1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
2 /* Copyright (C) 2015-2018 Netronome Systems, Inc. */
6 * Netronome network device driver: Common functions between PF and VF
7 * Authors: Jakub Kicinski <jakub.kicinski@netronome.com>
8 * Jason McMullan <jason.mcmullan@netronome.com>
9 * Rolf Neugebauer <rolf.neugebauer@netronome.com>
10 * Brad Petrus <brad.petrus@netronome.com>
11 * Chris Telfer <chris.telfer@netronome.com>
14 #include <linux/bitfield.h>
15 #include <linux/bpf.h>
16 #include <linux/bpf_trace.h>
17 #include <linux/module.h>
18 #include <linux/kernel.h>
19 #include <linux/init.h>
21 #include <linux/netdevice.h>
22 #include <linux/etherdevice.h>
23 #include <linux/interrupt.h>
25 #include <linux/ipv6.h>
27 #include <linux/overflow.h>
28 #include <linux/page_ref.h>
29 #include <linux/pci.h>
30 #include <linux/pci_regs.h>
31 #include <linux/msi.h>
32 #include <linux/ethtool.h>
33 #include <linux/log2.h>
34 #include <linux/if_vlan.h>
35 #include <linux/random.h>
36 #include <linux/vmalloc.h>
37 #include <linux/ktime.h>
40 #include <net/vxlan.h>
42 #include "nfpcore/nfp_nsp.h"
45 #include "nfp_net_ctrl.h"
47 #include "nfp_net_sriov.h"
49 #include "crypto/crypto.h"
50 #include "crypto/fw.h"
53 * nfp_net_get_fw_version() - Read and parse the FW version
54 * @fw_ver: Output fw_version structure to read to
55 * @ctrl_bar: Mapped address of the control BAR
57 void nfp_net_get_fw_version(struct nfp_net_fw_version *fw_ver,
58 void __iomem *ctrl_bar)
62 reg = readl(ctrl_bar + NFP_NET_CFG_VERSION);
63 put_unaligned_le32(reg, fw_ver);
66 static dma_addr_t nfp_net_dma_map_rx(struct nfp_net_dp *dp, void *frag)
68 return dma_map_single_attrs(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM,
69 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
70 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
74 nfp_net_dma_sync_dev_rx(const struct nfp_net_dp *dp, dma_addr_t dma_addr)
76 dma_sync_single_for_device(dp->dev, dma_addr,
77 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
81 static void nfp_net_dma_unmap_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr)
83 dma_unmap_single_attrs(dp->dev, dma_addr,
84 dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
85 dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
88 static void nfp_net_dma_sync_cpu_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr,
91 dma_sync_single_for_cpu(dp->dev, dma_addr - NFP_NET_RX_BUF_HEADROOM,
97 * Firmware reconfig may take a while so we have two versions of it -
98 * synchronous and asynchronous (posted). All synchronous callers are holding
99 * RTNL so we don't have to worry about serializing them.
101 static void nfp_net_reconfig_start(struct nfp_net *nn, u32 update)
103 nn_writel(nn, NFP_NET_CFG_UPDATE, update);
104 /* ensure update is written before pinging HW */
106 nfp_qcp_wr_ptr_add(nn->qcp_cfg, 1);
107 nn->reconfig_in_progress_update = update;
110 /* Pass 0 as update to run posted reconfigs. */
111 static void nfp_net_reconfig_start_async(struct nfp_net *nn, u32 update)
113 update |= nn->reconfig_posted;
114 nn->reconfig_posted = 0;
116 nfp_net_reconfig_start(nn, update);
118 nn->reconfig_timer_active = true;
119 mod_timer(&nn->reconfig_timer, jiffies + NFP_NET_POLL_TIMEOUT * HZ);
122 static bool nfp_net_reconfig_check_done(struct nfp_net *nn, bool last_check)
126 reg = nn_readl(nn, NFP_NET_CFG_UPDATE);
129 if (reg & NFP_NET_CFG_UPDATE_ERR) {
130 nn_err(nn, "Reconfig error (status: 0x%08x update: 0x%08x ctrl: 0x%08x)\n",
131 reg, nn->reconfig_in_progress_update,
132 nn_readl(nn, NFP_NET_CFG_CTRL));
134 } else if (last_check) {
135 nn_err(nn, "Reconfig timeout (status: 0x%08x update: 0x%08x ctrl: 0x%08x)\n",
136 reg, nn->reconfig_in_progress_update,
137 nn_readl(nn, NFP_NET_CFG_CTRL));
144 static bool __nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline)
146 bool timed_out = false;
149 /* Poll update field, waiting for NFP to ack the config.
150 * Do an opportunistic wait-busy loop, afterward sleep.
152 for (i = 0; i < 50; i++) {
153 if (nfp_net_reconfig_check_done(nn, false))
158 while (!nfp_net_reconfig_check_done(nn, timed_out)) {
159 usleep_range(250, 500);
160 timed_out = time_is_before_eq_jiffies(deadline);
166 static int nfp_net_reconfig_wait(struct nfp_net *nn, unsigned long deadline)
168 if (__nfp_net_reconfig_wait(nn, deadline))
171 if (nn_readl(nn, NFP_NET_CFG_UPDATE) & NFP_NET_CFG_UPDATE_ERR)
177 static void nfp_net_reconfig_timer(struct timer_list *t)
179 struct nfp_net *nn = from_timer(nn, t, reconfig_timer);
181 spin_lock_bh(&nn->reconfig_lock);
183 nn->reconfig_timer_active = false;
185 /* If sync caller is present it will take over from us */
186 if (nn->reconfig_sync_present)
189 /* Read reconfig status and report errors */
190 nfp_net_reconfig_check_done(nn, true);
192 if (nn->reconfig_posted)
193 nfp_net_reconfig_start_async(nn, 0);
195 spin_unlock_bh(&nn->reconfig_lock);
199 * nfp_net_reconfig_post() - Post async reconfig request
200 * @nn: NFP Net device to reconfigure
201 * @update: The value for the update field in the BAR config
203 * Record FW reconfiguration request. Reconfiguration will be kicked off
204 * whenever reconfiguration machinery is idle. Multiple requests can be
207 static void nfp_net_reconfig_post(struct nfp_net *nn, u32 update)
209 spin_lock_bh(&nn->reconfig_lock);
211 /* Sync caller will kick off async reconf when it's done, just post */
212 if (nn->reconfig_sync_present) {
213 nn->reconfig_posted |= update;
217 /* Opportunistically check if the previous command is done */
218 if (!nn->reconfig_timer_active ||
219 nfp_net_reconfig_check_done(nn, false))
220 nfp_net_reconfig_start_async(nn, update);
222 nn->reconfig_posted |= update;
224 spin_unlock_bh(&nn->reconfig_lock);
227 static void nfp_net_reconfig_sync_enter(struct nfp_net *nn)
229 bool cancelled_timer = false;
230 u32 pre_posted_requests;
232 spin_lock_bh(&nn->reconfig_lock);
234 WARN_ON(nn->reconfig_sync_present);
235 nn->reconfig_sync_present = true;
237 if (nn->reconfig_timer_active) {
238 nn->reconfig_timer_active = false;
239 cancelled_timer = true;
241 pre_posted_requests = nn->reconfig_posted;
242 nn->reconfig_posted = 0;
244 spin_unlock_bh(&nn->reconfig_lock);
246 if (cancelled_timer) {
247 del_timer_sync(&nn->reconfig_timer);
248 nfp_net_reconfig_wait(nn, nn->reconfig_timer.expires);
251 /* Run the posted reconfigs which were issued before we started */
252 if (pre_posted_requests) {
253 nfp_net_reconfig_start(nn, pre_posted_requests);
254 nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT);
258 static void nfp_net_reconfig_wait_posted(struct nfp_net *nn)
260 nfp_net_reconfig_sync_enter(nn);
262 spin_lock_bh(&nn->reconfig_lock);
263 nn->reconfig_sync_present = false;
264 spin_unlock_bh(&nn->reconfig_lock);
268 * __nfp_net_reconfig() - Reconfigure the firmware
269 * @nn: NFP Net device to reconfigure
270 * @update: The value for the update field in the BAR config
272 * Write the update word to the BAR and ping the reconfig queue. The
273 * poll until the firmware has acknowledged the update by zeroing the
276 * Return: Negative errno on error, 0 on success
278 int __nfp_net_reconfig(struct nfp_net *nn, u32 update)
282 nfp_net_reconfig_sync_enter(nn);
284 nfp_net_reconfig_start(nn, update);
285 ret = nfp_net_reconfig_wait(nn, jiffies + HZ * NFP_NET_POLL_TIMEOUT);
287 spin_lock_bh(&nn->reconfig_lock);
289 if (nn->reconfig_posted)
290 nfp_net_reconfig_start_async(nn, 0);
292 nn->reconfig_sync_present = false;
294 spin_unlock_bh(&nn->reconfig_lock);
299 int nfp_net_reconfig(struct nfp_net *nn, u32 update)
303 nn_ctrl_bar_lock(nn);
304 ret = __nfp_net_reconfig(nn, update);
305 nn_ctrl_bar_unlock(nn);
310 int nfp_net_mbox_lock(struct nfp_net *nn, unsigned int data_size)
312 if (nn->tlv_caps.mbox_len < NFP_NET_CFG_MBOX_SIMPLE_VAL + data_size) {
313 nn_err(nn, "mailbox too small for %u of data (%u)\n",
314 data_size, nn->tlv_caps.mbox_len);
318 nn_ctrl_bar_lock(nn);
323 * nfp_net_mbox_reconfig() - Reconfigure the firmware via the mailbox
324 * @nn: NFP Net device to reconfigure
325 * @mbox_cmd: The value for the mailbox command
327 * Helper function for mailbox updates
329 * Return: Negative errno on error, 0 on success
331 int nfp_net_mbox_reconfig(struct nfp_net *nn, u32 mbox_cmd)
333 u32 mbox = nn->tlv_caps.mbox_off;
336 nn_writeq(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_CMD, mbox_cmd);
338 ret = __nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MBOX);
340 nn_err(nn, "Mailbox update error\n");
344 return -nn_readl(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_RET);
347 void nfp_net_mbox_reconfig_post(struct nfp_net *nn, u32 mbox_cmd)
349 u32 mbox = nn->tlv_caps.mbox_off;
351 nn_writeq(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_CMD, mbox_cmd);
353 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_MBOX);
356 int nfp_net_mbox_reconfig_wait_posted(struct nfp_net *nn)
358 u32 mbox = nn->tlv_caps.mbox_off;
360 nfp_net_reconfig_wait_posted(nn);
362 return -nn_readl(nn, mbox + NFP_NET_CFG_MBOX_SIMPLE_RET);
365 int nfp_net_mbox_reconfig_and_unlock(struct nfp_net *nn, u32 mbox_cmd)
369 ret = nfp_net_mbox_reconfig(nn, mbox_cmd);
370 nn_ctrl_bar_unlock(nn);
374 /* Interrupt configuration and handling
378 * nfp_net_irq_unmask() - Unmask automasked interrupt
379 * @nn: NFP Network structure
380 * @entry_nr: MSI-X table entry
382 * Clear the ICR for the IRQ entry.
384 static void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr)
386 nn_writeb(nn, NFP_NET_CFG_ICR(entry_nr), NFP_NET_CFG_ICR_UNMASKED);
391 * nfp_net_irqs_alloc() - allocates MSI-X irqs
392 * @pdev: PCI device structure
393 * @irq_entries: Array to be initialized and used to hold the irq entries
394 * @min_irqs: Minimal acceptable number of interrupts
395 * @wanted_irqs: Target number of interrupts to allocate
397 * Return: Number of irqs obtained or 0 on error.
400 nfp_net_irqs_alloc(struct pci_dev *pdev, struct msix_entry *irq_entries,
401 unsigned int min_irqs, unsigned int wanted_irqs)
406 for (i = 0; i < wanted_irqs; i++)
407 irq_entries[i].entry = i;
409 got_irqs = pci_enable_msix_range(pdev, irq_entries,
410 min_irqs, wanted_irqs);
412 dev_err(&pdev->dev, "Failed to enable %d-%d MSI-X (err=%d)\n",
413 min_irqs, wanted_irqs, got_irqs);
417 if (got_irqs < wanted_irqs)
418 dev_warn(&pdev->dev, "Unable to allocate %d IRQs got only %d\n",
419 wanted_irqs, got_irqs);
425 * nfp_net_irqs_assign() - Assign interrupts allocated externally to netdev
426 * @nn: NFP Network structure
427 * @irq_entries: Table of allocated interrupts
428 * @n: Size of @irq_entries (number of entries to grab)
430 * After interrupts are allocated with nfp_net_irqs_alloc() this function
431 * should be called to assign them to a specific netdev (port).
434 nfp_net_irqs_assign(struct nfp_net *nn, struct msix_entry *irq_entries,
437 struct nfp_net_dp *dp = &nn->dp;
439 nn->max_r_vecs = n - NFP_NET_NON_Q_VECTORS;
440 dp->num_r_vecs = nn->max_r_vecs;
442 memcpy(nn->irq_entries, irq_entries, sizeof(*irq_entries) * n);
444 if (dp->num_rx_rings > dp->num_r_vecs ||
445 dp->num_tx_rings > dp->num_r_vecs)
446 dev_warn(nn->dp.dev, "More rings (%d,%d) than vectors (%d).\n",
447 dp->num_rx_rings, dp->num_tx_rings,
450 dp->num_rx_rings = min(dp->num_r_vecs, dp->num_rx_rings);
451 dp->num_tx_rings = min(dp->num_r_vecs, dp->num_tx_rings);
452 dp->num_stack_tx_rings = dp->num_tx_rings;
456 * nfp_net_irqs_disable() - Disable interrupts
457 * @pdev: PCI device structure
459 * Undoes what @nfp_net_irqs_alloc() does.
461 void nfp_net_irqs_disable(struct pci_dev *pdev)
463 pci_disable_msix(pdev);
467 * nfp_net_irq_rxtx() - Interrupt service routine for RX/TX rings.
469 * @data: Opaque data structure
471 * Return: Indicate if the interrupt has been handled.
473 static irqreturn_t nfp_net_irq_rxtx(int irq, void *data)
475 struct nfp_net_r_vector *r_vec = data;
477 napi_schedule_irqoff(&r_vec->napi);
479 /* The FW auto-masks any interrupt, either via the MASK bit in
480 * the MSI-X table or via the per entry ICR field. So there
481 * is no need to disable interrupts here.
486 static irqreturn_t nfp_ctrl_irq_rxtx(int irq, void *data)
488 struct nfp_net_r_vector *r_vec = data;
490 tasklet_schedule(&r_vec->tasklet);
496 * nfp_net_read_link_status() - Reread link status from control BAR
497 * @nn: NFP Network structure
499 static void nfp_net_read_link_status(struct nfp_net *nn)
505 spin_lock_irqsave(&nn->link_status_lock, flags);
507 sts = nn_readl(nn, NFP_NET_CFG_STS);
508 link_up = !!(sts & NFP_NET_CFG_STS_LINK);
510 if (nn->link_up == link_up)
513 nn->link_up = link_up;
515 set_bit(NFP_PORT_CHANGED, &nn->port->flags);
518 netif_carrier_on(nn->dp.netdev);
519 netdev_info(nn->dp.netdev, "NIC Link is Up\n");
521 netif_carrier_off(nn->dp.netdev);
522 netdev_info(nn->dp.netdev, "NIC Link is Down\n");
525 spin_unlock_irqrestore(&nn->link_status_lock, flags);
529 * nfp_net_irq_lsc() - Interrupt service routine for link state changes
531 * @data: Opaque data structure
533 * Return: Indicate if the interrupt has been handled.
535 static irqreturn_t nfp_net_irq_lsc(int irq, void *data)
537 struct nfp_net *nn = data;
538 struct msix_entry *entry;
540 entry = &nn->irq_entries[NFP_NET_IRQ_LSC_IDX];
542 nfp_net_read_link_status(nn);
544 nfp_net_irq_unmask(nn, entry->entry);
550 * nfp_net_irq_exn() - Interrupt service routine for exceptions
552 * @data: Opaque data structure
554 * Return: Indicate if the interrupt has been handled.
556 static irqreturn_t nfp_net_irq_exn(int irq, void *data)
558 struct nfp_net *nn = data;
560 nn_err(nn, "%s: UNIMPLEMENTED.\n", __func__);
561 /* XXX TO BE IMPLEMENTED */
566 * nfp_net_tx_ring_init() - Fill in the boilerplate for a TX ring
567 * @tx_ring: TX ring structure
568 * @r_vec: IRQ vector servicing this ring
570 * @is_xdp: Is this an XDP TX ring?
573 nfp_net_tx_ring_init(struct nfp_net_tx_ring *tx_ring,
574 struct nfp_net_r_vector *r_vec, unsigned int idx,
577 struct nfp_net *nn = r_vec->nfp_net;
580 tx_ring->r_vec = r_vec;
581 tx_ring->is_xdp = is_xdp;
582 u64_stats_init(&tx_ring->r_vec->tx_sync);
584 tx_ring->qcidx = tx_ring->idx * nn->stride_tx;
585 tx_ring->qcp_q = nn->tx_bar + NFP_QCP_QUEUE_OFF(tx_ring->qcidx);
589 * nfp_net_rx_ring_init() - Fill in the boilerplate for a RX ring
590 * @rx_ring: RX ring structure
591 * @r_vec: IRQ vector servicing this ring
595 nfp_net_rx_ring_init(struct nfp_net_rx_ring *rx_ring,
596 struct nfp_net_r_vector *r_vec, unsigned int idx)
598 struct nfp_net *nn = r_vec->nfp_net;
601 rx_ring->r_vec = r_vec;
602 u64_stats_init(&rx_ring->r_vec->rx_sync);
604 rx_ring->fl_qcidx = rx_ring->idx * nn->stride_rx;
605 rx_ring->qcp_fl = nn->rx_bar + NFP_QCP_QUEUE_OFF(rx_ring->fl_qcidx);
609 * nfp_net_aux_irq_request() - Request an auxiliary interrupt (LSC or EXN)
610 * @nn: NFP Network structure
611 * @ctrl_offset: Control BAR offset where IRQ configuration should be written
612 * @format: printf-style format to construct the interrupt name
613 * @name: Pointer to allocated space for interrupt name
614 * @name_sz: Size of space for interrupt name
615 * @vector_idx: Index of MSI-X vector used for this interrupt
616 * @handler: IRQ handler to register for this interrupt
619 nfp_net_aux_irq_request(struct nfp_net *nn, u32 ctrl_offset,
620 const char *format, char *name, size_t name_sz,
621 unsigned int vector_idx, irq_handler_t handler)
623 struct msix_entry *entry;
626 entry = &nn->irq_entries[vector_idx];
628 snprintf(name, name_sz, format, nfp_net_name(nn));
629 err = request_irq(entry->vector, handler, 0, name, nn);
631 nn_err(nn, "Failed to request IRQ %d (err=%d).\n",
635 nn_writeb(nn, ctrl_offset, entry->entry);
636 nfp_net_irq_unmask(nn, entry->entry);
642 * nfp_net_aux_irq_free() - Free an auxiliary interrupt (LSC or EXN)
643 * @nn: NFP Network structure
644 * @ctrl_offset: Control BAR offset where IRQ configuration should be written
645 * @vector_idx: Index of MSI-X vector used for this interrupt
647 static void nfp_net_aux_irq_free(struct nfp_net *nn, u32 ctrl_offset,
648 unsigned int vector_idx)
650 nn_writeb(nn, ctrl_offset, 0xff);
652 free_irq(nn->irq_entries[vector_idx].vector, nn);
657 * One queue controller peripheral queue is used for transmit. The
658 * driver en-queues packets for transmit by advancing the write
659 * pointer. The device indicates that packets have transmitted by
660 * advancing the read pointer. The driver maintains a local copy of
661 * the read and write pointer in @struct nfp_net_tx_ring. The driver
662 * keeps @wr_p in sync with the queue controller write pointer and can
663 * determine how many packets have been transmitted by comparing its
664 * copy of the read pointer @rd_p with the read pointer maintained by
665 * the queue controller peripheral.
669 * nfp_net_tx_full() - Check if the TX ring is full
670 * @tx_ring: TX ring to check
671 * @dcnt: Number of descriptors that need to be enqueued (must be >= 1)
673 * This function checks, based on the *host copy* of read/write
674 * pointer if a given TX ring is full. The real TX queue may have
675 * some newly made available slots.
677 * Return: True if the ring is full.
679 static int nfp_net_tx_full(struct nfp_net_tx_ring *tx_ring, int dcnt)
681 return (tx_ring->wr_p - tx_ring->rd_p) >= (tx_ring->cnt - dcnt);
684 /* Wrappers for deciding when to stop and restart TX queues */
685 static int nfp_net_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
687 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
690 static int nfp_net_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
692 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
696 * nfp_net_tx_ring_stop() - stop tx ring
697 * @nd_q: netdev queue
698 * @tx_ring: driver tx queue structure
700 * Safely stop TX ring. Remember that while we are running .start_xmit()
701 * someone else may be cleaning the TX ring completions so we need to be
702 * extra careful here.
704 static void nfp_net_tx_ring_stop(struct netdev_queue *nd_q,
705 struct nfp_net_tx_ring *tx_ring)
707 netif_tx_stop_queue(nd_q);
709 /* We can race with the TX completion out of NAPI so recheck */
711 if (unlikely(nfp_net_tx_ring_should_wake(tx_ring)))
712 netif_tx_start_queue(nd_q);
716 * nfp_net_tx_tso() - Set up Tx descriptor for LSO
717 * @r_vec: per-ring structure
718 * @txbuf: Pointer to driver soft TX descriptor
719 * @txd: Pointer to HW TX descriptor
720 * @skb: Pointer to SKB
721 * @md_bytes: Prepend length
723 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
724 * Return error on packet header greater than maximum supported LSO header size.
726 static void nfp_net_tx_tso(struct nfp_net_r_vector *r_vec,
727 struct nfp_net_tx_buf *txbuf,
728 struct nfp_net_tx_desc *txd, struct sk_buff *skb,
731 u32 l3_offset, l4_offset, hdrlen;
734 if (!skb_is_gso(skb))
737 if (!skb->encapsulation) {
738 l3_offset = skb_network_offset(skb);
739 l4_offset = skb_transport_offset(skb);
740 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
742 l3_offset = skb_inner_network_offset(skb);
743 l4_offset = skb_inner_transport_offset(skb);
744 hdrlen = skb_inner_transport_header(skb) - skb->data +
745 inner_tcp_hdrlen(skb);
748 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
749 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
751 mss = skb_shinfo(skb)->gso_size & PCIE_DESC_TX_MSS_MASK;
752 txd->l3_offset = l3_offset - md_bytes;
753 txd->l4_offset = l4_offset - md_bytes;
754 txd->lso_hdrlen = hdrlen - md_bytes;
755 txd->mss = cpu_to_le16(mss);
756 txd->flags |= PCIE_DESC_TX_LSO;
758 u64_stats_update_begin(&r_vec->tx_sync);
760 u64_stats_update_end(&r_vec->tx_sync);
764 * nfp_net_tx_csum() - Set TX CSUM offload flags in TX descriptor
765 * @dp: NFP Net data path struct
766 * @r_vec: per-ring structure
767 * @txbuf: Pointer to driver soft TX descriptor
768 * @txd: Pointer to TX descriptor
769 * @skb: Pointer to SKB
771 * This function sets the TX checksum flags in the TX descriptor based
772 * on the configuration and the protocol of the packet to be transmitted.
774 static void nfp_net_tx_csum(struct nfp_net_dp *dp,
775 struct nfp_net_r_vector *r_vec,
776 struct nfp_net_tx_buf *txbuf,
777 struct nfp_net_tx_desc *txd, struct sk_buff *skb)
779 struct ipv6hdr *ipv6h;
783 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
786 if (skb->ip_summed != CHECKSUM_PARTIAL)
789 txd->flags |= PCIE_DESC_TX_CSUM;
790 if (skb->encapsulation)
791 txd->flags |= PCIE_DESC_TX_ENCAP;
793 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
794 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
796 if (iph->version == 4) {
797 txd->flags |= PCIE_DESC_TX_IP4_CSUM;
798 l4_hdr = iph->protocol;
799 } else if (ipv6h->version == 6) {
800 l4_hdr = ipv6h->nexthdr;
802 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
808 txd->flags |= PCIE_DESC_TX_TCP_CSUM;
811 txd->flags |= PCIE_DESC_TX_UDP_CSUM;
814 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
818 u64_stats_update_begin(&r_vec->tx_sync);
819 if (skb->encapsulation)
820 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
822 r_vec->hw_csum_tx += txbuf->pkt_cnt;
823 u64_stats_update_end(&r_vec->tx_sync);
826 static struct sk_buff *
827 nfp_net_tls_tx(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
828 struct sk_buff *skb, u64 *tls_handle, int *nr_frags)
830 #ifdef CONFIG_TLS_DEVICE
831 struct nfp_net_tls_offload_ctx *ntls;
832 struct sk_buff *nskb;
836 if (likely(!dp->ktls_tx))
838 if (!skb->sk || !tls_is_sk_tx_device_offloaded(skb->sk))
841 datalen = skb->len - (skb_transport_offset(skb) + tcp_hdrlen(skb));
842 seq = ntohl(tcp_hdr(skb)->seq);
843 ntls = tls_driver_ctx(skb->sk, TLS_OFFLOAD_CTX_DIR_TX);
844 resync_pending = tls_offload_tx_resync_pending(skb->sk);
845 if (unlikely(resync_pending || ntls->next_seq != seq)) {
846 /* Pure ACK out of order already */
850 u64_stats_update_begin(&r_vec->tx_sync);
851 r_vec->tls_tx_fallback++;
852 u64_stats_update_end(&r_vec->tx_sync);
854 nskb = tls_encrypt_skb(skb);
856 u64_stats_update_begin(&r_vec->tx_sync);
857 r_vec->tls_tx_no_fallback++;
858 u64_stats_update_end(&r_vec->tx_sync);
861 /* encryption wasn't necessary */
864 /* we don't re-check ring space */
865 if (unlikely(skb_is_nonlinear(nskb))) {
866 nn_dp_warn(dp, "tls_encrypt_skb() produced fragmented frame\n");
867 u64_stats_update_begin(&r_vec->tx_sync);
869 u64_stats_update_end(&r_vec->tx_sync);
870 dev_kfree_skb_any(nskb);
874 /* jump forward, a TX may have gotten lost, need to sync TX */
875 if (!resync_pending && seq - ntls->next_seq < U32_MAX / 4)
876 tls_offload_tx_resync_request(nskb->sk, seq,
884 u64_stats_update_begin(&r_vec->tx_sync);
885 if (!skb_is_gso(skb))
888 r_vec->hw_tls_tx += skb_shinfo(skb)->gso_segs;
889 u64_stats_update_end(&r_vec->tx_sync);
892 memcpy(tls_handle, ntls->fw_handle, sizeof(ntls->fw_handle));
893 ntls->next_seq += datalen;
898 static void nfp_net_tls_tx_undo(struct sk_buff *skb, u64 tls_handle)
900 #ifdef CONFIG_TLS_DEVICE
901 struct nfp_net_tls_offload_ctx *ntls;
906 if (WARN_ON_ONCE(!skb->sk || !tls_is_sk_tx_device_offloaded(skb->sk)))
909 datalen = skb->len - (skb_transport_offset(skb) + tcp_hdrlen(skb));
910 seq = ntohl(tcp_hdr(skb)->seq);
912 ntls = tls_driver_ctx(skb->sk, TLS_OFFLOAD_CTX_DIR_TX);
913 if (ntls->next_seq == seq + datalen)
914 ntls->next_seq = seq;
920 static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring)
923 nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add);
924 tx_ring->wr_ptr_add = 0;
927 static int nfp_net_prep_tx_meta(struct sk_buff *skb, u64 tls_handle)
929 struct metadata_dst *md_dst = skb_metadata_dst(skb);
934 if (likely(!md_dst && !tls_handle))
936 if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX)) {
942 md_bytes = 4 + !!md_dst * 4 + !!tls_handle * 8;
944 if (unlikely(skb_cow_head(skb, md_bytes)))
948 data = skb_push(skb, md_bytes) + md_bytes;
951 put_unaligned_be32(md_dst->u.port_info.port_id, data);
952 meta_id = NFP_NET_META_PORTID;
955 /* conn handle is opaque, we just use u64 to be able to quickly
959 memcpy(data, &tls_handle, sizeof(tls_handle));
960 meta_id <<= NFP_NET_META_FIELD_SIZE;
961 meta_id |= NFP_NET_META_CONN_HANDLE;
965 put_unaligned_be32(meta_id, data);
971 * nfp_net_tx() - Main transmit entry point
972 * @skb: SKB to transmit
973 * @netdev: netdev structure
975 * Return: NETDEV_TX_OK on success.
977 static netdev_tx_t nfp_net_tx(struct sk_buff *skb, struct net_device *netdev)
979 struct nfp_net *nn = netdev_priv(netdev);
980 const skb_frag_t *frag;
981 int f, nr_frags, wr_idx, md_bytes;
982 struct nfp_net_tx_ring *tx_ring;
983 struct nfp_net_r_vector *r_vec;
984 struct nfp_net_tx_buf *txbuf;
985 struct nfp_net_tx_desc *txd;
986 struct netdev_queue *nd_q;
987 struct nfp_net_dp *dp;
994 qidx = skb_get_queue_mapping(skb);
995 tx_ring = &dp->tx_rings[qidx];
996 r_vec = tx_ring->r_vec;
998 nr_frags = skb_shinfo(skb)->nr_frags;
1000 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
1001 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
1002 qidx, tx_ring->wr_p, tx_ring->rd_p);
1003 nd_q = netdev_get_tx_queue(dp->netdev, qidx);
1004 netif_tx_stop_queue(nd_q);
1005 nfp_net_tx_xmit_more_flush(tx_ring);
1006 u64_stats_update_begin(&r_vec->tx_sync);
1008 u64_stats_update_end(&r_vec->tx_sync);
1009 return NETDEV_TX_BUSY;
1012 skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags);
1013 if (unlikely(!skb)) {
1014 nfp_net_tx_xmit_more_flush(tx_ring);
1015 return NETDEV_TX_OK;
1018 md_bytes = nfp_net_prep_tx_meta(skb, tls_handle);
1019 if (unlikely(md_bytes < 0))
1022 /* Start with the head skbuf */
1023 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1025 if (dma_mapping_error(dp->dev, dma_addr))
1028 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1030 /* Stash the soft descriptor of the head then initialize it */
1031 txbuf = &tx_ring->txbufs[wr_idx];
1033 txbuf->dma_addr = dma_addr;
1036 txbuf->real_len = skb->len;
1038 /* Build TX descriptor */
1039 txd = &tx_ring->txds[wr_idx];
1040 txd->offset_eop = (nr_frags ? 0 : PCIE_DESC_TX_EOP) | md_bytes;
1041 txd->dma_len = cpu_to_le16(skb_headlen(skb));
1042 nfp_desc_set_dma_addr(txd, dma_addr);
1043 txd->data_len = cpu_to_le16(skb->len);
1047 txd->lso_hdrlen = 0;
1049 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
1050 nfp_net_tx_tso(r_vec, txbuf, txd, skb, md_bytes);
1051 nfp_net_tx_csum(dp, r_vec, txbuf, txd, skb);
1052 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
1053 txd->flags |= PCIE_DESC_TX_VLAN;
1054 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
1061 /* all descs must match except for in addr, length and eop */
1062 second_half = txd->vals8[1];
1064 for (f = 0; f < nr_frags; f++) {
1065 frag = &skb_shinfo(skb)->frags[f];
1066 fsize = skb_frag_size(frag);
1068 dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
1069 fsize, DMA_TO_DEVICE);
1070 if (dma_mapping_error(dp->dev, dma_addr))
1073 wr_idx = D_IDX(tx_ring, wr_idx + 1);
1074 tx_ring->txbufs[wr_idx].skb = skb;
1075 tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
1076 tx_ring->txbufs[wr_idx].fidx = f;
1078 txd = &tx_ring->txds[wr_idx];
1079 txd->dma_len = cpu_to_le16(fsize);
1080 nfp_desc_set_dma_addr(txd, dma_addr);
1081 txd->offset_eop = md_bytes |
1082 ((f == nr_frags - 1) ? PCIE_DESC_TX_EOP : 0);
1083 txd->vals8[1] = second_half;
1086 u64_stats_update_begin(&r_vec->tx_sync);
1088 u64_stats_update_end(&r_vec->tx_sync);
1091 skb_tx_timestamp(skb);
1093 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
1095 tx_ring->wr_p += nr_frags + 1;
1096 if (nfp_net_tx_ring_should_stop(tx_ring))
1097 nfp_net_tx_ring_stop(nd_q, tx_ring);
1099 tx_ring->wr_ptr_add += nr_frags + 1;
1100 if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, netdev_xmit_more()))
1101 nfp_net_tx_xmit_more_flush(tx_ring);
1103 return NETDEV_TX_OK;
1107 frag = &skb_shinfo(skb)->frags[f];
1108 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
1109 skb_frag_size(frag), DMA_TO_DEVICE);
1110 tx_ring->txbufs[wr_idx].skb = NULL;
1111 tx_ring->txbufs[wr_idx].dma_addr = 0;
1112 tx_ring->txbufs[wr_idx].fidx = -2;
1113 wr_idx = wr_idx - 1;
1115 wr_idx += tx_ring->cnt;
1117 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
1118 skb_headlen(skb), DMA_TO_DEVICE);
1119 tx_ring->txbufs[wr_idx].skb = NULL;
1120 tx_ring->txbufs[wr_idx].dma_addr = 0;
1121 tx_ring->txbufs[wr_idx].fidx = -2;
1123 nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
1125 nfp_net_tx_xmit_more_flush(tx_ring);
1126 u64_stats_update_begin(&r_vec->tx_sync);
1128 u64_stats_update_end(&r_vec->tx_sync);
1129 nfp_net_tls_tx_undo(skb, tls_handle);
1130 dev_kfree_skb_any(skb);
1131 return NETDEV_TX_OK;
1135 * nfp_net_tx_complete() - Handled completed TX packets
1136 * @tx_ring: TX ring structure
1137 * @budget: NAPI budget (only used as bool to determine if in NAPI context)
1139 static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
1141 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
1142 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
1143 struct netdev_queue *nd_q;
1144 u32 done_pkts = 0, done_bytes = 0;
1148 if (tx_ring->wr_p == tx_ring->rd_p)
1151 /* Work out how many descriptors have been transmitted */
1152 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
1154 if (qcp_rd_p == tx_ring->qcp_rd_p)
1157 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
1160 const skb_frag_t *frag;
1161 struct nfp_net_tx_buf *tx_buf;
1162 struct sk_buff *skb;
1166 idx = D_IDX(tx_ring, tx_ring->rd_p++);
1167 tx_buf = &tx_ring->txbufs[idx];
1173 nr_frags = skb_shinfo(skb)->nr_frags;
1174 fidx = tx_buf->fidx;
1178 dma_unmap_single(dp->dev, tx_buf->dma_addr,
1179 skb_headlen(skb), DMA_TO_DEVICE);
1181 done_pkts += tx_buf->pkt_cnt;
1182 done_bytes += tx_buf->real_len;
1184 /* unmap fragment */
1185 frag = &skb_shinfo(skb)->frags[fidx];
1186 dma_unmap_page(dp->dev, tx_buf->dma_addr,
1187 skb_frag_size(frag), DMA_TO_DEVICE);
1190 /* check for last gather fragment */
1191 if (fidx == nr_frags - 1)
1192 napi_consume_skb(skb, budget);
1194 tx_buf->dma_addr = 0;
1199 tx_ring->qcp_rd_p = qcp_rd_p;
1201 u64_stats_update_begin(&r_vec->tx_sync);
1202 r_vec->tx_bytes += done_bytes;
1203 r_vec->tx_pkts += done_pkts;
1204 u64_stats_update_end(&r_vec->tx_sync);
1209 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
1210 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
1211 if (nfp_net_tx_ring_should_wake(tx_ring)) {
1212 /* Make sure TX thread will see updated tx_ring->rd_p */
1215 if (unlikely(netif_tx_queue_stopped(nd_q)))
1216 netif_tx_wake_queue(nd_q);
1219 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
1220 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
1221 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
1224 static bool nfp_net_xdp_complete(struct nfp_net_tx_ring *tx_ring)
1226 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
1227 u32 done_pkts = 0, done_bytes = 0;
1232 /* Work out how many descriptors have been transmitted */
1233 qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
1235 if (qcp_rd_p == tx_ring->qcp_rd_p)
1238 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
1240 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
1241 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
1243 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
1247 idx = D_IDX(tx_ring, tx_ring->rd_p);
1250 done_bytes += tx_ring->txbufs[idx].real_len;
1253 u64_stats_update_begin(&r_vec->tx_sync);
1254 r_vec->tx_bytes += done_bytes;
1255 r_vec->tx_pkts += done_pkts;
1256 u64_stats_update_end(&r_vec->tx_sync);
1258 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
1259 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
1260 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
1266 * nfp_net_tx_ring_reset() - Free any untransmitted buffers and reset pointers
1267 * @dp: NFP Net data path struct
1268 * @tx_ring: TX ring structure
1270 * Assumes that the device is stopped, must be idempotent.
1273 nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
1275 const skb_frag_t *frag;
1276 struct netdev_queue *nd_q;
1278 while (!tx_ring->is_xdp && tx_ring->rd_p != tx_ring->wr_p) {
1279 struct nfp_net_tx_buf *tx_buf;
1280 struct sk_buff *skb;
1283 idx = D_IDX(tx_ring, tx_ring->rd_p);
1284 tx_buf = &tx_ring->txbufs[idx];
1286 skb = tx_ring->txbufs[idx].skb;
1287 nr_frags = skb_shinfo(skb)->nr_frags;
1289 if (tx_buf->fidx == -1) {
1291 dma_unmap_single(dp->dev, tx_buf->dma_addr,
1292 skb_headlen(skb), DMA_TO_DEVICE);
1294 /* unmap fragment */
1295 frag = &skb_shinfo(skb)->frags[tx_buf->fidx];
1296 dma_unmap_page(dp->dev, tx_buf->dma_addr,
1297 skb_frag_size(frag), DMA_TO_DEVICE);
1300 /* check for last gather fragment */
1301 if (tx_buf->fidx == nr_frags - 1)
1302 dev_kfree_skb_any(skb);
1304 tx_buf->dma_addr = 0;
1308 tx_ring->qcp_rd_p++;
1312 memset(tx_ring->txds, 0, tx_ring->size);
1315 tx_ring->qcp_rd_p = 0;
1316 tx_ring->wr_ptr_add = 0;
1318 if (tx_ring->is_xdp || !dp->netdev)
1321 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
1322 netdev_tx_reset_queue(nd_q);
1325 static void nfp_net_tx_timeout(struct net_device *netdev, unsigned int txqueue)
1327 struct nfp_net *nn = netdev_priv(netdev);
1329 nn_warn(nn, "TX watchdog timeout on ring: %u\n", txqueue);
1332 /* Receive processing
1335 nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp)
1337 unsigned int fl_bufsz;
1339 fl_bufsz = NFP_NET_RX_BUF_HEADROOM;
1340 fl_bufsz += dp->rx_dma_off;
1341 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1342 fl_bufsz += NFP_NET_MAX_PREPEND;
1344 fl_bufsz += dp->rx_offset;
1345 fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu;
1347 fl_bufsz = SKB_DATA_ALIGN(fl_bufsz);
1348 fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1354 nfp_net_free_frag(void *frag, bool xdp)
1357 skb_free_frag(frag);
1359 __free_page(virt_to_page(frag));
1363 * nfp_net_rx_alloc_one() - Allocate and map page frag for RX
1364 * @dp: NFP Net data path struct
1365 * @dma_addr: Pointer to storage for DMA address (output param)
1367 * This function will allcate a new page frag, map it for DMA.
1369 * Return: allocated page frag or NULL on failure.
1371 static void *nfp_net_rx_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
1375 if (!dp->xdp_prog) {
1376 frag = netdev_alloc_frag(dp->fl_bufsz);
1380 page = alloc_page(GFP_KERNEL);
1381 frag = page ? page_address(page) : NULL;
1384 nn_dp_warn(dp, "Failed to alloc receive page frag\n");
1388 *dma_addr = nfp_net_dma_map_rx(dp, frag);
1389 if (dma_mapping_error(dp->dev, *dma_addr)) {
1390 nfp_net_free_frag(frag, dp->xdp_prog);
1391 nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
1398 static void *nfp_net_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
1402 if (!dp->xdp_prog) {
1403 frag = napi_alloc_frag(dp->fl_bufsz);
1404 if (unlikely(!frag))
1409 page = dev_alloc_page();
1410 if (unlikely(!page))
1412 frag = page_address(page);
1415 *dma_addr = nfp_net_dma_map_rx(dp, frag);
1416 if (dma_mapping_error(dp->dev, *dma_addr)) {
1417 nfp_net_free_frag(frag, dp->xdp_prog);
1418 nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
1426 * nfp_net_rx_give_one() - Put mapped skb on the software and hardware rings
1427 * @dp: NFP Net data path struct
1428 * @rx_ring: RX ring structure
1429 * @frag: page fragment buffer
1430 * @dma_addr: DMA address of skb mapping
1432 static void nfp_net_rx_give_one(const struct nfp_net_dp *dp,
1433 struct nfp_net_rx_ring *rx_ring,
1434 void *frag, dma_addr_t dma_addr)
1436 unsigned int wr_idx;
1438 wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
1440 nfp_net_dma_sync_dev_rx(dp, dma_addr);
1442 /* Stash SKB and DMA address away */
1443 rx_ring->rxbufs[wr_idx].frag = frag;
1444 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
1446 /* Fill freelist descriptor */
1447 rx_ring->rxds[wr_idx].fld.reserved = 0;
1448 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
1449 nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld,
1450 dma_addr + dp->rx_dma_off);
1453 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
1454 /* Update write pointer of the freelist queue. Make
1455 * sure all writes are flushed before telling the hardware.
1458 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
1463 * nfp_net_rx_ring_reset() - Reflect in SW state of freelist after disable
1464 * @rx_ring: RX ring structure
1466 * Assumes that the device is stopped, must be idempotent.
1468 static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring)
1470 unsigned int wr_idx, last_idx;
1472 /* wr_p == rd_p means ring was never fed FL bufs. RX rings are always
1473 * kept at cnt - 1 FL bufs.
1475 if (rx_ring->wr_p == 0 && rx_ring->rd_p == 0)
1478 /* Move the empty entry to the end of the list */
1479 wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
1480 last_idx = rx_ring->cnt - 1;
1481 rx_ring->rxbufs[wr_idx].dma_addr = rx_ring->rxbufs[last_idx].dma_addr;
1482 rx_ring->rxbufs[wr_idx].frag = rx_ring->rxbufs[last_idx].frag;
1483 rx_ring->rxbufs[last_idx].dma_addr = 0;
1484 rx_ring->rxbufs[last_idx].frag = NULL;
1486 memset(rx_ring->rxds, 0, rx_ring->size);
1492 * nfp_net_rx_ring_bufs_free() - Free any buffers currently on the RX ring
1493 * @dp: NFP Net data path struct
1494 * @rx_ring: RX ring to remove buffers from
1496 * Assumes that the device is stopped and buffers are in [0, ring->cnt - 1)
1497 * entries. After device is disabled nfp_net_rx_ring_reset() must be called
1498 * to restore required ring geometry.
1501 nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp,
1502 struct nfp_net_rx_ring *rx_ring)
1506 for (i = 0; i < rx_ring->cnt - 1; i++) {
1507 /* NULL skb can only happen when initial filling of the ring
1508 * fails to allocate enough buffers and calls here to free
1509 * already allocated ones.
1511 if (!rx_ring->rxbufs[i].frag)
1514 nfp_net_dma_unmap_rx(dp, rx_ring->rxbufs[i].dma_addr);
1515 nfp_net_free_frag(rx_ring->rxbufs[i].frag, dp->xdp_prog);
1516 rx_ring->rxbufs[i].dma_addr = 0;
1517 rx_ring->rxbufs[i].frag = NULL;
1522 * nfp_net_rx_ring_bufs_alloc() - Fill RX ring with buffers (don't give to FW)
1523 * @dp: NFP Net data path struct
1524 * @rx_ring: RX ring to remove buffers from
1527 nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp,
1528 struct nfp_net_rx_ring *rx_ring)
1530 struct nfp_net_rx_buf *rxbufs;
1533 rxbufs = rx_ring->rxbufs;
1535 for (i = 0; i < rx_ring->cnt - 1; i++) {
1536 rxbufs[i].frag = nfp_net_rx_alloc_one(dp, &rxbufs[i].dma_addr);
1537 if (!rxbufs[i].frag) {
1538 nfp_net_rx_ring_bufs_free(dp, rx_ring);
1547 * nfp_net_rx_ring_fill_freelist() - Give buffers from the ring to FW
1548 * @dp: NFP Net data path struct
1549 * @rx_ring: RX ring to fill
1552 nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp,
1553 struct nfp_net_rx_ring *rx_ring)
1557 for (i = 0; i < rx_ring->cnt - 1; i++)
1558 nfp_net_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
1559 rx_ring->rxbufs[i].dma_addr);
1563 * nfp_net_rx_csum_has_errors() - group check if rxd has any csum errors
1564 * @flags: RX descriptor flags field in CPU byte order
1566 static int nfp_net_rx_csum_has_errors(u16 flags)
1568 u16 csum_all_checked, csum_all_ok;
1570 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
1571 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
1573 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
1577 * nfp_net_rx_csum() - set SKB checksum field based on RX descriptor flags
1578 * @dp: NFP Net data path struct
1579 * @r_vec: per-ring structure
1580 * @rxd: Pointer to RX descriptor
1581 * @meta: Parsed metadata prepend
1582 * @skb: Pointer to SKB
1584 static void nfp_net_rx_csum(struct nfp_net_dp *dp,
1585 struct nfp_net_r_vector *r_vec,
1586 struct nfp_net_rx_desc *rxd,
1587 struct nfp_meta_parsed *meta, struct sk_buff *skb)
1589 skb_checksum_none_assert(skb);
1591 if (!(dp->netdev->features & NETIF_F_RXCSUM))
1594 if (meta->csum_type) {
1595 skb->ip_summed = meta->csum_type;
1596 skb->csum = meta->csum;
1597 u64_stats_update_begin(&r_vec->rx_sync);
1598 r_vec->hw_csum_rx_complete++;
1599 u64_stats_update_end(&r_vec->rx_sync);
1603 if (nfp_net_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
1604 u64_stats_update_begin(&r_vec->rx_sync);
1605 r_vec->hw_csum_rx_error++;
1606 u64_stats_update_end(&r_vec->rx_sync);
1610 /* Assume that the firmware will never report inner CSUM_OK unless outer
1611 * L4 headers were successfully parsed. FW will always report zero UDP
1612 * checksum as CSUM_OK.
1614 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
1615 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
1616 __skb_incr_checksum_unnecessary(skb);
1617 u64_stats_update_begin(&r_vec->rx_sync);
1618 r_vec->hw_csum_rx_ok++;
1619 u64_stats_update_end(&r_vec->rx_sync);
1622 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
1623 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
1624 __skb_incr_checksum_unnecessary(skb);
1625 u64_stats_update_begin(&r_vec->rx_sync);
1626 r_vec->hw_csum_rx_inner_ok++;
1627 u64_stats_update_end(&r_vec->rx_sync);
1632 nfp_net_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
1633 unsigned int type, __be32 *hash)
1635 if (!(netdev->features & NETIF_F_RXHASH))
1639 case NFP_NET_RSS_IPV4:
1640 case NFP_NET_RSS_IPV6:
1641 case NFP_NET_RSS_IPV6_EX:
1642 meta->hash_type = PKT_HASH_TYPE_L3;
1645 meta->hash_type = PKT_HASH_TYPE_L4;
1649 meta->hash = get_unaligned_be32(hash);
1653 nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
1654 void *data, struct nfp_net_rx_desc *rxd)
1656 struct nfp_net_rx_hash *rx_hash = data;
1658 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
1661 nfp_net_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
1666 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
1667 void *data, void *pkt, unsigned int pkt_len, int meta_len)
1671 meta_info = get_unaligned_be32(data);
1675 switch (meta_info & NFP_NET_META_FIELD_MASK) {
1676 case NFP_NET_META_HASH:
1677 meta_info >>= NFP_NET_META_FIELD_SIZE;
1678 nfp_net_set_hash(netdev, meta,
1679 meta_info & NFP_NET_META_FIELD_MASK,
1683 case NFP_NET_META_MARK:
1684 meta->mark = get_unaligned_be32(data);
1687 case NFP_NET_META_PORTID:
1688 meta->portid = get_unaligned_be32(data);
1691 case NFP_NET_META_CSUM:
1692 meta->csum_type = CHECKSUM_COMPLETE;
1694 (__force __wsum)__get_unaligned_cpu32(data);
1697 case NFP_NET_META_RESYNC_INFO:
1698 if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
1701 data += sizeof(struct nfp_net_tls_resync_req);
1707 meta_info >>= NFP_NET_META_FIELD_SIZE;
1714 nfp_net_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
1715 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
1716 struct sk_buff *skb)
1718 u64_stats_update_begin(&r_vec->rx_sync);
1720 /* If we have both skb and rxbuf the replacement buffer allocation
1721 * must have failed, count this as an alloc failure.
1724 r_vec->rx_replace_buf_alloc_fail++;
1725 u64_stats_update_end(&r_vec->rx_sync);
1727 /* skb is build based on the frag, free_skb() would free the frag
1728 * so to be able to reuse it we need an extra ref.
1730 if (skb && rxbuf && skb->head == rxbuf->frag)
1731 page_ref_inc(virt_to_head_page(rxbuf->frag));
1733 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
1735 dev_kfree_skb_any(skb);
1739 nfp_net_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
1740 struct nfp_net_tx_ring *tx_ring,
1741 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
1742 unsigned int pkt_len, bool *completed)
1744 unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
1745 struct nfp_net_tx_buf *txbuf;
1746 struct nfp_net_tx_desc *txd;
1749 /* Reject if xdp_adjust_tail grow packet beyond DMA area */
1750 if (pkt_len + dma_off > dma_map_sz)
1753 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1755 nfp_net_xdp_complete(tx_ring);
1759 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1760 nfp_net_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
1766 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1768 /* Stash the soft descriptor of the head then initialize it */
1769 txbuf = &tx_ring->txbufs[wr_idx];
1771 nfp_net_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
1773 txbuf->frag = rxbuf->frag;
1774 txbuf->dma_addr = rxbuf->dma_addr;
1777 txbuf->real_len = pkt_len;
1779 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
1780 pkt_len, DMA_BIDIRECTIONAL);
1782 /* Build TX descriptor */
1783 txd = &tx_ring->txds[wr_idx];
1784 txd->offset_eop = PCIE_DESC_TX_EOP;
1785 txd->dma_len = cpu_to_le16(pkt_len);
1786 nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off);
1787 txd->data_len = cpu_to_le16(pkt_len);
1791 txd->lso_hdrlen = 0;
1794 tx_ring->wr_ptr_add++;
1799 * nfp_net_rx() - receive up to @budget packets on @rx_ring
1800 * @rx_ring: RX ring to receive from
1801 * @budget: NAPI budget
1803 * Note, this function is separated out from the napi poll function to
1804 * more cleanly separate packet receive code from other bookkeeping
1805 * functions performed in the napi poll function.
1807 * Return: Number of packets received.
1809 static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
1811 struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
1812 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
1813 struct nfp_net_tx_ring *tx_ring;
1814 struct bpf_prog *xdp_prog;
1815 bool xdp_tx_cmpl = false;
1816 unsigned int true_bufsz;
1817 struct sk_buff *skb;
1818 int pkts_polled = 0;
1819 struct xdp_buff xdp;
1823 xdp_prog = READ_ONCE(dp->xdp_prog);
1824 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
1825 xdp.frame_sz = PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM;
1826 xdp.rxq = &rx_ring->xdp_rxq;
1827 tx_ring = r_vec->xdp_ring;
1829 while (pkts_polled < budget) {
1830 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1831 struct nfp_net_rx_buf *rxbuf;
1832 struct nfp_net_rx_desc *rxd;
1833 struct nfp_meta_parsed meta;
1834 bool redir_egress = false;
1835 struct net_device *netdev;
1836 dma_addr_t new_dma_addr;
1837 u32 meta_len_xdp = 0;
1840 idx = D_IDX(rx_ring, rx_ring->rd_p);
1842 rxd = &rx_ring->rxds[idx];
1843 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1846 /* Memory barrier to ensure that we won't do other reads
1847 * before the DD bit.
1851 memset(&meta, 0, sizeof(meta));
1856 rxbuf = &rx_ring->rxbufs[idx];
1858 * <-- [rx_offset] -->
1859 * ---------------------------------------------------------
1860 * | [XX] | metadata | packet | XXXX |
1861 * ---------------------------------------------------------
1862 * <---------------- data_len --------------->
1864 * The rx_offset is fixed for all packets, the meta_len can vary
1865 * on a packet by packet basis. If rx_offset is set to zero
1866 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
1867 * buffer and is immediately followed by the packet (no [XX]).
1869 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1870 data_len = le16_to_cpu(rxd->rxd.data_len);
1871 pkt_len = data_len - meta_len;
1873 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1874 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1875 pkt_off += meta_len;
1877 pkt_off += dp->rx_offset;
1878 meta_off = pkt_off - meta_len;
1881 u64_stats_update_begin(&r_vec->rx_sync);
1883 r_vec->rx_bytes += pkt_len;
1884 u64_stats_update_end(&r_vec->rx_sync);
1886 if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
1887 (dp->rx_offset && meta_len > dp->rx_offset))) {
1888 nn_dp_warn(dp, "oversized RX packet metadata %u\n",
1890 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1894 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
1897 if (!dp->chained_metadata_format) {
1898 nfp_net_set_hash_desc(dp->netdev, &meta,
1899 rxbuf->frag + meta_off, rxd);
1900 } else if (meta_len) {
1901 if (unlikely(nfp_net_parse_meta(dp->netdev, &meta,
1902 rxbuf->frag + meta_off,
1903 rxbuf->frag + pkt_off,
1904 pkt_len, meta_len))) {
1905 nn_dp_warn(dp, "invalid RX packet metadata\n");
1906 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf,
1912 if (xdp_prog && !meta.portid) {
1913 void *orig_data = rxbuf->frag + pkt_off;
1914 unsigned int dma_off;
1917 xdp.data_hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM;
1918 xdp.data = orig_data;
1919 xdp.data_meta = orig_data;
1920 xdp.data_end = orig_data + pkt_len;
1922 act = bpf_prog_run_xdp(xdp_prog, &xdp);
1924 pkt_len = xdp.data_end - xdp.data;
1925 pkt_off += xdp.data - orig_data;
1929 meta_len_xdp = xdp.data - xdp.data_meta;
1932 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
1933 if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring,
1938 trace_xdp_exception(dp->netdev,
1942 bpf_warn_invalid_xdp_action(act);
1945 trace_xdp_exception(dp->netdev, xdp_prog, act);
1948 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag,
1954 if (likely(!meta.portid)) {
1955 netdev = dp->netdev;
1956 } else if (meta.portid == NFP_META_PORT_ID_CTRL) {
1957 struct nfp_net *nn = netdev_priv(dp->netdev);
1959 nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
1961 nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag,
1967 nn = netdev_priv(dp->netdev);
1968 netdev = nfp_app_dev_get(nn->app, meta.portid,
1970 if (unlikely(!netdev)) {
1971 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf,
1976 if (nfp_netdev_is_nfp_repr(netdev))
1977 nfp_repr_inc_rx_stats(netdev, pkt_len);
1980 skb = build_skb(rxbuf->frag, true_bufsz);
1981 if (unlikely(!skb)) {
1982 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1985 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr);
1986 if (unlikely(!new_frag)) {
1987 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1991 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1993 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1995 skb_reserve(skb, pkt_off);
1996 skb_put(skb, pkt_len);
1998 skb->mark = meta.mark;
1999 skb_set_hash(skb, meta.hash, meta.hash_type);
2001 skb_record_rx_queue(skb, rx_ring->idx);
2002 skb->protocol = eth_type_trans(skb, netdev);
2004 nfp_net_rx_csum(dp, r_vec, rxd, &meta, skb);
2006 #ifdef CONFIG_TLS_DEVICE
2007 if (rxd->rxd.flags & PCIE_DESC_RX_DECRYPTED) {
2008 skb->decrypted = true;
2009 u64_stats_update_begin(&r_vec->rx_sync);
2011 u64_stats_update_end(&r_vec->rx_sync);
2015 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
2016 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
2017 le16_to_cpu(rxd->rxd.vlan));
2019 skb_metadata_set(skb, meta_len_xdp);
2021 if (likely(!redir_egress)) {
2022 napi_gro_receive(&rx_ring->r_vec->napi, skb);
2025 skb_reset_network_header(skb);
2026 __skb_push(skb, ETH_HLEN);
2027 dev_queue_xmit(skb);
2032 if (tx_ring->wr_ptr_add)
2033 nfp_net_tx_xmit_more_flush(tx_ring);
2034 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
2036 if (!nfp_net_xdp_complete(tx_ring))
2037 pkts_polled = budget;
2045 * nfp_net_poll() - napi poll function
2046 * @napi: NAPI structure
2047 * @budget: NAPI budget
2049 * Return: number of packets polled.
2051 static int nfp_net_poll(struct napi_struct *napi, int budget)
2053 struct nfp_net_r_vector *r_vec =
2054 container_of(napi, struct nfp_net_r_vector, napi);
2055 unsigned int pkts_polled = 0;
2058 nfp_net_tx_complete(r_vec->tx_ring, budget);
2060 pkts_polled = nfp_net_rx(r_vec->rx_ring, budget);
2062 if (pkts_polled < budget)
2063 if (napi_complete_done(napi, pkts_polled))
2064 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
2069 /* Control device data path
2073 nfp_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
2074 struct sk_buff *skb, bool old)
2076 unsigned int real_len = skb->len, meta_len = 0;
2077 struct nfp_net_tx_ring *tx_ring;
2078 struct nfp_net_tx_buf *txbuf;
2079 struct nfp_net_tx_desc *txd;
2080 struct nfp_net_dp *dp;
2081 dma_addr_t dma_addr;
2084 dp = &r_vec->nfp_net->dp;
2085 tx_ring = r_vec->tx_ring;
2087 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
2088 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
2092 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
2093 u64_stats_update_begin(&r_vec->tx_sync);
2095 u64_stats_update_end(&r_vec->tx_sync);
2097 __skb_queue_tail(&r_vec->queue, skb);
2099 __skb_queue_head(&r_vec->queue, skb);
2103 if (nfp_app_ctrl_has_meta(nn->app)) {
2104 if (unlikely(skb_headroom(skb) < 8)) {
2105 nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
2109 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
2110 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
2113 /* Start with the head skbuf */
2114 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
2116 if (dma_mapping_error(dp->dev, dma_addr))
2119 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
2121 /* Stash the soft descriptor of the head then initialize it */
2122 txbuf = &tx_ring->txbufs[wr_idx];
2124 txbuf->dma_addr = dma_addr;
2127 txbuf->real_len = real_len;
2129 /* Build TX descriptor */
2130 txd = &tx_ring->txds[wr_idx];
2131 txd->offset_eop = meta_len | PCIE_DESC_TX_EOP;
2132 txd->dma_len = cpu_to_le16(skb_headlen(skb));
2133 nfp_desc_set_dma_addr(txd, dma_addr);
2134 txd->data_len = cpu_to_le16(skb->len);
2138 txd->lso_hdrlen = 0;
2141 tx_ring->wr_ptr_add++;
2142 nfp_net_tx_xmit_more_flush(tx_ring);
2147 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
2149 u64_stats_update_begin(&r_vec->tx_sync);
2151 u64_stats_update_end(&r_vec->tx_sync);
2152 dev_kfree_skb_any(skb);
2156 bool __nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb)
2158 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0];
2160 return nfp_ctrl_tx_one(nn, r_vec, skb, false);
2163 bool nfp_ctrl_tx(struct nfp_net *nn, struct sk_buff *skb)
2165 struct nfp_net_r_vector *r_vec = &nn->r_vecs[0];
2168 spin_lock_bh(&r_vec->lock);
2169 ret = nfp_ctrl_tx_one(nn, r_vec, skb, false);
2170 spin_unlock_bh(&r_vec->lock);
2175 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
2177 struct sk_buff *skb;
2179 while ((skb = __skb_dequeue(&r_vec->queue)))
2180 if (nfp_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
2185 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
2187 u32 meta_type, meta_tag;
2189 if (!nfp_app_ctrl_has_meta(nn->app))
2195 meta_type = get_unaligned_be32(data);
2196 meta_tag = get_unaligned_be32(data + 4);
2198 return (meta_type == NFP_NET_META_PORTID &&
2199 meta_tag == NFP_META_PORT_ID_CTRL);
2203 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
2204 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
2206 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
2207 struct nfp_net_rx_buf *rxbuf;
2208 struct nfp_net_rx_desc *rxd;
2209 dma_addr_t new_dma_addr;
2210 struct sk_buff *skb;
2214 idx = D_IDX(rx_ring, rx_ring->rd_p);
2216 rxd = &rx_ring->rxds[idx];
2217 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
2220 /* Memory barrier to ensure that we won't do other reads
2221 * before the DD bit.
2227 rxbuf = &rx_ring->rxbufs[idx];
2228 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
2229 data_len = le16_to_cpu(rxd->rxd.data_len);
2230 pkt_len = data_len - meta_len;
2232 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
2233 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
2234 pkt_off += meta_len;
2236 pkt_off += dp->rx_offset;
2237 meta_off = pkt_off - meta_len;
2240 u64_stats_update_begin(&r_vec->rx_sync);
2242 r_vec->rx_bytes += pkt_len;
2243 u64_stats_update_end(&r_vec->rx_sync);
2245 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len);
2247 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
2248 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
2250 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
2254 skb = build_skb(rxbuf->frag, dp->fl_bufsz);
2255 if (unlikely(!skb)) {
2256 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
2259 new_frag = nfp_net_napi_alloc_one(dp, &new_dma_addr);
2260 if (unlikely(!new_frag)) {
2261 nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
2265 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
2267 nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
2269 skb_reserve(skb, pkt_off);
2270 skb_put(skb, pkt_len);
2272 nfp_app_ctrl_rx(nn->app, skb);
2277 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
2279 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
2280 struct nfp_net *nn = r_vec->nfp_net;
2281 struct nfp_net_dp *dp = &nn->dp;
2282 unsigned int budget = 512;
2284 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
2290 static void nfp_ctrl_poll(unsigned long arg)
2292 struct nfp_net_r_vector *r_vec = (void *)arg;
2294 spin_lock(&r_vec->lock);
2295 nfp_net_tx_complete(r_vec->tx_ring, 0);
2296 __nfp_ctrl_tx_queued(r_vec);
2297 spin_unlock(&r_vec->lock);
2299 if (nfp_ctrl_rx(r_vec)) {
2300 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
2302 tasklet_schedule(&r_vec->tasklet);
2303 nn_dp_warn(&r_vec->nfp_net->dp,
2304 "control message budget exceeded!\n");
2308 /* Setup and Configuration
2312 * nfp_net_vecs_init() - Assign IRQs and setup rvecs.
2313 * @nn: NFP Network structure
2315 static void nfp_net_vecs_init(struct nfp_net *nn)
2317 struct nfp_net_r_vector *r_vec;
2320 nn->lsc_handler = nfp_net_irq_lsc;
2321 nn->exn_handler = nfp_net_irq_exn;
2323 for (r = 0; r < nn->max_r_vecs; r++) {
2324 struct msix_entry *entry;
2326 entry = &nn->irq_entries[NFP_NET_NON_Q_VECTORS + r];
2328 r_vec = &nn->r_vecs[r];
2329 r_vec->nfp_net = nn;
2330 r_vec->irq_entry = entry->entry;
2331 r_vec->irq_vector = entry->vector;
2333 if (nn->dp.netdev) {
2334 r_vec->handler = nfp_net_irq_rxtx;
2336 r_vec->handler = nfp_ctrl_irq_rxtx;
2338 __skb_queue_head_init(&r_vec->queue);
2339 spin_lock_init(&r_vec->lock);
2340 tasklet_init(&r_vec->tasklet, nfp_ctrl_poll,
2341 (unsigned long)r_vec);
2342 tasklet_disable(&r_vec->tasklet);
2345 cpumask_set_cpu(r, &r_vec->affinity_mask);
2350 * nfp_net_tx_ring_free() - Free resources allocated to a TX ring
2351 * @tx_ring: TX ring to free
2353 static void nfp_net_tx_ring_free(struct nfp_net_tx_ring *tx_ring)
2355 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
2356 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
2358 kvfree(tx_ring->txbufs);
2361 dma_free_coherent(dp->dev, tx_ring->size,
2362 tx_ring->txds, tx_ring->dma);
2365 tx_ring->txbufs = NULL;
2366 tx_ring->txds = NULL;
2372 * nfp_net_tx_ring_alloc() - Allocate resource for a TX ring
2373 * @dp: NFP Net data path struct
2374 * @tx_ring: TX Ring structure to allocate
2376 * Return: 0 on success, negative errno otherwise.
2379 nfp_net_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
2381 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
2383 tx_ring->cnt = dp->txd_cnt;
2385 tx_ring->size = array_size(tx_ring->cnt, sizeof(*tx_ring->txds));
2386 tx_ring->txds = dma_alloc_coherent(dp->dev, tx_ring->size,
2388 GFP_KERNEL | __GFP_NOWARN);
2389 if (!tx_ring->txds) {
2390 netdev_warn(dp->netdev, "failed to allocate TX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n",
2395 tx_ring->txbufs = kvcalloc(tx_ring->cnt, sizeof(*tx_ring->txbufs),
2397 if (!tx_ring->txbufs)
2400 if (!tx_ring->is_xdp && dp->netdev)
2401 netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask,
2407 nfp_net_tx_ring_free(tx_ring);
2412 nfp_net_tx_ring_bufs_free(struct nfp_net_dp *dp,
2413 struct nfp_net_tx_ring *tx_ring)
2417 if (!tx_ring->is_xdp)
2420 for (i = 0; i < tx_ring->cnt; i++) {
2421 if (!tx_ring->txbufs[i].frag)
2424 nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr);
2425 __free_page(virt_to_page(tx_ring->txbufs[i].frag));
2430 nfp_net_tx_ring_bufs_alloc(struct nfp_net_dp *dp,
2431 struct nfp_net_tx_ring *tx_ring)
2433 struct nfp_net_tx_buf *txbufs = tx_ring->txbufs;
2436 if (!tx_ring->is_xdp)
2439 for (i = 0; i < tx_ring->cnt; i++) {
2440 txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr);
2441 if (!txbufs[i].frag) {
2442 nfp_net_tx_ring_bufs_free(dp, tx_ring);
2450 static int nfp_net_tx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp)
2454 dp->tx_rings = kcalloc(dp->num_tx_rings, sizeof(*dp->tx_rings),
2459 for (r = 0; r < dp->num_tx_rings; r++) {
2462 if (r >= dp->num_stack_tx_rings)
2463 bias = dp->num_stack_tx_rings;
2465 nfp_net_tx_ring_init(&dp->tx_rings[r], &nn->r_vecs[r - bias],
2468 if (nfp_net_tx_ring_alloc(dp, &dp->tx_rings[r]))
2471 if (nfp_net_tx_ring_bufs_alloc(dp, &dp->tx_rings[r]))
2479 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]);
2481 nfp_net_tx_ring_free(&dp->tx_rings[r]);
2483 kfree(dp->tx_rings);
2487 static void nfp_net_tx_rings_free(struct nfp_net_dp *dp)
2491 for (r = 0; r < dp->num_tx_rings; r++) {
2492 nfp_net_tx_ring_bufs_free(dp, &dp->tx_rings[r]);
2493 nfp_net_tx_ring_free(&dp->tx_rings[r]);
2496 kfree(dp->tx_rings);
2500 * nfp_net_rx_ring_free() - Free resources allocated to a RX ring
2501 * @rx_ring: RX ring to free
2503 static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring)
2505 struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
2506 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
2509 xdp_rxq_info_unreg(&rx_ring->xdp_rxq);
2510 kvfree(rx_ring->rxbufs);
2513 dma_free_coherent(dp->dev, rx_ring->size,
2514 rx_ring->rxds, rx_ring->dma);
2517 rx_ring->rxbufs = NULL;
2518 rx_ring->rxds = NULL;
2524 * nfp_net_rx_ring_alloc() - Allocate resource for a RX ring
2525 * @dp: NFP Net data path struct
2526 * @rx_ring: RX ring to allocate
2528 * Return: 0 on success, negative errno otherwise.
2531 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring)
2536 err = xdp_rxq_info_reg(&rx_ring->xdp_rxq, dp->netdev,
2542 rx_ring->cnt = dp->rxd_cnt;
2543 rx_ring->size = array_size(rx_ring->cnt, sizeof(*rx_ring->rxds));
2544 rx_ring->rxds = dma_alloc_coherent(dp->dev, rx_ring->size,
2546 GFP_KERNEL | __GFP_NOWARN);
2547 if (!rx_ring->rxds) {
2548 netdev_warn(dp->netdev, "failed to allocate RX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n",
2553 rx_ring->rxbufs = kvcalloc(rx_ring->cnt, sizeof(*rx_ring->rxbufs),
2555 if (!rx_ring->rxbufs)
2561 nfp_net_rx_ring_free(rx_ring);
2565 static int nfp_net_rx_rings_prepare(struct nfp_net *nn, struct nfp_net_dp *dp)
2569 dp->rx_rings = kcalloc(dp->num_rx_rings, sizeof(*dp->rx_rings),
2574 for (r = 0; r < dp->num_rx_rings; r++) {
2575 nfp_net_rx_ring_init(&dp->rx_rings[r], &nn->r_vecs[r], r);
2577 if (nfp_net_rx_ring_alloc(dp, &dp->rx_rings[r]))
2580 if (nfp_net_rx_ring_bufs_alloc(dp, &dp->rx_rings[r]))
2588 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]);
2590 nfp_net_rx_ring_free(&dp->rx_rings[r]);
2592 kfree(dp->rx_rings);
2596 static void nfp_net_rx_rings_free(struct nfp_net_dp *dp)
2600 for (r = 0; r < dp->num_rx_rings; r++) {
2601 nfp_net_rx_ring_bufs_free(dp, &dp->rx_rings[r]);
2602 nfp_net_rx_ring_free(&dp->rx_rings[r]);
2605 kfree(dp->rx_rings);
2609 nfp_net_vector_assign_rings(struct nfp_net_dp *dp,
2610 struct nfp_net_r_vector *r_vec, int idx)
2612 r_vec->rx_ring = idx < dp->num_rx_rings ? &dp->rx_rings[idx] : NULL;
2614 idx < dp->num_stack_tx_rings ? &dp->tx_rings[idx] : NULL;
2616 r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ?
2617 &dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL;
2621 nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
2628 netif_napi_add(nn->dp.netdev, &r_vec->napi,
2629 nfp_net_poll, NAPI_POLL_WEIGHT);
2631 tasklet_enable(&r_vec->tasklet);
2633 snprintf(r_vec->name, sizeof(r_vec->name),
2634 "%s-rxtx-%d", nfp_net_name(nn), idx);
2635 err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name,
2639 netif_napi_del(&r_vec->napi);
2641 tasklet_disable(&r_vec->tasklet);
2643 nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector);
2646 disable_irq(r_vec->irq_vector);
2648 irq_set_affinity_hint(r_vec->irq_vector, &r_vec->affinity_mask);
2650 nn_dbg(nn, "RV%02d: irq=%03d/%03d\n", idx, r_vec->irq_vector,
2657 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec)
2659 irq_set_affinity_hint(r_vec->irq_vector, NULL);
2661 netif_napi_del(&r_vec->napi);
2663 tasklet_disable(&r_vec->tasklet);
2665 free_irq(r_vec->irq_vector, r_vec);
2669 * nfp_net_rss_write_itbl() - Write RSS indirection table to device
2670 * @nn: NFP Net device to reconfigure
2672 void nfp_net_rss_write_itbl(struct nfp_net *nn)
2676 for (i = 0; i < NFP_NET_CFG_RSS_ITBL_SZ; i += 4)
2677 nn_writel(nn, NFP_NET_CFG_RSS_ITBL + i,
2678 get_unaligned_le32(nn->rss_itbl + i));
2682 * nfp_net_rss_write_key() - Write RSS hash key to device
2683 * @nn: NFP Net device to reconfigure
2685 void nfp_net_rss_write_key(struct nfp_net *nn)
2689 for (i = 0; i < nfp_net_rss_key_sz(nn); i += 4)
2690 nn_writel(nn, NFP_NET_CFG_RSS_KEY + i,
2691 get_unaligned_le32(nn->rss_key + i));
2695 * nfp_net_coalesce_write_cfg() - Write irq coalescence configuration to HW
2696 * @nn: NFP Net device to reconfigure
2698 void nfp_net_coalesce_write_cfg(struct nfp_net *nn)
2704 /* Compute factor used to convert coalesce '_usecs' parameters to
2705 * ME timestamp ticks. There are 16 ME clock cycles for each timestamp
2708 factor = nn->tlv_caps.me_freq_mhz / 16;
2710 /* copy RX interrupt coalesce parameters */
2711 value = (nn->rx_coalesce_max_frames << 16) |
2712 (factor * nn->rx_coalesce_usecs);
2713 for (i = 0; i < nn->dp.num_rx_rings; i++)
2714 nn_writel(nn, NFP_NET_CFG_RXR_IRQ_MOD(i), value);
2716 /* copy TX interrupt coalesce parameters */
2717 value = (nn->tx_coalesce_max_frames << 16) |
2718 (factor * nn->tx_coalesce_usecs);
2719 for (i = 0; i < nn->dp.num_tx_rings; i++)
2720 nn_writel(nn, NFP_NET_CFG_TXR_IRQ_MOD(i), value);
2724 * nfp_net_write_mac_addr() - Write mac address to the device control BAR
2725 * @nn: NFP Net device to reconfigure
2726 * @addr: MAC address to write
2728 * Writes the MAC address from the netdev to the device control BAR. Does not
2729 * perform the required reconfig. We do a bit of byte swapping dance because
2732 static void nfp_net_write_mac_addr(struct nfp_net *nn, const u8 *addr)
2734 nn_writel(nn, NFP_NET_CFG_MACADDR + 0, get_unaligned_be32(addr));
2735 nn_writew(nn, NFP_NET_CFG_MACADDR + 6, get_unaligned_be16(addr + 4));
2738 static void nfp_net_vec_clear_ring_data(struct nfp_net *nn, unsigned int idx)
2740 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), 0);
2741 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), 0);
2742 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), 0);
2744 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), 0);
2745 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), 0);
2746 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), 0);
2750 * nfp_net_clear_config_and_disable() - Clear control BAR and disable NFP
2751 * @nn: NFP Net device to reconfigure
2753 * Warning: must be fully idempotent.
2755 static void nfp_net_clear_config_and_disable(struct nfp_net *nn)
2757 u32 new_ctrl, update;
2761 new_ctrl = nn->dp.ctrl;
2762 new_ctrl &= ~NFP_NET_CFG_CTRL_ENABLE;
2763 update = NFP_NET_CFG_UPDATE_GEN;
2764 update |= NFP_NET_CFG_UPDATE_MSIX;
2765 update |= NFP_NET_CFG_UPDATE_RING;
2767 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG)
2768 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG;
2770 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0);
2771 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0);
2773 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2774 err = nfp_net_reconfig(nn, update);
2776 nn_err(nn, "Could not disable device: %d\n", err);
2778 for (r = 0; r < nn->dp.num_rx_rings; r++)
2779 nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]);
2780 for (r = 0; r < nn->dp.num_tx_rings; r++)
2781 nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]);
2782 for (r = 0; r < nn->dp.num_r_vecs; r++)
2783 nfp_net_vec_clear_ring_data(nn, r);
2785 nn->dp.ctrl = new_ctrl;
2789 nfp_net_rx_ring_hw_cfg_write(struct nfp_net *nn,
2790 struct nfp_net_rx_ring *rx_ring, unsigned int idx)
2792 /* Write the DMA address, size and MSI-X info to the device */
2793 nn_writeq(nn, NFP_NET_CFG_RXR_ADDR(idx), rx_ring->dma);
2794 nn_writeb(nn, NFP_NET_CFG_RXR_SZ(idx), ilog2(rx_ring->cnt));
2795 nn_writeb(nn, NFP_NET_CFG_RXR_VEC(idx), rx_ring->r_vec->irq_entry);
2799 nfp_net_tx_ring_hw_cfg_write(struct nfp_net *nn,
2800 struct nfp_net_tx_ring *tx_ring, unsigned int idx)
2802 nn_writeq(nn, NFP_NET_CFG_TXR_ADDR(idx), tx_ring->dma);
2803 nn_writeb(nn, NFP_NET_CFG_TXR_SZ(idx), ilog2(tx_ring->cnt));
2804 nn_writeb(nn, NFP_NET_CFG_TXR_VEC(idx), tx_ring->r_vec->irq_entry);
2808 * nfp_net_set_config_and_enable() - Write control BAR and enable NFP
2809 * @nn: NFP Net device to reconfigure
2811 static int nfp_net_set_config_and_enable(struct nfp_net *nn)
2813 u32 bufsz, new_ctrl, update = 0;
2817 new_ctrl = nn->dp.ctrl;
2819 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_RSS_ANY) {
2820 nfp_net_rss_write_key(nn);
2821 nfp_net_rss_write_itbl(nn);
2822 nn_writel(nn, NFP_NET_CFG_RSS_CTRL, nn->rss_cfg);
2823 update |= NFP_NET_CFG_UPDATE_RSS;
2826 if (nn->dp.ctrl & NFP_NET_CFG_CTRL_IRQMOD) {
2827 nfp_net_coalesce_write_cfg(nn);
2828 update |= NFP_NET_CFG_UPDATE_IRQMOD;
2831 for (r = 0; r < nn->dp.num_tx_rings; r++)
2832 nfp_net_tx_ring_hw_cfg_write(nn, &nn->dp.tx_rings[r], r);
2833 for (r = 0; r < nn->dp.num_rx_rings; r++)
2834 nfp_net_rx_ring_hw_cfg_write(nn, &nn->dp.rx_rings[r], r);
2836 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, nn->dp.num_tx_rings == 64 ?
2837 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_tx_rings) - 1);
2839 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, nn->dp.num_rx_rings == 64 ?
2840 0xffffffffffffffffULL : ((u64)1 << nn->dp.num_rx_rings) - 1);
2843 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr);
2845 nn_writel(nn, NFP_NET_CFG_MTU, nn->dp.mtu);
2847 bufsz = nn->dp.fl_bufsz - nn->dp.rx_dma_off - NFP_NET_RX_BUF_NON_DATA;
2848 nn_writel(nn, NFP_NET_CFG_FLBUFSZ, bufsz);
2851 new_ctrl |= NFP_NET_CFG_CTRL_ENABLE;
2852 update |= NFP_NET_CFG_UPDATE_GEN;
2853 update |= NFP_NET_CFG_UPDATE_MSIX;
2854 update |= NFP_NET_CFG_UPDATE_RING;
2855 if (nn->cap & NFP_NET_CFG_CTRL_RINGCFG)
2856 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG;
2858 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
2859 err = nfp_net_reconfig(nn, update);
2861 nfp_net_clear_config_and_disable(nn);
2865 nn->dp.ctrl = new_ctrl;
2867 for (r = 0; r < nn->dp.num_rx_rings; r++)
2868 nfp_net_rx_ring_fill_freelist(&nn->dp, &nn->dp.rx_rings[r]);
2874 * nfp_net_close_stack() - Quiesce the stack (part of close)
2875 * @nn: NFP Net device to reconfigure
2877 static void nfp_net_close_stack(struct nfp_net *nn)
2881 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2882 netif_carrier_off(nn->dp.netdev);
2883 nn->link_up = false;
2885 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2886 disable_irq(nn->r_vecs[r].irq_vector);
2887 napi_disable(&nn->r_vecs[r].napi);
2890 netif_tx_disable(nn->dp.netdev);
2894 * nfp_net_close_free_all() - Free all runtime resources
2895 * @nn: NFP Net device to reconfigure
2897 static void nfp_net_close_free_all(struct nfp_net *nn)
2901 nfp_net_tx_rings_free(&nn->dp);
2902 nfp_net_rx_rings_free(&nn->dp);
2904 for (r = 0; r < nn->dp.num_r_vecs; r++)
2905 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
2907 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2908 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX);
2912 * nfp_net_netdev_close() - Called when the device is downed
2913 * @netdev: netdev structure
2915 static int nfp_net_netdev_close(struct net_device *netdev)
2917 struct nfp_net *nn = netdev_priv(netdev);
2919 /* Step 1: Disable RX and TX rings from the Linux kernel perspective
2921 nfp_net_close_stack(nn);
2925 nfp_net_clear_config_and_disable(nn);
2926 nfp_port_configure(netdev, false);
2928 /* Step 3: Free resources
2930 nfp_net_close_free_all(nn);
2932 nn_dbg(nn, "%s down", netdev->name);
2936 void nfp_ctrl_close(struct nfp_net *nn)
2942 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2943 disable_irq(nn->r_vecs[r].irq_vector);
2944 tasklet_disable(&nn->r_vecs[r].tasklet);
2947 nfp_net_clear_config_and_disable(nn);
2949 nfp_net_close_free_all(nn);
2955 * nfp_net_open_stack() - Start the device from stack's perspective
2956 * @nn: NFP Net device to reconfigure
2958 static void nfp_net_open_stack(struct nfp_net *nn)
2962 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2963 napi_enable(&nn->r_vecs[r].napi);
2964 enable_irq(nn->r_vecs[r].irq_vector);
2967 netif_tx_wake_all_queues(nn->dp.netdev);
2969 enable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2970 nfp_net_read_link_status(nn);
2973 static int nfp_net_open_alloc_all(struct nfp_net *nn)
2977 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_EXN, "%s-exn",
2978 nn->exn_name, sizeof(nn->exn_name),
2979 NFP_NET_IRQ_EXN_IDX, nn->exn_handler);
2982 err = nfp_net_aux_irq_request(nn, NFP_NET_CFG_LSC, "%s-lsc",
2983 nn->lsc_name, sizeof(nn->lsc_name),
2984 NFP_NET_IRQ_LSC_IDX, nn->lsc_handler);
2987 disable_irq(nn->irq_entries[NFP_NET_IRQ_LSC_IDX].vector);
2989 for (r = 0; r < nn->dp.num_r_vecs; r++) {
2990 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r);
2992 goto err_cleanup_vec_p;
2995 err = nfp_net_rx_rings_prepare(nn, &nn->dp);
2997 goto err_cleanup_vec;
2999 err = nfp_net_tx_rings_prepare(nn, &nn->dp);
3001 goto err_free_rx_rings;
3003 for (r = 0; r < nn->max_r_vecs; r++)
3004 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r);
3009 nfp_net_rx_rings_free(&nn->dp);
3011 r = nn->dp.num_r_vecs;
3014 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3015 nfp_net_aux_irq_free(nn, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
3017 nfp_net_aux_irq_free(nn, NFP_NET_CFG_EXN, NFP_NET_IRQ_EXN_IDX);
3021 static int nfp_net_netdev_open(struct net_device *netdev)
3023 struct nfp_net *nn = netdev_priv(netdev);
3026 /* Step 1: Allocate resources for rings and the like
3027 * - Request interrupts
3028 * - Allocate RX and TX ring resources
3029 * - Setup initial RSS table
3031 err = nfp_net_open_alloc_all(nn);
3035 err = netif_set_real_num_tx_queues(netdev, nn->dp.num_stack_tx_rings);
3039 err = netif_set_real_num_rx_queues(netdev, nn->dp.num_rx_rings);
3043 /* Step 2: Configure the NFP
3044 * - Ifup the physical interface if it exists
3045 * - Enable rings from 0 to tx_rings/rx_rings - 1.
3046 * - Write MAC address (in case it changed)
3048 * - Set the Freelist buffer size
3051 err = nfp_port_configure(netdev, true);
3055 err = nfp_net_set_config_and_enable(nn);
3057 goto err_port_disable;
3059 /* Step 3: Enable for kernel
3060 * - put some freelist descriptors on each RX ring
3061 * - enable NAPI on each ring
3062 * - enable all TX queues
3065 nfp_net_open_stack(nn);
3070 nfp_port_configure(netdev, false);
3072 nfp_net_close_free_all(nn);
3076 int nfp_ctrl_open(struct nfp_net *nn)
3080 /* ring dumping depends on vNICs being opened/closed under rtnl */
3083 err = nfp_net_open_alloc_all(nn);
3087 err = nfp_net_set_config_and_enable(nn);
3091 for (r = 0; r < nn->dp.num_r_vecs; r++)
3092 enable_irq(nn->r_vecs[r].irq_vector);
3099 nfp_net_close_free_all(nn);
3105 static void nfp_net_set_rx_mode(struct net_device *netdev)
3107 struct nfp_net *nn = netdev_priv(netdev);
3110 new_ctrl = nn->dp.ctrl;
3112 if (!netdev_mc_empty(netdev) || netdev->flags & IFF_ALLMULTI)
3113 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_L2MC;
3115 new_ctrl &= ~NFP_NET_CFG_CTRL_L2MC;
3117 if (netdev->flags & IFF_PROMISC) {
3118 if (nn->cap & NFP_NET_CFG_CTRL_PROMISC)
3119 new_ctrl |= NFP_NET_CFG_CTRL_PROMISC;
3121 nn_warn(nn, "FW does not support promiscuous mode\n");
3123 new_ctrl &= ~NFP_NET_CFG_CTRL_PROMISC;
3126 if (new_ctrl == nn->dp.ctrl)
3129 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
3130 nfp_net_reconfig_post(nn, NFP_NET_CFG_UPDATE_GEN);
3132 nn->dp.ctrl = new_ctrl;
3135 static void nfp_net_rss_init_itbl(struct nfp_net *nn)
3139 for (i = 0; i < sizeof(nn->rss_itbl); i++)
3141 ethtool_rxfh_indir_default(i, nn->dp.num_rx_rings);
3144 static void nfp_net_dp_swap(struct nfp_net *nn, struct nfp_net_dp *dp)
3146 struct nfp_net_dp new_dp = *dp;
3151 nn->dp.netdev->mtu = new_dp.mtu;
3153 if (!netif_is_rxfh_configured(nn->dp.netdev))
3154 nfp_net_rss_init_itbl(nn);
3157 static int nfp_net_dp_swap_enable(struct nfp_net *nn, struct nfp_net_dp *dp)
3162 nfp_net_dp_swap(nn, dp);
3164 for (r = 0; r < nn->max_r_vecs; r++)
3165 nfp_net_vector_assign_rings(&nn->dp, &nn->r_vecs[r], r);
3167 err = netif_set_real_num_rx_queues(nn->dp.netdev, nn->dp.num_rx_rings);
3171 if (nn->dp.netdev->real_num_tx_queues != nn->dp.num_stack_tx_rings) {
3172 err = netif_set_real_num_tx_queues(nn->dp.netdev,
3173 nn->dp.num_stack_tx_rings);
3178 return nfp_net_set_config_and_enable(nn);
3181 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn)
3183 struct nfp_net_dp *new;
3185 new = kmalloc(sizeof(*new), GFP_KERNEL);
3191 /* Clear things which need to be recomputed */
3193 new->tx_rings = NULL;
3194 new->rx_rings = NULL;
3195 new->num_r_vecs = 0;
3196 new->num_stack_tx_rings = 0;
3202 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp,
3203 struct netlink_ext_ack *extack)
3205 /* XDP-enabled tests */
3208 if (dp->fl_bufsz > PAGE_SIZE) {
3209 NL_SET_ERR_MSG_MOD(extack, "MTU too large w/ XDP enabled");
3212 if (dp->num_tx_rings > nn->max_tx_rings) {
3213 NL_SET_ERR_MSG_MOD(extack, "Insufficient number of TX rings w/ XDP enabled");
3220 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp,
3221 struct netlink_ext_ack *extack)
3225 dp->fl_bufsz = nfp_net_calc_fl_bufsz(dp);
3227 dp->num_stack_tx_rings = dp->num_tx_rings;
3229 dp->num_stack_tx_rings -= dp->num_rx_rings;
3231 dp->num_r_vecs = max(dp->num_rx_rings, dp->num_stack_tx_rings);
3233 err = nfp_net_check_config(nn, dp, extack);
3237 if (!netif_running(dp->netdev)) {
3238 nfp_net_dp_swap(nn, dp);
3243 /* Prepare new rings */
3244 for (r = nn->dp.num_r_vecs; r < dp->num_r_vecs; r++) {
3245 err = nfp_net_prepare_vector(nn, &nn->r_vecs[r], r);
3248 goto err_cleanup_vecs;
3252 err = nfp_net_rx_rings_prepare(nn, dp);
3254 goto err_cleanup_vecs;
3256 err = nfp_net_tx_rings_prepare(nn, dp);
3260 /* Stop device, swap in new rings, try to start the firmware */
3261 nfp_net_close_stack(nn);
3262 nfp_net_clear_config_and_disable(nn);
3264 err = nfp_net_dp_swap_enable(nn, dp);
3268 nfp_net_clear_config_and_disable(nn);
3270 /* Try with old configuration and old rings */
3271 err2 = nfp_net_dp_swap_enable(nn, dp);
3273 nn_err(nn, "Can't restore ring config - FW communication failed (%d,%d)\n",
3276 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
3277 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3279 nfp_net_rx_rings_free(dp);
3280 nfp_net_tx_rings_free(dp);
3282 nfp_net_open_stack(nn);
3289 nfp_net_rx_rings_free(dp);
3291 for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
3292 nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
3297 static int nfp_net_change_mtu(struct net_device *netdev, int new_mtu)
3299 struct nfp_net *nn = netdev_priv(netdev);
3300 struct nfp_net_dp *dp;
3303 err = nfp_app_check_mtu(nn->app, netdev, new_mtu);
3307 dp = nfp_net_clone_dp(nn);
3313 return nfp_net_ring_reconfig(nn, dp, NULL);
3317 nfp_net_vlan_rx_add_vid(struct net_device *netdev, __be16 proto, u16 vid)
3319 const u32 cmd = NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_ADD;
3320 struct nfp_net *nn = netdev_priv(netdev);
3323 /* Priority tagged packets with vlan id 0 are processed by the
3324 * NFP as untagged packets
3329 err = nfp_net_mbox_lock(nn, NFP_NET_CFG_VLAN_FILTER_SZ);
3333 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid);
3334 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO,
3337 return nfp_net_mbox_reconfig_and_unlock(nn, cmd);
3341 nfp_net_vlan_rx_kill_vid(struct net_device *netdev, __be16 proto, u16 vid)
3343 const u32 cmd = NFP_NET_CFG_MBOX_CMD_CTAG_FILTER_KILL;
3344 struct nfp_net *nn = netdev_priv(netdev);
3347 /* Priority tagged packets with vlan id 0 are processed by the
3348 * NFP as untagged packets
3353 err = nfp_net_mbox_lock(nn, NFP_NET_CFG_VLAN_FILTER_SZ);
3357 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_VID, vid);
3358 nn_writew(nn, nn->tlv_caps.mbox_off + NFP_NET_CFG_VLAN_FILTER_PROTO,
3361 return nfp_net_mbox_reconfig_and_unlock(nn, cmd);
3364 static void nfp_net_stat64(struct net_device *netdev,
3365 struct rtnl_link_stats64 *stats)
3367 struct nfp_net *nn = netdev_priv(netdev);
3370 /* Collect software stats */
3371 for (r = 0; r < nn->max_r_vecs; r++) {
3372 struct nfp_net_r_vector *r_vec = &nn->r_vecs[r];
3377 start = u64_stats_fetch_begin(&r_vec->rx_sync);
3378 data[0] = r_vec->rx_pkts;
3379 data[1] = r_vec->rx_bytes;
3380 data[2] = r_vec->rx_drops;
3381 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
3382 stats->rx_packets += data[0];
3383 stats->rx_bytes += data[1];
3384 stats->rx_dropped += data[2];
3387 start = u64_stats_fetch_begin(&r_vec->tx_sync);
3388 data[0] = r_vec->tx_pkts;
3389 data[1] = r_vec->tx_bytes;
3390 data[2] = r_vec->tx_errors;
3391 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
3392 stats->tx_packets += data[0];
3393 stats->tx_bytes += data[1];
3394 stats->tx_errors += data[2];
3397 /* Add in device stats */
3398 stats->multicast += nn_readq(nn, NFP_NET_CFG_STATS_RX_MC_FRAMES);
3399 stats->rx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_RX_DISCARDS);
3400 stats->rx_errors += nn_readq(nn, NFP_NET_CFG_STATS_RX_ERRORS);
3402 stats->tx_dropped += nn_readq(nn, NFP_NET_CFG_STATS_TX_DISCARDS);
3403 stats->tx_errors += nn_readq(nn, NFP_NET_CFG_STATS_TX_ERRORS);
3406 static int nfp_net_set_features(struct net_device *netdev,
3407 netdev_features_t features)
3409 netdev_features_t changed = netdev->features ^ features;
3410 struct nfp_net *nn = netdev_priv(netdev);
3414 /* Assume this is not called with features we have not advertised */
3416 new_ctrl = nn->dp.ctrl;
3418 if (changed & NETIF_F_RXCSUM) {
3419 if (features & NETIF_F_RXCSUM)
3420 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY;
3422 new_ctrl &= ~NFP_NET_CFG_CTRL_RXCSUM_ANY;
3425 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3426 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM))
3427 new_ctrl |= NFP_NET_CFG_CTRL_TXCSUM;
3429 new_ctrl &= ~NFP_NET_CFG_CTRL_TXCSUM;
3432 if (changed & (NETIF_F_TSO | NETIF_F_TSO6)) {
3433 if (features & (NETIF_F_TSO | NETIF_F_TSO6))
3434 new_ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?:
3435 NFP_NET_CFG_CTRL_LSO;
3437 new_ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY;
3440 if (changed & NETIF_F_HW_VLAN_CTAG_RX) {
3441 if (features & NETIF_F_HW_VLAN_CTAG_RX)
3442 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
3444 new_ctrl &= ~NFP_NET_CFG_CTRL_RXVLAN;
3447 if (changed & NETIF_F_HW_VLAN_CTAG_TX) {
3448 if (features & NETIF_F_HW_VLAN_CTAG_TX)
3449 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
3451 new_ctrl &= ~NFP_NET_CFG_CTRL_TXVLAN;
3454 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) {
3455 if (features & NETIF_F_HW_VLAN_CTAG_FILTER)
3456 new_ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER;
3458 new_ctrl &= ~NFP_NET_CFG_CTRL_CTAG_FILTER;
3461 if (changed & NETIF_F_SG) {
3462 if (features & NETIF_F_SG)
3463 new_ctrl |= NFP_NET_CFG_CTRL_GATHER;
3465 new_ctrl &= ~NFP_NET_CFG_CTRL_GATHER;
3468 err = nfp_port_set_features(netdev, features);
3472 nn_dbg(nn, "Feature change 0x%llx -> 0x%llx (changed=0x%llx)\n",
3473 netdev->features, features, changed);
3475 if (new_ctrl == nn->dp.ctrl)
3478 nn_dbg(nn, "NIC ctrl: 0x%x -> 0x%x\n", nn->dp.ctrl, new_ctrl);
3479 nn_writel(nn, NFP_NET_CFG_CTRL, new_ctrl);
3480 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN);
3484 nn->dp.ctrl = new_ctrl;
3489 static netdev_features_t
3490 nfp_net_features_check(struct sk_buff *skb, struct net_device *dev,
3491 netdev_features_t features)
3495 /* We can't do TSO over double tagged packets (802.1AD) */
3496 features &= vlan_features_check(skb, features);
3498 if (!skb->encapsulation)
3501 /* Ensure that inner L4 header offset fits into TX descriptor field */
3502 if (skb_is_gso(skb)) {
3505 hdrlen = skb_inner_transport_header(skb) - skb->data +
3506 inner_tcp_hdrlen(skb);
3508 /* Assume worst case scenario of having longest possible
3509 * metadata prepend - 8B
3511 if (unlikely(hdrlen > NFP_NET_LSO_MAX_HDR_SZ - 8))
3512 features &= ~NETIF_F_GSO_MASK;
3515 /* VXLAN/GRE check */
3516 switch (vlan_get_protocol(skb)) {
3517 case htons(ETH_P_IP):
3518 l4_hdr = ip_hdr(skb)->protocol;
3520 case htons(ETH_P_IPV6):
3521 l4_hdr = ipv6_hdr(skb)->nexthdr;
3524 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3527 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
3528 skb->inner_protocol != htons(ETH_P_TEB) ||
3529 (l4_hdr != IPPROTO_UDP && l4_hdr != IPPROTO_GRE) ||
3530 (l4_hdr == IPPROTO_UDP &&
3531 (skb_inner_mac_header(skb) - skb_transport_header(skb) !=
3532 sizeof(struct udphdr) + sizeof(struct vxlanhdr))))
3533 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3539 nfp_net_get_phys_port_name(struct net_device *netdev, char *name, size_t len)
3541 struct nfp_net *nn = netdev_priv(netdev);
3544 /* If port is defined, devlink_port is registered and devlink core
3545 * is taking care of name formatting.
3550 if (nn->dp.is_vf || nn->vnic_no_name)
3553 n = snprintf(name, len, "n%d", nn->id);
3560 static int nfp_net_xdp_setup_drv(struct nfp_net *nn, struct netdev_bpf *bpf)
3562 struct bpf_prog *prog = bpf->prog;
3563 struct nfp_net_dp *dp;
3566 if (!xdp_attachment_flags_ok(&nn->xdp, bpf))
3569 if (!prog == !nn->dp.xdp_prog) {
3570 WRITE_ONCE(nn->dp.xdp_prog, prog);
3571 xdp_attachment_setup(&nn->xdp, bpf);
3575 dp = nfp_net_clone_dp(nn);
3579 dp->xdp_prog = prog;
3580 dp->num_tx_rings += prog ? nn->dp.num_rx_rings : -nn->dp.num_rx_rings;
3581 dp->rx_dma_dir = prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
3582 dp->rx_dma_off = prog ? XDP_PACKET_HEADROOM - nn->dp.rx_offset : 0;
3584 /* We need RX reconfig to remap the buffers (BIDIR vs FROM_DEV) */
3585 err = nfp_net_ring_reconfig(nn, dp, bpf->extack);
3589 xdp_attachment_setup(&nn->xdp, bpf);
3593 static int nfp_net_xdp_setup_hw(struct nfp_net *nn, struct netdev_bpf *bpf)
3597 if (!xdp_attachment_flags_ok(&nn->xdp_hw, bpf))
3600 err = nfp_app_xdp_offload(nn->app, nn, bpf->prog, bpf->extack);
3604 xdp_attachment_setup(&nn->xdp_hw, bpf);
3608 static int nfp_net_xdp(struct net_device *netdev, struct netdev_bpf *xdp)
3610 struct nfp_net *nn = netdev_priv(netdev);
3612 switch (xdp->command) {
3613 case XDP_SETUP_PROG:
3614 return nfp_net_xdp_setup_drv(nn, xdp);
3615 case XDP_SETUP_PROG_HW:
3616 return nfp_net_xdp_setup_hw(nn, xdp);
3618 return nfp_app_bpf(nn->app, nn, xdp);
3622 static int nfp_net_set_mac_address(struct net_device *netdev, void *addr)
3624 struct nfp_net *nn = netdev_priv(netdev);
3625 struct sockaddr *saddr = addr;
3628 err = eth_prepare_mac_addr_change(netdev, addr);
3632 nfp_net_write_mac_addr(nn, saddr->sa_data);
3634 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_MACADDR);
3638 eth_commit_mac_addr_change(netdev, addr);
3643 const struct net_device_ops nfp_net_netdev_ops = {
3644 .ndo_init = nfp_app_ndo_init,
3645 .ndo_uninit = nfp_app_ndo_uninit,
3646 .ndo_open = nfp_net_netdev_open,
3647 .ndo_stop = nfp_net_netdev_close,
3648 .ndo_start_xmit = nfp_net_tx,
3649 .ndo_get_stats64 = nfp_net_stat64,
3650 .ndo_vlan_rx_add_vid = nfp_net_vlan_rx_add_vid,
3651 .ndo_vlan_rx_kill_vid = nfp_net_vlan_rx_kill_vid,
3652 .ndo_set_vf_mac = nfp_app_set_vf_mac,
3653 .ndo_set_vf_vlan = nfp_app_set_vf_vlan,
3654 .ndo_set_vf_spoofchk = nfp_app_set_vf_spoofchk,
3655 .ndo_set_vf_trust = nfp_app_set_vf_trust,
3656 .ndo_get_vf_config = nfp_app_get_vf_config,
3657 .ndo_set_vf_link_state = nfp_app_set_vf_link_state,
3658 .ndo_setup_tc = nfp_port_setup_tc,
3659 .ndo_tx_timeout = nfp_net_tx_timeout,
3660 .ndo_set_rx_mode = nfp_net_set_rx_mode,
3661 .ndo_change_mtu = nfp_net_change_mtu,
3662 .ndo_set_mac_address = nfp_net_set_mac_address,
3663 .ndo_set_features = nfp_net_set_features,
3664 .ndo_features_check = nfp_net_features_check,
3665 .ndo_get_phys_port_name = nfp_net_get_phys_port_name,
3666 .ndo_udp_tunnel_add = udp_tunnel_nic_add_port,
3667 .ndo_udp_tunnel_del = udp_tunnel_nic_del_port,
3668 .ndo_bpf = nfp_net_xdp,
3669 .ndo_get_devlink_port = nfp_devlink_get_devlink_port,
3672 static int nfp_udp_tunnel_sync(struct net_device *netdev, unsigned int table)
3674 struct nfp_net *nn = netdev_priv(netdev);
3677 BUILD_BUG_ON(NFP_NET_N_VXLAN_PORTS & 1);
3678 for (i = 0; i < NFP_NET_N_VXLAN_PORTS; i += 2) {
3679 struct udp_tunnel_info ti0, ti1;
3681 udp_tunnel_nic_get_port(netdev, table, i, &ti0);
3682 udp_tunnel_nic_get_port(netdev, table, i + 1, &ti1);
3684 nn_writel(nn, NFP_NET_CFG_VXLAN_PORT + i * sizeof(ti0.port),
3685 be16_to_cpu(ti1.port) << 16 | be16_to_cpu(ti0.port));
3688 return nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_VXLAN);
3691 static const struct udp_tunnel_nic_info nfp_udp_tunnels = {
3692 .sync_table = nfp_udp_tunnel_sync,
3693 .flags = UDP_TUNNEL_NIC_INFO_MAY_SLEEP |
3694 UDP_TUNNEL_NIC_INFO_OPEN_ONLY,
3697 .n_entries = NFP_NET_N_VXLAN_PORTS,
3698 .tunnel_types = UDP_TUNNEL_TYPE_VXLAN,
3704 * nfp_net_info() - Print general info about the NIC
3705 * @nn: NFP Net device to reconfigure
3707 void nfp_net_info(struct nfp_net *nn)
3709 nn_info(nn, "Netronome NFP-6xxx %sNetdev: TxQs=%d/%d RxQs=%d/%d\n",
3710 nn->dp.is_vf ? "VF " : "",
3711 nn->dp.num_tx_rings, nn->max_tx_rings,
3712 nn->dp.num_rx_rings, nn->max_rx_rings);
3713 nn_info(nn, "VER: %d.%d.%d.%d, Maximum supported MTU: %d\n",
3714 nn->fw_ver.resv, nn->fw_ver.class,
3715 nn->fw_ver.major, nn->fw_ver.minor,
3717 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\n",
3719 nn->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "",
3720 nn->cap & NFP_NET_CFG_CTRL_L2BC ? "L2BCFILT " : "",
3721 nn->cap & NFP_NET_CFG_CTRL_L2MC ? "L2MCFILT " : "",
3722 nn->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "",
3723 nn->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "",
3724 nn->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "",
3725 nn->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "",
3726 nn->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "",
3727 nn->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "",
3728 nn->cap & NFP_NET_CFG_CTRL_LSO ? "TSO1 " : "",
3729 nn->cap & NFP_NET_CFG_CTRL_LSO2 ? "TSO2 " : "",
3730 nn->cap & NFP_NET_CFG_CTRL_RSS ? "RSS1 " : "",
3731 nn->cap & NFP_NET_CFG_CTRL_RSS2 ? "RSS2 " : "",
3732 nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER ? "CTAG_FILTER " : "",
3733 nn->cap & NFP_NET_CFG_CTRL_MSIXAUTO ? "AUTOMASK " : "",
3734 nn->cap & NFP_NET_CFG_CTRL_IRQMOD ? "IRQMOD " : "",
3735 nn->cap & NFP_NET_CFG_CTRL_VXLAN ? "VXLAN " : "",
3736 nn->cap & NFP_NET_CFG_CTRL_NVGRE ? "NVGRE " : "",
3737 nn->cap & NFP_NET_CFG_CTRL_CSUM_COMPLETE ?
3738 "RXCSUM_COMPLETE " : "",
3739 nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR ? "LIVE_ADDR " : "",
3740 nfp_app_extra_cap(nn->app, nn));
3744 * nfp_net_alloc() - Allocate netdev and related structure
3746 * @ctrl_bar: PCI IOMEM with vNIC config memory
3747 * @needs_netdev: Whether to allocate a netdev for this vNIC
3748 * @max_tx_rings: Maximum number of TX rings supported by device
3749 * @max_rx_rings: Maximum number of RX rings supported by device
3751 * This function allocates a netdev device and fills in the initial
3752 * part of the @struct nfp_net structure. In case of control device
3753 * nfp_net structure is allocated without the netdev.
3755 * Return: NFP Net device structure, or ERR_PTR on error.
3758 nfp_net_alloc(struct pci_dev *pdev, void __iomem *ctrl_bar, bool needs_netdev,
3759 unsigned int max_tx_rings, unsigned int max_rx_rings)
3765 struct net_device *netdev;
3767 netdev = alloc_etherdev_mqs(sizeof(struct nfp_net),
3768 max_tx_rings, max_rx_rings);
3770 return ERR_PTR(-ENOMEM);
3772 SET_NETDEV_DEV(netdev, &pdev->dev);
3773 nn = netdev_priv(netdev);
3774 nn->dp.netdev = netdev;
3776 nn = vzalloc(sizeof(*nn));
3778 return ERR_PTR(-ENOMEM);
3781 nn->dp.dev = &pdev->dev;
3782 nn->dp.ctrl_bar = ctrl_bar;
3785 nn->max_tx_rings = max_tx_rings;
3786 nn->max_rx_rings = max_rx_rings;
3788 nn->dp.num_tx_rings = min_t(unsigned int,
3789 max_tx_rings, num_online_cpus());
3790 nn->dp.num_rx_rings = min_t(unsigned int, max_rx_rings,
3791 netif_get_num_default_rss_queues());
3793 nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings);
3794 nn->dp.num_r_vecs = min_t(unsigned int,
3795 nn->dp.num_r_vecs, num_online_cpus());
3797 nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT;
3798 nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT;
3800 sema_init(&nn->bar_lock, 1);
3802 spin_lock_init(&nn->reconfig_lock);
3803 spin_lock_init(&nn->link_status_lock);
3805 timer_setup(&nn->reconfig_timer, nfp_net_reconfig_timer, 0);
3807 err = nfp_net_tlv_caps_parse(&nn->pdev->dev, nn->dp.ctrl_bar,
3812 err = nfp_ccm_mbox_alloc(nn);
3820 free_netdev(nn->dp.netdev);
3823 return ERR_PTR(err);
3827 * nfp_net_free() - Undo what @nfp_net_alloc() did
3828 * @nn: NFP Net device to reconfigure
3830 void nfp_net_free(struct nfp_net *nn)
3832 WARN_ON(timer_pending(&nn->reconfig_timer) || nn->reconfig_posted);
3833 nfp_ccm_mbox_free(nn);
3836 free_netdev(nn->dp.netdev);
3842 * nfp_net_rss_key_sz() - Get current size of the RSS key
3843 * @nn: NFP Net device instance
3845 * Return: size of the RSS key for currently selected hash function.
3847 unsigned int nfp_net_rss_key_sz(struct nfp_net *nn)
3849 switch (nn->rss_hfunc) {
3850 case ETH_RSS_HASH_TOP:
3851 return NFP_NET_CFG_RSS_KEY_SZ;
3852 case ETH_RSS_HASH_XOR:
3854 case ETH_RSS_HASH_CRC32:
3858 nn_warn(nn, "Unknown hash function: %u\n", nn->rss_hfunc);
3863 * nfp_net_rss_init() - Set the initial RSS parameters
3864 * @nn: NFP Net device to reconfigure
3866 static void nfp_net_rss_init(struct nfp_net *nn)
3868 unsigned long func_bit, rss_cap_hfunc;
3871 /* Read the RSS function capability and select first supported func */
3872 reg = nn_readl(nn, NFP_NET_CFG_RSS_CAP);
3873 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC, reg);
3875 rss_cap_hfunc = FIELD_GET(NFP_NET_CFG_RSS_CAP_HFUNC,
3876 NFP_NET_CFG_RSS_TOEPLITZ);
3878 func_bit = find_first_bit(&rss_cap_hfunc, NFP_NET_CFG_RSS_HFUNCS);
3879 if (func_bit == NFP_NET_CFG_RSS_HFUNCS) {
3880 dev_warn(nn->dp.dev,
3881 "Bad RSS config, defaulting to Toeplitz hash\n");
3882 func_bit = ETH_RSS_HASH_TOP_BIT;
3884 nn->rss_hfunc = 1 << func_bit;
3886 netdev_rss_key_fill(nn->rss_key, nfp_net_rss_key_sz(nn));
3888 nfp_net_rss_init_itbl(nn);
3890 /* Enable IPv4/IPv6 TCP by default */
3891 nn->rss_cfg = NFP_NET_CFG_RSS_IPV4_TCP |
3892 NFP_NET_CFG_RSS_IPV6_TCP |
3893 FIELD_PREP(NFP_NET_CFG_RSS_HFUNC, nn->rss_hfunc) |
3894 NFP_NET_CFG_RSS_MASK;
3898 * nfp_net_irqmod_init() - Set the initial IRQ moderation parameters
3899 * @nn: NFP Net device to reconfigure
3901 static void nfp_net_irqmod_init(struct nfp_net *nn)
3903 nn->rx_coalesce_usecs = 50;
3904 nn->rx_coalesce_max_frames = 64;
3905 nn->tx_coalesce_usecs = 50;
3906 nn->tx_coalesce_max_frames = 64;
3909 static void nfp_net_netdev_init(struct nfp_net *nn)
3911 struct net_device *netdev = nn->dp.netdev;
3913 nfp_net_write_mac_addr(nn, nn->dp.netdev->dev_addr);
3915 netdev->mtu = nn->dp.mtu;
3917 /* Advertise/enable offloads based on capabilities
3919 * Note: netdev->features show the currently enabled features
3920 * and netdev->hw_features advertises which features are
3921 * supported. By default we enable most features.
3923 if (nn->cap & NFP_NET_CFG_CTRL_LIVE_ADDR)
3924 netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
3926 netdev->hw_features = NETIF_F_HIGHDMA;
3927 if (nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY) {
3928 netdev->hw_features |= NETIF_F_RXCSUM;
3929 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RXCSUM_ANY;
3931 if (nn->cap & NFP_NET_CFG_CTRL_TXCSUM) {
3932 netdev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
3933 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXCSUM;
3935 if (nn->cap & NFP_NET_CFG_CTRL_GATHER) {
3936 netdev->hw_features |= NETIF_F_SG;
3937 nn->dp.ctrl |= NFP_NET_CFG_CTRL_GATHER;
3939 if ((nn->cap & NFP_NET_CFG_CTRL_LSO && nn->fw_ver.major > 2) ||
3940 nn->cap & NFP_NET_CFG_CTRL_LSO2) {
3941 netdev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6;
3942 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_LSO2 ?:
3943 NFP_NET_CFG_CTRL_LSO;
3945 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY)
3946 netdev->hw_features |= NETIF_F_RXHASH;
3947 if (nn->cap & NFP_NET_CFG_CTRL_VXLAN) {
3948 if (nn->cap & NFP_NET_CFG_CTRL_LSO)
3949 netdev->hw_features |= NETIF_F_GSO_UDP_TUNNEL;
3950 netdev->udp_tunnel_nic_info = &nfp_udp_tunnels;
3951 nn->dp.ctrl |= NFP_NET_CFG_CTRL_VXLAN;
3953 if (nn->cap & NFP_NET_CFG_CTRL_NVGRE) {
3954 if (nn->cap & NFP_NET_CFG_CTRL_LSO)
3955 netdev->hw_features |= NETIF_F_GSO_GRE;
3956 nn->dp.ctrl |= NFP_NET_CFG_CTRL_NVGRE;
3958 if (nn->cap & (NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE))
3959 netdev->hw_enc_features = netdev->hw_features;
3961 netdev->vlan_features = netdev->hw_features;
3963 if (nn->cap & NFP_NET_CFG_CTRL_RXVLAN) {
3964 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
3965 nn->dp.ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
3967 if (nn->cap & NFP_NET_CFG_CTRL_TXVLAN) {
3968 if (nn->cap & NFP_NET_CFG_CTRL_LSO2) {
3969 nn_warn(nn, "Device advertises both TSO2 and TXVLAN. Refusing to enable TXVLAN.\n");
3971 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX;
3972 nn->dp.ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
3975 if (nn->cap & NFP_NET_CFG_CTRL_CTAG_FILTER) {
3976 netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3977 nn->dp.ctrl |= NFP_NET_CFG_CTRL_CTAG_FILTER;
3980 netdev->features = netdev->hw_features;
3982 if (nfp_app_has_tc(nn->app) && nn->port)
3983 netdev->hw_features |= NETIF_F_HW_TC;
3985 /* Advertise but disable TSO by default. */
3986 netdev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6);
3987 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_LSO_ANY;
3989 /* Finalise the netdev setup */
3990 netdev->netdev_ops = &nfp_net_netdev_ops;
3991 netdev->watchdog_timeo = msecs_to_jiffies(5 * 1000);
3993 /* MTU range: 68 - hw-specific max */
3994 netdev->min_mtu = ETH_MIN_MTU;
3995 netdev->max_mtu = nn->max_mtu;
3997 netdev->gso_max_segs = NFP_NET_LSO_MAX_SEGS;
3999 netif_carrier_off(netdev);
4001 nfp_net_set_ethtool_ops(netdev);
4004 static int nfp_net_read_caps(struct nfp_net *nn)
4006 /* Get some of the read-only fields from the BAR */
4007 nn->cap = nn_readl(nn, NFP_NET_CFG_CAP);
4008 nn->max_mtu = nn_readl(nn, NFP_NET_CFG_MAX_MTU);
4010 /* ABI 4.x and ctrl vNIC always use chained metadata, in other cases
4011 * we allow use of non-chained metadata if RSS(v1) is the only
4012 * advertised capability requiring metadata.
4014 nn->dp.chained_metadata_format = nn->fw_ver.major == 4 ||
4016 !(nn->cap & NFP_NET_CFG_CTRL_RSS) ||
4017 nn->cap & NFP_NET_CFG_CTRL_CHAIN_META;
4018 /* RSS(v1) uses non-chained metadata format, except in ABI 4.x where
4019 * it has the same meaning as RSSv2.
4021 if (nn->dp.chained_metadata_format && nn->fw_ver.major != 4)
4022 nn->cap &= ~NFP_NET_CFG_CTRL_RSS;
4024 /* Determine RX packet/metadata boundary offset */
4025 if (nn->fw_ver.major >= 2) {
4028 reg = nn_readl(nn, NFP_NET_CFG_RX_OFFSET);
4029 if (reg > NFP_NET_MAX_PREPEND) {
4030 nn_err(nn, "Invalid rx offset: %d\n", reg);
4033 nn->dp.rx_offset = reg;
4035 nn->dp.rx_offset = NFP_NET_RX_OFFSET;
4038 /* For control vNICs mask out the capabilities app doesn't want. */
4040 nn->cap &= nn->app->type->ctrl_cap_mask;
4046 * nfp_net_init() - Initialise/finalise the nfp_net structure
4047 * @nn: NFP Net device structure
4049 * Return: 0 on success or negative errno on error.
4051 int nfp_net_init(struct nfp_net *nn)
4055 nn->dp.rx_dma_dir = DMA_FROM_DEVICE;
4057 err = nfp_net_read_caps(nn);
4061 /* Set default MTU and Freelist buffer size */
4062 if (!nfp_net_is_data_vnic(nn) && nn->app->ctrl_mtu) {
4063 nn->dp.mtu = min(nn->app->ctrl_mtu, nn->max_mtu);
4064 } else if (nn->max_mtu < NFP_NET_DEFAULT_MTU) {
4065 nn->dp.mtu = nn->max_mtu;
4067 nn->dp.mtu = NFP_NET_DEFAULT_MTU;
4069 nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp);
4071 if (nfp_app_ctrl_uses_data_vnics(nn->app))
4072 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_CMSG_DATA;
4074 if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) {
4075 nfp_net_rss_init(nn);
4076 nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?:
4077 NFP_NET_CFG_CTRL_RSS;
4080 /* Allow L2 Broadcast and Multicast through by default, if supported */
4081 if (nn->cap & NFP_NET_CFG_CTRL_L2BC)
4082 nn->dp.ctrl |= NFP_NET_CFG_CTRL_L2BC;
4084 /* Allow IRQ moderation, if supported */
4085 if (nn->cap & NFP_NET_CFG_CTRL_IRQMOD) {
4086 nfp_net_irqmod_init(nn);
4087 nn->dp.ctrl |= NFP_NET_CFG_CTRL_IRQMOD;
4090 /* Stash the re-configuration queue away. First odd queue in TX Bar */
4091 nn->qcp_cfg = nn->tx_bar + NFP_QCP_QUEUE_ADDR_SZ;
4093 /* Make sure the FW knows the netdev is supposed to be disabled here */
4094 nn_writel(nn, NFP_NET_CFG_CTRL, 0);
4095 nn_writeq(nn, NFP_NET_CFG_TXRS_ENABLE, 0);
4096 nn_writeq(nn, NFP_NET_CFG_RXRS_ENABLE, 0);
4097 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_RING |
4098 NFP_NET_CFG_UPDATE_GEN);
4102 if (nn->dp.netdev) {
4103 nfp_net_netdev_init(nn);
4105 err = nfp_ccm_mbox_init(nn);
4109 err = nfp_net_tls_init(nn);
4111 goto err_clean_mbox;
4114 nfp_net_vecs_init(nn);
4118 return register_netdev(nn->dp.netdev);
4121 nfp_ccm_mbox_clean(nn);
4126 * nfp_net_clean() - Undo what nfp_net_init() did.
4127 * @nn: NFP Net device structure
4129 void nfp_net_clean(struct nfp_net *nn)
4134 unregister_netdev(nn->dp.netdev);
4135 nfp_ccm_mbox_clean(nn);
4136 nfp_net_reconfig_wait_posted(nn);