1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2018 Intel Corporation. */
4 #include <linux/bpf_trace.h>
5 #include <net/xdp_sock_drv.h>
6 #include "i40e_txrx_common.h"
9 void i40e_clear_rx_bi_zc(struct i40e_ring *rx_ring)
11 memset(rx_ring->rx_bi_zc, 0,
12 sizeof(*rx_ring->rx_bi_zc) * rx_ring->count);
15 static struct xdp_buff **i40e_rx_bi(struct i40e_ring *rx_ring, u32 idx)
17 return &rx_ring->rx_bi_zc[idx];
21 * i40e_realloc_rx_xdp_bi - reallocate SW ring for either XSK or normal buffer
22 * @rx_ring: Current rx ring
23 * @pool_present: is pool for XSK present
25 * Try allocating memory and return ENOMEM, if failed to allocate.
26 * If allocation was successful, substitute buffer with allocated one.
27 * Returns 0 on success, negative on failure
29 static int i40e_realloc_rx_xdp_bi(struct i40e_ring *rx_ring, bool pool_present)
31 size_t elem_size = pool_present ? sizeof(*rx_ring->rx_bi_zc) :
32 sizeof(*rx_ring->rx_bi);
33 void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
39 kfree(rx_ring->rx_bi);
40 rx_ring->rx_bi = NULL;
41 rx_ring->rx_bi_zc = sw_ring;
43 kfree(rx_ring->rx_bi_zc);
44 rx_ring->rx_bi_zc = NULL;
45 rx_ring->rx_bi = sw_ring;
51 * i40e_realloc_rx_bi_zc - reallocate rx SW rings
53 * @zc: is zero copy set
55 * Reallocate buffer for rx_rings that might be used by XSK.
56 * XDP requires more memory, than rx_buf provides.
57 * Returns 0 on success, negative on failure
59 int i40e_realloc_rx_bi_zc(struct i40e_vsi *vsi, bool zc)
61 struct i40e_ring *rx_ring;
64 for_each_set_bit(q, vsi->af_xdp_zc_qps, vsi->alloc_queue_pairs) {
65 rx_ring = vsi->rx_rings[q];
66 if (i40e_realloc_rx_xdp_bi(rx_ring, zc))
73 * i40e_xsk_pool_enable - Enable/associate an AF_XDP buffer pool to a
77 * @qid: Rx ring to associate buffer pool with
79 * Returns 0 on success, <0 on failure
81 static int i40e_xsk_pool_enable(struct i40e_vsi *vsi,
82 struct xsk_buff_pool *pool,
85 struct net_device *netdev = vsi->netdev;
89 if (vsi->type != I40E_VSI_MAIN)
92 if (qid >= vsi->num_queue_pairs)
95 if (qid >= netdev->real_num_rx_queues ||
96 qid >= netdev->real_num_tx_queues)
99 err = xsk_pool_dma_map(pool, &vsi->back->pdev->dev, I40E_RX_DMA_ATTR);
103 set_bit(qid, vsi->af_xdp_zc_qps);
105 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
108 err = i40e_queue_pair_disable(vsi, qid);
112 err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], true);
116 err = i40e_queue_pair_enable(vsi, qid);
120 /* Kick start the NAPI context so that receiving will start */
121 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
130 * i40e_xsk_pool_disable - Disassociate an AF_XDP buffer pool from a
133 * @qid: Rx ring to associate buffer pool with
135 * Returns 0 on success, <0 on failure
137 static int i40e_xsk_pool_disable(struct i40e_vsi *vsi, u16 qid)
139 struct net_device *netdev = vsi->netdev;
140 struct xsk_buff_pool *pool;
144 pool = xsk_get_pool_from_qid(netdev, qid);
148 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
151 err = i40e_queue_pair_disable(vsi, qid);
156 clear_bit(qid, vsi->af_xdp_zc_qps);
157 xsk_pool_dma_unmap(pool, I40E_RX_DMA_ATTR);
160 err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], false);
163 err = i40e_queue_pair_enable(vsi, qid);
172 * i40e_xsk_pool_setup - Enable/disassociate an AF_XDP buffer pool to/from
175 * @pool: Buffer pool to enable/associate to a ring, or NULL to disable
176 * @qid: Rx ring to (dis)associate buffer pool (from)to
178 * This function enables or disables a buffer pool to a certain ring.
180 * Returns 0 on success, <0 on failure
182 int i40e_xsk_pool_setup(struct i40e_vsi *vsi, struct xsk_buff_pool *pool,
185 return pool ? i40e_xsk_pool_enable(vsi, pool, qid) :
186 i40e_xsk_pool_disable(vsi, qid);
190 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
192 * @xdp: xdp_buff used as input to the XDP program
193 * @xdp_prog: XDP program to run
195 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
197 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp,
198 struct bpf_prog *xdp_prog)
200 int err, result = I40E_XDP_PASS;
201 struct i40e_ring *xdp_ring;
204 act = bpf_prog_run_xdp(xdp_prog, xdp);
206 if (likely(act == XDP_REDIRECT)) {
207 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
209 return I40E_XDP_REDIR;
210 if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
211 result = I40E_XDP_EXIT;
213 result = I40E_XDP_CONSUMED;
221 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
222 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
223 if (result == I40E_XDP_CONSUMED)
227 result = I40E_XDP_CONSUMED;
230 bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
233 result = I40E_XDP_CONSUMED;
235 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
240 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
242 u16 ntu = rx_ring->next_to_use;
243 union i40e_rx_desc *rx_desc;
244 struct xdp_buff **xdp;
248 rx_desc = I40E_RX_DESC(rx_ring, ntu);
249 xdp = i40e_rx_bi(rx_ring, ntu);
251 nb_buffs = min_t(u16, count, rx_ring->count - ntu);
252 nb_buffs = xsk_buff_alloc_batch(rx_ring->xsk_pool, xdp, nb_buffs);
258 dma = xsk_buff_xdp_get_dma(*xdp);
259 rx_desc->read.pkt_addr = cpu_to_le64(dma);
260 rx_desc->read.hdr_addr = 0;
267 if (ntu == rx_ring->count) {
268 rx_desc = I40E_RX_DESC(rx_ring, 0);
272 /* clear the status bits for the next_to_use descriptor */
273 rx_desc->wb.qword1.status_error_len = 0;
274 i40e_release_rx_desc(rx_ring, ntu);
276 return count == nb_buffs;
280 * i40e_construct_skb_zc - Create skbuff from zero-copy Rx buffer
284 * This functions allocates a new skb from a zero-copy Rx buffer.
286 * Returns the skb, or NULL on failure.
288 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
289 struct xdp_buff *xdp)
291 unsigned int totalsize = xdp->data_end - xdp->data_meta;
292 unsigned int metasize = xdp->data - xdp->data_meta;
293 struct skb_shared_info *sinfo = NULL;
297 if (unlikely(xdp_buff_has_frags(xdp))) {
298 sinfo = xdp_get_shared_info_from_buff(xdp);
299 nr_frags = sinfo->nr_frags;
301 net_prefetch(xdp->data_meta);
303 /* allocate a skb to store the frags */
304 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
305 GFP_ATOMIC | __GFP_NOWARN);
309 memcpy(__skb_put(skb, totalsize), xdp->data_meta,
310 ALIGN(totalsize, sizeof(long)));
313 skb_metadata_set(skb, metasize);
314 __skb_pull(skb, metasize);
317 if (likely(!xdp_buff_has_frags(xdp)))
320 for (int i = 0; i < nr_frags; i++) {
321 struct skb_shared_info *skinfo = skb_shinfo(skb);
322 skb_frag_t *frag = &sinfo->frags[i];
326 page = dev_alloc_page();
331 addr = page_to_virt(page);
333 memcpy(addr, skb_frag_page(frag), skb_frag_size(frag));
335 __skb_fill_page_desc_noacc(skinfo, skinfo->nr_frags++,
336 addr, 0, skb_frag_size(frag));
344 static void i40e_handle_xdp_result_zc(struct i40e_ring *rx_ring,
345 struct xdp_buff *xdp_buff,
346 union i40e_rx_desc *rx_desc,
347 unsigned int *rx_packets,
348 unsigned int *rx_bytes,
349 unsigned int xdp_res,
355 *rx_bytes = xdp_get_buff_len(xdp_buff);
357 if (likely(xdp_res == I40E_XDP_REDIR) || xdp_res == I40E_XDP_TX)
360 if (xdp_res == I40E_XDP_EXIT) {
365 if (xdp_res == I40E_XDP_CONSUMED) {
366 xsk_buff_free(xdp_buff);
369 if (xdp_res == I40E_XDP_PASS) {
370 /* NB! We are not checking for errors using
371 * i40e_test_staterr with
372 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
373 * SBP is *not* set in PRT_SBPVSI (default not set).
375 skb = i40e_construct_skb_zc(rx_ring, xdp_buff);
377 rx_ring->rx_stats.alloc_buff_failed++;
383 if (eth_skb_pad(skb)) {
389 i40e_process_skb_fields(rx_ring, rx_desc, skb);
390 napi_gro_receive(&rx_ring->q_vector->napi, skb);
394 /* Should never get here, as all valid cases have been handled already.
400 i40e_add_xsk_frag(struct i40e_ring *rx_ring, struct xdp_buff *first,
401 struct xdp_buff *xdp, const unsigned int size)
403 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(first);
405 if (!xdp_buff_has_frags(first)) {
407 sinfo->xdp_frags_size = 0;
408 xdp_buff_set_frags_flag(first);
411 if (unlikely(sinfo->nr_frags == MAX_SKB_FRAGS)) {
412 xsk_buff_free(first);
416 __skb_fill_page_desc_noacc(sinfo, sinfo->nr_frags++,
417 virt_to_page(xdp->data_hard_start),
418 XDP_PACKET_HEADROOM, size);
419 sinfo->xdp_frags_size += size;
420 xsk_buff_add_frag(xdp);
426 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
428 * @budget: NAPI budget
430 * Returns amount of work completed
432 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
434 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
435 u16 next_to_process = rx_ring->next_to_process;
436 u16 next_to_clean = rx_ring->next_to_clean;
437 unsigned int xdp_res, xdp_xmit = 0;
438 struct xdp_buff *first = NULL;
439 u32 count = rx_ring->count;
440 struct bpf_prog *xdp_prog;
441 u32 entries_to_alloc;
442 bool failure = false;
444 if (next_to_process != next_to_clean)
445 first = *i40e_rx_bi(rx_ring, next_to_clean);
447 /* NB! xdp_prog will always be !NULL, due to the fact that
448 * this path is enabled by setting an XDP program.
450 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
452 while (likely(total_rx_packets < (unsigned int)budget)) {
453 union i40e_rx_desc *rx_desc;
454 unsigned int rx_packets;
455 unsigned int rx_bytes;
460 rx_desc = I40E_RX_DESC(rx_ring, next_to_process);
461 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
463 /* This memory barrier is needed to keep us from reading
464 * any other fields out of the rx_desc until we have
465 * verified the descriptor has been written back.
469 if (i40e_rx_is_programming_status(qword)) {
470 i40e_clean_programming_status(rx_ring,
471 rx_desc->raw.qword[0],
473 bi = *i40e_rx_bi(rx_ring, next_to_process);
475 if (++next_to_process == count)
480 size = FIELD_GET(I40E_RXD_QW1_LENGTH_PBUF_MASK, qword);
484 bi = *i40e_rx_bi(rx_ring, next_to_process);
485 xsk_buff_set_size(bi, size);
486 xsk_buff_dma_sync_for_cpu(bi, rx_ring->xsk_pool);
490 else if (i40e_add_xsk_frag(rx_ring, first, bi, size))
493 if (++next_to_process == count)
496 if (i40e_is_non_eop(rx_ring, rx_desc))
499 xdp_res = i40e_run_xdp_zc(rx_ring, first, xdp_prog);
500 i40e_handle_xdp_result_zc(rx_ring, first, rx_desc, &rx_packets,
501 &rx_bytes, xdp_res, &failure);
502 next_to_clean = next_to_process;
505 total_rx_packets += rx_packets;
506 total_rx_bytes += rx_bytes;
507 xdp_xmit |= xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR);
511 rx_ring->next_to_clean = next_to_clean;
512 rx_ring->next_to_process = next_to_process;
514 entries_to_alloc = I40E_DESC_UNUSED(rx_ring);
515 if (entries_to_alloc >= I40E_RX_BUFFER_WRITE)
516 failure |= !i40e_alloc_rx_buffers_zc(rx_ring, entries_to_alloc);
518 i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
519 i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
521 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
522 if (failure || next_to_clean == rx_ring->next_to_use)
523 xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
525 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
527 return (int)total_rx_packets;
529 return failure ? budget : (int)total_rx_packets;
532 static void i40e_xmit_pkt(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
533 unsigned int *total_bytes)
535 u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(desc);
536 struct i40e_tx_desc *tx_desc;
539 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
540 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
542 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
543 tx_desc->buffer_addr = cpu_to_le64(dma);
544 tx_desc->cmd_type_offset_bsz = build_ctob(cmd, 0, desc->len, 0);
546 *total_bytes += desc->len;
549 static void i40e_xmit_pkt_batch(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
550 unsigned int *total_bytes)
552 u16 ntu = xdp_ring->next_to_use;
553 struct i40e_tx_desc *tx_desc;
557 loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
558 u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(&desc[i]);
560 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc[i].addr);
561 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc[i].len);
563 tx_desc = I40E_TX_DESC(xdp_ring, ntu++);
564 tx_desc->buffer_addr = cpu_to_le64(dma);
565 tx_desc->cmd_type_offset_bsz = build_ctob(cmd, 0, desc[i].len, 0);
567 *total_bytes += desc[i].len;
570 xdp_ring->next_to_use = ntu;
573 static void i40e_fill_tx_hw_ring(struct i40e_ring *xdp_ring, struct xdp_desc *descs, u32 nb_pkts,
574 unsigned int *total_bytes)
576 u32 batched, leftover, i;
578 batched = nb_pkts & ~(PKTS_PER_BATCH - 1);
579 leftover = nb_pkts & (PKTS_PER_BATCH - 1);
580 for (i = 0; i < batched; i += PKTS_PER_BATCH)
581 i40e_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
582 for (i = batched; i < batched + leftover; i++)
583 i40e_xmit_pkt(xdp_ring, &descs[i], total_bytes);
586 static void i40e_set_rs_bit(struct i40e_ring *xdp_ring)
588 u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1;
589 struct i40e_tx_desc *tx_desc;
591 tx_desc = I40E_TX_DESC(xdp_ring, ntu);
592 tx_desc->cmd_type_offset_bsz |= cpu_to_le64(I40E_TX_DESC_CMD_RS << I40E_TXD_QW1_CMD_SHIFT);
596 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
597 * @xdp_ring: XDP Tx ring
598 * @budget: NAPI budget
600 * Returns true if the work is finished.
602 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
604 struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
605 u32 nb_pkts, nb_processed = 0;
606 unsigned int total_bytes = 0;
608 nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
612 if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
613 nb_processed = xdp_ring->count - xdp_ring->next_to_use;
614 i40e_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
615 xdp_ring->next_to_use = 0;
618 i40e_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
621 /* Request an interrupt for the last frame and bump tail ptr. */
622 i40e_set_rs_bit(xdp_ring);
623 i40e_xdp_ring_update_tail(xdp_ring);
625 i40e_update_tx_stats(xdp_ring, nb_pkts, total_bytes);
627 return nb_pkts < budget;
631 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
632 * @tx_ring: XDP Tx ring
633 * @tx_bi: Tx buffer info to clean
635 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
636 struct i40e_tx_buffer *tx_bi)
638 xdp_return_frame(tx_bi->xdpf);
639 tx_ring->xdp_tx_active--;
640 dma_unmap_single(tx_ring->dev,
641 dma_unmap_addr(tx_bi, dma),
642 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
643 dma_unmap_len_set(tx_bi, len, 0);
647 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
649 * @tx_ring: XDP Tx ring
651 * Returns true if cleanup/transmission is done.
653 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring)
655 struct xsk_buff_pool *bp = tx_ring->xsk_pool;
656 u32 i, completed_frames, xsk_frames = 0;
657 u32 head_idx = i40e_get_head(tx_ring);
658 struct i40e_tx_buffer *tx_bi;
661 if (head_idx < tx_ring->next_to_clean)
662 head_idx += tx_ring->count;
663 completed_frames = head_idx - tx_ring->next_to_clean;
665 if (completed_frames == 0)
668 if (likely(!tx_ring->xdp_tx_active)) {
669 xsk_frames = completed_frames;
673 ntc = tx_ring->next_to_clean;
675 for (i = 0; i < completed_frames; i++) {
676 tx_bi = &tx_ring->tx_bi[ntc];
679 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
685 if (++ntc >= tx_ring->count)
690 tx_ring->next_to_clean += completed_frames;
691 if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
692 tx_ring->next_to_clean -= tx_ring->count;
695 xsk_tx_completed(bp, xsk_frames);
697 i40e_arm_wb(tx_ring, vsi, completed_frames);
700 if (xsk_uses_need_wakeup(tx_ring->xsk_pool))
701 xsk_set_tx_need_wakeup(tx_ring->xsk_pool);
703 return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring));
707 * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
708 * @dev: the netdevice
709 * @queue_id: queue id to wake up
710 * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
712 * Returns <0 for errors, 0 otherwise.
714 int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
716 struct i40e_netdev_priv *np = netdev_priv(dev);
717 struct i40e_vsi *vsi = np->vsi;
718 struct i40e_pf *pf = vsi->back;
719 struct i40e_ring *ring;
721 if (test_bit(__I40E_CONFIG_BUSY, pf->state))
724 if (test_bit(__I40E_VSI_DOWN, vsi->state))
727 if (!i40e_enabled_xdp_vsi(vsi))
730 if (queue_id >= vsi->num_queue_pairs)
733 if (!vsi->xdp_rings[queue_id]->xsk_pool)
736 ring = vsi->xdp_rings[queue_id];
738 /* The idea here is that if NAPI is running, mark a miss, so
739 * it will run again. If not, trigger an interrupt and
740 * schedule the NAPI from interrupt context. If NAPI would be
741 * scheduled here, the interrupt affinity would not be
744 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
745 i40e_force_wb(vsi, ring->q_vector);
750 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
752 u16 ntc = rx_ring->next_to_clean;
753 u16 ntu = rx_ring->next_to_use;
756 struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, ntc);
758 xsk_buff_free(rx_bi);
760 if (ntc >= rx_ring->count)
766 * i40e_xsk_clean_tx_ring - Clean the XDP Tx ring on shutdown
767 * @tx_ring: XDP Tx ring
769 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
771 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
772 struct xsk_buff_pool *bp = tx_ring->xsk_pool;
773 struct i40e_tx_buffer *tx_bi;
777 tx_bi = &tx_ring->tx_bi[ntc];
780 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
787 if (ntc >= tx_ring->count)
792 xsk_tx_completed(bp, xsk_frames);
796 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have an AF_XDP
797 * buffer pool attached
800 * Returns true if any of the Rx rings has an AF_XDP buffer pool attached
802 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
804 struct net_device *netdev = vsi->netdev;
807 for (i = 0; i < vsi->num_queue_pairs; i++) {
808 if (xsk_get_pool_from_qid(netdev, i))