1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2018 Intel Corporation. */
4 #include <linux/bpf_trace.h>
5 #include <net/xdp_sock.h>
9 #include "i40e_txrx_common.h"
13 * i40e_xsk_umem_dma_map - DMA maps all UMEM memory for the netdev
15 * @umem: UMEM to DMA map
17 * Returns 0 on success, <0 on failure
19 static int i40e_xsk_umem_dma_map(struct i40e_vsi *vsi, struct xdp_umem *umem)
21 struct i40e_pf *pf = vsi->back;
27 for (i = 0; i < umem->npgs; i++) {
28 dma = dma_map_page_attrs(dev, umem->pgs[i], 0, PAGE_SIZE,
29 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
30 if (dma_mapping_error(dev, dma))
33 umem->pages[i].dma = dma;
39 for (j = 0; j < i; j++) {
40 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
41 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
42 umem->pages[i].dma = 0;
49 * i40e_xsk_umem_dma_unmap - DMA unmaps all UMEM memory for the netdev
51 * @umem: UMEM to DMA map
53 static void i40e_xsk_umem_dma_unmap(struct i40e_vsi *vsi, struct xdp_umem *umem)
55 struct i40e_pf *pf = vsi->back;
61 for (i = 0; i < umem->npgs; i++) {
62 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
63 DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
65 umem->pages[i].dma = 0;
70 * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
73 * @qid: Rx ring to associate UMEM to
75 * Returns 0 on success, <0 on failure
77 static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
80 struct net_device *netdev = vsi->netdev;
81 struct xdp_umem_fq_reuse *reuseq;
85 if (vsi->type != I40E_VSI_MAIN)
88 if (qid >= vsi->num_queue_pairs)
91 if (qid >= netdev->real_num_rx_queues ||
92 qid >= netdev->real_num_tx_queues)
95 reuseq = xsk_reuseq_prepare(vsi->rx_rings[0]->count);
99 xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
101 err = i40e_xsk_umem_dma_map(vsi, umem);
105 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
108 err = i40e_queue_pair_disable(vsi, qid);
112 err = i40e_queue_pair_enable(vsi, qid);
121 * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
123 * @qid: Rx ring to associate UMEM to
125 * Returns 0 on success, <0 on failure
127 static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
129 struct net_device *netdev = vsi->netdev;
130 struct xdp_umem *umem;
134 umem = xdp_get_umem_from_qid(netdev, qid);
138 if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
141 err = i40e_queue_pair_disable(vsi, qid);
146 i40e_xsk_umem_dma_unmap(vsi, umem);
149 err = i40e_queue_pair_enable(vsi, qid);
158 * i40e_xsk_umem_query - Queries a certain ring/qid for its UMEM
160 * @umem: UMEM associated to the ring, if any
161 * @qid: Rx ring to associate UMEM to
163 * This function will store, if any, the UMEM associated to certain ring.
165 * Returns 0 on success, <0 on failure
167 int i40e_xsk_umem_query(struct i40e_vsi *vsi, struct xdp_umem **umem,
170 struct net_device *netdev = vsi->netdev;
171 struct xdp_umem *queried_umem;
173 if (vsi->type != I40E_VSI_MAIN)
176 queried_umem = xdp_get_umem_from_qid(netdev, qid);
181 *umem = queried_umem;
186 * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
188 * @umem: UMEM to enable/associate to a ring, or NULL to disable
189 * @qid: Rx ring to (dis)associate UMEM (from)to
191 * This function enables or disables a UMEM to a certain ring.
193 * Returns 0 on success, <0 on failure
195 int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
198 return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
199 i40e_xsk_umem_disable(vsi, qid);
203 * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
205 * @xdp: xdp_buff used as input to the XDP program
207 * This function enables or disables a UMEM to a certain ring.
209 * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
211 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
213 int err, result = I40E_XDP_PASS;
214 struct i40e_ring *xdp_ring;
215 struct bpf_prog *xdp_prog;
219 /* NB! xdp_prog will always be !NULL, due to the fact that
220 * this path is enabled by setting an XDP program.
222 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
223 act = bpf_prog_run_xdp(xdp_prog, xdp);
224 xdp->handle += xdp->data - xdp->data_hard_start;
229 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
230 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
233 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
234 result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
237 bpf_warn_invalid_xdp_action(act);
239 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
240 /* fallthrough -- handle aborts by dropping packet */
242 result = I40E_XDP_CONSUMED;
250 * i40e_alloc_buffer_zc - Allocates an i40e_rx_buffer
252 * @bi: Rx buffer to populate
254 * This function allocates an Rx buffer. The buffer can come from fill
255 * queue, or via the recycle queue (next_to_alloc).
257 * Returns true for a successful allocation, false otherwise
259 static bool i40e_alloc_buffer_zc(struct i40e_ring *rx_ring,
260 struct i40e_rx_buffer *bi)
262 struct xdp_umem *umem = rx_ring->xsk_umem;
263 void *addr = bi->addr;
267 rx_ring->rx_stats.page_reuse_count++;
271 if (!xsk_umem_peek_addr(umem, &handle)) {
272 rx_ring->rx_stats.alloc_page_failed++;
276 hr = umem->headroom + XDP_PACKET_HEADROOM;
278 bi->dma = xdp_umem_get_dma(umem, handle);
281 bi->addr = xdp_umem_get_data(umem, handle);
284 bi->handle = handle + umem->headroom;
286 xsk_umem_discard_addr(umem);
291 * i40e_alloc_buffer_slow_zc - Allocates an i40e_rx_buffer
293 * @bi: Rx buffer to populate
295 * This function allocates an Rx buffer. The buffer can come from fill
296 * queue, or via the reuse queue.
298 * Returns true for a successful allocation, false otherwise
300 static bool i40e_alloc_buffer_slow_zc(struct i40e_ring *rx_ring,
301 struct i40e_rx_buffer *bi)
303 struct xdp_umem *umem = rx_ring->xsk_umem;
306 if (!xsk_umem_peek_addr_rq(umem, &handle)) {
307 rx_ring->rx_stats.alloc_page_failed++;
311 handle &= rx_ring->xsk_umem->chunk_mask;
313 hr = umem->headroom + XDP_PACKET_HEADROOM;
315 bi->dma = xdp_umem_get_dma(umem, handle);
318 bi->addr = xdp_umem_get_data(umem, handle);
321 bi->handle = handle + umem->headroom;
323 xsk_umem_discard_addr_rq(umem);
327 static __always_inline bool
328 __i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count,
329 bool alloc(struct i40e_ring *rx_ring,
330 struct i40e_rx_buffer *bi))
332 u16 ntu = rx_ring->next_to_use;
333 union i40e_rx_desc *rx_desc;
334 struct i40e_rx_buffer *bi;
337 rx_desc = I40E_RX_DESC(rx_ring, ntu);
338 bi = &rx_ring->rx_bi[ntu];
340 if (!alloc(rx_ring, bi)) {
345 dma_sync_single_range_for_device(rx_ring->dev, bi->dma, 0,
349 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
355 if (unlikely(ntu == rx_ring->count)) {
356 rx_desc = I40E_RX_DESC(rx_ring, 0);
361 rx_desc->wb.qword1.status_error_len = 0;
366 if (rx_ring->next_to_use != ntu)
367 i40e_release_rx_desc(rx_ring, ntu);
373 * i40e_alloc_rx_buffers_zc - Allocates a number of Rx buffers
375 * @count: The number of buffers to allocate
377 * This function allocates a number of Rx buffers from the reuse queue
378 * or fill ring and places them on the Rx ring.
380 * Returns true for a successful allocation, false otherwise
382 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
384 return __i40e_alloc_rx_buffers_zc(rx_ring, count,
385 i40e_alloc_buffer_slow_zc);
389 * i40e_alloc_rx_buffers_fast_zc - Allocates a number of Rx buffers
391 * @count: The number of buffers to allocate
393 * This function allocates a number of Rx buffers from the fill ring
394 * or the internal recycle mechanism and places them on the Rx ring.
396 * Returns true for a successful allocation, false otherwise
398 static bool i40e_alloc_rx_buffers_fast_zc(struct i40e_ring *rx_ring, u16 count)
400 return __i40e_alloc_rx_buffers_zc(rx_ring, count,
401 i40e_alloc_buffer_zc);
405 * i40e_get_rx_buffer_zc - Return the current Rx buffer
407 * @size: The size of the rx buffer (read from descriptor)
409 * This function returns the current, received Rx buffer, and also
410 * does DMA synchronization. the Rx ring.
412 * Returns the received Rx buffer
414 static struct i40e_rx_buffer *i40e_get_rx_buffer_zc(struct i40e_ring *rx_ring,
415 const unsigned int size)
417 struct i40e_rx_buffer *bi;
419 bi = &rx_ring->rx_bi[rx_ring->next_to_clean];
421 /* we are reusing so sync this buffer for CPU use */
422 dma_sync_single_range_for_cpu(rx_ring->dev,
431 * i40e_reuse_rx_buffer_zc - Recycle an Rx buffer
433 * @old_bi: The Rx buffer to recycle
435 * This function recycles a finished Rx buffer, and places it on the
436 * recycle queue (next_to_alloc).
438 static void i40e_reuse_rx_buffer_zc(struct i40e_ring *rx_ring,
439 struct i40e_rx_buffer *old_bi)
441 struct i40e_rx_buffer *new_bi = &rx_ring->rx_bi[rx_ring->next_to_alloc];
442 unsigned long mask = (unsigned long)rx_ring->xsk_umem->chunk_mask;
443 u64 hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
444 u16 nta = rx_ring->next_to_alloc;
446 /* update, and store next to alloc */
448 rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
450 /* transfer page from old buffer to new buffer */
451 new_bi->dma = old_bi->dma & mask;
454 new_bi->addr = (void *)((unsigned long)old_bi->addr & mask);
457 new_bi->handle = old_bi->handle & mask;
458 new_bi->handle += rx_ring->xsk_umem->headroom;
464 * i40e_zca_free - Free callback for MEM_TYPE_ZERO_COPY allocations
465 * @alloc: Zero-copy allocator
466 * @handle: Buffer handle
468 void i40e_zca_free(struct zero_copy_allocator *alloc, unsigned long handle)
470 struct i40e_rx_buffer *bi;
471 struct i40e_ring *rx_ring;
475 rx_ring = container_of(alloc, struct i40e_ring, zca);
476 hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
477 mask = rx_ring->xsk_umem->chunk_mask;
479 nta = rx_ring->next_to_alloc;
480 bi = &rx_ring->rx_bi[nta];
483 rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
487 bi->dma = xdp_umem_get_dma(rx_ring->xsk_umem, handle);
490 bi->addr = xdp_umem_get_data(rx_ring->xsk_umem, handle);
493 bi->handle = (u64)handle + rx_ring->xsk_umem->headroom;
497 * i40e_construct_skb_zc - Create skbufff from zero-copy Rx buffer
502 * This functions allocates a new skb from a zero-copy Rx buffer.
504 * Returns the skb, or NULL on failure.
506 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
507 struct i40e_rx_buffer *bi,
508 struct xdp_buff *xdp)
510 unsigned int metasize = xdp->data - xdp->data_meta;
511 unsigned int datasize = xdp->data_end - xdp->data;
514 /* allocate a skb to store the frags */
515 skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
516 xdp->data_end - xdp->data_hard_start,
517 GFP_ATOMIC | __GFP_NOWARN);
521 skb_reserve(skb, xdp->data - xdp->data_hard_start);
522 memcpy(__skb_put(skb, datasize), xdp->data, datasize);
524 skb_metadata_set(skb, metasize);
526 i40e_reuse_rx_buffer_zc(rx_ring, bi);
531 * i40e_inc_ntc: Advance the next_to_clean index
534 static void i40e_inc_ntc(struct i40e_ring *rx_ring)
536 u32 ntc = rx_ring->next_to_clean + 1;
538 ntc = (ntc < rx_ring->count) ? ntc : 0;
539 rx_ring->next_to_clean = ntc;
540 prefetch(I40E_RX_DESC(rx_ring, ntc));
544 * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
546 * @budget: NAPI budget
548 * Returns amount of work completed
550 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
552 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
553 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
554 unsigned int xdp_res, xdp_xmit = 0;
555 bool failure = false;
559 xdp.rxq = &rx_ring->xdp_rxq;
561 while (likely(total_rx_packets < (unsigned int)budget)) {
562 struct i40e_rx_buffer *bi;
563 union i40e_rx_desc *rx_desc;
567 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
569 !i40e_alloc_rx_buffers_fast_zc(rx_ring,
574 rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
575 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
577 /* This memory barrier is needed to keep us from reading
578 * any other fields out of the rx_desc until we have
579 * verified the descriptor has been written back.
583 bi = i40e_clean_programming_status(rx_ring, rx_desc,
586 i40e_reuse_rx_buffer_zc(rx_ring, bi);
591 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
592 I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
596 bi = i40e_get_rx_buffer_zc(rx_ring, size);
598 xdp.data_meta = xdp.data;
599 xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
600 xdp.data_end = xdp.data + size;
601 xdp.handle = bi->handle;
603 xdp_res = i40e_run_xdp_zc(rx_ring, &xdp);
605 if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) {
609 i40e_reuse_rx_buffer_zc(rx_ring, bi);
612 total_rx_bytes += size;
616 i40e_inc_ntc(rx_ring);
622 /* NB! We are not checking for errors using
623 * i40e_test_staterr with
624 * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
625 * SBP is *not* set in PRT_SBPVSI (default not set).
627 skb = i40e_construct_skb_zc(rx_ring, bi, &xdp);
629 rx_ring->rx_stats.alloc_buff_failed++;
634 i40e_inc_ntc(rx_ring);
636 if (eth_skb_pad(skb))
639 total_rx_bytes += skb->len;
642 i40e_process_skb_fields(rx_ring, rx_desc, skb);
643 napi_gro_receive(&rx_ring->q_vector->napi, skb);
646 i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
647 i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
648 return failure ? budget : (int)total_rx_packets;
652 * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
653 * @xdp_ring: XDP Tx ring
654 * @budget: NAPI budget
656 * Returns true if the work is finished.
658 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
660 struct i40e_tx_desc *tx_desc = NULL;
661 struct i40e_tx_buffer *tx_bi;
662 bool work_done = true;
666 while (budget-- > 0) {
667 if (!unlikely(I40E_DESC_UNUSED(xdp_ring))) {
668 xdp_ring->tx_stats.tx_busy++;
673 if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &dma, &len))
676 dma_sync_single_for_device(xdp_ring->dev, dma, len,
679 tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
680 tx_bi->bytecount = len;
682 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
683 tx_desc->buffer_addr = cpu_to_le64(dma);
684 tx_desc->cmd_type_offset_bsz =
685 build_ctob(I40E_TX_DESC_CMD_ICRC
686 | I40E_TX_DESC_CMD_EOP,
689 xdp_ring->next_to_use++;
690 if (xdp_ring->next_to_use == xdp_ring->count)
691 xdp_ring->next_to_use = 0;
695 /* Request an interrupt for the last frame and bump tail ptr. */
696 tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
697 I40E_TXD_QW1_CMD_SHIFT);
698 i40e_xdp_ring_update_tail(xdp_ring);
700 xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
703 return !!budget && work_done;
707 * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
708 * @tx_ring: XDP Tx ring
709 * @tx_bi: Tx buffer info to clean
711 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
712 struct i40e_tx_buffer *tx_bi)
714 xdp_return_frame(tx_bi->xdpf);
715 dma_unmap_single(tx_ring->dev,
716 dma_unmap_addr(tx_bi, dma),
717 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
718 dma_unmap_len_set(tx_bi, len, 0);
722 * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
723 * @tx_ring: XDP Tx ring
724 * @tx_bi: Tx buffer info to clean
726 * Returns true if cleanup/tranmission is done.
728 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi,
729 struct i40e_ring *tx_ring, int napi_budget)
731 unsigned int ntc, total_bytes = 0, budget = vsi->work_limit;
732 u32 i, completed_frames, frames_ready, xsk_frames = 0;
733 struct xdp_umem *umem = tx_ring->xsk_umem;
734 u32 head_idx = i40e_get_head(tx_ring);
735 bool work_done = true, xmit_done;
736 struct i40e_tx_buffer *tx_bi;
738 if (head_idx < tx_ring->next_to_clean)
739 head_idx += tx_ring->count;
740 frames_ready = head_idx - tx_ring->next_to_clean;
742 if (frames_ready == 0) {
744 } else if (frames_ready > budget) {
745 completed_frames = budget;
748 completed_frames = frames_ready;
751 ntc = tx_ring->next_to_clean;
753 for (i = 0; i < completed_frames; i++) {
754 tx_bi = &tx_ring->tx_bi[ntc];
757 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
762 total_bytes += tx_bi->bytecount;
764 if (++ntc >= tx_ring->count)
768 tx_ring->next_to_clean += completed_frames;
769 if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
770 tx_ring->next_to_clean -= tx_ring->count;
773 xsk_umem_complete_tx(umem, xsk_frames);
775 i40e_arm_wb(tx_ring, vsi, budget);
776 i40e_update_tx_stats(tx_ring, completed_frames, total_bytes);
779 xmit_done = i40e_xmit_zc(tx_ring, budget);
781 return work_done && xmit_done;
785 * i40e_xsk_async_xmit - Implements the ndo_xsk_async_xmit
786 * @dev: the netdevice
787 * @queue_id: queue id to wake up
789 * Returns <0 for errors, 0 otherwise.
791 int i40e_xsk_async_xmit(struct net_device *dev, u32 queue_id)
793 struct i40e_netdev_priv *np = netdev_priv(dev);
794 struct i40e_vsi *vsi = np->vsi;
795 struct i40e_ring *ring;
797 if (test_bit(__I40E_VSI_DOWN, vsi->state))
800 if (!i40e_enabled_xdp_vsi(vsi))
803 if (queue_id >= vsi->num_queue_pairs)
806 if (!vsi->xdp_rings[queue_id]->xsk_umem)
809 ring = vsi->xdp_rings[queue_id];
811 /* The idea here is that if NAPI is running, mark a miss, so
812 * it will run again. If not, trigger an interrupt and
813 * schedule the NAPI from interrupt context. If NAPI would be
814 * scheduled here, the interrupt affinity would not be
817 if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
818 i40e_force_wb(vsi, ring->q_vector);
823 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
827 for (i = 0; i < rx_ring->count; i++) {
828 struct i40e_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
833 xsk_umem_fq_reuse(rx_ring->xsk_umem, rx_bi->handle);
839 * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
840 * @xdp_ring: XDP Tx ring
842 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
844 u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
845 struct xdp_umem *umem = tx_ring->xsk_umem;
846 struct i40e_tx_buffer *tx_bi;
850 tx_bi = &tx_ring->tx_bi[ntc];
853 i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
860 if (ntc >= tx_ring->count)
865 xsk_umem_complete_tx(umem, xsk_frames);
869 * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
872 * Returns true if any of the Rx rings has an AF_XDP UMEM attached
874 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
876 struct net_device *netdev = vsi->netdev;
879 for (i = 0; i < vsi->num_queue_pairs; i++) {
880 if (xdp_get_umem_from_qid(netdev, i))