1 // SPDX-License-Identifier: (GPL-2.0+ OR BSD-3-Clause)
2 /* Copyright 2014-2016 Freescale Semiconductor Inc.
3 * Copyright 2016-2022 NXP
5 #include <linux/init.h>
6 #include <linux/module.h>
7 #include <linux/platform_device.h>
8 #include <linux/etherdevice.h>
9 #include <linux/of_net.h>
10 #include <linux/interrupt.h>
11 #include <linux/kthread.h>
12 #include <linux/iommu.h>
13 #include <linux/fsl/mc.h>
14 #include <linux/bpf.h>
15 #include <linux/bpf_trace.h>
16 #include <linux/fsl/ptp_qoriq.h>
17 #include <linux/ptp_classify.h>
18 #include <net/pkt_cls.h>
21 #include <net/xdp_sock_drv.h>
23 #include "dpaa2-eth.h"
25 /* CREATE_TRACE_POINTS only needs to be defined once. Other dpa files
26 * using trace events only need to #include <trace/events/sched.h>
28 #define CREATE_TRACE_POINTS
29 #include "dpaa2-eth-trace.h"
31 MODULE_LICENSE("Dual BSD/GPL");
32 MODULE_AUTHOR("Freescale Semiconductor, Inc");
33 MODULE_DESCRIPTION("Freescale DPAA2 Ethernet Driver");
35 struct ptp_qoriq *dpaa2_ptp;
36 EXPORT_SYMBOL(dpaa2_ptp);
38 static void dpaa2_eth_detect_features(struct dpaa2_eth_priv *priv)
42 if (dpaa2_eth_cmp_dpni_ver(priv, DPNI_PTP_ONESTEP_VER_MAJOR,
43 DPNI_PTP_ONESTEP_VER_MINOR) >= 0)
44 priv->features |= DPAA2_ETH_FEATURE_ONESTEP_CFG_DIRECT;
47 static void dpaa2_update_ptp_onestep_indirect(struct dpaa2_eth_priv *priv,
50 struct dpni_single_step_cfg cfg;
57 if (dpni_set_single_step_cfg(priv->mc_io, 0, priv->mc_token, &cfg))
58 WARN_ONCE(1, "Failed to set single step register");
61 static void dpaa2_update_ptp_onestep_direct(struct dpaa2_eth_priv *priv,
66 val = DPAA2_PTP_SINGLE_STEP_ENABLE |
67 DPAA2_PTP_SINGLE_CORRECTION_OFF(offset);
70 val |= DPAA2_PTP_SINGLE_STEP_CH;
72 if (priv->onestep_reg_base)
73 writel(val, priv->onestep_reg_base);
76 static void dpaa2_ptp_onestep_reg_update_method(struct dpaa2_eth_priv *priv)
78 struct device *dev = priv->net_dev->dev.parent;
79 struct dpni_single_step_cfg ptp_cfg;
81 priv->dpaa2_set_onestep_params_cb = dpaa2_update_ptp_onestep_indirect;
83 if (!(priv->features & DPAA2_ETH_FEATURE_ONESTEP_CFG_DIRECT))
86 if (dpni_get_single_step_cfg(priv->mc_io, 0,
87 priv->mc_token, &ptp_cfg)) {
88 dev_err(dev, "dpni_get_single_step_cfg cannot retrieve onestep reg, falling back to indirect update\n");
92 if (!ptp_cfg.ptp_onestep_reg_base) {
93 dev_err(dev, "1588 onestep reg not available, falling back to indirect update\n");
97 priv->onestep_reg_base = ioremap(ptp_cfg.ptp_onestep_reg_base,
99 if (!priv->onestep_reg_base) {
100 dev_err(dev, "1588 onestep reg cannot be mapped, falling back to indirect update\n");
104 priv->dpaa2_set_onestep_params_cb = dpaa2_update_ptp_onestep_direct;
107 void *dpaa2_iova_to_virt(struct iommu_domain *domain,
108 dma_addr_t iova_addr)
110 phys_addr_t phys_addr;
112 phys_addr = domain ? iommu_iova_to_phys(domain, iova_addr) : iova_addr;
114 return phys_to_virt(phys_addr);
117 static void dpaa2_eth_validate_rx_csum(struct dpaa2_eth_priv *priv,
121 skb_checksum_none_assert(skb);
123 /* HW checksum validation is disabled, nothing to do here */
124 if (!(priv->net_dev->features & NETIF_F_RXCSUM))
127 /* Read checksum validation bits */
128 if (!((fd_status & DPAA2_FAS_L3CV) &&
129 (fd_status & DPAA2_FAS_L4CV)))
132 /* Inform the stack there's no need to compute L3/L4 csum anymore */
133 skb->ip_summed = CHECKSUM_UNNECESSARY;
136 /* Free a received FD.
137 * Not to be used for Tx conf FDs or on any other paths.
139 static void dpaa2_eth_free_rx_fd(struct dpaa2_eth_priv *priv,
140 const struct dpaa2_fd *fd,
143 struct device *dev = priv->net_dev->dev.parent;
144 dma_addr_t addr = dpaa2_fd_get_addr(fd);
145 u8 fd_format = dpaa2_fd_get_format(fd);
146 struct dpaa2_sg_entry *sgt;
150 /* If single buffer frame, just free the data buffer */
151 if (fd_format == dpaa2_fd_single)
153 else if (fd_format != dpaa2_fd_sg)
154 /* We don't support any other format */
157 /* For S/G frames, we first need to free all SG entries
158 * except the first one, which was taken care of already
160 sgt = vaddr + dpaa2_fd_get_offset(fd);
161 for (i = 1; i < DPAA2_ETH_MAX_SG_ENTRIES; i++) {
162 addr = dpaa2_sg_get_addr(&sgt[i]);
163 sg_vaddr = dpaa2_iova_to_virt(priv->iommu_domain, addr);
164 dma_unmap_page(dev, addr, priv->rx_buf_size,
167 free_pages((unsigned long)sg_vaddr, 0);
168 if (dpaa2_sg_is_final(&sgt[i]))
173 free_pages((unsigned long)vaddr, 0);
176 /* Build a linear skb based on a single-buffer frame descriptor */
177 static struct sk_buff *dpaa2_eth_build_linear_skb(struct dpaa2_eth_channel *ch,
178 const struct dpaa2_fd *fd,
181 struct sk_buff *skb = NULL;
182 u16 fd_offset = dpaa2_fd_get_offset(fd);
183 u32 fd_length = dpaa2_fd_get_len(fd);
187 skb = build_skb(fd_vaddr, DPAA2_ETH_RX_BUF_RAW_SIZE);
191 skb_reserve(skb, fd_offset);
192 skb_put(skb, fd_length);
197 /* Build a non linear (fragmented) skb based on a S/G table */
198 static struct sk_buff *dpaa2_eth_build_frag_skb(struct dpaa2_eth_priv *priv,
199 struct dpaa2_eth_channel *ch,
200 struct dpaa2_sg_entry *sgt)
202 struct sk_buff *skb = NULL;
203 struct device *dev = priv->net_dev->dev.parent;
208 struct page *page, *head_page;
212 for (i = 0; i < DPAA2_ETH_MAX_SG_ENTRIES; i++) {
213 struct dpaa2_sg_entry *sge = &sgt[i];
215 /* NOTE: We only support SG entries in dpaa2_sg_single format,
216 * but this is the only format we may receive from HW anyway
219 /* Get the address and length from the S/G entry */
220 sg_addr = dpaa2_sg_get_addr(sge);
221 sg_vaddr = dpaa2_iova_to_virt(priv->iommu_domain, sg_addr);
222 dma_unmap_page(dev, sg_addr, priv->rx_buf_size,
225 sg_length = dpaa2_sg_get_len(sge);
228 /* We build the skb around the first data buffer */
229 skb = build_skb(sg_vaddr, DPAA2_ETH_RX_BUF_RAW_SIZE);
230 if (unlikely(!skb)) {
231 /* Free the first SG entry now, since we already
232 * unmapped it and obtained the virtual address
234 free_pages((unsigned long)sg_vaddr, 0);
236 /* We still need to subtract the buffers used
237 * by this FD from our software counter
239 while (!dpaa2_sg_is_final(&sgt[i]) &&
240 i < DPAA2_ETH_MAX_SG_ENTRIES)
245 sg_offset = dpaa2_sg_get_offset(sge);
246 skb_reserve(skb, sg_offset);
247 skb_put(skb, sg_length);
249 /* Rest of the data buffers are stored as skb frags */
250 page = virt_to_page(sg_vaddr);
251 head_page = virt_to_head_page(sg_vaddr);
253 /* Offset in page (which may be compound).
254 * Data in subsequent SG entries is stored from the
255 * beginning of the buffer, so we don't need to add the
258 page_offset = ((unsigned long)sg_vaddr &
260 (page_address(page) - page_address(head_page));
262 skb_add_rx_frag(skb, i - 1, head_page, page_offset,
263 sg_length, priv->rx_buf_size);
266 if (dpaa2_sg_is_final(sge))
270 WARN_ONCE(i == DPAA2_ETH_MAX_SG_ENTRIES, "Final bit not set in SGT");
272 /* Count all data buffers + SG table buffer */
273 ch->buf_count -= i + 2;
278 /* Free buffers acquired from the buffer pool or which were meant to
279 * be released in the pool
281 static void dpaa2_eth_free_bufs(struct dpaa2_eth_priv *priv, u64 *buf_array,
282 int count, bool xsk_zc)
284 struct device *dev = priv->net_dev->dev.parent;
285 struct dpaa2_eth_swa *swa;
286 struct xdp_buff *xdp_buff;
290 for (i = 0; i < count; i++) {
291 vaddr = dpaa2_iova_to_virt(priv->iommu_domain, buf_array[i]);
294 dma_unmap_page(dev, buf_array[i], priv->rx_buf_size,
296 free_pages((unsigned long)vaddr, 0);
298 swa = (struct dpaa2_eth_swa *)
299 (vaddr + DPAA2_ETH_RX_HWA_SIZE);
300 xdp_buff = swa->xsk.xdp_buff;
301 xsk_buff_free(xdp_buff);
306 void dpaa2_eth_recycle_buf(struct dpaa2_eth_priv *priv,
307 struct dpaa2_eth_channel *ch,
313 ch->recycled_bufs[ch->recycled_bufs_cnt++] = addr;
314 if (ch->recycled_bufs_cnt < DPAA2_ETH_BUFS_PER_CMD)
317 while ((err = dpaa2_io_service_release(ch->dpio, ch->bp->bpid,
319 ch->recycled_bufs_cnt)) == -EBUSY) {
320 if (retries++ >= DPAA2_ETH_SWP_BUSY_RETRIES)
326 dpaa2_eth_free_bufs(priv, ch->recycled_bufs,
327 ch->recycled_bufs_cnt, ch->xsk_zc);
328 ch->buf_count -= ch->recycled_bufs_cnt;
331 ch->recycled_bufs_cnt = 0;
334 static int dpaa2_eth_xdp_flush(struct dpaa2_eth_priv *priv,
335 struct dpaa2_eth_fq *fq,
336 struct dpaa2_eth_xdp_fds *xdp_fds)
338 int total_enqueued = 0, retries = 0, enqueued;
339 struct dpaa2_eth_drv_stats *percpu_extras;
340 int num_fds, err, max_retries;
341 struct dpaa2_fd *fds;
343 percpu_extras = this_cpu_ptr(priv->percpu_extras);
345 /* try to enqueue all the FDs until the max number of retries is hit */
347 num_fds = xdp_fds->num;
348 max_retries = num_fds * DPAA2_ETH_ENQUEUE_RETRIES;
349 while (total_enqueued < num_fds && retries < max_retries) {
350 err = priv->enqueue(priv, fq, &fds[total_enqueued],
351 0, num_fds - total_enqueued, &enqueued);
353 percpu_extras->tx_portal_busy += ++retries;
356 total_enqueued += enqueued;
360 return total_enqueued;
363 static void dpaa2_eth_xdp_tx_flush(struct dpaa2_eth_priv *priv,
364 struct dpaa2_eth_channel *ch,
365 struct dpaa2_eth_fq *fq)
367 struct rtnl_link_stats64 *percpu_stats;
368 struct dpaa2_fd *fds;
371 percpu_stats = this_cpu_ptr(priv->percpu_stats);
373 // enqueue the array of XDP_TX frames
374 enqueued = dpaa2_eth_xdp_flush(priv, fq, &fq->xdp_tx_fds);
376 /* update statistics */
377 percpu_stats->tx_packets += enqueued;
378 fds = fq->xdp_tx_fds.fds;
379 for (i = 0; i < enqueued; i++) {
380 percpu_stats->tx_bytes += dpaa2_fd_get_len(&fds[i]);
383 for (i = enqueued; i < fq->xdp_tx_fds.num; i++) {
384 dpaa2_eth_recycle_buf(priv, ch, dpaa2_fd_get_addr(&fds[i]));
385 percpu_stats->tx_errors++;
386 ch->stats.xdp_tx_err++;
388 fq->xdp_tx_fds.num = 0;
391 void dpaa2_eth_xdp_enqueue(struct dpaa2_eth_priv *priv,
392 struct dpaa2_eth_channel *ch,
394 void *buf_start, u16 queue_id)
396 struct dpaa2_faead *faead;
397 struct dpaa2_fd *dest_fd;
398 struct dpaa2_eth_fq *fq;
401 /* Mark the egress frame hardware annotation area as valid */
402 frc = dpaa2_fd_get_frc(fd);
403 dpaa2_fd_set_frc(fd, frc | DPAA2_FD_FRC_FAEADV);
404 dpaa2_fd_set_ctrl(fd, DPAA2_FD_CTRL_ASAL);
406 /* Instruct hardware to release the FD buffer directly into
407 * the buffer pool once transmission is completed, instead of
408 * sending a Tx confirmation frame to us
410 ctrl = DPAA2_FAEAD_A4V | DPAA2_FAEAD_A2V | DPAA2_FAEAD_EBDDV;
411 faead = dpaa2_get_faead(buf_start, false);
412 faead->ctrl = cpu_to_le32(ctrl);
413 faead->conf_fqid = 0;
415 fq = &priv->fq[queue_id];
416 dest_fd = &fq->xdp_tx_fds.fds[fq->xdp_tx_fds.num++];
417 memcpy(dest_fd, fd, sizeof(*dest_fd));
419 if (fq->xdp_tx_fds.num < DEV_MAP_BULK_SIZE)
422 dpaa2_eth_xdp_tx_flush(priv, ch, fq);
425 static u32 dpaa2_eth_run_xdp(struct dpaa2_eth_priv *priv,
426 struct dpaa2_eth_channel *ch,
427 struct dpaa2_eth_fq *rx_fq,
428 struct dpaa2_fd *fd, void *vaddr)
430 dma_addr_t addr = dpaa2_fd_get_addr(fd);
431 struct bpf_prog *xdp_prog;
433 u32 xdp_act = XDP_PASS;
436 xdp_prog = READ_ONCE(ch->xdp.prog);
440 offset = dpaa2_fd_get_offset(fd) - XDP_PACKET_HEADROOM;
441 xdp_init_buff(&xdp, DPAA2_ETH_RX_BUF_RAW_SIZE - offset, &ch->xdp_rxq);
442 xdp_prepare_buff(&xdp, vaddr + offset, XDP_PACKET_HEADROOM,
443 dpaa2_fd_get_len(fd), false);
445 xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
447 /* xdp.data pointer may have changed */
448 dpaa2_fd_set_offset(fd, xdp.data - vaddr);
449 dpaa2_fd_set_len(fd, xdp.data_end - xdp.data);
455 dpaa2_eth_xdp_enqueue(priv, ch, fd, vaddr, rx_fq->flowid);
458 bpf_warn_invalid_xdp_action(priv->net_dev, xdp_prog, xdp_act);
461 trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
464 dpaa2_eth_recycle_buf(priv, ch, addr);
465 ch->stats.xdp_drop++;
468 dma_unmap_page(priv->net_dev->dev.parent, addr,
469 priv->rx_buf_size, DMA_BIDIRECTIONAL);
472 /* Allow redirect use of full headroom */
473 xdp.data_hard_start = vaddr;
474 xdp.frame_sz = DPAA2_ETH_RX_BUF_RAW_SIZE;
476 err = xdp_do_redirect(priv->net_dev, &xdp, xdp_prog);
478 addr = dma_map_page(priv->net_dev->dev.parent,
479 virt_to_page(vaddr), 0,
480 priv->rx_buf_size, DMA_BIDIRECTIONAL);
481 if (unlikely(dma_mapping_error(priv->net_dev->dev.parent, addr))) {
482 free_pages((unsigned long)vaddr, 0);
485 dpaa2_eth_recycle_buf(priv, ch, addr);
487 ch->stats.xdp_drop++;
489 ch->stats.xdp_redirect++;
494 ch->xdp.res |= xdp_act;
499 struct sk_buff *dpaa2_eth_alloc_skb(struct dpaa2_eth_priv *priv,
500 struct dpaa2_eth_channel *ch,
501 const struct dpaa2_fd *fd, u32 fd_length,
504 u16 fd_offset = dpaa2_fd_get_offset(fd);
505 struct sk_buff *skb = NULL;
506 unsigned int skb_len;
508 skb_len = fd_length + dpaa2_eth_needed_headroom(NULL);
510 skb = napi_alloc_skb(&ch->napi, skb_len);
514 skb_reserve(skb, dpaa2_eth_needed_headroom(NULL));
515 skb_put(skb, fd_length);
517 memcpy(skb->data, fd_vaddr + fd_offset, fd_length);
522 static struct sk_buff *dpaa2_eth_copybreak(struct dpaa2_eth_channel *ch,
523 const struct dpaa2_fd *fd,
526 struct dpaa2_eth_priv *priv = ch->priv;
527 u32 fd_length = dpaa2_fd_get_len(fd);
529 if (fd_length > priv->rx_copybreak)
532 return dpaa2_eth_alloc_skb(priv, ch, fd, fd_length, fd_vaddr);
535 void dpaa2_eth_receive_skb(struct dpaa2_eth_priv *priv,
536 struct dpaa2_eth_channel *ch,
537 const struct dpaa2_fd *fd, void *vaddr,
538 struct dpaa2_eth_fq *fq,
539 struct rtnl_link_stats64 *percpu_stats,
542 struct dpaa2_fas *fas;
545 fas = dpaa2_get_fas(vaddr, false);
549 /* Get the timestamp value */
550 if (priv->rx_tstamp) {
551 struct skb_shared_hwtstamps *shhwtstamps = skb_hwtstamps(skb);
552 __le64 *ts = dpaa2_get_ts(vaddr, false);
555 memset(shhwtstamps, 0, sizeof(*shhwtstamps));
557 ns = DPAA2_PTP_CLK_PERIOD_NS * le64_to_cpup(ts);
558 shhwtstamps->hwtstamp = ns_to_ktime(ns);
561 /* Check if we need to validate the L4 csum */
562 if (likely(dpaa2_fd_get_frc(fd) & DPAA2_FD_FRC_FASV)) {
563 status = le32_to_cpu(fas->status);
564 dpaa2_eth_validate_rx_csum(priv, status, skb);
567 skb->protocol = eth_type_trans(skb, priv->net_dev);
568 skb_record_rx_queue(skb, fq->flowid);
570 percpu_stats->rx_packets++;
571 percpu_stats->rx_bytes += dpaa2_fd_get_len(fd);
572 ch->stats.bytes_per_cdan += dpaa2_fd_get_len(fd);
574 list_add_tail(&skb->list, ch->rx_list);
577 /* Main Rx frame processing routine */
578 void dpaa2_eth_rx(struct dpaa2_eth_priv *priv,
579 struct dpaa2_eth_channel *ch,
580 const struct dpaa2_fd *fd,
581 struct dpaa2_eth_fq *fq)
583 dma_addr_t addr = dpaa2_fd_get_addr(fd);
584 u8 fd_format = dpaa2_fd_get_format(fd);
587 struct rtnl_link_stats64 *percpu_stats;
588 struct dpaa2_eth_drv_stats *percpu_extras;
589 struct device *dev = priv->net_dev->dev.parent;
590 bool recycle_rx_buf = false;
595 trace_dpaa2_rx_fd(priv->net_dev, fd);
597 vaddr = dpaa2_iova_to_virt(priv->iommu_domain, addr);
598 dma_sync_single_for_cpu(dev, addr, priv->rx_buf_size,
601 buf_data = vaddr + dpaa2_fd_get_offset(fd);
604 percpu_stats = this_cpu_ptr(priv->percpu_stats);
605 percpu_extras = this_cpu_ptr(priv->percpu_extras);
607 if (fd_format == dpaa2_fd_single) {
608 xdp_act = dpaa2_eth_run_xdp(priv, ch, fq, (struct dpaa2_fd *)fd, vaddr);
609 if (xdp_act != XDP_PASS) {
610 percpu_stats->rx_packets++;
611 percpu_stats->rx_bytes += dpaa2_fd_get_len(fd);
615 skb = dpaa2_eth_copybreak(ch, fd, vaddr);
617 dma_unmap_page(dev, addr, priv->rx_buf_size,
619 skb = dpaa2_eth_build_linear_skb(ch, fd, vaddr);
621 recycle_rx_buf = true;
623 } else if (fd_format == dpaa2_fd_sg) {
624 WARN_ON(priv->xdp_prog);
626 dma_unmap_page(dev, addr, priv->rx_buf_size,
628 skb = dpaa2_eth_build_frag_skb(priv, ch, buf_data);
629 free_pages((unsigned long)vaddr, 0);
630 percpu_extras->rx_sg_frames++;
631 percpu_extras->rx_sg_bytes += dpaa2_fd_get_len(fd);
633 /* We don't support any other format */
634 goto err_frame_format;
640 dpaa2_eth_receive_skb(priv, ch, fd, vaddr, fq, percpu_stats, skb);
643 dpaa2_eth_recycle_buf(priv, ch, dpaa2_fd_get_addr(fd));
647 dpaa2_eth_free_rx_fd(priv, fd, vaddr);
649 percpu_stats->rx_dropped++;
652 /* Processing of Rx frames received on the error FQ
653 * We check and print the error bits and then free the frame
655 static void dpaa2_eth_rx_err(struct dpaa2_eth_priv *priv,
656 struct dpaa2_eth_channel *ch,
657 const struct dpaa2_fd *fd,
658 struct dpaa2_eth_fq *fq __always_unused)
660 struct device *dev = priv->net_dev->dev.parent;
661 dma_addr_t addr = dpaa2_fd_get_addr(fd);
662 u8 fd_format = dpaa2_fd_get_format(fd);
663 struct rtnl_link_stats64 *percpu_stats;
664 struct dpaa2_eth_trap_item *trap_item;
665 struct dpaa2_fapr *fapr;
670 vaddr = dpaa2_iova_to_virt(priv->iommu_domain, addr);
671 dma_sync_single_for_cpu(dev, addr, priv->rx_buf_size,
674 buf_data = vaddr + dpaa2_fd_get_offset(fd);
676 if (fd_format == dpaa2_fd_single) {
677 dma_unmap_page(dev, addr, priv->rx_buf_size,
679 skb = dpaa2_eth_build_linear_skb(ch, fd, vaddr);
680 } else if (fd_format == dpaa2_fd_sg) {
681 dma_unmap_page(dev, addr, priv->rx_buf_size,
683 skb = dpaa2_eth_build_frag_skb(priv, ch, buf_data);
684 free_pages((unsigned long)vaddr, 0);
686 /* We don't support any other format */
687 dpaa2_eth_free_rx_fd(priv, fd, vaddr);
688 goto err_frame_format;
691 fapr = dpaa2_get_fapr(vaddr, false);
692 trap_item = dpaa2_eth_dl_get_trap(priv, fapr);
694 devlink_trap_report(priv->devlink, skb, trap_item->trap_ctx,
695 &priv->devlink_port, NULL);
699 percpu_stats = this_cpu_ptr(priv->percpu_stats);
700 percpu_stats->rx_errors++;
704 /* Consume all frames pull-dequeued into the store. This is the simplest way to
705 * make sure we don't accidentally issue another volatile dequeue which would
706 * overwrite (leak) frames already in the store.
708 * Observance of NAPI budget is not our concern, leaving that to the caller.
710 static int dpaa2_eth_consume_frames(struct dpaa2_eth_channel *ch,
711 struct dpaa2_eth_fq **src)
713 struct dpaa2_eth_priv *priv = ch->priv;
714 struct dpaa2_eth_fq *fq = NULL;
716 const struct dpaa2_fd *fd;
717 int cleaned = 0, retries = 0;
721 dq = dpaa2_io_store_next(ch->store, &is_last);
723 /* If we're here, we *must* have placed a
724 * volatile dequeue comnmand, so keep reading through
725 * the store until we get some sort of valid response
726 * token (either a valid frame or an "empty dequeue")
728 if (retries++ >= DPAA2_ETH_SWP_BUSY_RETRIES) {
729 netdev_err_once(priv->net_dev,
730 "Unable to read a valid dequeue response\n");
736 fd = dpaa2_dq_fd(dq);
737 fq = (struct dpaa2_eth_fq *)(uintptr_t)dpaa2_dq_fqd_ctx(dq);
739 fq->consume(priv, ch, fd, fq);
747 fq->stats.frames += cleaned;
748 ch->stats.frames += cleaned;
749 ch->stats.frames_per_cdan += cleaned;
751 /* A dequeue operation only pulls frames from a single queue
752 * into the store. Return the frame queue as an out param.
760 static int dpaa2_eth_ptp_parse(struct sk_buff *skb,
761 u8 *msgtype, u8 *twostep, u8 *udp,
762 u16 *correction_offset,
763 u16 *origintimestamp_offset)
765 unsigned int ptp_class;
766 struct ptp_header *hdr;
770 ptp_class = ptp_classify_raw(skb);
771 if (ptp_class == PTP_CLASS_NONE)
774 hdr = ptp_parse_header(skb, ptp_class);
778 *msgtype = ptp_get_msgtype(hdr, ptp_class);
779 *twostep = hdr->flag_field[0] & 0x2;
781 type = ptp_class & PTP_CLASS_PMASK;
782 if (type == PTP_CLASS_IPV4 ||
783 type == PTP_CLASS_IPV6)
788 base = skb_mac_header(skb);
789 *correction_offset = (u8 *)&hdr->correction - base;
790 *origintimestamp_offset = (u8 *)hdr + sizeof(struct ptp_header) - base;
795 /* Configure the egress frame annotation for timestamp update */
796 static void dpaa2_eth_enable_tx_tstamp(struct dpaa2_eth_priv *priv,
801 struct ptp_tstamp origin_timestamp;
802 u8 msgtype, twostep, udp;
803 struct dpaa2_faead *faead;
804 struct dpaa2_fas *fas;
805 struct timespec64 ts;
806 u16 offset1, offset2;
811 /* Mark the egress frame annotation area as valid */
812 frc = dpaa2_fd_get_frc(fd);
813 dpaa2_fd_set_frc(fd, frc | DPAA2_FD_FRC_FAEADV);
815 /* Set hardware annotation size */
816 ctrl = dpaa2_fd_get_ctrl(fd);
817 dpaa2_fd_set_ctrl(fd, ctrl | DPAA2_FD_CTRL_ASAL);
819 /* enable UPD (update prepanded data) bit in FAEAD field of
820 * hardware frame annotation area
822 ctrl = DPAA2_FAEAD_A2V | DPAA2_FAEAD_UPDV | DPAA2_FAEAD_UPD;
823 faead = dpaa2_get_faead(buf_start, true);
824 faead->ctrl = cpu_to_le32(ctrl);
826 if (skb->cb[0] == TX_TSTAMP_ONESTEP_SYNC) {
827 if (dpaa2_eth_ptp_parse(skb, &msgtype, &twostep, &udp,
828 &offset1, &offset2) ||
829 msgtype != PTP_MSGTYPE_SYNC || twostep) {
830 WARN_ONCE(1, "Bad packet for one-step timestamping\n");
834 /* Mark the frame annotation status as valid */
835 frc = dpaa2_fd_get_frc(fd);
836 dpaa2_fd_set_frc(fd, frc | DPAA2_FD_FRC_FASV);
838 /* Mark the PTP flag for one step timestamping */
839 fas = dpaa2_get_fas(buf_start, true);
840 fas->status = cpu_to_le32(DPAA2_FAS_PTP);
842 dpaa2_ptp->caps.gettime64(&dpaa2_ptp->caps, &ts);
843 ns = dpaa2_get_ts(buf_start, true);
844 *ns = cpu_to_le64(timespec64_to_ns(&ts) /
845 DPAA2_PTP_CLK_PERIOD_NS);
847 /* Update current time to PTP message originTimestamp field */
848 ns_to_ptp_tstamp(&origin_timestamp, le64_to_cpup(ns));
849 data = skb_mac_header(skb);
850 *(__be16 *)(data + offset2) = htons(origin_timestamp.sec_msb);
851 *(__be32 *)(data + offset2 + 2) =
852 htonl(origin_timestamp.sec_lsb);
853 *(__be32 *)(data + offset2 + 6) = htonl(origin_timestamp.nsec);
855 if (priv->ptp_correction_off == offset1)
858 priv->dpaa2_set_onestep_params_cb(priv, offset1, udp);
859 priv->ptp_correction_off = offset1;
864 void *dpaa2_eth_sgt_get(struct dpaa2_eth_priv *priv)
866 struct dpaa2_eth_sgt_cache *sgt_cache;
867 void *sgt_buf = NULL;
870 sgt_cache = this_cpu_ptr(priv->sgt_cache);
871 sgt_buf_size = priv->tx_data_offset +
872 DPAA2_ETH_SG_ENTRIES_MAX * sizeof(struct dpaa2_sg_entry);
874 if (sgt_cache->count == 0)
875 sgt_buf = napi_alloc_frag_align(sgt_buf_size, DPAA2_ETH_TX_BUF_ALIGN);
877 sgt_buf = sgt_cache->buf[--sgt_cache->count];
881 memset(sgt_buf, 0, sgt_buf_size);
886 void dpaa2_eth_sgt_recycle(struct dpaa2_eth_priv *priv, void *sgt_buf)
888 struct dpaa2_eth_sgt_cache *sgt_cache;
890 sgt_cache = this_cpu_ptr(priv->sgt_cache);
891 if (sgt_cache->count >= DPAA2_ETH_SGT_CACHE_SIZE)
892 skb_free_frag(sgt_buf);
894 sgt_cache->buf[sgt_cache->count++] = sgt_buf;
897 /* Create a frame descriptor based on a fragmented skb */
898 static int dpaa2_eth_build_sg_fd(struct dpaa2_eth_priv *priv,
903 struct device *dev = priv->net_dev->dev.parent;
904 void *sgt_buf = NULL;
906 int nr_frags = skb_shinfo(skb)->nr_frags;
907 struct dpaa2_sg_entry *sgt;
910 struct scatterlist *scl, *crt_scl;
913 struct dpaa2_eth_swa *swa;
915 /* Create and map scatterlist.
916 * We don't advertise NETIF_F_FRAGLIST, so skb_to_sgvec() will not have
917 * to go beyond nr_frags+1.
918 * Note: We don't support chained scatterlists
920 if (unlikely(PAGE_SIZE / sizeof(struct scatterlist) < nr_frags + 1))
923 scl = kmalloc_array(nr_frags + 1, sizeof(struct scatterlist), GFP_ATOMIC);
927 sg_init_table(scl, nr_frags + 1);
928 num_sg = skb_to_sgvec(skb, scl, 0, skb->len);
929 if (unlikely(num_sg < 0)) {
931 goto dma_map_sg_failed;
933 num_dma_bufs = dma_map_sg(dev, scl, num_sg, DMA_BIDIRECTIONAL);
934 if (unlikely(!num_dma_bufs)) {
936 goto dma_map_sg_failed;
939 /* Prepare the HW SGT structure */
940 sgt_buf_size = priv->tx_data_offset +
941 sizeof(struct dpaa2_sg_entry) * num_dma_bufs;
942 sgt_buf = dpaa2_eth_sgt_get(priv);
943 if (unlikely(!sgt_buf)) {
945 goto sgt_buf_alloc_failed;
948 sgt = (struct dpaa2_sg_entry *)(sgt_buf + priv->tx_data_offset);
950 /* Fill in the HW SGT structure.
952 * sgt_buf is zeroed out, so the following fields are implicit
953 * in all sgt entries:
955 * - format is 'dpaa2_sg_single'
957 for_each_sg(scl, crt_scl, num_dma_bufs, i) {
958 dpaa2_sg_set_addr(&sgt[i], sg_dma_address(crt_scl));
959 dpaa2_sg_set_len(&sgt[i], sg_dma_len(crt_scl));
961 dpaa2_sg_set_final(&sgt[i - 1], true);
963 /* Store the skb backpointer in the SGT buffer.
964 * Fit the scatterlist and the number of buffers alongside the
965 * skb backpointer in the software annotation area. We'll need
966 * all of them on Tx Conf.
968 *swa_addr = (void *)sgt_buf;
969 swa = (struct dpaa2_eth_swa *)sgt_buf;
970 swa->type = DPAA2_ETH_SWA_SG;
973 swa->sg.num_sg = num_sg;
974 swa->sg.sgt_size = sgt_buf_size;
976 /* Separately map the SGT buffer */
977 addr = dma_map_single(dev, sgt_buf, sgt_buf_size, DMA_BIDIRECTIONAL);
978 if (unlikely(dma_mapping_error(dev, addr))) {
980 goto dma_map_single_failed;
982 memset(fd, 0, sizeof(struct dpaa2_fd));
983 dpaa2_fd_set_offset(fd, priv->tx_data_offset);
984 dpaa2_fd_set_format(fd, dpaa2_fd_sg);
985 dpaa2_fd_set_addr(fd, addr);
986 dpaa2_fd_set_len(fd, skb->len);
987 dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
991 dma_map_single_failed:
992 dpaa2_eth_sgt_recycle(priv, sgt_buf);
993 sgt_buf_alloc_failed:
994 dma_unmap_sg(dev, scl, num_sg, DMA_BIDIRECTIONAL);
1000 /* Create a SG frame descriptor based on a linear skb.
1002 * This function is used on the Tx path when the skb headroom is not large
1003 * enough for the HW requirements, thus instead of realloc-ing the skb we
1004 * create a SG frame descriptor with only one entry.
1006 static int dpaa2_eth_build_sg_fd_single_buf(struct dpaa2_eth_priv *priv,
1007 struct sk_buff *skb,
1008 struct dpaa2_fd *fd,
1011 struct device *dev = priv->net_dev->dev.parent;
1012 struct dpaa2_sg_entry *sgt;
1013 struct dpaa2_eth_swa *swa;
1014 dma_addr_t addr, sgt_addr;
1015 void *sgt_buf = NULL;
1019 /* Prepare the HW SGT structure */
1020 sgt_buf_size = priv->tx_data_offset + sizeof(struct dpaa2_sg_entry);
1021 sgt_buf = dpaa2_eth_sgt_get(priv);
1022 if (unlikely(!sgt_buf))
1024 sgt = (struct dpaa2_sg_entry *)(sgt_buf + priv->tx_data_offset);
1026 addr = dma_map_single(dev, skb->data, skb->len, DMA_BIDIRECTIONAL);
1027 if (unlikely(dma_mapping_error(dev, addr))) {
1029 goto data_map_failed;
1032 /* Fill in the HW SGT structure */
1033 dpaa2_sg_set_addr(sgt, addr);
1034 dpaa2_sg_set_len(sgt, skb->len);
1035 dpaa2_sg_set_final(sgt, true);
1037 /* Store the skb backpointer in the SGT buffer */
1038 *swa_addr = (void *)sgt_buf;
1039 swa = (struct dpaa2_eth_swa *)sgt_buf;
1040 swa->type = DPAA2_ETH_SWA_SINGLE;
1041 swa->single.skb = skb;
1042 swa->single.sgt_size = sgt_buf_size;
1044 /* Separately map the SGT buffer */
1045 sgt_addr = dma_map_single(dev, sgt_buf, sgt_buf_size, DMA_BIDIRECTIONAL);
1046 if (unlikely(dma_mapping_error(dev, sgt_addr))) {
1048 goto sgt_map_failed;
1051 memset(fd, 0, sizeof(struct dpaa2_fd));
1052 dpaa2_fd_set_offset(fd, priv->tx_data_offset);
1053 dpaa2_fd_set_format(fd, dpaa2_fd_sg);
1054 dpaa2_fd_set_addr(fd, sgt_addr);
1055 dpaa2_fd_set_len(fd, skb->len);
1056 dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
1061 dma_unmap_single(dev, addr, skb->len, DMA_BIDIRECTIONAL);
1063 dpaa2_eth_sgt_recycle(priv, sgt_buf);
1068 /* Create a frame descriptor based on a linear skb */
1069 static int dpaa2_eth_build_single_fd(struct dpaa2_eth_priv *priv,
1070 struct sk_buff *skb,
1071 struct dpaa2_fd *fd,
1074 struct device *dev = priv->net_dev->dev.parent;
1075 u8 *buffer_start, *aligned_start;
1076 struct dpaa2_eth_swa *swa;
1079 buffer_start = skb->data - dpaa2_eth_needed_headroom(skb);
1080 aligned_start = PTR_ALIGN(buffer_start - DPAA2_ETH_TX_BUF_ALIGN,
1081 DPAA2_ETH_TX_BUF_ALIGN);
1082 if (aligned_start >= skb->head)
1083 buffer_start = aligned_start;
1087 /* Store a backpointer to the skb at the beginning of the buffer
1088 * (in the private data area) such that we can release it
1091 *swa_addr = (void *)buffer_start;
1092 swa = (struct dpaa2_eth_swa *)buffer_start;
1093 swa->type = DPAA2_ETH_SWA_SINGLE;
1094 swa->single.skb = skb;
1096 addr = dma_map_single(dev, buffer_start,
1097 skb_tail_pointer(skb) - buffer_start,
1099 if (unlikely(dma_mapping_error(dev, addr)))
1102 memset(fd, 0, sizeof(struct dpaa2_fd));
1103 dpaa2_fd_set_addr(fd, addr);
1104 dpaa2_fd_set_offset(fd, (u16)(skb->data - buffer_start));
1105 dpaa2_fd_set_len(fd, skb->len);
1106 dpaa2_fd_set_format(fd, dpaa2_fd_single);
1107 dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
1112 /* FD freeing routine on the Tx path
1114 * DMA-unmap and free FD and possibly SGT buffer allocated on Tx. The skb
1115 * back-pointed to is also freed.
1116 * This can be called either from dpaa2_eth_tx_conf() or on the error path of
1119 void dpaa2_eth_free_tx_fd(struct dpaa2_eth_priv *priv,
1120 struct dpaa2_eth_channel *ch,
1121 struct dpaa2_eth_fq *fq,
1122 const struct dpaa2_fd *fd, bool in_napi)
1124 struct device *dev = priv->net_dev->dev.parent;
1125 dma_addr_t fd_addr, sg_addr;
1126 struct sk_buff *skb = NULL;
1127 unsigned char *buffer_start;
1128 struct dpaa2_eth_swa *swa;
1129 u8 fd_format = dpaa2_fd_get_format(fd);
1130 u32 fd_len = dpaa2_fd_get_len(fd);
1131 struct dpaa2_sg_entry *sgt;
1132 int should_free_skb = 1;
1136 fd_addr = dpaa2_fd_get_addr(fd);
1137 buffer_start = dpaa2_iova_to_virt(priv->iommu_domain, fd_addr);
1138 swa = (struct dpaa2_eth_swa *)buffer_start;
1140 if (fd_format == dpaa2_fd_single) {
1141 if (swa->type == DPAA2_ETH_SWA_SINGLE) {
1142 skb = swa->single.skb;
1143 /* Accessing the skb buffer is safe before dma unmap,
1144 * because we didn't map the actual skb shell.
1146 dma_unmap_single(dev, fd_addr,
1147 skb_tail_pointer(skb) - buffer_start,
1150 WARN_ONCE(swa->type != DPAA2_ETH_SWA_XDP, "Wrong SWA type");
1151 dma_unmap_single(dev, fd_addr, swa->xdp.dma_size,
1154 } else if (fd_format == dpaa2_fd_sg) {
1155 if (swa->type == DPAA2_ETH_SWA_SG) {
1158 /* Unmap the scatterlist */
1159 dma_unmap_sg(dev, swa->sg.scl, swa->sg.num_sg,
1163 /* Unmap the SGT buffer */
1164 dma_unmap_single(dev, fd_addr, swa->sg.sgt_size,
1166 } else if (swa->type == DPAA2_ETH_SWA_SW_TSO) {
1169 sgt = (struct dpaa2_sg_entry *)(buffer_start +
1170 priv->tx_data_offset);
1172 /* Unmap the SGT buffer */
1173 dma_unmap_single(dev, fd_addr, swa->tso.sgt_size,
1176 /* Unmap and free the header */
1177 tso_hdr = dpaa2_iova_to_virt(priv->iommu_domain, dpaa2_sg_get_addr(sgt));
1178 dma_unmap_single(dev, dpaa2_sg_get_addr(sgt), TSO_HEADER_SIZE,
1182 /* Unmap the other SG entries for the data */
1183 for (i = 1; i < swa->tso.num_sg; i++)
1184 dma_unmap_single(dev, dpaa2_sg_get_addr(&sgt[i]),
1185 dpaa2_sg_get_len(&sgt[i]), DMA_TO_DEVICE);
1187 if (!swa->tso.is_last_fd)
1188 should_free_skb = 0;
1189 } else if (swa->type == DPAA2_ETH_SWA_XSK) {
1190 /* Unmap the SGT Buffer */
1191 dma_unmap_single(dev, fd_addr, swa->xsk.sgt_size,
1194 skb = swa->single.skb;
1196 /* Unmap the SGT Buffer */
1197 dma_unmap_single(dev, fd_addr, swa->single.sgt_size,
1200 sgt = (struct dpaa2_sg_entry *)(buffer_start +
1201 priv->tx_data_offset);
1202 sg_addr = dpaa2_sg_get_addr(sgt);
1203 dma_unmap_single(dev, sg_addr, skb->len, DMA_BIDIRECTIONAL);
1206 netdev_dbg(priv->net_dev, "Invalid FD format\n");
1210 if (swa->type == DPAA2_ETH_SWA_XSK) {
1211 ch->xsk_tx_pkts_sent++;
1212 dpaa2_eth_sgt_recycle(priv, buffer_start);
1216 if (swa->type != DPAA2_ETH_SWA_XDP && in_napi) {
1218 fq->dq_bytes += fd_len;
1221 if (swa->type == DPAA2_ETH_SWA_XDP) {
1222 xdp_return_frame(swa->xdp.xdpf);
1226 /* Get the timestamp value */
1227 if (swa->type != DPAA2_ETH_SWA_SW_TSO) {
1228 if (skb->cb[0] == TX_TSTAMP) {
1229 struct skb_shared_hwtstamps shhwtstamps;
1230 __le64 *ts = dpaa2_get_ts(buffer_start, true);
1233 memset(&shhwtstamps, 0, sizeof(shhwtstamps));
1235 ns = DPAA2_PTP_CLK_PERIOD_NS * le64_to_cpup(ts);
1236 shhwtstamps.hwtstamp = ns_to_ktime(ns);
1237 skb_tstamp_tx(skb, &shhwtstamps);
1238 } else if (skb->cb[0] == TX_TSTAMP_ONESTEP_SYNC) {
1239 mutex_unlock(&priv->onestep_tstamp_lock);
1243 /* Free SGT buffer allocated on tx */
1244 if (fd_format != dpaa2_fd_single)
1245 dpaa2_eth_sgt_recycle(priv, buffer_start);
1247 /* Move on with skb release. If we are just confirming multiple FDs
1248 * from the same TSO skb then only the last one will need to free the
1251 if (should_free_skb)
1252 napi_consume_skb(skb, in_napi);
1255 static int dpaa2_eth_build_gso_fd(struct dpaa2_eth_priv *priv,
1256 struct sk_buff *skb, struct dpaa2_fd *fd,
1257 int *num_fds, u32 *total_fds_len)
1259 struct device *dev = priv->net_dev->dev.parent;
1260 int hdr_len, total_len, data_left, fd_len;
1261 int num_sge, err, i, sgt_buf_size;
1262 struct dpaa2_fd *fd_start = fd;
1263 struct dpaa2_sg_entry *sgt;
1264 struct dpaa2_eth_swa *swa;
1265 dma_addr_t sgt_addr, addr;
1266 dma_addr_t tso_hdr_dma;
1267 unsigned int index = 0;
1272 /* Initialize the TSO handler, and prepare the first payload */
1273 hdr_len = tso_start(skb, &tso);
1276 total_len = skb->len - hdr_len;
1277 while (total_len > 0) {
1278 /* Prepare the HW SGT structure for this frame */
1279 sgt_buf = dpaa2_eth_sgt_get(priv);
1280 if (unlikely(!sgt_buf)) {
1281 netdev_err(priv->net_dev, "dpaa2_eth_sgt_get() failed\n");
1285 sgt = (struct dpaa2_sg_entry *)(sgt_buf + priv->tx_data_offset);
1287 /* Determine the data length of this frame */
1288 data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
1289 total_len -= data_left;
1290 fd_len = data_left + hdr_len;
1292 /* Prepare packet headers: MAC + IP + TCP */
1293 tso_hdr = kmalloc(TSO_HEADER_SIZE, GFP_ATOMIC);
1296 goto err_alloc_tso_hdr;
1299 tso_build_hdr(skb, tso_hdr, &tso, data_left, total_len == 0);
1300 tso_hdr_dma = dma_map_single(dev, tso_hdr, TSO_HEADER_SIZE, DMA_TO_DEVICE);
1301 if (dma_mapping_error(dev, tso_hdr_dma)) {
1302 netdev_err(priv->net_dev, "dma_map_single(tso_hdr) failed\n");
1304 goto err_map_tso_hdr;
1307 /* Setup the SG entry for the header */
1308 dpaa2_sg_set_addr(sgt, tso_hdr_dma);
1309 dpaa2_sg_set_len(sgt, hdr_len);
1310 dpaa2_sg_set_final(sgt, data_left <= 0);
1312 /* Compose the SG entries for each fragment of data */
1314 while (data_left > 0) {
1317 /* Move to the next SG entry */
1319 size = min_t(int, tso.size, data_left);
1321 addr = dma_map_single(dev, tso.data, size, DMA_TO_DEVICE);
1322 if (dma_mapping_error(dev, addr)) {
1323 netdev_err(priv->net_dev, "dma_map_single(tso.data) failed\n");
1327 dpaa2_sg_set_addr(sgt, addr);
1328 dpaa2_sg_set_len(sgt, size);
1329 dpaa2_sg_set_final(sgt, size == data_left);
1333 /* Build the data for the __next__ fragment */
1335 tso_build_data(skb, &tso, size);
1338 /* Store the skb backpointer in the SGT buffer */
1339 sgt_buf_size = priv->tx_data_offset + num_sge * sizeof(struct dpaa2_sg_entry);
1340 swa = (struct dpaa2_eth_swa *)sgt_buf;
1341 swa->type = DPAA2_ETH_SWA_SW_TSO;
1343 swa->tso.num_sg = num_sge;
1344 swa->tso.sgt_size = sgt_buf_size;
1345 swa->tso.is_last_fd = total_len == 0 ? 1 : 0;
1347 /* Separately map the SGT buffer */
1348 sgt_addr = dma_map_single(dev, sgt_buf, sgt_buf_size, DMA_BIDIRECTIONAL);
1349 if (unlikely(dma_mapping_error(dev, sgt_addr))) {
1350 netdev_err(priv->net_dev, "dma_map_single(sgt_buf) failed\n");
1355 /* Setup the frame descriptor */
1356 memset(fd, 0, sizeof(struct dpaa2_fd));
1357 dpaa2_fd_set_offset(fd, priv->tx_data_offset);
1358 dpaa2_fd_set_format(fd, dpaa2_fd_sg);
1359 dpaa2_fd_set_addr(fd, sgt_addr);
1360 dpaa2_fd_set_len(fd, fd_len);
1361 dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
1363 *total_fds_len += fd_len;
1364 /* Advance to the next frame descriptor */
1375 /* Unmap all the data S/G entries for the current FD */
1376 sgt = (struct dpaa2_sg_entry *)(sgt_buf + priv->tx_data_offset);
1377 for (i = 1; i < num_sge; i++)
1378 dma_unmap_single(dev, dpaa2_sg_get_addr(&sgt[i]),
1379 dpaa2_sg_get_len(&sgt[i]), DMA_TO_DEVICE);
1381 /* Unmap the header entry */
1382 dma_unmap_single(dev, tso_hdr_dma, TSO_HEADER_SIZE, DMA_TO_DEVICE);
1386 dpaa2_eth_sgt_recycle(priv, sgt_buf);
1388 /* Free all the other FDs that were already fully created */
1389 for (i = 0; i < index; i++)
1390 dpaa2_eth_free_tx_fd(priv, NULL, NULL, &fd_start[i], false);
1395 static netdev_tx_t __dpaa2_eth_tx(struct sk_buff *skb,
1396 struct net_device *net_dev)
1398 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
1399 int total_enqueued = 0, retries = 0, enqueued;
1400 struct dpaa2_eth_drv_stats *percpu_extras;
1401 struct rtnl_link_stats64 *percpu_stats;
1402 unsigned int needed_headroom;
1403 int num_fds = 1, max_retries;
1404 struct dpaa2_eth_fq *fq;
1405 struct netdev_queue *nq;
1406 struct dpaa2_fd *fd;
1413 percpu_stats = this_cpu_ptr(priv->percpu_stats);
1414 percpu_extras = this_cpu_ptr(priv->percpu_extras);
1415 fd = (this_cpu_ptr(priv->fd))->array;
1417 needed_headroom = dpaa2_eth_needed_headroom(skb);
1419 /* We'll be holding a back-reference to the skb until Tx Confirmation;
1420 * we don't want that overwritten by a concurrent Tx with a cloned skb.
1422 skb = skb_unshare(skb, GFP_ATOMIC);
1423 if (unlikely(!skb)) {
1424 /* skb_unshare() has already freed the skb */
1425 percpu_stats->tx_dropped++;
1426 return NETDEV_TX_OK;
1429 /* Setup the FD fields */
1431 if (skb_is_gso(skb)) {
1432 err = dpaa2_eth_build_gso_fd(priv, skb, fd, &num_fds, &fd_len);
1433 percpu_extras->tx_sg_frames += num_fds;
1434 percpu_extras->tx_sg_bytes += fd_len;
1435 percpu_extras->tx_tso_frames += num_fds;
1436 percpu_extras->tx_tso_bytes += fd_len;
1437 } else if (skb_is_nonlinear(skb)) {
1438 err = dpaa2_eth_build_sg_fd(priv, skb, fd, &swa);
1439 percpu_extras->tx_sg_frames++;
1440 percpu_extras->tx_sg_bytes += skb->len;
1441 fd_len = dpaa2_fd_get_len(fd);
1442 } else if (skb_headroom(skb) < needed_headroom) {
1443 err = dpaa2_eth_build_sg_fd_single_buf(priv, skb, fd, &swa);
1444 percpu_extras->tx_sg_frames++;
1445 percpu_extras->tx_sg_bytes += skb->len;
1446 percpu_extras->tx_converted_sg_frames++;
1447 percpu_extras->tx_converted_sg_bytes += skb->len;
1448 fd_len = dpaa2_fd_get_len(fd);
1450 err = dpaa2_eth_build_single_fd(priv, skb, fd, &swa);
1451 fd_len = dpaa2_fd_get_len(fd);
1454 if (unlikely(err)) {
1455 percpu_stats->tx_dropped++;
1459 if (swa && skb->cb[0])
1460 dpaa2_eth_enable_tx_tstamp(priv, fd, swa, skb);
1463 for (i = 0; i < num_fds; i++)
1464 trace_dpaa2_tx_fd(net_dev, &fd[i]);
1466 /* TxConf FQ selection relies on queue id from the stack.
1467 * In case of a forwarded frame from another DPNI interface, we choose
1468 * a queue affined to the same core that processed the Rx frame
1470 queue_mapping = skb_get_queue_mapping(skb);
1472 if (net_dev->num_tc) {
1473 prio = netdev_txq_to_tc(net_dev, queue_mapping);
1474 /* Hardware interprets priority level 0 as being the highest,
1475 * so we need to do a reverse mapping to the netdev tc index
1477 prio = net_dev->num_tc - prio - 1;
1478 /* We have only one FQ array entry for all Tx hardware queues
1479 * with the same flow id (but different priority levels)
1481 queue_mapping %= dpaa2_eth_queue_count(priv);
1483 fq = &priv->fq[queue_mapping];
1484 nq = netdev_get_tx_queue(net_dev, queue_mapping);
1485 netdev_tx_sent_queue(nq, fd_len);
1487 /* Everything that happens after this enqueues might race with
1488 * the Tx confirmation callback for this frame
1490 max_retries = num_fds * DPAA2_ETH_ENQUEUE_RETRIES;
1491 while (total_enqueued < num_fds && retries < max_retries) {
1492 err = priv->enqueue(priv, fq, &fd[total_enqueued],
1493 prio, num_fds - total_enqueued, &enqueued);
1494 if (err == -EBUSY) {
1499 total_enqueued += enqueued;
1501 percpu_extras->tx_portal_busy += retries;
1503 if (unlikely(err < 0)) {
1504 percpu_stats->tx_errors++;
1505 /* Clean up everything, including freeing the skb */
1506 dpaa2_eth_free_tx_fd(priv, NULL, fq, fd, false);
1507 netdev_tx_completed_queue(nq, 1, fd_len);
1509 percpu_stats->tx_packets += total_enqueued;
1510 percpu_stats->tx_bytes += fd_len;
1513 return NETDEV_TX_OK;
1518 return NETDEV_TX_OK;
1521 static void dpaa2_eth_tx_onestep_tstamp(struct work_struct *work)
1523 struct dpaa2_eth_priv *priv = container_of(work, struct dpaa2_eth_priv,
1525 struct sk_buff *skb;
1528 skb = skb_dequeue(&priv->tx_skbs);
1532 /* Lock just before TX one-step timestamping packet,
1533 * and release the lock in dpaa2_eth_free_tx_fd when
1534 * confirm the packet has been sent on hardware, or
1535 * when clean up during transmit failure.
1537 mutex_lock(&priv->onestep_tstamp_lock);
1538 __dpaa2_eth_tx(skb, priv->net_dev);
1542 static netdev_tx_t dpaa2_eth_tx(struct sk_buff *skb, struct net_device *net_dev)
1544 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
1545 u8 msgtype, twostep, udp;
1546 u16 offset1, offset2;
1548 /* Utilize skb->cb[0] for timestamping request per skb */
1551 if ((skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) && dpaa2_ptp) {
1552 if (priv->tx_tstamp_type == HWTSTAMP_TX_ON)
1553 skb->cb[0] = TX_TSTAMP;
1554 else if (priv->tx_tstamp_type == HWTSTAMP_TX_ONESTEP_SYNC)
1555 skb->cb[0] = TX_TSTAMP_ONESTEP_SYNC;
1558 /* TX for one-step timestamping PTP Sync packet */
1559 if (skb->cb[0] == TX_TSTAMP_ONESTEP_SYNC) {
1560 if (!dpaa2_eth_ptp_parse(skb, &msgtype, &twostep, &udp,
1561 &offset1, &offset2))
1562 if (msgtype == PTP_MSGTYPE_SYNC && twostep == 0) {
1563 skb_queue_tail(&priv->tx_skbs, skb);
1564 queue_work(priv->dpaa2_ptp_wq,
1565 &priv->tx_onestep_tstamp);
1566 return NETDEV_TX_OK;
1568 /* Use two-step timestamping if not one-step timestamping
1571 skb->cb[0] = TX_TSTAMP;
1574 /* TX for other packets */
1575 return __dpaa2_eth_tx(skb, net_dev);
1578 /* Tx confirmation frame processing routine */
1579 static void dpaa2_eth_tx_conf(struct dpaa2_eth_priv *priv,
1580 struct dpaa2_eth_channel *ch,
1581 const struct dpaa2_fd *fd,
1582 struct dpaa2_eth_fq *fq)
1584 struct rtnl_link_stats64 *percpu_stats;
1585 struct dpaa2_eth_drv_stats *percpu_extras;
1586 u32 fd_len = dpaa2_fd_get_len(fd);
1590 trace_dpaa2_tx_conf_fd(priv->net_dev, fd);
1592 percpu_extras = this_cpu_ptr(priv->percpu_extras);
1593 percpu_extras->tx_conf_frames++;
1594 percpu_extras->tx_conf_bytes += fd_len;
1595 ch->stats.bytes_per_cdan += fd_len;
1597 /* Check frame errors in the FD field */
1598 fd_errors = dpaa2_fd_get_ctrl(fd) & DPAA2_FD_TX_ERR_MASK;
1599 dpaa2_eth_free_tx_fd(priv, ch, fq, fd, true);
1601 if (likely(!fd_errors))
1604 if (net_ratelimit())
1605 netdev_dbg(priv->net_dev, "TX frame FD error: 0x%08x\n",
1608 percpu_stats = this_cpu_ptr(priv->percpu_stats);
1609 /* Tx-conf logically pertains to the egress path. */
1610 percpu_stats->tx_errors++;
1613 static int dpaa2_eth_set_rx_vlan_filtering(struct dpaa2_eth_priv *priv,
1618 err = dpni_enable_vlan_filter(priv->mc_io, 0, priv->mc_token, enable);
1621 netdev_err(priv->net_dev,
1622 "dpni_enable_vlan_filter failed\n");
1629 static int dpaa2_eth_set_rx_csum(struct dpaa2_eth_priv *priv, bool enable)
1633 err = dpni_set_offload(priv->mc_io, 0, priv->mc_token,
1634 DPNI_OFF_RX_L3_CSUM, enable);
1636 netdev_err(priv->net_dev,
1637 "dpni_set_offload(RX_L3_CSUM) failed\n");
1641 err = dpni_set_offload(priv->mc_io, 0, priv->mc_token,
1642 DPNI_OFF_RX_L4_CSUM, enable);
1644 netdev_err(priv->net_dev,
1645 "dpni_set_offload(RX_L4_CSUM) failed\n");
1652 static int dpaa2_eth_set_tx_csum(struct dpaa2_eth_priv *priv, bool enable)
1656 err = dpni_set_offload(priv->mc_io, 0, priv->mc_token,
1657 DPNI_OFF_TX_L3_CSUM, enable);
1659 netdev_err(priv->net_dev, "dpni_set_offload(TX_L3_CSUM) failed\n");
1663 err = dpni_set_offload(priv->mc_io, 0, priv->mc_token,
1664 DPNI_OFF_TX_L4_CSUM, enable);
1666 netdev_err(priv->net_dev, "dpni_set_offload(TX_L4_CSUM) failed\n");
1673 /* Perform a single release command to add buffers
1674 * to the specified buffer pool
1676 static int dpaa2_eth_add_bufs(struct dpaa2_eth_priv *priv,
1677 struct dpaa2_eth_channel *ch)
1679 struct xdp_buff *xdp_buffs[DPAA2_ETH_BUFS_PER_CMD];
1680 struct device *dev = priv->net_dev->dev.parent;
1681 u64 buf_array[DPAA2_ETH_BUFS_PER_CMD];
1682 struct dpaa2_eth_swa *swa;
1689 /* Allocate buffers visible to WRIOP */
1691 for (i = 0; i < DPAA2_ETH_BUFS_PER_CMD; i++) {
1692 /* Also allocate skb shared info and alignment padding.
1693 * There is one page for each Rx buffer. WRIOP sees
1694 * the entire page except for a tailroom reserved for
1697 page = dev_alloc_pages(0);
1701 addr = dma_map_page(dev, page, 0, priv->rx_buf_size,
1703 if (unlikely(dma_mapping_error(dev, addr)))
1706 buf_array[i] = addr;
1709 trace_dpaa2_eth_buf_seed(priv->net_dev,
1711 DPAA2_ETH_RX_BUF_RAW_SIZE,
1712 addr, priv->rx_buf_size,
1715 } else if (xsk_buff_can_alloc(ch->xsk_pool, DPAA2_ETH_BUFS_PER_CMD)) {
1716 /* Allocate XSK buffers for AF_XDP fast path in batches
1717 * of DPAA2_ETH_BUFS_PER_CMD. Bail out if the UMEM cannot
1718 * provide enough buffers at the moment
1720 batch = xsk_buff_alloc_batch(ch->xsk_pool, xdp_buffs,
1721 DPAA2_ETH_BUFS_PER_CMD);
1725 for (i = 0; i < batch; i++) {
1726 swa = (struct dpaa2_eth_swa *)(xdp_buffs[i]->data_hard_start +
1727 DPAA2_ETH_RX_HWA_SIZE);
1728 swa->xsk.xdp_buff = xdp_buffs[i];
1730 addr = xsk_buff_xdp_get_frame_dma(xdp_buffs[i]);
1731 if (unlikely(dma_mapping_error(dev, addr)))
1734 buf_array[i] = addr;
1736 trace_dpaa2_xsk_buf_seed(priv->net_dev,
1737 xdp_buffs[i]->data_hard_start,
1738 DPAA2_ETH_RX_BUF_RAW_SIZE,
1739 addr, priv->rx_buf_size,
1745 /* In case the portal is busy, retry until successful */
1746 while ((err = dpaa2_io_service_release(ch->dpio, ch->bp->bpid,
1747 buf_array, i)) == -EBUSY) {
1748 if (retries++ >= DPAA2_ETH_SWP_BUSY_RETRIES)
1753 /* If release command failed, clean up and bail out;
1754 * not much else we can do about it
1757 dpaa2_eth_free_bufs(priv, buf_array, i, ch->xsk_zc);
1765 __free_pages(page, 0);
1767 for (; i < batch; i++)
1768 xsk_buff_free(xdp_buffs[i]);
1771 /* If we managed to allocate at least some buffers,
1772 * release them to hardware
1780 static int dpaa2_eth_seed_pool(struct dpaa2_eth_priv *priv,
1781 struct dpaa2_eth_channel *ch)
1786 for (i = 0; i < DPAA2_ETH_NUM_BUFS; i += DPAA2_ETH_BUFS_PER_CMD) {
1787 new_count = dpaa2_eth_add_bufs(priv, ch);
1788 ch->buf_count += new_count;
1790 if (new_count < DPAA2_ETH_BUFS_PER_CMD)
1797 static void dpaa2_eth_seed_pools(struct dpaa2_eth_priv *priv)
1799 struct net_device *net_dev = priv->net_dev;
1800 struct dpaa2_eth_channel *channel;
1803 for (i = 0; i < priv->num_channels; i++) {
1804 channel = priv->channel[i];
1806 err = dpaa2_eth_seed_pool(priv, channel);
1808 /* Not much to do; the buffer pool, though not filled up,
1809 * may still contain some buffers which would enable us
1813 netdev_err(net_dev, "Buffer seeding failed for DPBP %d (bpid=%d)\n",
1814 channel->bp->dev->obj_desc.id,
1820 * Drain the specified number of buffers from one of the DPNI's private buffer
1822 * @count must not exceeed DPAA2_ETH_BUFS_PER_CMD
1824 static void dpaa2_eth_drain_bufs(struct dpaa2_eth_priv *priv, int bpid,
1827 u64 buf_array[DPAA2_ETH_BUFS_PER_CMD];
1828 bool xsk_zc = false;
1832 for (i = 0; i < priv->num_channels; i++)
1833 if (priv->channel[i]->bp->bpid == bpid)
1834 xsk_zc = priv->channel[i]->xsk_zc;
1837 ret = dpaa2_io_service_acquire(NULL, bpid, buf_array, count);
1839 if (ret == -EBUSY &&
1840 retries++ < DPAA2_ETH_SWP_BUSY_RETRIES)
1842 netdev_err(priv->net_dev, "dpaa2_io_service_acquire() failed\n");
1845 dpaa2_eth_free_bufs(priv, buf_array, ret, xsk_zc);
1850 static void dpaa2_eth_drain_pool(struct dpaa2_eth_priv *priv, int bpid)
1854 /* Drain the buffer pool */
1855 dpaa2_eth_drain_bufs(priv, bpid, DPAA2_ETH_BUFS_PER_CMD);
1856 dpaa2_eth_drain_bufs(priv, bpid, 1);
1858 /* Setup to zero the buffer count of all channels which were
1859 * using this buffer pool.
1861 for (i = 0; i < priv->num_channels; i++)
1862 if (priv->channel[i]->bp->bpid == bpid)
1863 priv->channel[i]->buf_count = 0;
1866 static void dpaa2_eth_drain_pools(struct dpaa2_eth_priv *priv)
1870 for (i = 0; i < priv->num_bps; i++)
1871 dpaa2_eth_drain_pool(priv, priv->bp[i]->bpid);
1874 /* Function is called from softirq context only, so we don't need to guard
1875 * the access to percpu count
1877 static int dpaa2_eth_refill_pool(struct dpaa2_eth_priv *priv,
1878 struct dpaa2_eth_channel *ch)
1882 if (likely(ch->buf_count >= DPAA2_ETH_REFILL_THRESH))
1886 new_count = dpaa2_eth_add_bufs(priv, ch);
1887 if (unlikely(!new_count)) {
1888 /* Out of memory; abort for now, we'll try later on */
1891 ch->buf_count += new_count;
1892 } while (ch->buf_count < DPAA2_ETH_NUM_BUFS);
1894 if (unlikely(ch->buf_count < DPAA2_ETH_NUM_BUFS))
1900 static void dpaa2_eth_sgt_cache_drain(struct dpaa2_eth_priv *priv)
1902 struct dpaa2_eth_sgt_cache *sgt_cache;
1906 for_each_possible_cpu(k) {
1907 sgt_cache = per_cpu_ptr(priv->sgt_cache, k);
1908 count = sgt_cache->count;
1910 for (i = 0; i < count; i++)
1911 skb_free_frag(sgt_cache->buf[i]);
1912 sgt_cache->count = 0;
1916 static int dpaa2_eth_pull_channel(struct dpaa2_eth_channel *ch)
1921 /* Retry while portal is busy */
1923 err = dpaa2_io_service_pull_channel(ch->dpio, ch->ch_id,
1927 } while (err == -EBUSY && dequeues < DPAA2_ETH_SWP_BUSY_RETRIES);
1929 ch->stats.dequeue_portal_busy += dequeues;
1931 ch->stats.pull_err++;
1936 /* NAPI poll routine
1938 * Frames are dequeued from the QMan channel associated with this NAPI context.
1939 * Rx, Tx confirmation and (if configured) Rx error frames all count
1940 * towards the NAPI budget.
1942 static int dpaa2_eth_poll(struct napi_struct *napi, int budget)
1944 struct dpaa2_eth_channel *ch;
1945 struct dpaa2_eth_priv *priv;
1946 int rx_cleaned = 0, txconf_cleaned = 0;
1947 struct dpaa2_eth_fq *fq, *txc_fq = NULL;
1948 struct netdev_queue *nq;
1949 int store_cleaned, work_done;
1950 bool work_done_zc = false;
1951 struct list_head rx_list;
1956 ch = container_of(napi, struct dpaa2_eth_channel, napi);
1960 INIT_LIST_HEAD(&rx_list);
1961 ch->rx_list = &rx_list;
1964 work_done_zc = dpaa2_xsk_tx(priv, ch);
1965 /* If we reached the XSK Tx per NAPI threshold, we're done */
1973 err = dpaa2_eth_pull_channel(ch);
1977 /* Refill pool if appropriate */
1978 dpaa2_eth_refill_pool(priv, ch);
1980 store_cleaned = dpaa2_eth_consume_frames(ch, &fq);
1981 if (store_cleaned <= 0)
1983 if (fq->type == DPAA2_RX_FQ) {
1984 rx_cleaned += store_cleaned;
1985 flowid = fq->flowid;
1987 txconf_cleaned += store_cleaned;
1988 /* We have a single Tx conf FQ on this channel */
1992 /* If we either consumed the whole NAPI budget with Rx frames
1993 * or we reached the Tx confirmations threshold, we're done.
1995 if (rx_cleaned >= budget ||
1996 txconf_cleaned >= DPAA2_ETH_TXCONF_PER_NAPI) {
1998 if (ch->xdp.res & XDP_REDIRECT)
2002 } while (store_cleaned);
2004 if (ch->xdp.res & XDP_REDIRECT)
2007 /* Update NET DIM with the values for this CDAN */
2008 dpaa2_io_update_net_dim(ch->dpio, ch->stats.frames_per_cdan,
2009 ch->stats.bytes_per_cdan);
2010 ch->stats.frames_per_cdan = 0;
2011 ch->stats.bytes_per_cdan = 0;
2013 /* We didn't consume the entire budget, so finish napi and
2014 * re-enable data availability notifications
2016 napi_complete_done(napi, rx_cleaned);
2018 err = dpaa2_io_service_rearm(ch->dpio, &ch->nctx);
2020 } while (err == -EBUSY && retries++ < DPAA2_ETH_SWP_BUSY_RETRIES);
2021 WARN_ONCE(err, "CDAN notifications rearm failed on core %d",
2022 ch->nctx.desired_cpu);
2024 work_done = max(rx_cleaned, 1);
2027 netif_receive_skb_list(ch->rx_list);
2029 if (ch->xsk_tx_pkts_sent) {
2030 xsk_tx_completed(ch->xsk_pool, ch->xsk_tx_pkts_sent);
2031 ch->xsk_tx_pkts_sent = 0;
2034 if (txc_fq && txc_fq->dq_frames) {
2035 nq = netdev_get_tx_queue(priv->net_dev, txc_fq->flowid);
2036 netdev_tx_completed_queue(nq, txc_fq->dq_frames,
2038 txc_fq->dq_frames = 0;
2039 txc_fq->dq_bytes = 0;
2042 if (rx_cleaned && ch->xdp.res & XDP_TX)
2043 dpaa2_eth_xdp_tx_flush(priv, ch, &priv->fq[flowid]);
2048 static void dpaa2_eth_enable_ch_napi(struct dpaa2_eth_priv *priv)
2050 struct dpaa2_eth_channel *ch;
2053 for (i = 0; i < priv->num_channels; i++) {
2054 ch = priv->channel[i];
2055 napi_enable(&ch->napi);
2059 static void dpaa2_eth_disable_ch_napi(struct dpaa2_eth_priv *priv)
2061 struct dpaa2_eth_channel *ch;
2064 for (i = 0; i < priv->num_channels; i++) {
2065 ch = priv->channel[i];
2066 napi_disable(&ch->napi);
2070 void dpaa2_eth_set_rx_taildrop(struct dpaa2_eth_priv *priv,
2071 bool tx_pause, bool pfc)
2073 struct dpni_taildrop td = {0};
2074 struct dpaa2_eth_fq *fq;
2077 /* FQ taildrop: threshold is in bytes, per frame queue. Enabled if
2078 * flow control is disabled (as it might interfere with either the
2079 * buffer pool depletion trigger for pause frames or with the group
2080 * congestion trigger for PFC frames)
2082 td.enable = !tx_pause;
2083 if (priv->rx_fqtd_enabled == td.enable)
2086 td.threshold = DPAA2_ETH_FQ_TAILDROP_THRESH;
2087 td.units = DPNI_CONGESTION_UNIT_BYTES;
2089 for (i = 0; i < priv->num_fqs; i++) {
2091 if (fq->type != DPAA2_RX_FQ)
2093 err = dpni_set_taildrop(priv->mc_io, 0, priv->mc_token,
2094 DPNI_CP_QUEUE, DPNI_QUEUE_RX,
2095 fq->tc, fq->flowid, &td);
2097 netdev_err(priv->net_dev,
2098 "dpni_set_taildrop(FQ) failed\n");
2103 priv->rx_fqtd_enabled = td.enable;
2106 /* Congestion group taildrop: threshold is in frames, per group
2107 * of FQs belonging to the same traffic class
2108 * Enabled if general Tx pause disabled or if PFCs are enabled
2109 * (congestion group threhsold for PFC generation is lower than the
2110 * CG taildrop threshold, so it won't interfere with it; we also
2111 * want frames in non-PFC enabled traffic classes to be kept in check)
2113 td.enable = !tx_pause || pfc;
2114 if (priv->rx_cgtd_enabled == td.enable)
2117 td.threshold = DPAA2_ETH_CG_TAILDROP_THRESH(priv);
2118 td.units = DPNI_CONGESTION_UNIT_FRAMES;
2119 for (i = 0; i < dpaa2_eth_tc_count(priv); i++) {
2120 err = dpni_set_taildrop(priv->mc_io, 0, priv->mc_token,
2121 DPNI_CP_GROUP, DPNI_QUEUE_RX,
2124 netdev_err(priv->net_dev,
2125 "dpni_set_taildrop(CG) failed\n");
2130 priv->rx_cgtd_enabled = td.enable;
2133 static int dpaa2_eth_link_state_update(struct dpaa2_eth_priv *priv)
2135 struct dpni_link_state state = {0};
2139 err = dpni_get_link_state(priv->mc_io, 0, priv->mc_token, &state);
2140 if (unlikely(err)) {
2141 netdev_err(priv->net_dev,
2142 "dpni_get_link_state() failed\n");
2146 /* If Tx pause frame settings have changed, we need to update
2147 * Rx FQ taildrop configuration as well. We configure taildrop
2148 * only when pause frame generation is disabled.
2150 tx_pause = dpaa2_eth_tx_pause_enabled(state.options);
2151 dpaa2_eth_set_rx_taildrop(priv, tx_pause, priv->pfc_enabled);
2153 /* When we manage the MAC/PHY using phylink there is no need
2154 * to manually update the netif_carrier.
2155 * We can avoid locking because we are called from the "link changed"
2156 * IRQ handler, which is the same as the "endpoint changed" IRQ handler
2157 * (the writer to priv->mac), so we cannot race with it.
2159 if (dpaa2_mac_is_type_phy(priv->mac))
2162 /* Chech link state; speed / duplex changes are not treated yet */
2163 if (priv->link_state.up == state.up)
2167 netif_carrier_on(priv->net_dev);
2168 netif_tx_start_all_queues(priv->net_dev);
2170 netif_tx_stop_all_queues(priv->net_dev);
2171 netif_carrier_off(priv->net_dev);
2174 netdev_info(priv->net_dev, "Link Event: state %s\n",
2175 state.up ? "up" : "down");
2178 priv->link_state = state;
2183 static int dpaa2_eth_open(struct net_device *net_dev)
2185 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2188 dpaa2_eth_seed_pools(priv);
2190 mutex_lock(&priv->mac_lock);
2192 if (!dpaa2_eth_is_type_phy(priv)) {
2193 /* We'll only start the txqs when the link is actually ready;
2194 * make sure we don't race against the link up notification,
2195 * which may come immediately after dpni_enable();
2197 netif_tx_stop_all_queues(net_dev);
2199 /* Also, explicitly set carrier off, otherwise
2200 * netif_carrier_ok() will return true and cause 'ip link show'
2201 * to report the LOWER_UP flag, even though the link
2202 * notification wasn't even received.
2204 netif_carrier_off(net_dev);
2206 dpaa2_eth_enable_ch_napi(priv);
2208 err = dpni_enable(priv->mc_io, 0, priv->mc_token);
2210 mutex_unlock(&priv->mac_lock);
2211 netdev_err(net_dev, "dpni_enable() failed\n");
2215 if (dpaa2_eth_is_type_phy(priv))
2216 dpaa2_mac_start(priv->mac);
2218 mutex_unlock(&priv->mac_lock);
2223 dpaa2_eth_disable_ch_napi(priv);
2224 dpaa2_eth_drain_pools(priv);
2228 /* Total number of in-flight frames on ingress queues */
2229 static u32 dpaa2_eth_ingress_fq_count(struct dpaa2_eth_priv *priv)
2231 struct dpaa2_eth_fq *fq;
2232 u32 fcnt = 0, bcnt = 0, total = 0;
2235 for (i = 0; i < priv->num_fqs; i++) {
2237 err = dpaa2_io_query_fq_count(NULL, fq->fqid, &fcnt, &bcnt);
2239 netdev_warn(priv->net_dev, "query_fq_count failed");
2248 static void dpaa2_eth_wait_for_ingress_fq_empty(struct dpaa2_eth_priv *priv)
2254 pending = dpaa2_eth_ingress_fq_count(priv);
2257 } while (pending && --retries);
2260 #define DPNI_TX_PENDING_VER_MAJOR 7
2261 #define DPNI_TX_PENDING_VER_MINOR 13
2262 static void dpaa2_eth_wait_for_egress_fq_empty(struct dpaa2_eth_priv *priv)
2264 union dpni_statistics stats;
2268 if (dpaa2_eth_cmp_dpni_ver(priv, DPNI_TX_PENDING_VER_MAJOR,
2269 DPNI_TX_PENDING_VER_MINOR) < 0)
2273 err = dpni_get_statistics(priv->mc_io, 0, priv->mc_token, 6,
2277 if (stats.page_6.tx_pending_frames == 0)
2279 } while (--retries);
2285 static int dpaa2_eth_stop(struct net_device *net_dev)
2287 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2288 int dpni_enabled = 0;
2291 mutex_lock(&priv->mac_lock);
2293 if (dpaa2_eth_is_type_phy(priv)) {
2294 dpaa2_mac_stop(priv->mac);
2296 netif_tx_stop_all_queues(net_dev);
2297 netif_carrier_off(net_dev);
2300 mutex_unlock(&priv->mac_lock);
2302 /* On dpni_disable(), the MC firmware will:
2303 * - stop MAC Rx and wait for all Rx frames to be enqueued to software
2304 * - cut off WRIOP dequeues from egress FQs and wait until transmission
2305 * of all in flight Tx frames is finished (and corresponding Tx conf
2306 * frames are enqueued back to software)
2308 * Before calling dpni_disable(), we wait for all Tx frames to arrive
2309 * on WRIOP. After it finishes, wait until all remaining frames on Rx
2310 * and Tx conf queues are consumed on NAPI poll.
2312 dpaa2_eth_wait_for_egress_fq_empty(priv);
2315 dpni_disable(priv->mc_io, 0, priv->mc_token);
2316 dpni_is_enabled(priv->mc_io, 0, priv->mc_token, &dpni_enabled);
2318 /* Allow the hardware some slack */
2320 } while (dpni_enabled && --retries);
2322 netdev_warn(net_dev, "Retry count exceeded disabling DPNI\n");
2323 /* Must go on and disable NAPI nonetheless, so we don't crash at
2324 * the next "ifconfig up"
2328 dpaa2_eth_wait_for_ingress_fq_empty(priv);
2329 dpaa2_eth_disable_ch_napi(priv);
2331 /* Empty the buffer pool */
2332 dpaa2_eth_drain_pools(priv);
2334 /* Empty the Scatter-Gather Buffer cache */
2335 dpaa2_eth_sgt_cache_drain(priv);
2340 static int dpaa2_eth_set_addr(struct net_device *net_dev, void *addr)
2342 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2343 struct device *dev = net_dev->dev.parent;
2346 err = eth_mac_addr(net_dev, addr);
2348 dev_err(dev, "eth_mac_addr() failed (%d)\n", err);
2352 err = dpni_set_primary_mac_addr(priv->mc_io, 0, priv->mc_token,
2355 dev_err(dev, "dpni_set_primary_mac_addr() failed (%d)\n", err);
2362 /** Fill in counters maintained by the GPP driver. These may be different from
2363 * the hardware counters obtained by ethtool.
2365 static void dpaa2_eth_get_stats(struct net_device *net_dev,
2366 struct rtnl_link_stats64 *stats)
2368 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2369 struct rtnl_link_stats64 *percpu_stats;
2371 u64 *netstats = (u64 *)stats;
2373 int num = sizeof(struct rtnl_link_stats64) / sizeof(u64);
2375 for_each_possible_cpu(i) {
2376 percpu_stats = per_cpu_ptr(priv->percpu_stats, i);
2377 cpustats = (u64 *)percpu_stats;
2378 for (j = 0; j < num; j++)
2379 netstats[j] += cpustats[j];
2383 /* Copy mac unicast addresses from @net_dev to @priv.
2384 * Its sole purpose is to make dpaa2_eth_set_rx_mode() more readable.
2386 static void dpaa2_eth_add_uc_hw_addr(const struct net_device *net_dev,
2387 struct dpaa2_eth_priv *priv)
2389 struct netdev_hw_addr *ha;
2392 netdev_for_each_uc_addr(ha, net_dev) {
2393 err = dpni_add_mac_addr(priv->mc_io, 0, priv->mc_token,
2396 netdev_warn(priv->net_dev,
2397 "Could not add ucast MAC %pM to the filtering table (err %d)\n",
2402 /* Copy mac multicast addresses from @net_dev to @priv
2403 * Its sole purpose is to make dpaa2_eth_set_rx_mode() more readable.
2405 static void dpaa2_eth_add_mc_hw_addr(const struct net_device *net_dev,
2406 struct dpaa2_eth_priv *priv)
2408 struct netdev_hw_addr *ha;
2411 netdev_for_each_mc_addr(ha, net_dev) {
2412 err = dpni_add_mac_addr(priv->mc_io, 0, priv->mc_token,
2415 netdev_warn(priv->net_dev,
2416 "Could not add mcast MAC %pM to the filtering table (err %d)\n",
2421 static int dpaa2_eth_rx_add_vid(struct net_device *net_dev,
2422 __be16 vlan_proto, u16 vid)
2424 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2427 err = dpni_add_vlan_id(priv->mc_io, 0, priv->mc_token,
2431 netdev_warn(priv->net_dev,
2432 "Could not add the vlan id %u\n",
2440 static int dpaa2_eth_rx_kill_vid(struct net_device *net_dev,
2441 __be16 vlan_proto, u16 vid)
2443 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2446 err = dpni_remove_vlan_id(priv->mc_io, 0, priv->mc_token, vid);
2449 netdev_warn(priv->net_dev,
2450 "Could not remove the vlan id %u\n",
2458 static void dpaa2_eth_set_rx_mode(struct net_device *net_dev)
2460 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2461 int uc_count = netdev_uc_count(net_dev);
2462 int mc_count = netdev_mc_count(net_dev);
2463 u8 max_mac = priv->dpni_attrs.mac_filter_entries;
2464 u32 options = priv->dpni_attrs.options;
2465 u16 mc_token = priv->mc_token;
2466 struct fsl_mc_io *mc_io = priv->mc_io;
2469 /* Basic sanity checks; these probably indicate a misconfiguration */
2470 if (options & DPNI_OPT_NO_MAC_FILTER && max_mac != 0)
2471 netdev_info(net_dev,
2472 "mac_filter_entries=%d, DPNI_OPT_NO_MAC_FILTER option must be disabled\n",
2475 /* Force promiscuous if the uc or mc counts exceed our capabilities. */
2476 if (uc_count > max_mac) {
2477 netdev_info(net_dev,
2478 "Unicast addr count reached %d, max allowed is %d; forcing promisc\n",
2482 if (mc_count + uc_count > max_mac) {
2483 netdev_info(net_dev,
2484 "Unicast + multicast addr count reached %d, max allowed is %d; forcing promisc\n",
2485 uc_count + mc_count, max_mac);
2486 goto force_mc_promisc;
2489 /* Adjust promisc settings due to flag combinations */
2490 if (net_dev->flags & IFF_PROMISC)
2492 if (net_dev->flags & IFF_ALLMULTI) {
2493 /* First, rebuild unicast filtering table. This should be done
2494 * in promisc mode, in order to avoid frame loss while we
2495 * progressively add entries to the table.
2496 * We don't know whether we had been in promisc already, and
2497 * making an MC call to find out is expensive; so set uc promisc
2500 err = dpni_set_unicast_promisc(mc_io, 0, mc_token, 1);
2502 netdev_warn(net_dev, "Can't set uc promisc\n");
2504 /* Actual uc table reconstruction. */
2505 err = dpni_clear_mac_filters(mc_io, 0, mc_token, 1, 0);
2507 netdev_warn(net_dev, "Can't clear uc filters\n");
2508 dpaa2_eth_add_uc_hw_addr(net_dev, priv);
2510 /* Finally, clear uc promisc and set mc promisc as requested. */
2511 err = dpni_set_unicast_promisc(mc_io, 0, mc_token, 0);
2513 netdev_warn(net_dev, "Can't clear uc promisc\n");
2514 goto force_mc_promisc;
2517 /* Neither unicast, nor multicast promisc will be on... eventually.
2518 * For now, rebuild mac filtering tables while forcing both of them on.
2520 err = dpni_set_unicast_promisc(mc_io, 0, mc_token, 1);
2522 netdev_warn(net_dev, "Can't set uc promisc (%d)\n", err);
2523 err = dpni_set_multicast_promisc(mc_io, 0, mc_token, 1);
2525 netdev_warn(net_dev, "Can't set mc promisc (%d)\n", err);
2527 /* Actual mac filtering tables reconstruction */
2528 err = dpni_clear_mac_filters(mc_io, 0, mc_token, 1, 1);
2530 netdev_warn(net_dev, "Can't clear mac filters\n");
2531 dpaa2_eth_add_mc_hw_addr(net_dev, priv);
2532 dpaa2_eth_add_uc_hw_addr(net_dev, priv);
2534 /* Now we can clear both ucast and mcast promisc, without risking
2535 * to drop legitimate frames anymore.
2537 err = dpni_set_unicast_promisc(mc_io, 0, mc_token, 0);
2539 netdev_warn(net_dev, "Can't clear ucast promisc\n");
2540 err = dpni_set_multicast_promisc(mc_io, 0, mc_token, 0);
2542 netdev_warn(net_dev, "Can't clear mcast promisc\n");
2547 err = dpni_set_unicast_promisc(mc_io, 0, mc_token, 1);
2549 netdev_warn(net_dev, "Can't set ucast promisc\n");
2551 err = dpni_set_multicast_promisc(mc_io, 0, mc_token, 1);
2553 netdev_warn(net_dev, "Can't set mcast promisc\n");
2556 static int dpaa2_eth_set_features(struct net_device *net_dev,
2557 netdev_features_t features)
2559 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2560 netdev_features_t changed = features ^ net_dev->features;
2564 if (changed & NETIF_F_HW_VLAN_CTAG_FILTER) {
2565 enable = !!(features & NETIF_F_HW_VLAN_CTAG_FILTER);
2566 err = dpaa2_eth_set_rx_vlan_filtering(priv, enable);
2571 if (changed & NETIF_F_RXCSUM) {
2572 enable = !!(features & NETIF_F_RXCSUM);
2573 err = dpaa2_eth_set_rx_csum(priv, enable);
2578 if (changed & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
2579 enable = !!(features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM));
2580 err = dpaa2_eth_set_tx_csum(priv, enable);
2588 static int dpaa2_eth_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
2590 struct dpaa2_eth_priv *priv = netdev_priv(dev);
2591 struct hwtstamp_config config;
2596 if (copy_from_user(&config, rq->ifr_data, sizeof(config)))
2599 switch (config.tx_type) {
2600 case HWTSTAMP_TX_OFF:
2601 case HWTSTAMP_TX_ON:
2602 case HWTSTAMP_TX_ONESTEP_SYNC:
2603 priv->tx_tstamp_type = config.tx_type;
2609 if (config.rx_filter == HWTSTAMP_FILTER_NONE) {
2610 priv->rx_tstamp = false;
2612 priv->rx_tstamp = true;
2613 /* TS is set for all frame types, not only those requested */
2614 config.rx_filter = HWTSTAMP_FILTER_ALL;
2617 if (priv->tx_tstamp_type == HWTSTAMP_TX_ONESTEP_SYNC)
2618 dpaa2_ptp_onestep_reg_update_method(priv);
2620 return copy_to_user(rq->ifr_data, &config, sizeof(config)) ?
2624 static int dpaa2_eth_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
2626 struct dpaa2_eth_priv *priv = netdev_priv(dev);
2629 if (cmd == SIOCSHWTSTAMP)
2630 return dpaa2_eth_ts_ioctl(dev, rq, cmd);
2632 mutex_lock(&priv->mac_lock);
2634 if (dpaa2_eth_is_type_phy(priv)) {
2635 err = phylink_mii_ioctl(priv->mac->phylink, rq, cmd);
2636 mutex_unlock(&priv->mac_lock);
2640 mutex_unlock(&priv->mac_lock);
2645 static bool xdp_mtu_valid(struct dpaa2_eth_priv *priv, int mtu)
2647 int mfl, linear_mfl;
2649 mfl = DPAA2_ETH_L2_MAX_FRM(mtu);
2650 linear_mfl = priv->rx_buf_size - DPAA2_ETH_RX_HWA_SIZE -
2651 dpaa2_eth_rx_head_room(priv) - XDP_PACKET_HEADROOM;
2653 if (mfl > linear_mfl) {
2654 netdev_warn(priv->net_dev, "Maximum MTU for XDP is %d\n",
2655 linear_mfl - VLAN_ETH_HLEN);
2662 static int dpaa2_eth_set_rx_mfl(struct dpaa2_eth_priv *priv, int mtu, bool has_xdp)
2666 /* We enforce a maximum Rx frame length based on MTU only if we have
2667 * an XDP program attached (in order to avoid Rx S/G frames).
2668 * Otherwise, we accept all incoming frames as long as they are not
2669 * larger than maximum size supported in hardware
2672 mfl = DPAA2_ETH_L2_MAX_FRM(mtu);
2674 mfl = DPAA2_ETH_MFL;
2676 err = dpni_set_max_frame_length(priv->mc_io, 0, priv->mc_token, mfl);
2678 netdev_err(priv->net_dev, "dpni_set_max_frame_length failed\n");
2685 static int dpaa2_eth_change_mtu(struct net_device *dev, int new_mtu)
2687 struct dpaa2_eth_priv *priv = netdev_priv(dev);
2690 if (!priv->xdp_prog)
2693 if (!xdp_mtu_valid(priv, new_mtu))
2696 err = dpaa2_eth_set_rx_mfl(priv, new_mtu, true);
2705 static int dpaa2_eth_update_rx_buffer_headroom(struct dpaa2_eth_priv *priv, bool has_xdp)
2707 struct dpni_buffer_layout buf_layout = {0};
2710 err = dpni_get_buffer_layout(priv->mc_io, 0, priv->mc_token,
2711 DPNI_QUEUE_RX, &buf_layout);
2713 netdev_err(priv->net_dev, "dpni_get_buffer_layout failed\n");
2717 /* Reserve extra headroom for XDP header size changes */
2718 buf_layout.data_head_room = dpaa2_eth_rx_head_room(priv) +
2719 (has_xdp ? XDP_PACKET_HEADROOM : 0);
2720 buf_layout.options = DPNI_BUF_LAYOUT_OPT_DATA_HEAD_ROOM;
2721 err = dpni_set_buffer_layout(priv->mc_io, 0, priv->mc_token,
2722 DPNI_QUEUE_RX, &buf_layout);
2724 netdev_err(priv->net_dev, "dpni_set_buffer_layout failed\n");
2731 static int dpaa2_eth_setup_xdp(struct net_device *dev, struct bpf_prog *prog)
2733 struct dpaa2_eth_priv *priv = netdev_priv(dev);
2734 struct dpaa2_eth_channel *ch;
2735 struct bpf_prog *old;
2736 bool up, need_update;
2739 if (prog && !xdp_mtu_valid(priv, dev->mtu))
2743 bpf_prog_add(prog, priv->num_channels);
2745 up = netif_running(dev);
2746 need_update = (!!priv->xdp_prog != !!prog);
2751 /* While in xdp mode, enforce a maximum Rx frame size based on MTU.
2752 * Also, when switching between xdp/non-xdp modes we need to reconfigure
2753 * our Rx buffer layout. Buffer pool was drained on dpaa2_eth_stop,
2754 * so we are sure no old format buffers will be used from now on.
2757 err = dpaa2_eth_set_rx_mfl(priv, dev->mtu, !!prog);
2760 err = dpaa2_eth_update_rx_buffer_headroom(priv, !!prog);
2765 old = xchg(&priv->xdp_prog, prog);
2769 for (i = 0; i < priv->num_channels; i++) {
2770 ch = priv->channel[i];
2771 old = xchg(&ch->xdp.prog, prog);
2777 err = dev_open(dev, NULL);
2786 bpf_prog_sub(prog, priv->num_channels);
2788 dev_open(dev, NULL);
2793 static int dpaa2_eth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2795 switch (xdp->command) {
2796 case XDP_SETUP_PROG:
2797 return dpaa2_eth_setup_xdp(dev, xdp->prog);
2798 case XDP_SETUP_XSK_POOL:
2799 return dpaa2_xsk_setup_pool(dev, xdp->xsk.pool, xdp->xsk.queue_id);
2807 static int dpaa2_eth_xdp_create_fd(struct net_device *net_dev,
2808 struct xdp_frame *xdpf,
2809 struct dpaa2_fd *fd)
2811 struct device *dev = net_dev->dev.parent;
2812 unsigned int needed_headroom;
2813 struct dpaa2_eth_swa *swa;
2814 void *buffer_start, *aligned_start;
2817 /* We require a minimum headroom to be able to transmit the frame.
2818 * Otherwise return an error and let the original net_device handle it
2820 needed_headroom = dpaa2_eth_needed_headroom(NULL);
2821 if (xdpf->headroom < needed_headroom)
2824 /* Setup the FD fields */
2825 memset(fd, 0, sizeof(*fd));
2827 /* Align FD address, if possible */
2828 buffer_start = xdpf->data - needed_headroom;
2829 aligned_start = PTR_ALIGN(buffer_start - DPAA2_ETH_TX_BUF_ALIGN,
2830 DPAA2_ETH_TX_BUF_ALIGN);
2831 if (aligned_start >= xdpf->data - xdpf->headroom)
2832 buffer_start = aligned_start;
2834 swa = (struct dpaa2_eth_swa *)buffer_start;
2835 /* fill in necessary fields here */
2836 swa->type = DPAA2_ETH_SWA_XDP;
2837 swa->xdp.dma_size = xdpf->data + xdpf->len - buffer_start;
2838 swa->xdp.xdpf = xdpf;
2840 addr = dma_map_single(dev, buffer_start,
2843 if (unlikely(dma_mapping_error(dev, addr)))
2846 dpaa2_fd_set_addr(fd, addr);
2847 dpaa2_fd_set_offset(fd, xdpf->data - buffer_start);
2848 dpaa2_fd_set_len(fd, xdpf->len);
2849 dpaa2_fd_set_format(fd, dpaa2_fd_single);
2850 dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
2855 static int dpaa2_eth_xdp_xmit(struct net_device *net_dev, int n,
2856 struct xdp_frame **frames, u32 flags)
2858 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2859 struct dpaa2_eth_xdp_fds *xdp_redirect_fds;
2860 struct rtnl_link_stats64 *percpu_stats;
2861 struct dpaa2_eth_fq *fq;
2862 struct dpaa2_fd *fds;
2863 int enqueued, i, err;
2865 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
2868 if (!netif_running(net_dev))
2871 fq = &priv->fq[smp_processor_id()];
2872 xdp_redirect_fds = &fq->xdp_redirect_fds;
2873 fds = xdp_redirect_fds->fds;
2875 percpu_stats = this_cpu_ptr(priv->percpu_stats);
2877 /* create a FD for each xdp_frame in the list received */
2878 for (i = 0; i < n; i++) {
2879 err = dpaa2_eth_xdp_create_fd(net_dev, frames[i], &fds[i]);
2883 xdp_redirect_fds->num = i;
2885 /* enqueue all the frame descriptors */
2886 enqueued = dpaa2_eth_xdp_flush(priv, fq, xdp_redirect_fds);
2888 /* update statistics */
2889 percpu_stats->tx_packets += enqueued;
2890 for (i = 0; i < enqueued; i++)
2891 percpu_stats->tx_bytes += dpaa2_fd_get_len(&fds[i]);
2896 static int update_xps(struct dpaa2_eth_priv *priv)
2898 struct net_device *net_dev = priv->net_dev;
2899 struct cpumask xps_mask;
2900 struct dpaa2_eth_fq *fq;
2901 int i, num_queues, netdev_queues;
2904 num_queues = dpaa2_eth_queue_count(priv);
2905 netdev_queues = (net_dev->num_tc ? : 1) * num_queues;
2907 /* The first <num_queues> entries in priv->fq array are Tx/Tx conf
2908 * queues, so only process those
2910 for (i = 0; i < netdev_queues; i++) {
2911 fq = &priv->fq[i % num_queues];
2913 cpumask_clear(&xps_mask);
2914 cpumask_set_cpu(fq->target_cpu, &xps_mask);
2916 err = netif_set_xps_queue(net_dev, &xps_mask, i);
2918 netdev_warn_once(net_dev, "Error setting XPS queue\n");
2926 static int dpaa2_eth_setup_mqprio(struct net_device *net_dev,
2927 struct tc_mqprio_qopt *mqprio)
2929 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2930 u8 num_tc, num_queues;
2933 mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
2934 num_queues = dpaa2_eth_queue_count(priv);
2935 num_tc = mqprio->num_tc;
2937 if (num_tc == net_dev->num_tc)
2940 if (num_tc > dpaa2_eth_tc_count(priv)) {
2941 netdev_err(net_dev, "Max %d traffic classes supported\n",
2942 dpaa2_eth_tc_count(priv));
2947 netdev_reset_tc(net_dev);
2948 netif_set_real_num_tx_queues(net_dev, num_queues);
2952 netdev_set_num_tc(net_dev, num_tc);
2953 netif_set_real_num_tx_queues(net_dev, num_tc * num_queues);
2955 for (i = 0; i < num_tc; i++)
2956 netdev_set_tc_queue(net_dev, i, num_queues, i * num_queues);
2964 #define bps_to_mbits(rate) (div_u64((rate), 1000000) * 8)
2966 static int dpaa2_eth_setup_tbf(struct net_device *net_dev, struct tc_tbf_qopt_offload *p)
2968 struct tc_tbf_qopt_offload_replace_params *cfg = &p->replace_params;
2969 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
2970 struct dpni_tx_shaping_cfg tx_cr_shaper = { 0 };
2971 struct dpni_tx_shaping_cfg tx_er_shaper = { 0 };
2974 if (p->command == TC_TBF_STATS)
2977 /* Only per port Tx shaping */
2978 if (p->parent != TC_H_ROOT)
2981 if (p->command == TC_TBF_REPLACE) {
2982 if (cfg->max_size > DPAA2_ETH_MAX_BURST_SIZE) {
2983 netdev_err(net_dev, "burst size cannot be greater than %d\n",
2984 DPAA2_ETH_MAX_BURST_SIZE);
2988 tx_cr_shaper.max_burst_size = cfg->max_size;
2989 /* The TBF interface is in bytes/s, whereas DPAA2 expects the
2992 tx_cr_shaper.rate_limit = bps_to_mbits(cfg->rate.rate_bytes_ps);
2995 err = dpni_set_tx_shaping(priv->mc_io, 0, priv->mc_token, &tx_cr_shaper,
2998 netdev_err(net_dev, "dpni_set_tx_shaping() = %d\n", err);
3005 static int dpaa2_eth_setup_tc(struct net_device *net_dev,
3006 enum tc_setup_type type, void *type_data)
3009 case TC_SETUP_QDISC_MQPRIO:
3010 return dpaa2_eth_setup_mqprio(net_dev, type_data);
3011 case TC_SETUP_QDISC_TBF:
3012 return dpaa2_eth_setup_tbf(net_dev, type_data);
3018 static const struct net_device_ops dpaa2_eth_ops = {
3019 .ndo_open = dpaa2_eth_open,
3020 .ndo_start_xmit = dpaa2_eth_tx,
3021 .ndo_stop = dpaa2_eth_stop,
3022 .ndo_set_mac_address = dpaa2_eth_set_addr,
3023 .ndo_get_stats64 = dpaa2_eth_get_stats,
3024 .ndo_set_rx_mode = dpaa2_eth_set_rx_mode,
3025 .ndo_set_features = dpaa2_eth_set_features,
3026 .ndo_eth_ioctl = dpaa2_eth_ioctl,
3027 .ndo_change_mtu = dpaa2_eth_change_mtu,
3028 .ndo_bpf = dpaa2_eth_xdp,
3029 .ndo_xdp_xmit = dpaa2_eth_xdp_xmit,
3030 .ndo_xsk_wakeup = dpaa2_xsk_wakeup,
3031 .ndo_setup_tc = dpaa2_eth_setup_tc,
3032 .ndo_vlan_rx_add_vid = dpaa2_eth_rx_add_vid,
3033 .ndo_vlan_rx_kill_vid = dpaa2_eth_rx_kill_vid
3036 static void dpaa2_eth_cdan_cb(struct dpaa2_io_notification_ctx *ctx)
3038 struct dpaa2_eth_channel *ch;
3040 ch = container_of(ctx, struct dpaa2_eth_channel, nctx);
3042 /* Update NAPI statistics */
3045 /* NAPI can also be scheduled from the AF_XDP Tx path. Mark a missed
3046 * so that it can be rescheduled again.
3048 if (!napi_if_scheduled_mark_missed(&ch->napi))
3049 napi_schedule(&ch->napi);
3052 /* Allocate and configure a DPCON object */
3053 static struct fsl_mc_device *dpaa2_eth_setup_dpcon(struct dpaa2_eth_priv *priv)
3055 struct fsl_mc_device *dpcon;
3056 struct device *dev = priv->net_dev->dev.parent;
3059 err = fsl_mc_object_allocate(to_fsl_mc_device(dev),
3060 FSL_MC_POOL_DPCON, &dpcon);
3062 if (err == -ENXIO) {
3063 dev_dbg(dev, "Waiting for DPCON\n");
3064 err = -EPROBE_DEFER;
3066 dev_info(dev, "Not enough DPCONs, will go on as-is\n");
3068 return ERR_PTR(err);
3071 err = dpcon_open(priv->mc_io, 0, dpcon->obj_desc.id, &dpcon->mc_handle);
3073 dev_err(dev, "dpcon_open() failed\n");
3077 err = dpcon_reset(priv->mc_io, 0, dpcon->mc_handle);
3079 dev_err(dev, "dpcon_reset() failed\n");
3083 err = dpcon_enable(priv->mc_io, 0, dpcon->mc_handle);
3085 dev_err(dev, "dpcon_enable() failed\n");
3092 dpcon_close(priv->mc_io, 0, dpcon->mc_handle);
3094 fsl_mc_object_free(dpcon);
3096 return ERR_PTR(err);
3099 static void dpaa2_eth_free_dpcon(struct dpaa2_eth_priv *priv,
3100 struct fsl_mc_device *dpcon)
3102 dpcon_disable(priv->mc_io, 0, dpcon->mc_handle);
3103 dpcon_close(priv->mc_io, 0, dpcon->mc_handle);
3104 fsl_mc_object_free(dpcon);
3107 static struct dpaa2_eth_channel *dpaa2_eth_alloc_channel(struct dpaa2_eth_priv *priv)
3109 struct dpaa2_eth_channel *channel;
3110 struct dpcon_attr attr;
3111 struct device *dev = priv->net_dev->dev.parent;
3114 channel = kzalloc(sizeof(*channel), GFP_KERNEL);
3118 channel->dpcon = dpaa2_eth_setup_dpcon(priv);
3119 if (IS_ERR(channel->dpcon)) {
3120 err = PTR_ERR(channel->dpcon);
3124 err = dpcon_get_attributes(priv->mc_io, 0, channel->dpcon->mc_handle,
3127 dev_err(dev, "dpcon_get_attributes() failed\n");
3131 channel->dpcon_id = attr.id;
3132 channel->ch_id = attr.qbman_ch_id;
3133 channel->priv = priv;
3138 dpaa2_eth_free_dpcon(priv, channel->dpcon);
3141 return ERR_PTR(err);
3144 static void dpaa2_eth_free_channel(struct dpaa2_eth_priv *priv,
3145 struct dpaa2_eth_channel *channel)
3147 dpaa2_eth_free_dpcon(priv, channel->dpcon);
3151 /* DPIO setup: allocate and configure QBMan channels, setup core affinity
3152 * and register data availability notifications
3154 static int dpaa2_eth_setup_dpio(struct dpaa2_eth_priv *priv)
3156 struct dpaa2_io_notification_ctx *nctx;
3157 struct dpaa2_eth_channel *channel;
3158 struct dpcon_notification_cfg dpcon_notif_cfg;
3159 struct device *dev = priv->net_dev->dev.parent;
3162 /* We want the ability to spread ingress traffic (RX, TX conf) to as
3163 * many cores as possible, so we need one channel for each core
3164 * (unless there's fewer queues than cores, in which case the extra
3165 * channels would be wasted).
3166 * Allocate one channel per core and register it to the core's
3167 * affine DPIO. If not enough channels are available for all cores
3168 * or if some cores don't have an affine DPIO, there will be no
3169 * ingress frame processing on those cores.
3171 cpumask_clear(&priv->dpio_cpumask);
3172 for_each_online_cpu(i) {
3173 /* Try to allocate a channel */
3174 channel = dpaa2_eth_alloc_channel(priv);
3175 if (IS_ERR_OR_NULL(channel)) {
3176 err = PTR_ERR_OR_ZERO(channel);
3177 if (err == -EPROBE_DEFER)
3178 dev_dbg(dev, "waiting for affine channel\n");
3181 "No affine channel for cpu %d and above\n", i);
3185 priv->channel[priv->num_channels] = channel;
3187 nctx = &channel->nctx;
3189 nctx->cb = dpaa2_eth_cdan_cb;
3190 nctx->id = channel->ch_id;
3191 nctx->desired_cpu = i;
3193 /* Register the new context */
3194 channel->dpio = dpaa2_io_service_select(i);
3195 err = dpaa2_io_service_register(channel->dpio, nctx, dev);
3197 dev_dbg(dev, "No affine DPIO for cpu %d\n", i);
3198 /* If no affine DPIO for this core, there's probably
3199 * none available for next cores either. Signal we want
3200 * to retry later, in case the DPIO devices weren't
3203 err = -EPROBE_DEFER;
3204 goto err_service_reg;
3207 /* Register DPCON notification with MC */
3208 dpcon_notif_cfg.dpio_id = nctx->dpio_id;
3209 dpcon_notif_cfg.priority = 0;
3210 dpcon_notif_cfg.user_ctx = nctx->qman64;
3211 err = dpcon_set_notification(priv->mc_io, 0,
3212 channel->dpcon->mc_handle,
3215 dev_err(dev, "dpcon_set_notification failed()\n");
3219 /* If we managed to allocate a channel and also found an affine
3220 * DPIO for this core, add it to the final mask
3222 cpumask_set_cpu(i, &priv->dpio_cpumask);
3223 priv->num_channels++;
3225 /* Stop if we already have enough channels to accommodate all
3226 * RX and TX conf queues
3228 if (priv->num_channels == priv->dpni_attrs.num_queues)
3235 dpaa2_io_service_deregister(channel->dpio, nctx, dev);
3237 dpaa2_eth_free_channel(priv, channel);
3239 if (err == -EPROBE_DEFER) {
3240 for (i = 0; i < priv->num_channels; i++) {
3241 channel = priv->channel[i];
3242 nctx = &channel->nctx;
3243 dpaa2_io_service_deregister(channel->dpio, nctx, dev);
3244 dpaa2_eth_free_channel(priv, channel);
3246 priv->num_channels = 0;
3250 if (cpumask_empty(&priv->dpio_cpumask)) {
3251 dev_err(dev, "No cpu with an affine DPIO/DPCON\n");
3255 dev_info(dev, "Cores %*pbl available for processing ingress traffic\n",
3256 cpumask_pr_args(&priv->dpio_cpumask));
3261 static void dpaa2_eth_free_dpio(struct dpaa2_eth_priv *priv)
3263 struct device *dev = priv->net_dev->dev.parent;
3264 struct dpaa2_eth_channel *ch;
3267 /* deregister CDAN notifications and free channels */
3268 for (i = 0; i < priv->num_channels; i++) {
3269 ch = priv->channel[i];
3270 dpaa2_io_service_deregister(ch->dpio, &ch->nctx, dev);
3271 dpaa2_eth_free_channel(priv, ch);
3275 static struct dpaa2_eth_channel *dpaa2_eth_get_affine_channel(struct dpaa2_eth_priv *priv,
3278 struct device *dev = priv->net_dev->dev.parent;
3281 for (i = 0; i < priv->num_channels; i++)
3282 if (priv->channel[i]->nctx.desired_cpu == cpu)
3283 return priv->channel[i];
3285 /* We should never get here. Issue a warning and return
3286 * the first channel, because it's still better than nothing
3288 dev_warn(dev, "No affine channel found for cpu %d\n", cpu);
3290 return priv->channel[0];
3293 static void dpaa2_eth_set_fq_affinity(struct dpaa2_eth_priv *priv)
3295 struct device *dev = priv->net_dev->dev.parent;
3296 struct dpaa2_eth_fq *fq;
3297 int rx_cpu, txc_cpu;
3300 /* For each FQ, pick one channel/CPU to deliver frames to.
3301 * This may well change at runtime, either through irqbalance or
3302 * through direct user intervention.
3304 rx_cpu = txc_cpu = cpumask_first(&priv->dpio_cpumask);
3306 for (i = 0; i < priv->num_fqs; i++) {
3310 case DPAA2_RX_ERR_FQ:
3311 fq->target_cpu = rx_cpu;
3312 rx_cpu = cpumask_next(rx_cpu, &priv->dpio_cpumask);
3313 if (rx_cpu >= nr_cpu_ids)
3314 rx_cpu = cpumask_first(&priv->dpio_cpumask);
3316 case DPAA2_TX_CONF_FQ:
3317 fq->target_cpu = txc_cpu;
3318 txc_cpu = cpumask_next(txc_cpu, &priv->dpio_cpumask);
3319 if (txc_cpu >= nr_cpu_ids)
3320 txc_cpu = cpumask_first(&priv->dpio_cpumask);
3323 dev_err(dev, "Unknown FQ type: %d\n", fq->type);
3325 fq->channel = dpaa2_eth_get_affine_channel(priv, fq->target_cpu);
3331 static void dpaa2_eth_setup_fqs(struct dpaa2_eth_priv *priv)
3335 /* We have one TxConf FQ per Tx flow.
3336 * The number of Tx and Rx queues is the same.
3337 * Tx queues come first in the fq array.
3339 for (i = 0; i < dpaa2_eth_queue_count(priv); i++) {
3340 priv->fq[priv->num_fqs].type = DPAA2_TX_CONF_FQ;
3341 priv->fq[priv->num_fqs].consume = dpaa2_eth_tx_conf;
3342 priv->fq[priv->num_fqs++].flowid = (u16)i;
3345 for (j = 0; j < dpaa2_eth_tc_count(priv); j++) {
3346 for (i = 0; i < dpaa2_eth_queue_count(priv); i++) {
3347 priv->fq[priv->num_fqs].type = DPAA2_RX_FQ;
3348 priv->fq[priv->num_fqs].consume = dpaa2_eth_rx;
3349 priv->fq[priv->num_fqs].tc = (u8)j;
3350 priv->fq[priv->num_fqs++].flowid = (u16)i;
3354 /* We have exactly one Rx error queue per DPNI */
3355 priv->fq[priv->num_fqs].type = DPAA2_RX_ERR_FQ;
3356 priv->fq[priv->num_fqs++].consume = dpaa2_eth_rx_err;
3358 /* For each FQ, decide on which core to process incoming frames */
3359 dpaa2_eth_set_fq_affinity(priv);
3362 /* Allocate and configure a buffer pool */
3363 struct dpaa2_eth_bp *dpaa2_eth_allocate_dpbp(struct dpaa2_eth_priv *priv)
3365 struct device *dev = priv->net_dev->dev.parent;
3366 struct fsl_mc_device *dpbp_dev;
3367 struct dpbp_attr dpbp_attrs;
3368 struct dpaa2_eth_bp *bp;
3371 err = fsl_mc_object_allocate(to_fsl_mc_device(dev), FSL_MC_POOL_DPBP,
3375 err = -EPROBE_DEFER;
3377 dev_err(dev, "DPBP device allocation failed\n");
3378 return ERR_PTR(err);
3381 bp = kzalloc(sizeof(*bp), GFP_KERNEL);
3387 err = dpbp_open(priv->mc_io, 0, dpbp_dev->obj_desc.id,
3388 &dpbp_dev->mc_handle);
3390 dev_err(dev, "dpbp_open() failed\n");
3394 err = dpbp_reset(priv->mc_io, 0, dpbp_dev->mc_handle);
3396 dev_err(dev, "dpbp_reset() failed\n");
3400 err = dpbp_enable(priv->mc_io, 0, dpbp_dev->mc_handle);
3402 dev_err(dev, "dpbp_enable() failed\n");
3406 err = dpbp_get_attributes(priv->mc_io, 0, dpbp_dev->mc_handle,
3409 dev_err(dev, "dpbp_get_attributes() failed\n");
3414 bp->bpid = dpbp_attrs.bpid;
3419 dpbp_disable(priv->mc_io, 0, dpbp_dev->mc_handle);
3422 dpbp_close(priv->mc_io, 0, dpbp_dev->mc_handle);
3426 fsl_mc_object_free(dpbp_dev);
3428 return ERR_PTR(err);
3431 static int dpaa2_eth_setup_default_dpbp(struct dpaa2_eth_priv *priv)
3433 struct dpaa2_eth_bp *bp;
3436 bp = dpaa2_eth_allocate_dpbp(priv);
3440 priv->bp[DPAA2_ETH_DEFAULT_BP_IDX] = bp;
3443 for (i = 0; i < priv->num_channels; i++)
3444 priv->channel[i]->bp = bp;
3449 void dpaa2_eth_free_dpbp(struct dpaa2_eth_priv *priv, struct dpaa2_eth_bp *bp)
3453 /* Find the index at which this BP is stored */
3454 for (idx_bp = 0; idx_bp < priv->num_bps; idx_bp++)
3455 if (priv->bp[idx_bp] == bp)
3458 /* Drain the pool and disable the associated MC object */
3459 dpaa2_eth_drain_pool(priv, bp->bpid);
3460 dpbp_disable(priv->mc_io, 0, bp->dev->mc_handle);
3461 dpbp_close(priv->mc_io, 0, bp->dev->mc_handle);
3462 fsl_mc_object_free(bp->dev);
3465 /* Move the last in use DPBP over in this position */
3466 priv->bp[idx_bp] = priv->bp[priv->num_bps - 1];
3470 static void dpaa2_eth_free_dpbps(struct dpaa2_eth_priv *priv)
3474 for (i = 0; i < priv->num_bps; i++)
3475 dpaa2_eth_free_dpbp(priv, priv->bp[i]);
3478 static int dpaa2_eth_set_buffer_layout(struct dpaa2_eth_priv *priv)
3480 struct device *dev = priv->net_dev->dev.parent;
3481 struct dpni_buffer_layout buf_layout = {0};
3485 /* We need to check for WRIOP version 1.0.0, but depending on the MC
3486 * version, this number is not always provided correctly on rev1.
3487 * We need to check for both alternatives in this situation.
3489 if (priv->dpni_attrs.wriop_version == DPAA2_WRIOP_VERSION(0, 0, 0) ||
3490 priv->dpni_attrs.wriop_version == DPAA2_WRIOP_VERSION(1, 0, 0))
3491 rx_buf_align = DPAA2_ETH_RX_BUF_ALIGN_REV1;
3493 rx_buf_align = DPAA2_ETH_RX_BUF_ALIGN;
3495 /* We need to ensure that the buffer size seen by WRIOP is a multiple
3496 * of 64 or 256 bytes depending on the WRIOP version.
3498 priv->rx_buf_size = ALIGN_DOWN(DPAA2_ETH_RX_BUF_SIZE, rx_buf_align);
3501 buf_layout.private_data_size = DPAA2_ETH_SWA_SIZE;
3502 buf_layout.pass_timestamp = true;
3503 buf_layout.pass_frame_status = true;
3504 buf_layout.options = DPNI_BUF_LAYOUT_OPT_PRIVATE_DATA_SIZE |
3505 DPNI_BUF_LAYOUT_OPT_TIMESTAMP |
3506 DPNI_BUF_LAYOUT_OPT_FRAME_STATUS;
3507 err = dpni_set_buffer_layout(priv->mc_io, 0, priv->mc_token,
3508 DPNI_QUEUE_TX, &buf_layout);
3510 dev_err(dev, "dpni_set_buffer_layout(TX) failed\n");
3514 /* tx-confirm buffer */
3515 buf_layout.options = DPNI_BUF_LAYOUT_OPT_TIMESTAMP |
3516 DPNI_BUF_LAYOUT_OPT_FRAME_STATUS;
3517 err = dpni_set_buffer_layout(priv->mc_io, 0, priv->mc_token,
3518 DPNI_QUEUE_TX_CONFIRM, &buf_layout);
3520 dev_err(dev, "dpni_set_buffer_layout(TX_CONF) failed\n");
3524 /* Now that we've set our tx buffer layout, retrieve the minimum
3525 * required tx data offset.
3527 err = dpni_get_tx_data_offset(priv->mc_io, 0, priv->mc_token,
3528 &priv->tx_data_offset);
3530 dev_err(dev, "dpni_get_tx_data_offset() failed\n");
3534 if ((priv->tx_data_offset % 64) != 0)
3535 dev_warn(dev, "Tx data offset (%d) not a multiple of 64B\n",
3536 priv->tx_data_offset);
3539 buf_layout.pass_frame_status = true;
3540 buf_layout.pass_parser_result = true;
3541 buf_layout.data_align = rx_buf_align;
3542 buf_layout.data_head_room = dpaa2_eth_rx_head_room(priv);
3543 buf_layout.private_data_size = 0;
3544 buf_layout.options = DPNI_BUF_LAYOUT_OPT_PARSER_RESULT |
3545 DPNI_BUF_LAYOUT_OPT_FRAME_STATUS |
3546 DPNI_BUF_LAYOUT_OPT_DATA_ALIGN |
3547 DPNI_BUF_LAYOUT_OPT_DATA_HEAD_ROOM |
3548 DPNI_BUF_LAYOUT_OPT_TIMESTAMP;
3549 err = dpni_set_buffer_layout(priv->mc_io, 0, priv->mc_token,
3550 DPNI_QUEUE_RX, &buf_layout);
3552 dev_err(dev, "dpni_set_buffer_layout(RX) failed\n");
3559 #define DPNI_ENQUEUE_FQID_VER_MAJOR 7
3560 #define DPNI_ENQUEUE_FQID_VER_MINOR 9
3562 static inline int dpaa2_eth_enqueue_qd(struct dpaa2_eth_priv *priv,
3563 struct dpaa2_eth_fq *fq,
3564 struct dpaa2_fd *fd, u8 prio,
3565 u32 num_frames __always_unused,
3566 int *frames_enqueued)
3570 err = dpaa2_io_service_enqueue_qd(fq->channel->dpio,
3571 priv->tx_qdid, prio,
3573 if (!err && frames_enqueued)
3574 *frames_enqueued = 1;
3578 static inline int dpaa2_eth_enqueue_fq_multiple(struct dpaa2_eth_priv *priv,
3579 struct dpaa2_eth_fq *fq,
3580 struct dpaa2_fd *fd,
3581 u8 prio, u32 num_frames,
3582 int *frames_enqueued)
3586 err = dpaa2_io_service_enqueue_multiple_fq(fq->channel->dpio,
3593 if (frames_enqueued)
3594 *frames_enqueued = err;
3598 static void dpaa2_eth_set_enqueue_mode(struct dpaa2_eth_priv *priv)
3600 if (dpaa2_eth_cmp_dpni_ver(priv, DPNI_ENQUEUE_FQID_VER_MAJOR,
3601 DPNI_ENQUEUE_FQID_VER_MINOR) < 0)
3602 priv->enqueue = dpaa2_eth_enqueue_qd;
3604 priv->enqueue = dpaa2_eth_enqueue_fq_multiple;
3607 static int dpaa2_eth_set_pause(struct dpaa2_eth_priv *priv)
3609 struct device *dev = priv->net_dev->dev.parent;
3610 struct dpni_link_cfg link_cfg = {0};
3613 /* Get the default link options so we don't override other flags */
3614 err = dpni_get_link_cfg(priv->mc_io, 0, priv->mc_token, &link_cfg);
3616 dev_err(dev, "dpni_get_link_cfg() failed\n");
3620 /* By default, enable both Rx and Tx pause frames */
3621 link_cfg.options |= DPNI_LINK_OPT_PAUSE;
3622 link_cfg.options &= ~DPNI_LINK_OPT_ASYM_PAUSE;
3623 err = dpni_set_link_cfg(priv->mc_io, 0, priv->mc_token, &link_cfg);
3625 dev_err(dev, "dpni_set_link_cfg() failed\n");
3629 priv->link_state.options = link_cfg.options;
3634 static void dpaa2_eth_update_tx_fqids(struct dpaa2_eth_priv *priv)
3636 struct dpni_queue_id qid = {0};
3637 struct dpaa2_eth_fq *fq;
3638 struct dpni_queue queue;
3641 /* We only use Tx FQIDs for FQID-based enqueue, so check
3642 * if DPNI version supports it before updating FQIDs
3644 if (dpaa2_eth_cmp_dpni_ver(priv, DPNI_ENQUEUE_FQID_VER_MAJOR,
3645 DPNI_ENQUEUE_FQID_VER_MINOR) < 0)
3648 for (i = 0; i < priv->num_fqs; i++) {
3650 if (fq->type != DPAA2_TX_CONF_FQ)
3652 for (j = 0; j < dpaa2_eth_tc_count(priv); j++) {
3653 err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
3654 DPNI_QUEUE_TX, j, fq->flowid,
3659 fq->tx_fqid[j] = qid.fqid;
3660 if (fq->tx_fqid[j] == 0)
3665 priv->enqueue = dpaa2_eth_enqueue_fq_multiple;
3670 netdev_info(priv->net_dev,
3671 "Error reading Tx FQID, fallback to QDID-based enqueue\n");
3672 priv->enqueue = dpaa2_eth_enqueue_qd;
3675 /* Configure ingress classification based on VLAN PCP */
3676 static int dpaa2_eth_set_vlan_qos(struct dpaa2_eth_priv *priv)
3678 struct device *dev = priv->net_dev->dev.parent;
3679 struct dpkg_profile_cfg kg_cfg = {0};
3680 struct dpni_qos_tbl_cfg qos_cfg = {0};
3681 struct dpni_rule_cfg key_params;
3682 void *dma_mem, *key, *mask;
3683 u8 key_size = 2; /* VLAN TCI field */
3686 /* VLAN-based classification only makes sense if we have multiple
3688 * Also, we need to extract just the 3-bit PCP field from the VLAN
3689 * header and we can only do that by using a mask
3691 if (dpaa2_eth_tc_count(priv) == 1 || !dpaa2_eth_fs_mask_enabled(priv)) {
3692 dev_dbg(dev, "VLAN-based QoS classification not supported\n");
3696 dma_mem = kzalloc(DPAA2_CLASSIFIER_DMA_SIZE, GFP_KERNEL);
3700 kg_cfg.num_extracts = 1;
3701 kg_cfg.extracts[0].type = DPKG_EXTRACT_FROM_HDR;
3702 kg_cfg.extracts[0].extract.from_hdr.prot = NET_PROT_VLAN;
3703 kg_cfg.extracts[0].extract.from_hdr.type = DPKG_FULL_FIELD;
3704 kg_cfg.extracts[0].extract.from_hdr.field = NH_FLD_VLAN_TCI;
3706 err = dpni_prepare_key_cfg(&kg_cfg, dma_mem);
3708 dev_err(dev, "dpni_prepare_key_cfg failed\n");
3713 qos_cfg.default_tc = 0;
3714 qos_cfg.discard_on_miss = 0;
3715 qos_cfg.key_cfg_iova = dma_map_single(dev, dma_mem,
3716 DPAA2_CLASSIFIER_DMA_SIZE,
3718 if (dma_mapping_error(dev, qos_cfg.key_cfg_iova)) {
3719 dev_err(dev, "QoS table DMA mapping failed\n");
3724 err = dpni_set_qos_table(priv->mc_io, 0, priv->mc_token, &qos_cfg);
3726 dev_err(dev, "dpni_set_qos_table failed\n");
3730 /* Add QoS table entries */
3731 key = kzalloc(key_size * 2, GFP_KERNEL);
3736 mask = key + key_size;
3737 *(__be16 *)mask = cpu_to_be16(VLAN_PRIO_MASK);
3739 key_params.key_iova = dma_map_single(dev, key, key_size * 2,
3741 if (dma_mapping_error(dev, key_params.key_iova)) {
3742 dev_err(dev, "Qos table entry DMA mapping failed\n");
3747 key_params.mask_iova = key_params.key_iova + key_size;
3748 key_params.key_size = key_size;
3750 /* We add rules for PCP-based distribution starting with highest
3751 * priority (VLAN PCP = 7). If this DPNI doesn't have enough traffic
3752 * classes to accommodate all priority levels, the lowest ones end up
3753 * on TC 0 which was configured as default
3755 for (i = dpaa2_eth_tc_count(priv) - 1, pcp = 7; i >= 0; i--, pcp--) {
3756 *(__be16 *)key = cpu_to_be16(pcp << VLAN_PRIO_SHIFT);
3757 dma_sync_single_for_device(dev, key_params.key_iova,
3758 key_size * 2, DMA_TO_DEVICE);
3760 err = dpni_add_qos_entry(priv->mc_io, 0, priv->mc_token,
3763 dev_err(dev, "dpni_add_qos_entry failed\n");
3764 dpni_clear_qos_table(priv->mc_io, 0, priv->mc_token);
3769 priv->vlan_cls_enabled = true;
3771 /* Table and key memory is not persistent, clean everything up after
3772 * configuration is finished
3775 dma_unmap_single(dev, key_params.key_iova, key_size * 2, DMA_TO_DEVICE);
3779 dma_unmap_single(dev, qos_cfg.key_cfg_iova, DPAA2_CLASSIFIER_DMA_SIZE,
3787 /* Configure the DPNI object this interface is associated with */
3788 static int dpaa2_eth_setup_dpni(struct fsl_mc_device *ls_dev)
3790 struct device *dev = &ls_dev->dev;
3791 struct dpaa2_eth_priv *priv;
3792 struct net_device *net_dev;
3795 net_dev = dev_get_drvdata(dev);
3796 priv = netdev_priv(net_dev);
3798 /* get a handle for the DPNI object */
3799 err = dpni_open(priv->mc_io, 0, ls_dev->obj_desc.id, &priv->mc_token);
3801 dev_err(dev, "dpni_open() failed\n");
3805 /* Check if we can work with this DPNI object */
3806 err = dpni_get_api_version(priv->mc_io, 0, &priv->dpni_ver_major,
3807 &priv->dpni_ver_minor);
3809 dev_err(dev, "dpni_get_api_version() failed\n");
3812 if (dpaa2_eth_cmp_dpni_ver(priv, DPNI_VER_MAJOR, DPNI_VER_MINOR) < 0) {
3813 dev_err(dev, "DPNI version %u.%u not supported, need >= %u.%u\n",
3814 priv->dpni_ver_major, priv->dpni_ver_minor,
3815 DPNI_VER_MAJOR, DPNI_VER_MINOR);
3820 ls_dev->mc_io = priv->mc_io;
3821 ls_dev->mc_handle = priv->mc_token;
3823 err = dpni_reset(priv->mc_io, 0, priv->mc_token);
3825 dev_err(dev, "dpni_reset() failed\n");
3829 err = dpni_get_attributes(priv->mc_io, 0, priv->mc_token,
3832 dev_err(dev, "dpni_get_attributes() failed (err=%d)\n", err);
3836 err = dpaa2_eth_set_buffer_layout(priv);
3840 dpaa2_eth_set_enqueue_mode(priv);
3842 /* Enable pause frame support */
3843 if (dpaa2_eth_has_pause_support(priv)) {
3844 err = dpaa2_eth_set_pause(priv);
3849 err = dpaa2_eth_set_vlan_qos(priv);
3850 if (err && err != -EOPNOTSUPP)
3853 priv->cls_rules = devm_kcalloc(dev, dpaa2_eth_fs_count(priv),
3854 sizeof(struct dpaa2_eth_cls_rule),
3856 if (!priv->cls_rules) {
3864 dpni_close(priv->mc_io, 0, priv->mc_token);
3869 static void dpaa2_eth_free_dpni(struct dpaa2_eth_priv *priv)
3873 err = dpni_reset(priv->mc_io, 0, priv->mc_token);
3875 netdev_warn(priv->net_dev, "dpni_reset() failed (err %d)\n",
3878 dpni_close(priv->mc_io, 0, priv->mc_token);
3881 static int dpaa2_eth_setup_rx_flow(struct dpaa2_eth_priv *priv,
3882 struct dpaa2_eth_fq *fq)
3884 struct device *dev = priv->net_dev->dev.parent;
3885 struct dpni_queue queue;
3886 struct dpni_queue_id qid;
3889 err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
3890 DPNI_QUEUE_RX, fq->tc, fq->flowid, &queue, &qid);
3892 dev_err(dev, "dpni_get_queue(RX) failed\n");
3896 fq->fqid = qid.fqid;
3898 queue.destination.id = fq->channel->dpcon_id;
3899 queue.destination.type = DPNI_DEST_DPCON;
3900 queue.destination.priority = 1;
3901 queue.user_context = (u64)(uintptr_t)fq;
3902 err = dpni_set_queue(priv->mc_io, 0, priv->mc_token,
3903 DPNI_QUEUE_RX, fq->tc, fq->flowid,
3904 DPNI_QUEUE_OPT_USER_CTX | DPNI_QUEUE_OPT_DEST,
3907 dev_err(dev, "dpni_set_queue(RX) failed\n");
3912 /* only once for each channel */
3916 err = xdp_rxq_info_reg(&fq->channel->xdp_rxq, priv->net_dev,
3919 dev_err(dev, "xdp_rxq_info_reg failed\n");
3923 err = xdp_rxq_info_reg_mem_model(&fq->channel->xdp_rxq,
3924 MEM_TYPE_PAGE_ORDER0, NULL);
3926 dev_err(dev, "xdp_rxq_info_reg_mem_model failed\n");
3933 static int dpaa2_eth_setup_tx_flow(struct dpaa2_eth_priv *priv,
3934 struct dpaa2_eth_fq *fq)
3936 struct device *dev = priv->net_dev->dev.parent;
3937 struct dpni_queue queue;
3938 struct dpni_queue_id qid;
3941 for (i = 0; i < dpaa2_eth_tc_count(priv); i++) {
3942 err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
3943 DPNI_QUEUE_TX, i, fq->flowid,
3946 dev_err(dev, "dpni_get_queue(TX) failed\n");
3949 fq->tx_fqid[i] = qid.fqid;
3952 /* All Tx queues belonging to the same flowid have the same qdbin */
3953 fq->tx_qdbin = qid.qdbin;
3955 err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
3956 DPNI_QUEUE_TX_CONFIRM, 0, fq->flowid,
3959 dev_err(dev, "dpni_get_queue(TX_CONF) failed\n");
3963 fq->fqid = qid.fqid;
3965 queue.destination.id = fq->channel->dpcon_id;
3966 queue.destination.type = DPNI_DEST_DPCON;
3967 queue.destination.priority = 0;
3968 queue.user_context = (u64)(uintptr_t)fq;
3969 err = dpni_set_queue(priv->mc_io, 0, priv->mc_token,
3970 DPNI_QUEUE_TX_CONFIRM, 0, fq->flowid,
3971 DPNI_QUEUE_OPT_USER_CTX | DPNI_QUEUE_OPT_DEST,
3974 dev_err(dev, "dpni_set_queue(TX_CONF) failed\n");
3981 static int setup_rx_err_flow(struct dpaa2_eth_priv *priv,
3982 struct dpaa2_eth_fq *fq)
3984 struct device *dev = priv->net_dev->dev.parent;
3985 struct dpni_queue q = { { 0 } };
3986 struct dpni_queue_id qid;
3987 u8 q_opt = DPNI_QUEUE_OPT_USER_CTX | DPNI_QUEUE_OPT_DEST;
3990 err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
3991 DPNI_QUEUE_RX_ERR, 0, 0, &q, &qid);
3993 dev_err(dev, "dpni_get_queue() failed (%d)\n", err);
3997 fq->fqid = qid.fqid;
3999 q.destination.id = fq->channel->dpcon_id;
4000 q.destination.type = DPNI_DEST_DPCON;
4001 q.destination.priority = 1;
4002 q.user_context = (u64)(uintptr_t)fq;
4003 err = dpni_set_queue(priv->mc_io, 0, priv->mc_token,
4004 DPNI_QUEUE_RX_ERR, 0, 0, q_opt, &q);
4006 dev_err(dev, "dpni_set_queue() failed (%d)\n", err);
4013 /* Supported header fields for Rx hash distribution key */
4014 static const struct dpaa2_eth_dist_fields dist_fields[] = {
4017 .rxnfc_field = RXH_L2DA,
4018 .cls_prot = NET_PROT_ETH,
4019 .cls_field = NH_FLD_ETH_DA,
4020 .id = DPAA2_ETH_DIST_ETHDST,
4023 .cls_prot = NET_PROT_ETH,
4024 .cls_field = NH_FLD_ETH_SA,
4025 .id = DPAA2_ETH_DIST_ETHSRC,
4028 /* This is the last ethertype field parsed:
4029 * depending on frame format, it can be the MAC ethertype
4030 * or the VLAN etype.
4032 .cls_prot = NET_PROT_ETH,
4033 .cls_field = NH_FLD_ETH_TYPE,
4034 .id = DPAA2_ETH_DIST_ETHTYPE,
4038 .rxnfc_field = RXH_VLAN,
4039 .cls_prot = NET_PROT_VLAN,
4040 .cls_field = NH_FLD_VLAN_TCI,
4041 .id = DPAA2_ETH_DIST_VLAN,
4045 .rxnfc_field = RXH_IP_SRC,
4046 .cls_prot = NET_PROT_IP,
4047 .cls_field = NH_FLD_IP_SRC,
4048 .id = DPAA2_ETH_DIST_IPSRC,
4051 .rxnfc_field = RXH_IP_DST,
4052 .cls_prot = NET_PROT_IP,
4053 .cls_field = NH_FLD_IP_DST,
4054 .id = DPAA2_ETH_DIST_IPDST,
4057 .rxnfc_field = RXH_L3_PROTO,
4058 .cls_prot = NET_PROT_IP,
4059 .cls_field = NH_FLD_IP_PROTO,
4060 .id = DPAA2_ETH_DIST_IPPROTO,
4063 /* Using UDP ports, this is functionally equivalent to raw
4064 * byte pairs from L4 header.
4066 .rxnfc_field = RXH_L4_B_0_1,
4067 .cls_prot = NET_PROT_UDP,
4068 .cls_field = NH_FLD_UDP_PORT_SRC,
4069 .id = DPAA2_ETH_DIST_L4SRC,
4072 .rxnfc_field = RXH_L4_B_2_3,
4073 .cls_prot = NET_PROT_UDP,
4074 .cls_field = NH_FLD_UDP_PORT_DST,
4075 .id = DPAA2_ETH_DIST_L4DST,
4080 /* Configure the Rx hash key using the legacy API */
4081 static int dpaa2_eth_config_legacy_hash_key(struct dpaa2_eth_priv *priv, dma_addr_t key)
4083 struct device *dev = priv->net_dev->dev.parent;
4084 struct dpni_rx_tc_dist_cfg dist_cfg;
4087 memset(&dist_cfg, 0, sizeof(dist_cfg));
4089 dist_cfg.key_cfg_iova = key;
4090 dist_cfg.dist_size = dpaa2_eth_queue_count(priv);
4091 dist_cfg.dist_mode = DPNI_DIST_MODE_HASH;
4093 for (i = 0; i < dpaa2_eth_tc_count(priv); i++) {
4094 err = dpni_set_rx_tc_dist(priv->mc_io, 0, priv->mc_token,
4097 dev_err(dev, "dpni_set_rx_tc_dist failed\n");
4105 /* Configure the Rx hash key using the new API */
4106 static int dpaa2_eth_config_hash_key(struct dpaa2_eth_priv *priv, dma_addr_t key)
4108 struct device *dev = priv->net_dev->dev.parent;
4109 struct dpni_rx_dist_cfg dist_cfg;
4112 memset(&dist_cfg, 0, sizeof(dist_cfg));
4114 dist_cfg.key_cfg_iova = key;
4115 dist_cfg.dist_size = dpaa2_eth_queue_count(priv);
4116 dist_cfg.enable = 1;
4118 for (i = 0; i < dpaa2_eth_tc_count(priv); i++) {
4120 err = dpni_set_rx_hash_dist(priv->mc_io, 0, priv->mc_token,
4123 dev_err(dev, "dpni_set_rx_hash_dist failed\n");
4127 /* If the flow steering / hashing key is shared between all
4128 * traffic classes, install it just once
4130 if (priv->dpni_attrs.options & DPNI_OPT_SHARED_FS)
4137 /* Configure the Rx flow classification key */
4138 static int dpaa2_eth_config_cls_key(struct dpaa2_eth_priv *priv, dma_addr_t key)
4140 struct device *dev = priv->net_dev->dev.parent;
4141 struct dpni_rx_dist_cfg dist_cfg;
4144 memset(&dist_cfg, 0, sizeof(dist_cfg));
4146 dist_cfg.key_cfg_iova = key;
4147 dist_cfg.dist_size = dpaa2_eth_queue_count(priv);
4148 dist_cfg.enable = 1;
4150 for (i = 0; i < dpaa2_eth_tc_count(priv); i++) {
4152 err = dpni_set_rx_fs_dist(priv->mc_io, 0, priv->mc_token,
4155 dev_err(dev, "dpni_set_rx_fs_dist failed\n");
4159 /* If the flow steering / hashing key is shared between all
4160 * traffic classes, install it just once
4162 if (priv->dpni_attrs.options & DPNI_OPT_SHARED_FS)
4169 /* Size of the Rx flow classification key */
4170 int dpaa2_eth_cls_key_size(u64 fields)
4174 for (i = 0; i < ARRAY_SIZE(dist_fields); i++) {
4175 if (!(fields & dist_fields[i].id))
4177 size += dist_fields[i].size;
4183 /* Offset of header field in Rx classification key */
4184 int dpaa2_eth_cls_fld_off(int prot, int field)
4188 for (i = 0; i < ARRAY_SIZE(dist_fields); i++) {
4189 if (dist_fields[i].cls_prot == prot &&
4190 dist_fields[i].cls_field == field)
4192 off += dist_fields[i].size;
4195 WARN_ONCE(1, "Unsupported header field used for Rx flow cls\n");
4199 /* Prune unused fields from the classification rule.
4200 * Used when masking is not supported
4202 void dpaa2_eth_cls_trim_rule(void *key_mem, u64 fields)
4204 int off = 0, new_off = 0;
4207 for (i = 0; i < ARRAY_SIZE(dist_fields); i++) {
4208 size = dist_fields[i].size;
4209 if (dist_fields[i].id & fields) {
4210 memcpy(key_mem + new_off, key_mem + off, size);
4217 /* Set Rx distribution (hash or flow classification) key
4218 * flags is a combination of RXH_ bits
4220 static int dpaa2_eth_set_dist_key(struct net_device *net_dev,
4221 enum dpaa2_eth_rx_dist type, u64 flags)
4223 struct device *dev = net_dev->dev.parent;
4224 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
4225 struct dpkg_profile_cfg cls_cfg;
4226 u32 rx_hash_fields = 0;
4227 dma_addr_t key_iova;
4232 memset(&cls_cfg, 0, sizeof(cls_cfg));
4234 for (i = 0; i < ARRAY_SIZE(dist_fields); i++) {
4235 struct dpkg_extract *key =
4236 &cls_cfg.extracts[cls_cfg.num_extracts];
4238 /* For both Rx hashing and classification keys
4239 * we set only the selected fields.
4241 if (!(flags & dist_fields[i].id))
4243 if (type == DPAA2_ETH_RX_DIST_HASH)
4244 rx_hash_fields |= dist_fields[i].rxnfc_field;
4246 if (cls_cfg.num_extracts >= DPKG_MAX_NUM_OF_EXTRACTS) {
4247 dev_err(dev, "error adding key extraction rule, too many rules?\n");
4251 key->type = DPKG_EXTRACT_FROM_HDR;
4252 key->extract.from_hdr.prot = dist_fields[i].cls_prot;
4253 key->extract.from_hdr.type = DPKG_FULL_FIELD;
4254 key->extract.from_hdr.field = dist_fields[i].cls_field;
4255 cls_cfg.num_extracts++;
4258 dma_mem = kzalloc(DPAA2_CLASSIFIER_DMA_SIZE, GFP_KERNEL);
4262 err = dpni_prepare_key_cfg(&cls_cfg, dma_mem);
4264 dev_err(dev, "dpni_prepare_key_cfg error %d\n", err);
4268 /* Prepare for setting the rx dist */
4269 key_iova = dma_map_single(dev, dma_mem, DPAA2_CLASSIFIER_DMA_SIZE,
4271 if (dma_mapping_error(dev, key_iova)) {
4272 dev_err(dev, "DMA mapping failed\n");
4277 if (type == DPAA2_ETH_RX_DIST_HASH) {
4278 if (dpaa2_eth_has_legacy_dist(priv))
4279 err = dpaa2_eth_config_legacy_hash_key(priv, key_iova);
4281 err = dpaa2_eth_config_hash_key(priv, key_iova);
4283 err = dpaa2_eth_config_cls_key(priv, key_iova);
4286 dma_unmap_single(dev, key_iova, DPAA2_CLASSIFIER_DMA_SIZE,
4288 if (!err && type == DPAA2_ETH_RX_DIST_HASH)
4289 priv->rx_hash_fields = rx_hash_fields;
4296 int dpaa2_eth_set_hash(struct net_device *net_dev, u64 flags)
4298 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
4302 if (!dpaa2_eth_hash_enabled(priv))
4305 for (i = 0; i < ARRAY_SIZE(dist_fields); i++)
4306 if (dist_fields[i].rxnfc_field & flags)
4307 key |= dist_fields[i].id;
4309 return dpaa2_eth_set_dist_key(net_dev, DPAA2_ETH_RX_DIST_HASH, key);
4312 int dpaa2_eth_set_cls(struct net_device *net_dev, u64 flags)
4314 return dpaa2_eth_set_dist_key(net_dev, DPAA2_ETH_RX_DIST_CLS, flags);
4317 static int dpaa2_eth_set_default_cls(struct dpaa2_eth_priv *priv)
4319 struct device *dev = priv->net_dev->dev.parent;
4322 /* Check if we actually support Rx flow classification */
4323 if (dpaa2_eth_has_legacy_dist(priv)) {
4324 dev_dbg(dev, "Rx cls not supported by current MC version\n");
4328 if (!dpaa2_eth_fs_enabled(priv)) {
4329 dev_dbg(dev, "Rx cls disabled in DPNI options\n");
4333 if (!dpaa2_eth_hash_enabled(priv)) {
4334 dev_dbg(dev, "Rx cls disabled for single queue DPNIs\n");
4338 /* If there is no support for masking in the classification table,
4339 * we don't set a default key, as it will depend on the rules
4340 * added by the user at runtime.
4342 if (!dpaa2_eth_fs_mask_enabled(priv))
4345 err = dpaa2_eth_set_cls(priv->net_dev, DPAA2_ETH_DIST_ALL);
4350 priv->rx_cls_enabled = 1;
4355 /* Bind the DPNI to its needed objects and resources: buffer pool, DPIOs,
4356 * frame queues and channels
4358 static int dpaa2_eth_bind_dpni(struct dpaa2_eth_priv *priv)
4360 struct dpaa2_eth_bp *bp = priv->bp[DPAA2_ETH_DEFAULT_BP_IDX];
4361 struct net_device *net_dev = priv->net_dev;
4362 struct dpni_pools_cfg pools_params = { 0 };
4363 struct device *dev = net_dev->dev.parent;
4364 struct dpni_error_cfg err_cfg;
4368 pools_params.num_dpbp = 1;
4369 pools_params.pools[0].dpbp_id = bp->dev->obj_desc.id;
4370 pools_params.pools[0].backup_pool = 0;
4371 pools_params.pools[0].buffer_size = priv->rx_buf_size;
4372 err = dpni_set_pools(priv->mc_io, 0, priv->mc_token, &pools_params);
4374 dev_err(dev, "dpni_set_pools() failed\n");
4378 /* have the interface implicitly distribute traffic based on
4379 * the default hash key
4381 err = dpaa2_eth_set_hash(net_dev, DPAA2_RXH_DEFAULT);
4382 if (err && err != -EOPNOTSUPP)
4383 dev_err(dev, "Failed to configure hashing\n");
4385 /* Configure the flow classification key; it includes all
4386 * supported header fields and cannot be modified at runtime
4388 err = dpaa2_eth_set_default_cls(priv);
4389 if (err && err != -EOPNOTSUPP)
4390 dev_err(dev, "Failed to configure Rx classification key\n");
4392 /* Configure handling of error frames */
4393 err_cfg.errors = DPAA2_FAS_RX_ERR_MASK;
4394 err_cfg.set_frame_annotation = 1;
4395 err_cfg.error_action = DPNI_ERROR_ACTION_DISCARD;
4396 err = dpni_set_errors_behavior(priv->mc_io, 0, priv->mc_token,
4399 dev_err(dev, "dpni_set_errors_behavior failed\n");
4403 /* Configure Rx and Tx conf queues to generate CDANs */
4404 for (i = 0; i < priv->num_fqs; i++) {
4405 switch (priv->fq[i].type) {
4407 err = dpaa2_eth_setup_rx_flow(priv, &priv->fq[i]);
4409 case DPAA2_TX_CONF_FQ:
4410 err = dpaa2_eth_setup_tx_flow(priv, &priv->fq[i]);
4412 case DPAA2_RX_ERR_FQ:
4413 err = setup_rx_err_flow(priv, &priv->fq[i]);
4416 dev_err(dev, "Invalid FQ type %d\n", priv->fq[i].type);
4423 err = dpni_get_qdid(priv->mc_io, 0, priv->mc_token,
4424 DPNI_QUEUE_TX, &priv->tx_qdid);
4426 dev_err(dev, "dpni_get_qdid() failed\n");
4433 /* Allocate rings for storing incoming frame descriptors */
4434 static int dpaa2_eth_alloc_rings(struct dpaa2_eth_priv *priv)
4436 struct net_device *net_dev = priv->net_dev;
4437 struct device *dev = net_dev->dev.parent;
4440 for (i = 0; i < priv->num_channels; i++) {
4441 priv->channel[i]->store =
4442 dpaa2_io_store_create(DPAA2_ETH_STORE_SIZE, dev);
4443 if (!priv->channel[i]->store) {
4444 netdev_err(net_dev, "dpaa2_io_store_create() failed\n");
4452 for (i = 0; i < priv->num_channels; i++) {
4453 if (!priv->channel[i]->store)
4455 dpaa2_io_store_destroy(priv->channel[i]->store);
4461 static void dpaa2_eth_free_rings(struct dpaa2_eth_priv *priv)
4465 for (i = 0; i < priv->num_channels; i++)
4466 dpaa2_io_store_destroy(priv->channel[i]->store);
4469 static int dpaa2_eth_set_mac_addr(struct dpaa2_eth_priv *priv)
4471 struct net_device *net_dev = priv->net_dev;
4472 struct device *dev = net_dev->dev.parent;
4473 u8 mac_addr[ETH_ALEN], dpni_mac_addr[ETH_ALEN];
4476 /* Get firmware address, if any */
4477 err = dpni_get_port_mac_addr(priv->mc_io, 0, priv->mc_token, mac_addr);
4479 dev_err(dev, "dpni_get_port_mac_addr() failed\n");
4483 /* Get DPNI attributes address, if any */
4484 err = dpni_get_primary_mac_addr(priv->mc_io, 0, priv->mc_token,
4487 dev_err(dev, "dpni_get_primary_mac_addr() failed\n");
4491 /* First check if firmware has any address configured by bootloader */
4492 if (!is_zero_ether_addr(mac_addr)) {
4493 /* If the DPMAC addr != DPNI addr, update it */
4494 if (!ether_addr_equal(mac_addr, dpni_mac_addr)) {
4495 err = dpni_set_primary_mac_addr(priv->mc_io, 0,
4499 dev_err(dev, "dpni_set_primary_mac_addr() failed\n");
4503 eth_hw_addr_set(net_dev, mac_addr);
4504 } else if (is_zero_ether_addr(dpni_mac_addr)) {
4505 /* No MAC address configured, fill in net_dev->dev_addr
4508 eth_hw_addr_random(net_dev);
4509 dev_dbg_once(dev, "device(s) have all-zero hwaddr, replaced with random\n");
4511 err = dpni_set_primary_mac_addr(priv->mc_io, 0, priv->mc_token,
4514 dev_err(dev, "dpni_set_primary_mac_addr() failed\n");
4518 /* Override NET_ADDR_RANDOM set by eth_hw_addr_random(); for all
4519 * practical purposes, this will be our "permanent" mac address,
4520 * at least until the next reboot. This move will also permit
4521 * register_netdevice() to properly fill up net_dev->perm_addr.
4523 net_dev->addr_assign_type = NET_ADDR_PERM;
4525 /* NET_ADDR_PERM is default, all we have to do is
4526 * fill in the device addr.
4528 eth_hw_addr_set(net_dev, dpni_mac_addr);
4534 static int dpaa2_eth_netdev_init(struct net_device *net_dev)
4536 struct device *dev = net_dev->dev.parent;
4537 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
4538 u32 options = priv->dpni_attrs.options;
4539 u64 supported = 0, not_supported = 0;
4540 u8 bcast_addr[ETH_ALEN];
4544 net_dev->netdev_ops = &dpaa2_eth_ops;
4545 net_dev->ethtool_ops = &dpaa2_ethtool_ops;
4547 err = dpaa2_eth_set_mac_addr(priv);
4551 /* Explicitly add the broadcast address to the MAC filtering table */
4552 eth_broadcast_addr(bcast_addr);
4553 err = dpni_add_mac_addr(priv->mc_io, 0, priv->mc_token, bcast_addr);
4555 dev_err(dev, "dpni_add_mac_addr() failed\n");
4559 /* Set MTU upper limit; lower limit is 68B (default value) */
4560 net_dev->max_mtu = DPAA2_ETH_MAX_MTU;
4561 err = dpni_set_max_frame_length(priv->mc_io, 0, priv->mc_token,
4564 dev_err(dev, "dpni_set_max_frame_length() failed\n");
4568 /* Set actual number of queues in the net device */
4569 num_queues = dpaa2_eth_queue_count(priv);
4570 err = netif_set_real_num_tx_queues(net_dev, num_queues);
4572 dev_err(dev, "netif_set_real_num_tx_queues() failed\n");
4575 err = netif_set_real_num_rx_queues(net_dev, num_queues);
4577 dev_err(dev, "netif_set_real_num_rx_queues() failed\n");
4581 dpaa2_eth_detect_features(priv);
4583 /* Capabilities listing */
4584 supported |= IFF_LIVE_ADDR_CHANGE;
4586 if (options & DPNI_OPT_NO_MAC_FILTER)
4587 not_supported |= IFF_UNICAST_FLT;
4589 supported |= IFF_UNICAST_FLT;
4591 net_dev->priv_flags |= supported;
4592 net_dev->priv_flags &= ~not_supported;
4595 net_dev->features = NETIF_F_RXCSUM |
4596 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
4597 NETIF_F_SG | NETIF_F_HIGHDMA |
4598 NETIF_F_LLTX | NETIF_F_HW_TC | NETIF_F_TSO;
4599 net_dev->gso_max_segs = DPAA2_ETH_ENQUEUE_MAX_FDS;
4600 net_dev->hw_features = net_dev->features;
4601 net_dev->xdp_features = NETDEV_XDP_ACT_BASIC |
4602 NETDEV_XDP_ACT_REDIRECT |
4603 NETDEV_XDP_ACT_NDO_XMIT;
4604 if (priv->dpni_attrs.wriop_version >= DPAA2_WRIOP_VERSION(3, 0, 0) &&
4605 priv->dpni_attrs.num_queues <= 8)
4606 net_dev->xdp_features |= NETDEV_XDP_ACT_XSK_ZEROCOPY;
4608 if (priv->dpni_attrs.vlan_filter_entries)
4609 net_dev->hw_features |= NETIF_F_HW_VLAN_CTAG_FILTER;
4614 static int dpaa2_eth_poll_link_state(void *arg)
4616 struct dpaa2_eth_priv *priv = (struct dpaa2_eth_priv *)arg;
4619 while (!kthread_should_stop()) {
4620 err = dpaa2_eth_link_state_update(priv);
4624 msleep(DPAA2_ETH_LINK_STATE_REFRESH);
4630 static int dpaa2_eth_connect_mac(struct dpaa2_eth_priv *priv)
4632 struct fsl_mc_device *dpni_dev, *dpmac_dev;
4633 struct dpaa2_mac *mac;
4636 dpni_dev = to_fsl_mc_device(priv->net_dev->dev.parent);
4637 dpmac_dev = fsl_mc_get_endpoint(dpni_dev, 0);
4639 if (PTR_ERR(dpmac_dev) == -EPROBE_DEFER) {
4640 netdev_dbg(priv->net_dev, "waiting for mac\n");
4641 return PTR_ERR(dpmac_dev);
4644 if (IS_ERR(dpmac_dev) || dpmac_dev->dev.type != &fsl_mc_bus_dpmac_type)
4647 mac = kzalloc(sizeof(struct dpaa2_mac), GFP_KERNEL);
4651 mac->mc_dev = dpmac_dev;
4652 mac->mc_io = priv->mc_io;
4653 mac->net_dev = priv->net_dev;
4655 err = dpaa2_mac_open(mac);
4659 if (dpaa2_mac_is_type_phy(mac)) {
4660 err = dpaa2_mac_connect(mac);
4662 if (err == -EPROBE_DEFER)
4663 netdev_dbg(priv->net_dev,
4664 "could not connect to MAC\n");
4666 netdev_err(priv->net_dev,
4667 "Error connecting to the MAC endpoint: %pe",
4673 mutex_lock(&priv->mac_lock);
4675 mutex_unlock(&priv->mac_lock);
4680 dpaa2_mac_close(mac);
4686 static void dpaa2_eth_disconnect_mac(struct dpaa2_eth_priv *priv)
4688 struct dpaa2_mac *mac;
4690 mutex_lock(&priv->mac_lock);
4693 mutex_unlock(&priv->mac_lock);
4698 if (dpaa2_mac_is_type_phy(mac))
4699 dpaa2_mac_disconnect(mac);
4701 dpaa2_mac_close(mac);
4705 static irqreturn_t dpni_irq0_handler_thread(int irq_num, void *arg)
4708 struct device *dev = (struct device *)arg;
4709 struct fsl_mc_device *dpni_dev = to_fsl_mc_device(dev);
4710 struct net_device *net_dev = dev_get_drvdata(dev);
4711 struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
4715 err = dpni_get_irq_status(dpni_dev->mc_io, 0, dpni_dev->mc_handle,
4716 DPNI_IRQ_INDEX, &status);
4717 if (unlikely(err)) {
4718 netdev_err(net_dev, "Can't get irq status (err %d)\n", err);
4722 if (status & DPNI_IRQ_EVENT_LINK_CHANGED)
4723 dpaa2_eth_link_state_update(netdev_priv(net_dev));
4725 if (status & DPNI_IRQ_EVENT_ENDPOINT_CHANGED) {
4726 dpaa2_eth_set_mac_addr(netdev_priv(net_dev));
4727 dpaa2_eth_update_tx_fqids(priv);
4729 /* We can avoid locking because the "endpoint changed" IRQ
4730 * handler is the only one who changes priv->mac at runtime,
4731 * so we are not racing with anyone.
4733 had_mac = !!priv->mac;
4735 dpaa2_eth_disconnect_mac(priv);
4737 dpaa2_eth_connect_mac(priv);
4743 static int dpaa2_eth_setup_irqs(struct fsl_mc_device *ls_dev)
4746 struct fsl_mc_device_irq *irq;
4748 err = fsl_mc_allocate_irqs(ls_dev);
4750 dev_err(&ls_dev->dev, "MC irqs allocation failed\n");
4754 irq = ls_dev->irqs[0];
4755 err = devm_request_threaded_irq(&ls_dev->dev, irq->virq,
4756 NULL, dpni_irq0_handler_thread,
4757 IRQF_NO_SUSPEND | IRQF_ONESHOT,
4758 dev_name(&ls_dev->dev), &ls_dev->dev);
4760 dev_err(&ls_dev->dev, "devm_request_threaded_irq(): %d\n", err);
4764 err = dpni_set_irq_mask(ls_dev->mc_io, 0, ls_dev->mc_handle,
4765 DPNI_IRQ_INDEX, DPNI_IRQ_EVENT_LINK_CHANGED |
4766 DPNI_IRQ_EVENT_ENDPOINT_CHANGED);
4768 dev_err(&ls_dev->dev, "dpni_set_irq_mask(): %d\n", err);
4772 err = dpni_set_irq_enable(ls_dev->mc_io, 0, ls_dev->mc_handle,
4775 dev_err(&ls_dev->dev, "dpni_set_irq_enable(): %d\n", err);
4782 devm_free_irq(&ls_dev->dev, irq->virq, &ls_dev->dev);
4784 fsl_mc_free_irqs(ls_dev);
4789 static void dpaa2_eth_add_ch_napi(struct dpaa2_eth_priv *priv)
4792 struct dpaa2_eth_channel *ch;
4794 for (i = 0; i < priv->num_channels; i++) {
4795 ch = priv->channel[i];
4796 /* NAPI weight *MUST* be a multiple of DPAA2_ETH_STORE_SIZE */
4797 netif_napi_add(priv->net_dev, &ch->napi, dpaa2_eth_poll);
4801 static void dpaa2_eth_del_ch_napi(struct dpaa2_eth_priv *priv)
4804 struct dpaa2_eth_channel *ch;
4806 for (i = 0; i < priv->num_channels; i++) {
4807 ch = priv->channel[i];
4808 netif_napi_del(&ch->napi);
4812 static int dpaa2_eth_probe(struct fsl_mc_device *dpni_dev)
4815 struct net_device *net_dev = NULL;
4816 struct dpaa2_eth_priv *priv = NULL;
4819 dev = &dpni_dev->dev;
4822 net_dev = alloc_etherdev_mq(sizeof(*priv), DPAA2_ETH_MAX_NETDEV_QUEUES);
4824 dev_err(dev, "alloc_etherdev_mq() failed\n");
4828 SET_NETDEV_DEV(net_dev, dev);
4829 dev_set_drvdata(dev, net_dev);
4831 priv = netdev_priv(net_dev);
4832 priv->net_dev = net_dev;
4833 SET_NETDEV_DEVLINK_PORT(net_dev, &priv->devlink_port);
4835 mutex_init(&priv->mac_lock);
4837 priv->iommu_domain = iommu_get_domain_for_dev(dev);
4839 priv->tx_tstamp_type = HWTSTAMP_TX_OFF;
4840 priv->rx_tstamp = false;
4842 priv->dpaa2_ptp_wq = alloc_workqueue("dpaa2_ptp_wq", 0, 0);
4843 if (!priv->dpaa2_ptp_wq) {
4848 INIT_WORK(&priv->tx_onestep_tstamp, dpaa2_eth_tx_onestep_tstamp);
4849 mutex_init(&priv->onestep_tstamp_lock);
4850 skb_queue_head_init(&priv->tx_skbs);
4852 priv->rx_copybreak = DPAA2_ETH_DEFAULT_COPYBREAK;
4854 /* Obtain a MC portal */
4855 err = fsl_mc_portal_allocate(dpni_dev, FSL_MC_IO_ATOMIC_CONTEXT_PORTAL,
4858 if (err == -ENXIO) {
4859 dev_dbg(dev, "waiting for MC portal\n");
4860 err = -EPROBE_DEFER;
4862 dev_err(dev, "MC portal allocation failed\n");
4864 goto err_portal_alloc;
4867 /* MC objects initialization and configuration */
4868 err = dpaa2_eth_setup_dpni(dpni_dev);
4870 goto err_dpni_setup;
4872 err = dpaa2_eth_setup_dpio(priv);
4874 goto err_dpio_setup;
4876 dpaa2_eth_setup_fqs(priv);
4878 err = dpaa2_eth_setup_default_dpbp(priv);
4880 goto err_dpbp_setup;
4882 err = dpaa2_eth_bind_dpni(priv);
4886 /* Add a NAPI context for each channel */
4887 dpaa2_eth_add_ch_napi(priv);
4889 /* Percpu statistics */
4890 priv->percpu_stats = alloc_percpu(*priv->percpu_stats);
4891 if (!priv->percpu_stats) {
4892 dev_err(dev, "alloc_percpu(percpu_stats) failed\n");
4894 goto err_alloc_percpu_stats;
4896 priv->percpu_extras = alloc_percpu(*priv->percpu_extras);
4897 if (!priv->percpu_extras) {
4898 dev_err(dev, "alloc_percpu(percpu_extras) failed\n");
4900 goto err_alloc_percpu_extras;
4903 priv->sgt_cache = alloc_percpu(*priv->sgt_cache);
4904 if (!priv->sgt_cache) {
4905 dev_err(dev, "alloc_percpu(sgt_cache) failed\n");
4907 goto err_alloc_sgt_cache;
4910 priv->fd = alloc_percpu(*priv->fd);
4912 dev_err(dev, "alloc_percpu(fds) failed\n");
4917 err = dpaa2_eth_netdev_init(net_dev);
4919 goto err_netdev_init;
4921 /* Configure checksum offload based on current interface flags */
4922 err = dpaa2_eth_set_rx_csum(priv, !!(net_dev->features & NETIF_F_RXCSUM));
4926 err = dpaa2_eth_set_tx_csum(priv,
4927 !!(net_dev->features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)));
4931 err = dpaa2_eth_alloc_rings(priv);
4933 goto err_alloc_rings;
4935 #ifdef CONFIG_FSL_DPAA2_ETH_DCB
4936 if (dpaa2_eth_has_pause_support(priv) && priv->vlan_cls_enabled) {
4937 priv->dcbx_mode = DCB_CAP_DCBX_HOST | DCB_CAP_DCBX_VER_IEEE;
4938 net_dev->dcbnl_ops = &dpaa2_eth_dcbnl_ops;
4940 dev_dbg(dev, "PFC not supported\n");
4944 err = dpaa2_eth_connect_mac(priv);
4946 goto err_connect_mac;
4948 err = dpaa2_eth_setup_irqs(dpni_dev);
4950 netdev_warn(net_dev, "Failed to set link interrupt, fall back to polling\n");
4951 priv->poll_thread = kthread_run(dpaa2_eth_poll_link_state, priv,
4952 "%s_poll_link", net_dev->name);
4953 if (IS_ERR(priv->poll_thread)) {
4954 dev_err(dev, "Error starting polling thread\n");
4955 goto err_poll_thread;
4957 priv->do_link_poll = true;
4960 err = dpaa2_eth_dl_alloc(priv);
4962 goto err_dl_register;
4964 err = dpaa2_eth_dl_traps_register(priv);
4966 goto err_dl_trap_register;
4968 err = dpaa2_eth_dl_port_add(priv);
4970 goto err_dl_port_add;
4972 net_dev->needed_headroom = DPAA2_ETH_SWA_SIZE + DPAA2_ETH_TX_BUF_ALIGN;
4974 err = register_netdev(net_dev);
4976 dev_err(dev, "register_netdev() failed\n");
4977 goto err_netdev_reg;
4980 #ifdef CONFIG_DEBUG_FS
4981 dpaa2_dbg_add(priv);
4984 dpaa2_eth_dl_register(priv);
4985 dev_info(dev, "Probed interface %s\n", net_dev->name);
4989 dpaa2_eth_dl_port_del(priv);
4991 dpaa2_eth_dl_traps_unregister(priv);
4992 err_dl_trap_register:
4993 dpaa2_eth_dl_free(priv);
4995 if (priv->do_link_poll)
4996 kthread_stop(priv->poll_thread);
4998 fsl_mc_free_irqs(dpni_dev);
5000 dpaa2_eth_disconnect_mac(priv);
5002 dpaa2_eth_free_rings(priv);
5006 free_percpu(priv->fd);
5008 free_percpu(priv->sgt_cache);
5009 err_alloc_sgt_cache:
5010 free_percpu(priv->percpu_extras);
5011 err_alloc_percpu_extras:
5012 free_percpu(priv->percpu_stats);
5013 err_alloc_percpu_stats:
5014 dpaa2_eth_del_ch_napi(priv);
5016 dpaa2_eth_free_dpbps(priv);
5018 dpaa2_eth_free_dpio(priv);
5020 dpaa2_eth_free_dpni(priv);
5022 fsl_mc_portal_free(priv->mc_io);
5024 destroy_workqueue(priv->dpaa2_ptp_wq);
5026 dev_set_drvdata(dev, NULL);
5027 free_netdev(net_dev);
5032 static void dpaa2_eth_remove(struct fsl_mc_device *ls_dev)
5035 struct net_device *net_dev;
5036 struct dpaa2_eth_priv *priv;
5039 net_dev = dev_get_drvdata(dev);
5040 priv = netdev_priv(net_dev);
5042 dpaa2_eth_dl_unregister(priv);
5044 #ifdef CONFIG_DEBUG_FS
5045 dpaa2_dbg_remove(priv);
5048 unregister_netdev(net_dev);
5050 dpaa2_eth_dl_port_del(priv);
5051 dpaa2_eth_dl_traps_unregister(priv);
5052 dpaa2_eth_dl_free(priv);
5054 if (priv->do_link_poll)
5055 kthread_stop(priv->poll_thread);
5057 fsl_mc_free_irqs(ls_dev);
5059 dpaa2_eth_disconnect_mac(priv);
5060 dpaa2_eth_free_rings(priv);
5061 free_percpu(priv->fd);
5062 free_percpu(priv->sgt_cache);
5063 free_percpu(priv->percpu_stats);
5064 free_percpu(priv->percpu_extras);
5066 dpaa2_eth_del_ch_napi(priv);
5067 dpaa2_eth_free_dpbps(priv);
5068 dpaa2_eth_free_dpio(priv);
5069 dpaa2_eth_free_dpni(priv);
5070 if (priv->onestep_reg_base)
5071 iounmap(priv->onestep_reg_base);
5073 fsl_mc_portal_free(priv->mc_io);
5075 destroy_workqueue(priv->dpaa2_ptp_wq);
5077 dev_dbg(net_dev->dev.parent, "Removed interface %s\n", net_dev->name);
5079 free_netdev(net_dev);
5082 static const struct fsl_mc_device_id dpaa2_eth_match_id_table[] = {
5084 .vendor = FSL_MC_VENDOR_FREESCALE,
5089 MODULE_DEVICE_TABLE(fslmc, dpaa2_eth_match_id_table);
5091 static struct fsl_mc_driver dpaa2_eth_driver = {
5093 .name = KBUILD_MODNAME,
5095 .probe = dpaa2_eth_probe,
5096 .remove = dpaa2_eth_remove,
5097 .match_id_table = dpaa2_eth_match_id_table
5100 static int __init dpaa2_eth_driver_init(void)
5104 dpaa2_eth_dbg_init();
5105 err = fsl_mc_driver_register(&dpaa2_eth_driver);
5107 dpaa2_eth_dbg_exit();
5114 static void __exit dpaa2_eth_driver_exit(void)
5116 dpaa2_eth_dbg_exit();
5117 fsl_mc_driver_unregister(&dpaa2_eth_driver);
5120 module_init(dpaa2_eth_driver_init);
5121 module_exit(dpaa2_eth_driver_exit);