX descriptor */
844 txd = &tx_ring->txds[wr_idx];
845 txd->offset_eop = (nr_frags ? 0 : PCIE_DESC_TX_EOP) | md_bytes;
846 txd->dma_len = cpu_to_le16(skb_headlen(skb));
847 nfp_desc_set_dma_addr(txd, dma_addr);
848 txd->data_len = cpu_to_le16(skb->len);
854 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
855 nfp_net_tx_tso(r_vec, txbuf, txd, skb);
856 nfp_net_tx_csum(dp, r_vec, txbuf, txd, skb);
857 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
858 txd->flags |= PCIE_DESC_TX_VLAN;
859 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
864 /* all descs must match except for in addr, length and eop */
867 for (f = 0; f < nr_frags; f++) {
868 frag = &skb_shinfo(skb)->frags[f];
869 fsize = skb_frag_size(frag);
871 dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
872 fsize, DMA_TO_DEVICE);
873 if (dma_mapping_error(dp->dev, dma_addr))
876 wr_idx = D_IDX(tx_ring, wr_idx + 1);
877 tx_ring->txbufs[wr_idx].skb = skb;
878 tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
879 tx_ring->txbufs[wr_idx].fidx = f;
881 txd = &tx_ring->txds[wr_idx];
883 txd->dma_len = cpu_to_le16(fsize);
884 nfp_desc_set_dma_addr(txd, dma_addr);
886 (f == nr_frags - 1) ? PCIE_DESC_TX_EOP : 0;
889 u64_stats_update_begin(&r_vec->tx_sync);
891 u64_stats_update_end(&r_vec->tx_sync);
894 netdev_tx_sent_queue(nd_q, txbuf->real_len);
896 tx_ring->wr_p += nr_frags + 1;
897 if (nfp_net_tx_ring_should_stop(tx_ring))
898 nfp_net_tx_ring_stop(nd_q, tx_ring);
900 tx_ring->wr_ptr_add += nr_frags + 1;
901 if (!skb->xmit_more || netif_xmit_stopped(nd_q))
902 nfp_net_tx_xmit_more_flush(tx_ring);
904 skb_tx_timestamp(skb);
911 frag = &skb_shinfo(skb)->frags[f];
912 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
913 skb_frag_size(frag), DMA_TO_DEVICE);
914 tx_ring->txbufs[wr_idx].skb = NULL;
915 tx_ring->txbufs[wr_idx].dma_addr = 0;
916 tx_ring->txbufs[wr_idx].fidx = -2;
919 wr_idx += tx_ring->cnt;
921 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
922 skb_headlen(skb), 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;
927 nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
928 nfp_net_tx_xmit_more_flush(tx_ring);
929 u64_stats_update_begin(&r_vec->tx_sync);
931 u64_stats_update_end(&r_vec->tx_sync);
932 dev_kfree_skb_any(skb);
937 * nfp_net_tx_complete() - Handled completed TX packets
938 * @tx_ring: TX ring structure
940 * Return: Number of completed TX descriptors
942 static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring)
944 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
945 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
946 const struct skb_frag_struct *frag;
947 struct netdev_queue *nd_q;
948 u32 done_pkts = 0, done_bytes = 0;
955 if (tx_ring->wr_p == tx_ring->rd_p)
958 /* Work out how many descriptors have been transmitted */
959 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
961 if (qcp_rd_p == tx_ring->qcp_rd_p)
964 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
967 idx = D_IDX(tx_ring, tx_ring->rd_p++);
969 skb = tx_ring->txbufs[idx].skb;
973 nr_frags = skb_shinfo(skb)->nr_frags;
974 fidx = tx_ring->txbufs[idx].fidx;
978 dma_unmap_single(dp->dev, tx_ring->txbufs[idx].dma_addr,
979 skb_headlen(skb), DMA_TO_DEVICE);
981 done_pkts += tx_ring->txbufs[idx].pkt_cnt;
982 done_bytes += tx_ring->txbufs[idx].real_len;
985 frag = &skb_shinfo(skb)->frags[fidx];
986 dma_unmap_page(dp->dev, tx_ring->txbufs[idx].dma_addr,
987 skb_frag_size(frag), DMA_TO_DEVICE);
990 /* check for last gather fragment */
991 if (fidx == nr_frags - 1)
992 dev_kfree_skb_any(skb);
994 tx_ring->txbufs[idx].dma_addr = 0;
995 tx_ring->txbufs[idx].skb = NULL;
996 tx_ring->txbufs[idx].fidx = -2;
999 tx_ring->qcp_rd_p = qcp_rd_p;
1001 u64_stats_update_begin(&r_vec->tx_sync);
1002 r_vec->tx_bytes += done_bytes;
1003 r_vec->tx_pkts += done_pkts;
1004 u64_stats_update_end(&r_vec->tx_sync);
1009 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
1010 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
1011 if (nfp_net_tx_ring_should_wake(tx_ring)) {
1012 /* Make sure TX thread will see updated tx_ring->rd_p */
1015 if (unlikely(netif_tx_queue_stopped(nd_q)))
1016 netif_tx_wake_queue(nd_q);
1019 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
1020 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
1021 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
1024 static bool nfp_net_xdp_complete(struct nfp_net_tx_ring *tx_ring)
1026 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
1027 u32 done_pkts = 0, done_bytes = 0;
1032 /* Work out how many descriptors have been transmitted */
1033 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
1035 if (qcp_rd_p == tx_ring->qcp_rd_p)
1038 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
1040 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
1041 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
1043 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
1047 idx = D_IDX(tx_ring, tx_ring->rd_p);
1050 done_bytes += tx_ring->txbufs[idx].real_len;
1053 u64_stats_update_begin(&r_vec->tx_sync);
1054 r_vec->tx_bytes += done_bytes;
1055 r_vec->tx_pkts += done_pkts;
1056 u64_stats_update_end(&r_vec->tx_sync);
1058 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
1059 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
1060 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
1066 * nfp_net_tx_ring_reset() - Free any untransmitted buffers and reset pointers
1067 * @dp: NFP Net data path struct
1068 * @tx_ring: TX ring structure
1070 * Assumes that the device is stopped
1073 nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
1075 const struct skb_frag_struct *frag;
1076 struct netdev_queue *nd_q;
1078 while (!tx_ring->is_xdp && tx_ring->rd_p != tx_ring->wr_p) {
1079 struct nfp_net_tx_buf *tx_buf;
1080 struct sk_buff *skb;
1083 idx = D_IDX(tx_ring, tx_ring->rd_p);
1084 tx_buf = &tx_ring->txbufs[idx];
1086 skb = tx_ring->txbufs[idx].skb;
1087 nr_frags = skb_shinfo(skb)->nr_frags;
1089 if (tx_buf->fidx == -1) {
1091 dma_unmap_single(dp->dev, tx_buf->dma_addr,
1092 skb_headlen(skb), DMA_TO_DEVICE);
1094 /* unmap fragment */
1095 frag = &skb_shinfo(skb)->frags[tx_buf->fidx];
1096 dma_unmap_page(dp->dev, tx_buf->dma_addr,
1097 skb_frag_size(frag), DMA_TO_DEVICE);
1100 /* check for last gather fragment */
1101 if (tx_buf->fidx == nr_frags - 1)
1102 dev_kfree_skb_any(skb);
1104 tx_buf->dma_addr = 0;
1108 tx_ring->qcp_rd_p++;
1112 memset(tx_ring->txds, 0, sizeof(*tx_ring->txds) * tx_ring->cnt);
1115 tx_ring->qcp_rd_p = 0;
1116 tx_ring->wr_ptr_add = 0;
1118 if (tx_ring->is_xdp || !dp->netdev)
1121 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
1122 netdev_tx_reset_queue(nd_q);
1125 static void nfp_net_tx_timeout(struct net_device *netdev)
1127 struct nfp_net *nn = netdev_priv(netdev);
1130 for (i = 0; i < nn->dp.netdev->real_num_tx_queues; i++) {
1131 if (!netif_tx_queue_stopped(netdev_get_tx_queue(netdev, i)))
1133 nn_warn(nn, "TX timeout on ring: %d\n", i);
1135 nn_warn(nn, "TX watchdog timeout\n");
1138 /* Receive processing
1141 nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp)
1143 unsigned int fl_bufsz;
1145 fl_bufsz = NFP_NET_RX_BUF_HEADROOM;
1146 fl_bufsz += dp->rx_dma_off;
1147 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1148 fl_bufsz += NFP_NET_MAX_PREPEND;
1150 fl_bufsz += dp->rx_offset;
1151 fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu;
1153 fl_bufsz = SKB_DATA_ALIGN(fl_bufsz);
1154 fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1160 nfp_net_free_frag(void *frag, bool xdp)
1163 skb_free_frag(frag);
1165 __free_page(virt_to_page(frag));
1169 * nfp_net_rx_alloc_one() - Allocate and map page frag for RX
1170 * @dp: NFP Net data path struct
1171 * @dma_addr: Pointer to storage for DMA address (output param)
1173 * This function will allcate a new page frag, map it for DMA.
1175 * Return: allocated page frag or NULL on failure.
1177 static void *nfp_net_rx_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
1182 frag = netdev_alloc_frag(dp->fl_bufsz);
1184 frag = page_address(alloc_page(GFP_KERNEL | __GFP_COLD));
1186 nn_dp_warn(dp, "Failed to alloc receive page frag\n");
1190 *dma_addr = nfp_net_dma_map_rx(dp, frag);
1191 if (dma_mapping_error(dp->dev, *dma_addr)) {
1192 nfp_net_free_frag(frag, dp->xdp_prog);
1193 nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
1200 static void *nfp_net_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
1205 frag = napi_alloc_frag(dp->fl_bufsz);
1207 frag = page_address(alloc_page(GFP_ATOMIC | __GFP_COLD));
1209 nn_dp_warn(dp, "Failed to alloc receive page frag\n");
1213 *dma_addr = nfp_net_dma_map_rx(dp, frag);
1214 if (dma_mapping_error(dp->dev, *dma_addr)) {
1215 nfp_net_free_frag(frag, dp->xdp_prog);
1216 nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
1224 * nfp_net_rx_give_one() - Put mapped skb on the software and hardware rings
1225 * @dp: NFP Net data path struct
1226 * @rx_ring: RX ring structure
1227 * @frag: page fragment buffer
1228 * @dma_addr: DMA address of skb mapping
1230 static void nfp_net_rx_give_one(const struct nfp_net_dp *dp,
1231 struct nfp_net_rx_ring *rx_ring,
1232 void *frag, dma_addr_t dma_addr)
1234 unsigned int wr_idx;
1236 wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
1238 nfp_net_dma_sync_dev_rx(dp, dma_addr);
1240 /* Stash SKB and DMA address away */
1241 rx_ring->rxbufs[wr_idx].frag = frag;
1242 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
1244 /* Fill freelist descriptor */
1245 rx_ring->rxds[wr_idx].fld.reserved = 0;
1246 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
1247 nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld,
1248 dma_addr + dp->rx_dma_off);
1251 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
1252 /* Update write pointer of the freelist queue. Make
1253 * sure all writes are flushed before telling the hardware.
1256 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
1261 * nfp_net_rx_ring_reset() - Reflect in SW state of freelist after disable
1262 * @rx_ring: RX ring structure
1264 * Warning: Do *not* call if ring buffers were never put on the FW freelist
1265 * (i.e. device was not enabled)!
1267 static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring)
1269 unsigned int wr_idx, last_idx;
1271 /* Move the empty entry to the end of the list */
1272 wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
1273 last_idx = rx_ring->cnt - 1;
1274 rx_ring->rxbufs[wr_idx].dma_addr = rx_ring->rxbufs[last_idx].dma_addr;
1275 rx_ring->rxbufs[wr_idx].frag = rx_ring->rxbufs[last_idx].frag;
1276 rx_ring->rxbufs[last_idx].dma_addr = 0;
1277 rx_ring->rxbufs[last_idx].frag = NULL;
1279 memset(rx_ring->rxds, 0, sizeof(*rx_ring->rxds) * rx_ring->cnt);
1285 * nfp_net_rx_ring_bufs_free() - Free any buffers currently on the RX ring
1286 * @dp: NFP Net data path struct
1287 * @rx_ring: RX ring to remove buffers from
1289 * Assumes that the device is stopped and buffers are in [0, ring->cnt - 1)
1290 * entries. After device is disabled nfp_net_rx_ring_reset() must be called
1291 * to restore required ring geometry.
1294 nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp,
1295 struct nfp_net_rx_ring *rx_ring)
1299 for (i = 0; i < rx_ring->cnt - 1; i++) {
1300 /* NULL skb can only happen when initial filling of the ring
1301 * fails to allocate enough buffers and calls here to free
1302 * already allocated ones.
1304 if (!rx_ring->rxbufs[i].frag)
1307 nfp_net_dma_unmap_rx(dp, rx_ring->rxbufs[i].dma_addr);
1308 nfp_net_free_frag(rx_ring->rxbufs[i].frag, dp->xdp_prog);
1309 rx_ring->rxbufs[i].dma_addr = 0;
1310 rx_ring->rxbufs[i].frag = NULL;
1315 * nfp_net_rx_ring_bufs_alloc() - Fill RX ring with buffers (don't give to FW)
1316 * @dp: NFP Net data path struct
1317 * @rx_ring: RX ring to remove buffers from
1320 nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp,
1321 struct nfp_net_rx_ring *rx_ring)
1323 struct nfp_net_rx_buf *rxbufs;
1326 rxbufs = rx_ring->rxbufs;
1328 for (i = 0; i < rx_ring->cnt - 1; i++) {
1329 rxbufs[i].frag = nfp_net_rx_alloc_one(dp, &rxbufs[i].dma_addr);
1330 if (!rxbufs[i].frag) {
1331 nfp_net_rx_ring_bufs_free(dp, rx_ring);
1340 * nfp_net_rx_ring_fill_freelist() - Give buffers from the ring to FW
1341 * @dp: NFP Net data path struct
1342 * @rx_ring: RX ring to fill
1345 nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp,
1346 struct nfp_net_rx_ring *rx_ring)
1350 for (i = 0; i < rx_ring->cnt - 1; i++)
1351 nfp_net_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
1352 rx_ring->rxbufs[i].dma_addr);
1356 * nfp_net_rx_csum_has_errors() - group check if rxd has any csum errors
1357 * @flags: RX descriptor flags field in CPU byte order
1359 static int nfp_net_rx_csum_has_errors(u16 flags)
1361 u16 csum_all_checked, csum_all_ok;
1363 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
1364 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
1366 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
1370 * nfp_net_rx_csum() - set SKB checksum field based on RX descriptor flags
1371 * @dp: NFP Net data path struct
1372 * @r_vec: per-ring structure
1373 * @rxd: Pointer to RX descriptor
1374 * @meta: Parsed metadata prepend
1375 * @skb: Pointer to SKB
1377 static void nfp_net_rx_csum(struct nfp_net_dp *dp,
1378 struct nfp_net_r_vector *r_vec,
1379 struct nfp_net_rx_desc *rxd,
1380 struct nfp_meta_parsed *meta, struct sk_buff *skb)
1382 skb_checksum_none_assert(skb);
1384 if (!(dp->netdev->features & NETIF_F_RXCSUM))
1387 if (meta->csum_type) {
1388 skb->ip_summed = meta->csum_type;
1389 skb->csum = meta->csum;
1390 u64_stats_update_begin(&r_vec->rx_sync);
1391 r_vec->hw_csum_rx_ok++;
1392 u64_stats_update_end(&r_vec->rx_sync);
1396 if (nfp_net_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
1397 u64_stats_update_begin(&r_vec->rx_sync);
1398 r_vec->hw_csum_rx_error++;
1399 u64_stats_update_end(&r_vec->rx_sync);
1403 /* Assume that the firmware will never report inner CSUM_OK unless outer
1404 * L4 headers were successfully parsed. FW will always report zero UDP
1405 * checksum as CSUM_OK.
1407 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
1408 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
1409 __skb_incr_checksum_unnecessary(skb);
1410 u64_stats_update_begin(&r_vec->rx_sync);
1411 r_vec->hw_csum_rx_ok++;
1412 u64_stats_update_end(&r_vec->rx_sync);
1415 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
1416 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
1417 __skb_incr_checksum_unnecessary(skb);
1418 u64_stats_update_begin(&r_vec->rx_sync);
1419 r_vec->hw_csum_rx_inner_ok++;
1420 u64_stats_update_end(&r_vec->rx_sync);
1425 nfp_net_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
1426 unsigned int type, __be32 *hash)
1428 if (!(netdev->features & NETIF_F_RXHASH))
1432 case NFP_NET_RSS_IPV4:
1433 case NFP_NET_RSS_IPV6:
1434 case NFP_NET_RSS_IPV6_EX:
1435 meta->hash_type = PKT_HASH_TYPE_L3;
1438 meta->hash_type = PKT_HASH_TYPE_L4;
1442 meta->hash = get_unaligned_be32(hash);
1446 nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
1447 void *data, struct nfp_net_rx_desc *rxd)
1449 struct nfp_net_rx_hash *rx_hash = data;
1451 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
1454 nfp_net_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
1459 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
1460 void *data, int meta_len)
1464 meta_info = get_unaligned_be32(data);
1468 switch (meta_info & NFP_NET_META_FIELD_MASK) {
1469 case NFP_NET_META_HASH:
1470 meta_info >>= NFP_NET_META_FIELD_SIZE;
1471 nfp_net_set_hash(netdev, meta,
1472 meta_info & NFP_NET_META_FIELD_MASK,
1476 case NFP_NET_META_MARK:
1477 meta->mark = get_unaligned_be32(data);
1480 case NFP_NET_META_PORTID:
1481 meta->portid = get_unaligned_be32(data);
1484 case NFP_NET_META_CSUM:
1485 meta->csum_type = CHECKSUM_COMPLETE;
1487 (__force __wsum)__get_unaligned_cpu32(data);
1494 meta_info >>= NFP_NET_META_FIELD_SIZE;
1501 nfp_net_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
1502 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
1503 struct sk_buff *skb)
1505 u64_stats_update_begin(&r_vec->rx_sync);
1507 u64_stats_update_end(&r_vec->rx_sync);
1509 /* skb is build based on the frag, free_skb() would free the frag
1510 * so to be able to reuse it we need an extra ref.
1512 if (skb && rxbuf && skb->head == rxbuf->frag)
1513 page_ref_inc(virt_to_head_page(rxbuf->frag));
1515 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
1517 dev_kfree_skb_any(skb);
1521 nfp_net_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
1522 struct nfp_net_tx_ring *tx_ring,
1523 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
1524 unsigned int pkt_len, bool *completed)
1526 struct nfp_net_tx_buf *txbuf;
1527 struct nfp_net_tx_desc *txd;
1530 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1532 nfp_net_xdp_complete(tx_ring);
1536 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1537 nfp_net_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
1543 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1545 /* Stash the soft descriptor of the head then initialize it */
1546 txbuf = &tx_ring->txbufs[wr_idx];
1548 nfp_net_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
1550 txbuf->frag = rxbuf->frag;
1551 txbuf->dma_addr = rxbuf->dma_addr;
1554 txbuf->real_len = pkt_len;
1556 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
1557 pkt_len, DMA_BIDIRECTIONAL);
1559 /* Build TX descriptor */
1560 txd = &tx_ring->txds[wr_idx];
1561 txd->offset_eop = PCIE_DESC_TX_EOP;
1562 txd->dma_len = cpu_to_le16(pkt_len);
1563 nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off);
1564 txd->data_len = cpu_to_le16(pkt_len);
1568 txd->lso_hdrlen = 0;
1571 tx_ring->wr_ptr_add++;
1575 static int nfp_net_run_xdp(struct bpf_prog *prog, void *data, void *hard_start,
1576 unsigned int *off, unsigned int *len)
1578 struct xdp_buff xdp;
1582 xdp.data_hard_start = hard_start;
1583 xdp.data = data + *off;
1584 xdp.data_end = data + *off + *len;
1586 orig_data = xdp.data;
1587 ret = bpf_prog_run_xdp(prog, &xdp);
1589 *len -= xdp.data - orig_data;
1590 *off += xdp.data - orig_data;
1596 * nfp_net_rx() - receive up to @budget packets on @rx_ring
1597 * @rx_ring: RX ring to receive from
1598 * @budget: NAPI budget
1600 * Note, this function is separated out from the napi poll function to
1601 * more cleanly separate packet receive code from other bookkeeping
1602 * functions performed in the napi poll function.
1604 * Return: Number of packets received.
1606 static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
1608 struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
1609 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
1610 struct nfp_net_tx_ring *tx_ring;
1611 struct bpf_prog *xdp_prog;
1612 bool xdp_tx_cmpl = false;
1613 unsigned int true_bufsz;
1614 struct sk_buff *skb;
1615 int pkts_polled = 0;
1619 xdp_prog = READ_ONCE(dp->xdp_prog);
1620 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
1621 tx_ring = r_vec->xdp_ring;
1623 while (pkts_polled < budget) {
1624 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1625 struct nfp_net_rx_buf *rxbuf;
1626 struct nfp_net_rx_desc *rxd;
1627 struct nfp_meta_parsed meta;
1628 struct net_device *netdev;
1629 dma_addr_t new_dma_addr;
1632 idx = D_IDX(rx_ring, rx_ring->rd_p);
1634 rxd = &rx_ring->rxds[idx];
1635 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1638 /* Memory barrier to ensure that we won't do other reads
1639 * before the DD bit.
1643 memset(&meta, 0, sizeof(meta));
1648 rxbuf = &rx_ring->rxbufs[idx];
1650 * <-- [rx_offset] -->
1651 * ---------------------------------------------------------
1652 * | [XX] | metadata | packet | XXXX |
1653 * ---------------------------------------------------------
1654 * <---------------- data_len --------------->
1656 * The rx_offset is fixed for all packets, the meta_len can vary
1657 * on a packet by packet basis. If rx_offset is set to zero
1658 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
1659 * buffer and is immediately followed by the packet (no [XX]).
1661 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1662 data_len = le16_to_cpu(rxd->rxd.data_len);
1663 pkt_len = data_len - meta_len;
1665 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1666 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1667 pkt_off += meta_len;
1669 pkt_off += dp->rx_offset;
1670 meta_off = pkt_off - meta_len;
1673 u64_stats_update_begin(&r_vec->rx_sync);
1675 r_vec->rx_bytes += pkt_len;
1676 u64_stats_update_end(&r_vec->rx_sync);
1678 if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
1679 (dp->rx_offset && meta_len > dp->rx_offset))) {
1680 nn_dp_warn(dp, "oversized RX packet metadata %u\n",
1682 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1686 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
1689 if (!dp->chained_metadata_format) {
1690 nfp_net_set_hash_desc(dp->netdev, &meta,
1691 rxbuf->frag + meta_off, rxd);
1692 } else if (meta_len) {
1695 end = nfp_net_parse_meta(dp->netdev, &meta,
1696 rxbuf->frag + meta_off,
1698 if (unlikely(end != rxbuf->frag + pkt_off)) {
1699 nn_dp_warn(dp, "invalid RX packet metadata\n");
1700 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf,
1706 if (xdp_prog && !(rxd->rxd.flags & PCIE_DESC_RX_BPF &&
1707 dp->bpf_offload_xdp) && !meta.portid) {
1708 unsigned int dma_off;
1712 hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM;
1714 act = nfp_net_run_xdp(xdp_prog, rxbuf->frag, hard_start,
1715 &pkt_off, &pkt_len);
1720 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
1721 if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring,
1726 trace_xdp_exception(dp->netdev,
1730 bpf_warn_invalid_xdp_action(act);
1733 trace_xdp_exception(dp->netdev, xdp_prog, act);
1736 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag,
1742 skb = build_skb(rxbuf->frag, true_bufsz);
1743 if (unlikely(!skb)) {
1744 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1747 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr);
1748 if (unlikely(!new_frag)) {
1749 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1753 if (likely(!meta.portid)) {
1754 netdev = dp->netdev;
1758 nn = netdev_priv(dp->netdev);
1759 netdev = nfp_app_repr_get(nn->app, meta.portid);
1760 if (unlikely(!netdev)) {
1761 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1764 nfp_repr_inc_rx_stats(netdev, pkt_len);
1767 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1769 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1771 skb_reserve(skb, pkt_off);
1772 skb_put(skb, pkt_len);
1774 skb->mark = meta.mark;
1775 skb_set_hash(skb, meta.hash, meta.hash_type);
1777 skb_record_rx_queue(skb, rx_ring->idx);
1778 skb->protocol = eth_type_trans(skb, netdev);
1780 nfp_net_rx_csum(dp, r_vec, rxd, &meta, skb);
1782 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
1783 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
1784 le16_to_cpu(rxd->rxd.vlan));
1786 napi_gro_receive(&rx_ring->r_vec->napi, skb);
1790 if (tx_ring->wr_ptr_add)
1791 nfp_net_tx_xmit_more_flush(tx_ring);
1792 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1794 if (!nfp_net_xdp_complete(tx_ring))
1795 pkts_polled = budget;
1803 * nfp_net_poll() - napi poll function
1804 * @napi: NAPI structure
1805 * @budget: NAPI budget
1807 * Return: number of packets polled.
1809 static int nfp_net_poll(struct napi_struct *napi, int budget)
1811 struct nfp_net_r_vector *r_vec =
1812 container_of(napi, struct nfp_net_r_vector, napi);
1813 unsigned int pkts_polled = 0;
1816 nfp_net_tx_complete(r_vec->tx_ring);
1818 pkts_polled = nfp_net_rx(r_vec->rx_ring, budget);
1820 if (pkts_polled < budget)
1821 if (napi_complete_done(napi, pkts_polled))
1822 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1827 /* Control device data path
1831 nfp_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1832 struct sk_buff *skb, bool old)
1834 unsigned int real_len = skb->len, meta_len = 0;
1835 struct nfp_net_tx_ring *tx_ring;
1836 struct nfp_net_tx_buf *txbuf;
1837 struct nfp_net_tx_desc *txd;
1838 struct nfp_net_dp *dp;
1839 dma_addr_t dma_addr;
1842 dp = &r_vec->nfp_net->dp;
1843 tx_ring = r_vec->tx_ring;
1845 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1846 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1850 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1851 u64_stats_update_begin(&r_vec->tx_sync);
1853 u64_stats_update_end(&r_vec->tx_sync);
1855 __skb_queue_tail(&r_vec->queue, skb);
1857 __skb_queue_head(&r_vec->queue, skb);
1861 if (nfp_app_ctrl_has_meta(nn->app)) {
1862 if (unlikely(skb_headroom(skb) < 8)) {
1863 nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1867 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1868 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
1871 /* Start with the head skbuf */
1872 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1874 if (dma_mapping_error(dp->dev, dma_addr))
1877 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1879 /* Stash the soft descriptor of the head then initialize it */
1880 txbuf = &tx_ring->txbufs[wr_idx];
1882 txbuf->dma_addr = dma_addr;
1885 txbuf->real_len = real_len;
1887 /* Build TX descriptor */
1888 txd = &tx_ring->txds[wr_idx];
1889 txd->offset_eop = meta_len | PCIE_DESC_TX_EOP;
1890 txd->dma_len = cpu_to_le16(skb_headlen(skb));
1891 nfp_desc_set_dma_addr(txd, dma_addr);
1892 txd->data_len = cpu_to_le16(skb->len);
1896 txd->lso_hdrlen = 0;
1899 tx_ring->wr_ptr_add++;
1900 nfp_net_tx_xmit_more_flush(tx_ring);
1905 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
1907 u64_stats_update_begin(&r_vec->tx_sync);
1909 u64_stats_update_end(&r_vec->tx_sync);
1910 dev_kfree_skb_any(skb);
1914 bool nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb)
1916 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0];
1919 spin_lock_bh(&r_vec->lock);
1920 ret = nfp_ctrl_tx_one(nn, r_vec, skb, false);
1921 spin_unlock_bh(&r_vec->lock);
1926 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1928 struct sk_buff *skb;
1930 while ((skb = __skb_dequeue(&r_vec->queue)))
1931 if (nfp_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1936 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1938 u32 meta_type, meta_tag;
1940 if (!nfp_app_ctrl_has_meta(nn->app))
1946 meta_type = get_unaligned_be32(data);
1947 meta_tag = get_unaligned_be32(data + 4);
1949 return (meta_type == NFP_NET_META_PORTID &&
1950 meta_tag == NFP_META_PORT_ID_CTRL);
1954 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1955 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1957 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1958 struct nfp_net_rx_buf *rxbuf;
1959 struct nfp_net_rx_desc *rxd;
1960 dma_addr_t new_dma_addr;
1961 struct sk_buff *skb;
1965 idx = D_IDX(rx_ring, rx_ring->rd_p);
1967 rxd = &rx_ring->rxds[idx];
1968 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1971 /* Memory barrier to ensure that we won't do other reads
1972 * before the DD bit.
1978 rxbuf = &rx_ring->rxbufs[idx];
1979 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1980 data_len = le16_to_cpu(rxd->rxd.data_len);
1981 pkt_len = data_len - meta_len;
1983 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1984 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1985 pkt_off += meta_len;
1987 pkt_off += dp->rx_offset;
1988 meta_off = pkt_off - meta_len;
1991 u64_stats_update_begin(&r_vec->rx_sync);
1993 r_vec->rx_bytes += pkt_len;
1994 u64_stats_update_end(&r_vec->rx_sync);
1996 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len);
1998 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1999 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
2001 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
2005 skb = build_skb(rxbuf->frag, dp->fl_bufsz);
2006 if (unlikely(!skb)) {
2007 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
2010 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr);
2011 if (unlikely(!new_frag)) {
2012 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
2016 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
2018 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
2020 skb_reserve(skb, pkt_off);
2021 skb_put(skb, pkt_len);
2023 nfp_app_ctrl_rx(nn->app, skb);
2028 static void nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
2030 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
2031 struct nfp_net *nn = r_vec->nfp_net;
2032 struct nfp_net_dp *dp = &nn->dp;
2034 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring))
2038 static void nfp_ctrl_poll(unsigned long arg)
2040 struct nfp_net_r_vector *r_vec = (void *)arg;
2042 spin_lock_bh(&r_vec->lock);
2043 nfp_net_tx_complete(r_vec->tx_ring);
2044 __nfp_ctrl_tx_queued(r_vec);
2045 spin_unlock_bh(&r_vec->lock);
2049 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
2052 /* Setup and Configuration
2056 * nfp_net_vecs_init() - Assign IRQs and setup rvecs.
2057 * @nn: NFP Network structure
2059 static void nfp_net_vecs_init(struct nfp_net *nn)
2061 struct nfp_net_r_vector *r_vec;
2064 nn->lsc_handler = nfp_net_irq_lsc;
2065 nn->exn_handler = nfp_net_irq_exn;
2067 for (r = 0; r < nn->max_r_vecs; r++) {
2068 struct msix_entry *entry;
2070 entry = &nn->irq_entries[NFP_NET_NON_Q_VECTORS + r];
2072 r_vec = &nn->r_vecs[r];
2073 r_vec->nfp_net = nn;
2074 r_vec->irq_entry = entry->entry;
2075 r_vec->irq_vector = entry->vector;
2077 if (nn->dp.netdev) {
2078 r_vec->handler = nfp_net_irq_rxtx;
2080 r_vec->handler = nfp_ctrl_irq_rxtx;
2082 __skb_queue_head_init(&r_vec->queue);
2083 spin_lock_init(&r_vec->lock);
2084 tasklet_init(&r_vec->tasklet, nfp_ctrl_poll,
2085 (unsigned long)r_vec);
2086 tasklet_disable(&r_vec->tasklet);
2089 cpumask_set_cpu(r, &r_vec->affinity_mask);
2094 * nfp_net_tx_ring_free() - Free resources allocated to a TX ring
2095 * @tx_ring: TX ring to free
2097 static void nfp_net_tx_ring_free(struct nfp_net_tx_ring *tx_ring)
2099 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
2100 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
2102 kfree(tx_ring->txbufs);
2105 dma_free_coherent(dp->dev, tx_ring->size,
2106 tx_ring->txds, tx_ring->dma);
2109 tx_ring->txbufs = NULL;
2110 tx_ring->txds = NULL;
2116 * nfp_net_tx_ring_alloc() - Allocate resource for a TX ring
2117 * @dp: NFP Net data path struct
2118 * @tx_ring: TX Ring structure to allocate
2120 * Return: 0 on success, negative errno otherwise.
2123 nfp_net_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
2125 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
2128 tx_ring->cnt = dp->txd_cnt;
2130 tx_ring->size = sizeof(*tx_ring->txds) * tx_ring->cnt;
2131 tx_ring->txds = dma_zalloc_coherent(dp->dev, tx_ring->size,
2132 &tx_ring->dma, GFP_KERNEL);
2136 sz = sizeof(*tx_ring->txbufs) * tx_ring->cnt;
2137 tx_ring->txbufs = kzalloc(sz, GFP_KERNEL);
2138 if (!tx_ring->txbufs)
2141 if (!tx_ring->is_xdp && dp->netdev)
2142 netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask,
2148 nfp_net_tx_ring_free(tx_ring);
2153 nfp_net_tx_ring_bufs_free(struct nfp_net_dp *dp,
2154 struct nfp_net_tx_ring *tx_ring)
2158 if (!tx_ring->is_xdp)
2161 for (i = 0; i < tx_ring->cnt; i++) {
2162 if (!tx_ring->txbufs[i].frag)
2165 nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr);
2166 __free_page(virt_to_page(tx_ring->txbufs[i].frag));
2171 nfp_net_tx_ring_bufs_alloc(struct nfp_net_dp *dp,
2172 struct nfp_net_tx_ring *tx_ring)
2174 struct nfp_net_tx_buf *txbufs = tx_ring->txbufs;
2177 if (!tx_ring->is_xdp)
2180 for (i = 0; i < tx_ring->cnt; i++) {
2181 txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr);
2182 if (!txbufs[i].frag) {
2183 nfp_net_tx_ring_bufs_free(dp, tx_ring);
2191 static int nfp_net_tx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp)
2195 dp->tx_rings = kcalloc(dp->num_tx_rings, sizeof(*dp->tx_rings),
2200 for (r = 0; r < dp->num_tx_rings; r++) {
2203 if (r >= dp->num_stack_tx_rings)
2204 bias = dp->num_stack_tx_rings;
2206 nfp_net_tx_ring_init(&dp->tx_rings[r], &nn->r_vecs[r - bias],
2209 if (nfp_net_tx_ring_alloc(dp, &dp->tx_rings[r]))
2212 if (nfp_net_tx_ring_bufs_alloc(dp, &dp->tx_rings[r]))
2220 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]);
2222 nfp_net_tx_ring_free(&dp->tx_rings[r]);
2224 kfree(dp->tx_rings);
2228 static void nfp_net_tx_rings_free(struct nfp_net_dp *dp)
2232 for (r = 0; r < dp->num_tx_rings; r++) {
2233 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]);
2234 nfp_net_tx_ring_free(&dp->tx_rings[r]);
2237 kfree(dp->tx_rings);
2241 * nfp_net_rx_ring_free() - Free resources allocated to a RX ring
2242 * @rx_ring: RX ring to free
2244 static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring)
2246 struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
2247 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
2249 kfree(rx_ring->rxbufs);
2252 dma_free_coherent(dp->dev, rx_ring->size,
2253 rx_ring->rxds, rx_ring->dma);
2256 rx_ring->rxbufs = NULL;
2257 rx_ring->rxds = NULL;
2263 * nfp_net_rx_ring_alloc() - Allocate resource for a RX ring
2264 * @dp: NFP Net data path struct
2265 * @rx_ring: RX ring to allocate
2267 * Return: 0 on success, negative errno otherwise.
2270 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring)
2274 rx_ring->cnt = dp->rxd_cnt;
2275 rx_ring->size = sizeof(*rx_ring->rxds) * rx_ring->cnt;
2276 rx_ring->rxds = dma_zalloc_coherent(dp->dev, rx_ring->size,
2277 &rx_ring->dma, GFP_KERNEL);
2281 sz = sizeof(*rx_ring->rxbufs) * rx_ring->cnt;
2282 rx_ring->rxbufs = kzalloc(sz, GFP_KERNEL);
2283 if (!rx_ring->rxbufs)
2289 nfp_net_rx_ring_free(rx_ring);
2293 static int nfp_net_rx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp)
2297 dp->rx_rings = kcalloc(dp->num_rx_rings, sizeof(*dp->rx_rings),
2302 for (r = 0; r < dp->num_rx_rings; r++) {
2303 nfp_net_rx_ring_init(&dp->rx_rings[r], &nn->r_vecs[r], r);
2305 if (nfp_net_rx_ring_alloc(dp, &dp->rx_rings[r]))
2308 if (nfp_net_rx_ring_bufs_alloc(dp, &dp->rx_rings[r]))
2316 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]);
2318 nfp_net_rx_ring_free(&dp->rx_rings[r]);
2320 kfree(dp->rx_rings);
2324 static void nfp_net_rx_rings_free(struct nfp_net_dp *dp)
2328 for (r = 0; r < dp->num_rx_rings; r++) {
2329 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]);
2330 nfp_net_rx_ring_free(&dp->rx_rings[r]);
2333 kfree(dp->rx_rings);
2337 nfp_net_vector_assign_rings(struct nfp_net_dp *dp,
2338 struct nfp_net_r_vector *r_vec, int idx)
2340 r_vec->rx_ring = idx < dp->num_rx_rings ? &dp->rx_rings[idx] : NULL;
2342 idx < dp->num_stack_tx_rings ? &dp->tx_rings[idx] : NULL;
2344 r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ?
2345 &dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL;
2349 nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
2356 netif_napi_add(nn->dp.netdev, &r_vec->napi,
2357 nfp_net_poll, NAPI_POLL_WEIGHT);
2359 tasklet_enable(&r_vec->tasklet);
2361 snprintf(r_vec->name, sizeof(r_vec->name),
2362 "%s-rxtx-%d", nfp_net_name(nn), idx);
2363 err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name,
2367 netif_napi_del(&r_vec->napi);
2369 tasklet_disable(&r_vec->tasklet);
2371 nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector);
2374 disable_irq(r_vec->irq_vector);
2376 irq_set_affinity_hint(r_vec->irq_vector, &r_vec->affinity_mask);
2378 nn_dbg(nn, "RV%02d: irq=%03d/%03d\n", idx, r_vec->irq_vector,
2385 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec)
2387 irq_set_affinity_hint(r_vec->irq_vector, NULL);
2389 netif_napi_del(&r_vec->napi);
2391 tasklet_disable(&r_vec->tasklet);
2393 free_irq(r_vec->irq_vector, r_vec);
2397 * nfp_net_rss_write_itbl() - Write RSS indirection table to device
2398 * @nn: NFP Net device to reconfigure
2400 void nfp_net_rss_write_itbl(struct nfp_net *nn)
2404 for (i = 0; i < NFP_NET_CFG_RSS_ITBL_SZ; i += 4)
2405 nn_writel(nn, NFP_NET_CFG_RSS_ITBL + i,
2406 get_unaligned_le32(nn->rss_itbl + i));
2410 * nfp_net_rss_write_key() - Write RSS hash key to device
2411 * @nn: NFP Net device to reconfigure
2413 void nfp_net_rss_write_key(struct nfp_net *nn)
2417 for (i = 0; i < nfp_net_rss_key_sz(nn); i += 4)
2418 nn_writel(nn, NFP_NET_CFG_RSS_KEY + i,
2419 get_unaligned_le32(nn->rss_key + i));
2423 * nfp_net_coalesce_write_cfg() - Write irq coalescence configuration to HW
2424 * @nn: NFP Net device to reconfigure
2426 void nfp_net_coalesce_write_cfg(struct nfp_net *nn)
2432 /* Compute factor used to convert coalesce '_usecs' parameters to
2433 * ME timestamp ticks. There are 16 ME clock cycles for each timestamp
2436 factor = nn->me_freq_mhz / 16;
2438 /* copy RX interrupt coalesce parameters */
2439 value = (nn->rx_coalesce_max_frames << 16) |
2440 (factor * nn->rx_coalesce_usecs);
2441 for (i = 0; i < nn->dp.num_rx_rings; i++)
2442 nn_writel(nn, NFP_NET_CFG_RXR_IRQ_MOD(i), value);
2444 /* copy TX interrupt coalesce parameters */
2445 value = (nn->tx_coalesce_max_frames << 16) |
2446 (factor * nn->tx_coalesce_usecs);
2447 for (i = 0; i < nn->dp.num_tx_rings; i++)
2448 nn_writel(nn, NFP_NET_CFG_TXR_IRQ_MOD(i), value);
2452 * nfp_net_write_mac_addr() - Write mac address to the device control BAR
2453 * @nn: NFP Net device to reconfigure
2454 * @addr: MAC address to write
2456 * Writes the MAC address from the netdev to the device control BAR. Does not
2457 * perform the required reconfig. We do a bit of byte swapping dance because
2460 static void nfp_net_write_mac_addr(struct nfp_net *nn, const u8 *addr)
2462 nn_writel(nn, NFP_NET_CFG_MACADDR + 0, get_unaligned_be32(addr));
2463 nn_writew(nn, NFP_NET_CFG_MACADDR + 6, get_unaligned_be16(addr + 4));
2466 static void nfp_net_vec_clear_ring_data(struct nfp_net *nn, unsigned int idx)
2468 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), 0);
2469 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), 0);
2470 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), 0);
2472 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), 0);
2473 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), 0);
2474 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), 0);
2478 * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP
2479 * @nn: NFP Net device to reconfigure
2481 static void nfp_net_clear_config_and_disable(struct nfp_net *nn)
2483 u32 new_ctrl, update;
2487 new_ctrl = nn->dp.ctrl;
2488 new_ctrl &= ~NFP_NET_CFG_CTRL_ENABLE;
2489 update = NFP_NET_CFG_UPDATE_GEN;
2490 update |= NFP_NET_CFG_UPDATE_MSIX;
2491 update |= NFP_NET_CFG_UPDATE_RING;
2493 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG)
2494 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG;
2496 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0);
2497 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0);
2499 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2500 err = nfp_net_reconfig(nn, update);
2502 nn_err(nn, "Could not disable device: %d\n", err);
2504 for (r = 0; r < nn->dp.num_rx_rings; r++)
2505 nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]);
2506 for (r = 0; r < nn->dp.num_tx_rings; r++)
2507 nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]);
2508 for (r = 0; r < nn->dp.num_r_vecs; r++)
2509 nfp_net_vec_clear_ring_data(nn, r);
2511 nn->dp.ctrl = new_ctrl;
2515 nfp_net_rx_ring_hw_cfg_write(struct nfp_net *nn,
2516 struct nfp_net_rx_ring *rx_ring, unsigned int idx)
2518 /* Write the DMA address, size and MSI-X info to the device */
2519 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), rx_ring->dma);
2520 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), ilog2(rx_ring->cnt));
2521 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), rx_ring->r_vec->irq_entry);
2525 nfp_net_tx_ring_hw_cfg_write(struct nfp_net *nn,
2526 struct nfp_net_tx_ring *tx_ring, unsigned int idx)
2528 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), tx_ring->dma);
2529 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), ilog2(tx_ring->cnt));
2530 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), tx_ring->r_vec->irq_entry);
2534 * nfp_net_set_config_and_enable() - Write control BAR and enable NFP
2535 * @nn: NFP Net device to reconfigure
2537 static int nfp_net_set_config_and_enable(struct nfp_net *nn)
2539 u32 bufsz, new_ctrl, update = 0;
2543 new_ctrl = nn->dp.ctrl;
2545 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_RSS_ANY) {
2546 nfp_net_rss_write_key(nn);
2547 nfp_net_rss_write_itbl(nn);
2548 nn_writel(nn, NFP_NET_CFG_RSS_CTRL, nn->rss_cfg);
2549 update |= NFP_NET_CFG_UPDATE_RSS;
2552 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_IRQMOD) {
2553 nfp_net_coalesce_write_cfg(nn);
2554 update |= NFP_NET_CFG_UPDATE_IRQMOD;
2557 for (r = 0; r < nn->dp.num_tx_rings; r++)
2558 nfp_net_tx_ring_hw_cfg_write(nn, &nn->dp.tx_rings[r], r);
2559 for (r = 0; r < nn->dp.num_rx_rings; r++)
2560 nfp_net_rx_ring_hw_cfg_write(nn, &nn->dp.rx_rings[r], r);
2562 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, nn->dp.num_tx_rings == 64 ?
2563 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_tx_rings) - 1);
2565 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, nn->dp.num_rx_rings == 64 ?
2566 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_rx_rings) - 1);
2569 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr);
2571 nn_writel(nn, NFP_NET_CFG_MTU, nn->dp.mtu);
2573 bufsz = nn->dp.fl_bufsz - nn->dp.rx_dma_off - NFP_NET_RX_BUF_NON_DATA;
2574 nn_writel(nn, NFP_NET_CFG_FLBUFSZ, bufsz);
2577 new_ctrl |= NFP_NET_CFG_CTRL_ENABLE;
2578 update |= NFP_NET_CFG_UPDATE_GEN;
2579 update |= NFP_NET_CFG_UPDATE_MSIX;
2580 update |= NFP_NET_CFG_UPDATE_RING;
2581 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG)
2582 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG;
2584 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2585 err = nfp_net_reconfig(nn, update);
2587 nfp_net_clear_config_and_disable(nn);
2591 nn->dp.ctrl = new_ctrl;
2593 for (r = 0; r < nn->dp.num_rx_rings; r++)
2594 nfp_net_rx_ring_fill_freelist(&nn->dp, &nn->dp.rx_rings[r]);
2596 /* Since reconfiguration requests while NFP is down are ignored we
2597 * have to wipe the entire VXLAN configuration and reinitialize it.
2599 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN) {
2600 memset(&nn->vxlan_ports, 0, sizeof(nn->vxlan_ports));
2601 memset(&nn->vxlan_usecnt, 0, sizeof(nn->vxlan_usecnt));
2602 udp_tunnel_get_rx_info(nn->dp.netdev);
2609 * nfp_net_close_stack() - Quiesce the stack (part of close)
2610 * @nn: NFP Net device to reconfigure
2612 static void nfp_net_close_stack(struct nfp_net *nn)
2616 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2617 netif_carrier_off(nn->dp.netdev);
2618 nn->link_up = false;
2620 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2621 disable_irq(nn->r_vecs[r].irq_vector);
2622 napi_disable(&nn->r_vecs[r].napi);
2625 netif_tx_disable(nn->dp.netdev);
2629 * nfp_net_close_free_all() - Free all runtime resources
2630 * @nn: NFP Net device to reconfigure
2632 static void nfp_net_close_free_all(struct nfp_net *nn)
2636 nfp_net_tx_rings_free(&nn->dp);
2637 nfp_net_rx_rings_free(&nn->dp);
2639 for (r = 0; r < nn->dp.num_r_vecs; r++)
2640 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
2642 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2643 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX);
2647 * nfp_net_netdev_close() - Called when the device is downed
2648 * @netdev: netdev structure
2650 static int nfp_net_netdev_close(struct net_device *netdev)
2652 struct nfp_net *nn = netdev_priv(netdev);
2654 /* Step 1: Disable RX and TX rings from the Linux kernel perspective
2656 nfp_net_close_stack(nn);
2660 nfp_net_clear_config_and_disable(nn);
2662 /* Step 3: Free resources
2664 nfp_net_close_free_all(nn);
2666 nn_dbg(nn, "%s down", netdev->name);
2670 void nfp_ctrl_close(struct nfp_net *nn)
2676 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2677 disable_irq(nn->r_vecs[r].irq_vector);
2678 tasklet_disable(&nn->r_vecs[r].tasklet);
2681 nfp_net_clear_config_and_disable(nn);
2683 nfp_net_close_free_all(nn);
2689 * nfp_net_open_stack() - Start the device from stack's perspective
2690 * @nn: NFP Net device to reconfigure
2692 static void nfp_net_open_stack(struct nfp_net *nn)
2696 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2697 napi_enable(&nn->r_vecs[r].napi);
2698 enable_irq(nn->r_vecs[r].irq_vector);
2701 netif_tx_wake_all_queues(nn->dp.netdev);
2703 enable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2704 nfp_net_read_link_status(nn);
2707 static int nfp_net_open_alloc_all(struct nfp_net *nn)
2711 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_EXN, "%s-exn",
2712 nn->exn_name, sizeof(nn->exn_name),
2713 NFP_NET_IRQ_EXN_IDX, nn->exn_handler);
2716 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_LSC, "%s-lsc",
2717 nn->lsc_name, sizeof(nn->lsc_name),
2718 NFP_NET_IRQ_LSC_IDX, nn->lsc_handler);
2721 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2723 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2724 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r);
2726 goto err_cleanup_vec_p;
2729 err = nfp_net_rx_rings_prepare(nn, &nn->dp);
2731 goto err_cleanup_vec;
2733 err = nfp_net_tx_rings_prepare(nn, &nn->dp);
2735 goto err_free_rx_rings;
2737 for (r = 0; r < nn->max_r_vecs; r++)
2738 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r);
2743 nfp_net_rx_rings_free(&nn->dp);
2745 r = nn->dp.num_r_vecs;
2748 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
2749 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2751 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX);
2755 static int nfp_net_netdev_open(struct net_device *netdev)
2757 struct nfp_net *nn = netdev_priv(netdev);
2760 /* Step 1: Allocate resources for rings and the like
2761 * - Request interrupts
2762 * - Allocate RX and TX ring resources
2763 * - Setup initial RSS table
2765 err = nfp_net_open_alloc_all(nn);
2769 err = netif_set_real_num_tx_queues(netdev, nn->dp.num_stack_tx_rings);
2773 err = netif_set_real_num_rx_queues(netdev, nn->dp.num_rx_rings);
2777 /* Step 2: Configure the NFP
2778 * - Enable rings from 0 to tx_rings/rx_rings - 1.
2779 * - Write MAC address (in case it changed)
2781 * - Set the Freelist buffer size
2784 err = nfp_net_set_config_and_enable(nn);
2788 /* Step 3: Enable for kernel
2789 * - put some freelist descriptors on each RX ring
2790 * - enable NAPI on each ring
2791 * - enable all TX queues
2794 nfp_net_open_stack(nn);
2799 nfp_net_close_free_all(nn);
2803 int nfp_ctrl_open(struct nfp_net *nn)
2807 /* ring dumping depends on vNICs being opened/closed under rtnl */
2810 err = nfp_net_open_alloc_all(nn);
2814 err = nfp_net_set_config_and_enable(nn);
2818 for (r = 0; r < nn->dp.num_r_vecs; r++)
2819 enable_irq(nn->r_vecs[r].irq_vector);
2826 nfp_net_close_free_all(nn);
2832 static void nfp_net_set_rx_mode(struct net_device *netdev)
2834 struct nfp_net *nn = netdev_priv(netdev);
2837 new_ctrl = nn->dp.ctrl;
2839 if (netdev->flags & IFF_PROMISC) {
2840 if (nn->cap & NFP_NET_CFG_CTRL_PROMISC)
2841 new_ctrl |= NFP_NET_CFG_CTRL_PROMISC;
2843 nn_warn(nn, "FW does not support promiscuous mode\n");
2845 new_ctrl &= ~NFP_NET_CFG_CTRL_PROMISC;
2848 if (new_ctrl == nn->dp.ctrl)
2851 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2852 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_GEN);
2854 nn->dp.ctrl = new_ctrl;
2857 static void nfp_net_rss_init_itbl(struct nfp_net *nn)
2861 for (i = 0; i < sizeof(nn->rss_itbl); i++)
2863 ethtool_rxfh_indir_default(i, nn->dp.num_rx_rings);
2866 static void nfp_net_dp_swap(struct nfp_net *nn, struct nfp_net_dp *dp)
2868 struct nfp_net_dp new_dp = *dp;
2873 nn->dp.netdev->mtu = new_dp.mtu;
2875 if (!netif_is_rxfh_configured(nn->dp.netdev))
2876 nfp_net_rss_init_itbl(nn);
2879 static int nfp_net_dp_swap_enable(struct nfp_net *nn, struct nfp_net_dp *dp)
2884 nfp_net_dp_swap(nn, dp);
2886 for (r = 0; r < nn->max_r_vecs; r++)
2887 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r);
2889 err = netif_set_real_num_rx_queues(nn->dp.netdev, nn->dp.num_rx_rings);
2893 if (nn->dp.netdev->real_num_tx_queues != nn->dp.num_stack_tx_rings) {
2894 err = netif_set_real_num_tx_queues(nn->dp.netdev,
2895 nn->dp.num_stack_tx_rings);
2900 return nfp_net_set_config_and_enable(nn);
2903 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn)
2905 struct nfp_net_dp *new;
2907 new = kmalloc(sizeof(*new), GFP_KERNEL);
2913 /* Clear things which need to be recomputed */
2915 new->tx_rings = NULL;
2916 new->rx_rings = NULL;
2917 new->num_r_vecs = 0;
2918 new->num_stack_tx_rings = 0;
2924 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp,
2925 struct netlink_ext_ack *extack)
2927 /* XDP-enabled tests */
2930 if (dp->fl_bufsz > PAGE_SIZE) {
2931 NL_SET_ERR_MSG_MOD(extack, "MTU too large w/ XDP enabled");
2934 if (dp->num_tx_rings > nn->max_tx_rings) {
2935 NL_SET_ERR_MSG_MOD(extack, "Insufficient number of TX rings w/ XDP enabled");
2942 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp,
2943 struct netlink_ext_ack *extack)
2947 dp->fl_bufsz = nfp_net_calc_fl_bufsz(dp);
2949 dp->num_stack_tx_rings = dp->num_tx_rings;
2951 dp->num_stack_tx_rings -= dp->num_rx_rings;
2953 dp->num_r_vecs = max(dp->num_rx_rings, dp->num_stack_tx_rings);
2955 err = nfp_net_check_config(nn, dp, extack);
2959 if (!netif_running(dp->netdev)) {
2960 nfp_net_dp_swap(nn, dp);
2965 /* Prepare new rings */
2966 for (r = nn->dp.num_r_vecs; r < dp->num_r_vecs; r++) {
2967 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r);
2970 goto err_cleanup_vecs;
2974 err = nfp_net_rx_rings_prepare(nn, dp);
2976 goto err_cleanup_vecs;
2978 err = nfp_net_tx_rings_prepare(nn, dp);
2982 /* Stop device, swap in new rings, try to start the firmware */
2983 nfp_net_close_stack(nn);
2984 nfp_net_clear_config_and_disable(nn);
2986 err = nfp_net_dp_swap_enable(nn, dp);
2990 nfp_net_clear_config_and_disable(nn);
2992 /* Try with old configuration and old rings */
2993 err2 = nfp_net_dp_swap_enable(nn, dp);
2995 nn_err(nn, "Can't restore ring config - FW communication failed (%d,%d)\n",
2998 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
2999 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3001 nfp_net_rx_rings_free(dp);
3002 nfp_net_tx_rings_free(dp);
3004 nfp_net_open_stack(nn);
3011 nfp_net_rx_rings_free(dp);
3013 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
3014 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3019 static int nfp_net_change_mtu(struct net_device *netdev, int new_mtu)
3021 struct nfp_net *nn = netdev_priv(netdev);
3022 struct nfp_net_dp *dp;
3024 dp = nfp_net_clone_dp(nn);
3030 return nfp_net_ring_reconfig(nn, dp, NULL);
3034 nfp_net_vlan_rx_add_vid(struct net_device *netdev, __be16 proto, u16 vid)
3036 struct nfp_net *nn = netdev_priv(netdev);
3038 /* Priority tagged packets with vlan id 0 are processed by the
3039 * NFP as untagged packets
3044 nn_writew(nn, NFP_NET_CFG_VLAN_FILTER_VID, vid);
3045 nn_writew(nn, NFP_NET_CFG_VLAN_FILTER_PROTO, ETH_P_8021Q);
3047 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_ADD);
3051 nfp_net_vlan_rx_kill_vid(struct net_device *netdev, __be16 proto, u16 vid)
3053 struct nfp_net *nn = netdev_priv(netdev);
3055 /* Priority tagged packets with vlan id 0 are processed by the
3056 * NFP as untagged packets
3061 nn_writew(nn, NFP_NET_CFG_VLAN_FILTER_VID, vid);
3062 nn_writew(nn, NFP_NET_CFG_VLAN_FILTER_PROTO, ETH_P_8021Q);
3064 return nfp_net_reconfig_mbox(nn, NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_KILL);
3067 static void nfp_net_stat64(struct net_device *netdev,
3068 struct rtnl_link_stats64 *stats)
3070 struct nfp_net *nn = netdev_priv(netdev);
3073 for (r = 0; r < nn->dp.num_r_vecs; r++) {
3074 struct nfp_net_r_vector *r_vec = &nn->r_vecs[r];
3079 start = u64_stats_fetch_begin(&r_vec->rx_sync);
3080 data[0] = r_vec->rx_pkts;
3081 data[1] = r_vec->rx_bytes;
3082 data[2] = r_vec->rx_drops;
3083 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
3084 stats->rx_packets += data[0];
3085 stats->rx_bytes += data[1];
3086 stats->rx_dropped += data[2];
3089 start = u64_stats_fetch_begin(&r_vec->tx_sync);
3090 data[0] = r_vec->tx_pkts;
3091 data[1] = r_vec->tx_bytes;
3092 data[2] = r_vec->tx_errors;
3093 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
3094 stats->tx_packets += data[0];
3095 stats->tx_bytes += data[1];
3096 stats->tx_errors += data[2];
3100 static int nfp_net_set_features(struct net_device *netdev,
3101 netdev_features_t features)
3103 netdev_features_t changed = netdev->features ^ features;
3104 struct nfp_net *nn = netdev_priv(netdev);
3108 /* Assume this is not called with features we have not advertised */
3110 new_ctrl = nn->dp.ctrl;
3112 if (changed & NETIF_F_RXCSUM) {
3113 if (features & NETIF_F_RXCSUM)
3114 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY;
3116 new_ctrl &= ~NFP_NET_CFG_CTRL_RXCSUM_ANY;
3119 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3120 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM))
3121 new_ctrl |= NFP_NET_CFG_CTRL_TXCSUM;
3123 new_ctrl &= ~NFP_NET_CFG_CTRL_TXCSUM;
3126 if (changed & (NETIF_F_TSO | NETIF_F_TSO6)) {
3127 if (features & (NETIF_F_TSO | NETIF_F_TSO6))
3128 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?:
3129 NFP_NET_CFG_CTRL_LSO;
3131 new_ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY;
3134 if (changed & NETIF_F_HW_VLAN_CTAG_RX) {
3135 if (features & NETIF_F_HW_VLAN_CTAG_RX)
3136 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
3138 new_ctrl &= ~NFP_NET_CFG_CTRL_RXVLAN;
3141 if (changed & NETIF_F_HW_VLAN_CTAG_TX) {
3142 if (features & NETIF_F_HW_VLAN_CTAG_TX)
3143 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
3145 new_ctrl &= ~NFP_NET_CFG_CTRL_TXVLAN;
3148 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) {
3149 if (features & NETIF_F_HW_VLAN_CTAG_FILTER)
3150 new_ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER;
3152 new_ctrl &= ~NFP_NET_CFG_CTRL_CTAG_FILTER;
3155 if (changed & NETIF_F_SG) {
3156 if (features & NETIF_F_SG)
3157 new_ctrl |= NFP_NET_CFG_CTRL_GATHER;
3159 new_ctrl &= ~NFP_NET_CFG_CTRL_GATHER;
3162 if (changed & NETIF_F_HW_TC && nfp_app_tc_busy(nn->app, nn)) {
3163 nn_err(nn, "Cannot disable HW TC offload while in use\n");
3167 nn_dbg(nn, "Feature change 0x%llx -> 0x%llx (changed=0x%llx)\n",
3168 netdev->features, features, changed);
3170 if (new_ctrl == nn->dp.ctrl)
3173 nn_dbg(nn, "NIC ctrl: 0x%x -> 0x%x\n", nn->dp.ctrl, new_ctrl);
3174 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
3175 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN);
3179 nn->dp.ctrl = new_ctrl;
3184 static netdev_features_t
3185 nfp_net_features_check(struct sk_buff *skb, struct net_device *dev,
3186 netdev_features_t features)
3190 /* We can't do TSO over double tagged packets (802.1AD) */
3191 features &= vlan_features_check(skb, features);
3193 if (!skb->encapsulation)
3196 /* Ensure that inner L4 header offset fits into TX descriptor field */
3197 if (skb_is_gso(skb)) {
3200 hdrlen = skb_inner_transport_header(skb) - skb->data +
3201 inner_tcp_hdrlen(skb);
3203 if (unlikely(hdrlen > NFP_NET_LSO_MAX_HDR_SZ))
3204 features &= ~NETIF_F_GSO_MASK;
3207 /* VXLAN/GRE check */
3208 switch (vlan_get_protocol(skb)) {
3209 case htons(ETH_P_IP):
3210 l4_hdr = ip_hdr(skb)->protocol;
3212 case htons(ETH_P_IPV6):
3213 l4_hdr = ipv6_hdr(skb)->nexthdr;
3216 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3219 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
3220 skb->inner_protocol != htons(ETH_P_TEB) ||
3221 (l4_hdr != IPPROTO_UDP && l4_hdr != IPPROTO_GRE) ||
3222 (l4_hdr == IPPROTO_UDP &&
3223 (skb_inner_mac_header(skb) - skb_transport_header(skb) !=
3224 sizeof(struct udphdr) + sizeof(struct vxlanhdr))))
3225 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3231 * nfp_net_set_vxlan_port() - set vxlan port in SW and reconfigure HW
3232 * @nn: NFP Net device to reconfigure
3233 * @idx: Index into the port table where new port should be written
3234 * @port: UDP port to configure (pass zero to remove VXLAN port)
3236 static void nfp_net_set_vxlan_port(struct nfp_net *nn, int idx, __be16 port)
3240 nn->vxlan_ports[idx] = port;
3242 if (!(nn->dp.ctrl & NFP_NET_CFG_CTRL_VXLAN))
3245 BUILD_BUG_ON(NFP_NET_N_VXLAN_PORTS & 1);
3246 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i += 2)
3247 nn_writel(nn, NFP_NET_CFG_VXLAN_PORT + i * sizeof(port),
3248 be16_to_cpu(nn->vxlan_ports[i + 1]) << 16 |
3249 be16_to_cpu(nn->vxlan_ports[i]));
3251 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_VXLAN);
3255 * nfp_net_find_vxlan_idx() - find table entry of the port or a free one
3256 * @nn: NFP Network structure
3257 * @port: UDP port to look for
3259 * Return: if the port is already in the table -- it's position;
3260 * if the port is not in the table -- free position to use;
3261 * if the table is full -- -ENOSPC.
3263 static int nfp_net_find_vxlan_idx(struct nfp_net *nn, __be16 port)
3265 int i, free_idx = -ENOSPC;
3267 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i++) {
3268 if (nn->vxlan_ports[i] == port)
3270 if (!nn->vxlan_usecnt[i])
3277 static void nfp_net_add_vxlan_port(struct net_device *netdev,
3278 struct udp_tunnel_info *ti)
3280 struct nfp_net *nn = netdev_priv(netdev);
3283 if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3286 idx = nfp_net_find_vxlan_idx(nn, ti->port);
3290 if (!nn->vxlan_usecnt[idx]++)
3291 nfp_net_set_vxlan_port(nn, idx, ti->port);
3294 static void nfp_net_del_vxlan_port(struct net_device *netdev,
3295 struct udp_tunnel_info *ti)
3297 struct nfp_net *nn = netdev_priv(netdev);
3300 if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3303 idx = nfp_net_find_vxlan_idx(nn, ti->port);
3304 if (idx == -ENOSPC || !nn->vxlan_usecnt[idx])
3307 if (!--nn->vxlan_usecnt[idx])
3308 nfp_net_set_vxlan_port(nn, idx, 0);
3312 nfp_net_xdp_setup_drv(struct nfp_net *nn, struct bpf_prog *prog,
3313 struct netlink_ext_ack *extack)
3315 struct nfp_net_dp *dp;
3317 if (!prog == !nn->dp.xdp_prog) {
3318 WRITE_ONCE(nn->dp.xdp_prog, prog);
3322 dp = nfp_net_clone_dp(nn);
3326 dp->xdp_prog = prog;
3327 dp->num_tx_rings += prog ? nn->dp.num_rx_rings : -nn->dp.num_rx_rings;
3328 dp->rx_dma_dir = prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
3329 dp->rx_dma_off = prog ? XDP_PACKET_HEADROOM - nn->dp.rx_offset : 0;
3331 /* We need RX reconfig to remap the buffers (BIDIR vs FROM_DEV) */
3332 return nfp_net_ring_reconfig(nn, dp, extack);
3336 nfp_net_xdp_setup(struct nfp_net *nn, struct bpf_prog *prog, u32 flags,
3337 struct netlink_ext_ack *extack)
3339 struct bpf_prog *drv_prog, *offload_prog;
3342 if (nn->xdp_prog && (flags ^ nn->xdp_flags) & XDP_FLAGS_MODES)
3345 /* Load both when no flags set to allow easy activation of driver path
3346 * when program is replaced by one which can't be offloaded.
3348 drv_prog = flags & XDP_FLAGS_HW_MODE ? NULL : prog;
3349 offload_prog = flags & XDP_FLAGS_DRV_MODE ? NULL : prog;
3351 err = nfp_net_xdp_setup_drv(nn, drv_prog, extack);
3355 err = nfp_app_xdp_offload(nn->app, nn, offload_prog);
3356 if (err && flags & XDP_FLAGS_HW_MODE)
3360 bpf_prog_put(nn->xdp_prog);
3361 nn->xdp_prog = prog;
3362 nn->xdp_flags = flags;
3367 static int nfp_net_xdp(struct net_device *netdev, struct netdev_xdp *xdp)
3369 struct nfp_net *nn = netdev_priv(netdev);
3371 switch (xdp->command) {
3372 case XDP_SETUP_PROG:
3373 case XDP_SETUP_PROG_HW:
3374 return nfp_net_xdp_setup(nn, xdp->prog, xdp->flags,
3376 case XDP_QUERY_PROG:
3377 xdp->prog_attached = !!nn->xdp_prog;
3378 if (nn->dp.bpf_offload_xdp)
3379 xdp->prog_attached = XDP_ATTACHED_HW;
3380 xdp->prog_id = nn->xdp_prog ? nn->xdp_prog->aux->id : 0;
3387 static int nfp_net_set_mac_address(struct net_device *netdev, void *addr)
3389 struct nfp_net *nn = netdev_priv(netdev);
3390 struct sockaddr *saddr = addr;
3393 err = eth_prepare_mac_addr_change(netdev, addr);
3397 nfp_net_write_mac_addr(nn, saddr->sa_data);
3399 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MACADDR);
3403 eth_commit_mac_addr_change(netdev, addr);
3408 const struct net_device_ops nfp_net_netdev_ops = {
3409 .ndo_open = nfp_net_netdev_open,
3410 .ndo_stop = nfp_net_netdev_close,
3411 .ndo_start_xmit = nfp_net_tx,
3412 .ndo_get_stats64 = nfp_net_stat64,
3413 .ndo_vlan_rx_add_vid = nfp_net_vlan_rx_add_vid,
3414 .ndo_vlan_rx_kill_vid = nfp_net_vlan_rx_kill_vid,
3415 .ndo_setup_tc = nfp_port_setup_tc,
3416 .ndo_tx_timeout = nfp_net_tx_timeout,
3417 .ndo_set_rx_mode = nfp_net_set_rx_mode,
3418 .ndo_change_mtu = nfp_net_change_mtu,
3419 .ndo_set_mac_address = nfp_net_set_mac_address,
3420 .ndo_set_features = nfp_net_set_features,
3421 .ndo_features_check = nfp_net_features_check,
3422 .ndo_get_phys_port_name = nfp_port_get_phys_port_name,
3423 .ndo_udp_tunnel_add = nfp_net_add_vxlan_port,
3424 .ndo_udp_tunnel_del = nfp_net_del_vxlan_port,
3425 .ndo_xdp = nfp_net_xdp,
3429 * nfp_net_info() - Print general info about the NIC
3430 * @nn: NFP Net device to reconfigure
3432 void nfp_net_info(struct nfp_net *nn)
3434 nn_info(nn, "Netronome NFP-6xxx %sNetdev: TxQs=%d/%d RxQs=%d/%d\n",
3435 nn->dp.is_vf ? "VF " : "",
3436 nn->dp.num_tx_rings, nn->max_tx_rings,
3437 nn->dp.num_rx_rings, nn->max_rx_rings);
3438 nn_info(nn, "VER: %d.%d.%d.%d, Maximum supported MTU: %d\n",
3439 nn->fw_ver.resv, nn->fw_ver.class,
3440 nn->fw_ver.major, nn->fw_ver.minor,
3442 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",
3444 nn->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "",
3445 nn->cap & NFP_NET_CFG_CTRL_L2BC ? "L2BCFILT " : "",
3446 nn->cap & NFP_NET_CFG_CTRL_L2MC ? "L2MCFILT " : "",
3447 nn->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "",
3448 nn->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "",
3449 nn->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "",
3450 nn->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "",
3451 nn->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "",
3452 nn->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "",
3453 nn->cap & NFP_NET_CFG_CTRL_LSO ? "TSO1 " : "",
3454 nn->cap & NFP_NET_CFG_CTRL_LSO2 ? "TSO2 " : "",
3455 nn->cap & NFP_NET_CFG_CTRL_RSS ? "RSS1 " : "",
3456 nn->cap & NFP_NET_CFG_CTRL_RSS2 ? "RSS2 " : "",
3457 nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER ? "CTAG_FILTER " : "",
3458 nn->cap & NFP_NET_CFG_CTRL_L2SWITCH ? "L2SWITCH " : "",
3459 nn->cap & NFP_NET_CFG_CTRL_MSIXAUTO ? "AUTOMASK " : "",
3460 nn->cap & NFP_NET_CFG_CTRL_IRQMOD ? "IRQMOD " : "",
3461 nn->cap & NFP_NET_CFG_CTRL_VXLAN ? "VXLAN " : "",
3462 nn->cap & NFP_NET_CFG_CTRL_NVGRE ? "NVGRE " : "",
3463 nn->cap & NFP_NET_CFG_CTRL_CSUM_COMPLETE ?
3464 "RXCSUM_COMPLETE " : "",
3465 nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR ? "LIVE_ADDR " : "",
3466 nfp_app_extra_cap(nn->app, nn));
3470 * nfp_net_alloc() - Allocate netdev and related structure
3472 * @needs_netdev: Whether to allocate a netdev for this vNIC
3473 * @max_tx_rings: Maximum number of TX rings supported by device
3474 * @max_rx_rings: Maximum number of RX rings supported by device
3476 * This function allocates a netdev device and fills in the initial
3477 * part of the @struct nfp_net structure. In case of control device
3478 * nfp_net structure is allocated without the netdev.
3480 * Return: NFP Net device structure, or ERR_PTR on error.
3482 struct nfp_net *nfp_net_alloc(struct pci_dev *pdev, bool needs_netdev,
3483 unsigned int max_tx_rings,
3484 unsigned int max_rx_rings)
3489 struct net_device *netdev;
3491 netdev = alloc_etherdev_mqs(sizeof(struct nfp_net),
3492 max_tx_rings, max_rx_rings);
3494 return ERR_PTR(-ENOMEM);
3496 SET_NETDEV_DEV(netdev, &pdev->dev);
3497 nn = netdev_priv(netdev);
3498 nn->dp.netdev = netdev;
3500 nn = vzalloc(sizeof(*nn));
3502 return ERR_PTR(-ENOMEM);
3505 nn->dp.dev = &pdev->dev;
3508 nn->max_tx_rings = max_tx_rings;
3509 nn->max_rx_rings = max_rx_rings;
3511 nn->dp.num_tx_rings = min_t(unsigned int,
3512 max_tx_rings, num_online_cpus());
3513 nn->dp.num_rx_rings = min_t(unsigned int, max_rx_rings,
3514 netif_get_num_default_rss_queues());
3516 nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings);
3517 nn->dp.num_r_vecs = min_t(unsigned int,
3518 nn->dp.num_r_vecs, num_online_cpus());
3520 nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT;
3521 nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT;
3523 spin_lock_init(&nn->reconfig_lock);
3524 spin_lock_init(&nn->link_status_lock);
3526 setup_timer(&nn->reconfig_timer,
3527 nfp_net_reconfig_timer, (unsigned long)nn);
3533 * nfp_net_free() - Undo what @nfp_net_alloc() did
3534 * @nn: NFP Net device to reconfigure
3536 void nfp_net_free(struct nfp_net *nn)
3539 bpf_prog_put(nn->xdp_prog);
3542 free_netdev(nn->dp.netdev);
3548 * nfp_net_rss_key_sz() - Get current size of the RSS key
3549 * @nn: NFP Net device instance
3551 * Return: size of the RSS key for currently selected hash function.
3553 unsigned int nfp_net_rss_key_sz(struct nfp_net *nn)
3555 switch (nn->rss_hfunc) {
3556 case ETH_RSS_HASH_TOP:
3557 return NFP_NET_CFG_RSS_KEY_SZ;
3558 case ETH_RSS_HASH_XOR:
3560 case ETH_RSS_HASH_CRC32:
3564 nn_warn(nn, "Unknown hash function: %u\n", nn->rss_hfunc);
3569 * nfp_net_rss_init() - Set the initial RSS parameters
3570 * @nn: NFP Net device to reconfigure
3572 static void nfp_net_rss_init(struct nfp_net *nn)
3574 unsigned long func_bit, rss_cap_hfunc;
3577 /* Read the RSS function capability and select first supported func */
3578 reg = nn_readl(nn, NFP_NET_CFG_RSS_CAP);
3579 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, reg);
3581 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC,
3582 NFP_NET_CFG_RSS_TOEPLITZ);
3584 func_bit = find_first_bit(&rss_cap_hfunc, NFP_NET_CFG_RSS_HFUNCS);
3585 if (func_bit == NFP_NET_CFG_RSS_HFUNCS) {
3586 dev_warn(nn->dp.dev,
3587 "Bad RSS config, defaulting to Toeplitz hash\n");
3588 func_bit = ETH_RSS_HASH_TOP_BIT;
3590 nn->rss_hfunc = 1 << func_bit;
3592 netdev_rss_key_fill(nn->rss_key, nfp_net_rss_key_sz(nn));
3594 nfp_net_rss_init_itbl(nn);
3596 /* Enable IPv4/IPv6 TCP by default */
3597 nn->rss_cfg = NFP_NET_CFG_RSS_IPV4_TCP |
3598 NFP_NET_CFG_RSS_IPV6_TCP |
3599 FIELD_PREP(NFP_NET_CFG_RSS_HFUNC, nn->rss_hfunc) |
3600 NFP_NET_CFG_RSS_MASK;
3604 * nfp_net_irqmod_init() - Set the initial IRQ moderation parameters
3605 * @nn: NFP Net device to reconfigure
3607 static void nfp_net_irqmod_init(struct nfp_net *nn)
3609 nn->rx_coalesce_usecs = 50;
3610 nn->rx_coalesce_max_frames = 64;
3611 nn->tx_coalesce_usecs = 50;
3612 nn->tx_coalesce_max_frames = 64;
3615 static void nfp_net_netdev_init(struct nfp_net *nn)
3617 struct net_device *netdev = nn->dp.netdev;
3619 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr);
3621 netdev->mtu = nn->dp.mtu;
3623 /* Advertise/enable offloads based on capabilities
3625 * Note: netdev->features show the currently enabled features
3626 * and netdev->hw_features advertises which features are
3627 * supported. By default we enable most features.
3629 if (nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR)
3630 netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
3632 netdev->hw_features = NETIF_F_HIGHDMA;
3633 if (nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY) {
3634 netdev->hw_features |= NETIF_F_RXCSUM;
3635 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY;
3637 if (nn->cap & NFP_NET_CFG_CTRL_TXCSUM) {
3638 netdev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
3639 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXCSUM;
3641 if (nn->cap & NFP_NET_CFG_CTRL_GATHER) {
3642 netdev->hw_features |= NETIF_F_SG;
3643 nn->dp.ctrl |= NFP_NET_CFG_CTRL_GATHER;
3645 if ((nn->cap & NFP_NET_CFG_CTRL_LSO && nn->fw_ver.major > 2) ||
3646 nn->cap & NFP_NET_CFG_CTRL_LSO2) {
3647 netdev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
3648 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?:
3649 NFP_NET_CFG_CTRL_LSO;
3651 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY)
3652 netdev->hw_features |= NETIF_F_RXHASH;
3653 if (nn->cap & NFP_NET_CFG_CTRL_VXLAN &&
3654 nn->cap & NFP_NET_CFG_CTRL_NVGRE) {
3655 if (nn->cap & NFP_NET_CFG_CTRL_LSO)
3656 netdev->hw_features |= NETIF_F_GSO_GRE |
3657 NETIF_F_GSO_UDP_TUNNEL;
3658 nn->dp.ctrl |= NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE;
3660 netdev->hw_enc_features = netdev->hw_features;
3663 netdev->vlan_features = netdev->hw_features;
3665 if (nn->cap & NFP_NET_CFG_CTRL_RXVLAN) {
3666 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
3667 nn->dp.ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
3669 if (nn->cap & NFP_NET_CFG_CTRL_TXVLAN) {
3670 if (nn->cap & NFP_NET_CFG_CTRL_LSO2) {
3671 nn_warn(nn, "Device advertises both TSO2 and TXVLAN. Refusing to enable TXVLAN.\n");
3673 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX;
3674 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
3677 if (nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER) {
3678 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3679 nn->dp.ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER;
3682 netdev->features = netdev->hw_features;
3684 if (nfp_app_has_tc(nn->app))
3685 netdev->hw_features |= NETIF_F_HW_TC;
3687 /* Advertise but disable TSO by default. */
3688 netdev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6);
3689 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY;
3691 /* Finalise the netdev setup */
3692 netdev->netdev_ops = &nfp_net_netdev_ops;
3693 netdev->watchdog_timeo = msecs_to_jiffies(5 * 1000);
3695 SWITCHDEV_SET_OPS(netdev, &nfp_port_switchdev_ops);
3697 /* MTU range: 68 - hw-specific max */
3698 netdev->min_mtu = ETH_MIN_MTU;
3699 netdev->max_mtu = nn->max_mtu;
3701 netif_carrier_off(netdev);
3703 nfp_net_set_ethtool_ops(netdev);
3707 * nfp_net_init() - Initialise/finalise the nfp_net structure
3708 * @nn: NFP Net device structure
3710 * Return: 0 on success or negative errno on error.
3712 int nfp_net_init(struct nfp_net *nn)
3716 nn->dp.rx_dma_dir = DMA_FROM_DEVICE;
3718 /* Get some of the read-only fields from the BAR */
3719 nn->cap = nn_readl(nn, NFP_NET_CFG_CAP);
3720 nn->max_mtu = nn_readl(nn, NFP_NET_CFG_MAX_MTU);
3722 /* ABI 4.x and ctrl vNIC always use chained metadata, in other cases
3723 * we allow use of non-chained metadata if RSS(v1) is the only
3724 * advertised capability requiring metadata.
3726 nn->dp.chained_metadata_format = nn->fw_ver.major == 4 ||
3728 !(nn->cap & NFP_NET_CFG_CTRL_RSS) ||
3729 nn->cap & NFP_NET_CFG_CTRL_CHAIN_META;
3730 /* RSS(v1) uses non-chained metadata format, except in ABI 4.x where
3731 * it has the same meaning as RSSv2.
3733 if (nn->dp.chained_metadata_format && nn->fw_ver.major != 4)
3734 nn->cap &= ~NFP_NET_CFG_CTRL_RSS;
3736 /* Determine RX packet/metadata boundary offset */
3737 if (nn->fw_ver.major >= 2) {
3740 reg = nn_readl(nn, NFP_NET_CFG_RX_OFFSET);
3741 if (reg > NFP_NET_MAX_PREPEND) {
3742 nn_err(nn, "Invalid rx offset: %d\n", reg);
3745 nn->dp.rx_offset = reg;
3747 nn->dp.rx_offset = NFP_NET_RX_OFFSET;
3750 /* Set default MTU and Freelist buffer size */
3751 if (nn->max_mtu < NFP_NET_DEFAULT_MTU)
3752 nn->dp.mtu = nn->max_mtu;
3754 nn->dp.mtu = NFP_NET_DEFAULT_MTU;
3755 nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp);
3757 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) {
3758 nfp_net_rss_init(nn);
3759 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?:
3760 NFP_NET_CFG_CTRL_RSS;
3763 /* Allow L2 Broadcast and Multicast through by default, if supported */
3764 if (nn->cap & NFP_NET_CFG_CTRL_L2BC)
3765 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2BC;
3766 if (nn->cap & NFP_NET_CFG_CTRL_L2MC)
3767 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2MC;
3769 /* Allow IRQ moderation, if supported */
3770 if (nn->cap & NFP_NET_CFG_CTRL_IRQMOD) {
3771 nfp_net_irqmod_init(nn);
3772 nn->dp.ctrl |= NFP_NET_CFG_CTRL_IRQMOD;
3776 nfp_net_netdev_init(nn);
3778 /* Stash the re-configuration queue away. First odd queue in TX Bar */
3779 nn->qcp_cfg = nn->tx_bar + NFP_QCP_QUEUE_ADDR_SZ;
3781 /* Make sure the FW knows the netdev is supposed to be disabled here */
3782 nn_writel(nn, NFP_NET_CFG_CTRL, 0);
3783 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0);
3784 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0);
3785 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_RING |
3786 NFP_NET_CFG_UPDATE_GEN);
3790 nfp_net_vecs_init(nn);
3794 return register_netdev(nn->dp.netdev);
3798 * nfp_net_clean() - Undo what nfp_net_init() did.
3799 * @nn: NFP Net device structure
3801 void nfp_net_clean(struct nfp_net *nn)
3806 unregister_netdev(nn->dp.netdev);