1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2013 - 2019 Intel Corporation. */
4 #include <linux/types.h>
5 #include <linux/module.h>
9 #include <linux/if_macvlan.h>
10 #include <linux/prefetch.h>
14 #define DRV_SUMMARY "Intel(R) Ethernet Switch Host Interface Driver"
15 char fm10k_driver_name[] = "fm10k";
16 static const char fm10k_driver_string[] = DRV_SUMMARY;
17 static const char fm10k_copyright[] =
18 "Copyright(c) 2013 - 2019 Intel Corporation.";
21 MODULE_DESCRIPTION(DRV_SUMMARY);
22 MODULE_LICENSE("GPL v2");
24 /* single workqueue for entire fm10k driver */
25 struct workqueue_struct *fm10k_workqueue;
28 * fm10k_init_module - Driver Registration Routine
30 * fm10k_init_module is the first routine called when the driver is
31 * loaded. All it does is register with the PCI subsystem.
33 static int __init fm10k_init_module(void)
35 pr_info("%s\n", fm10k_driver_string);
36 pr_info("%s\n", fm10k_copyright);
38 /* create driver workqueue */
39 fm10k_workqueue = alloc_workqueue("%s", WQ_MEM_RECLAIM, 0,
46 return fm10k_register_pci_driver();
48 module_init(fm10k_init_module);
51 * fm10k_exit_module - Driver Exit Cleanup Routine
53 * fm10k_exit_module is called just before the driver is removed
56 static void __exit fm10k_exit_module(void)
58 fm10k_unregister_pci_driver();
62 /* destroy driver workqueue */
63 destroy_workqueue(fm10k_workqueue);
65 module_exit(fm10k_exit_module);
67 static bool fm10k_alloc_mapped_page(struct fm10k_ring *rx_ring,
68 struct fm10k_rx_buffer *bi)
70 struct page *page = bi->page;
73 /* Only page will be NULL if buffer was consumed */
77 /* alloc new page for storage */
78 page = dev_alloc_page();
79 if (unlikely(!page)) {
80 rx_ring->rx_stats.alloc_failed++;
84 /* map page for use */
85 dma = dma_map_page(rx_ring->dev, page, 0, PAGE_SIZE, DMA_FROM_DEVICE);
87 /* if mapping failed free memory back to system since
88 * there isn't much point in holding memory we can't use
90 if (dma_mapping_error(rx_ring->dev, dma)) {
93 rx_ring->rx_stats.alloc_failed++;
105 * fm10k_alloc_rx_buffers - Replace used receive buffers
106 * @rx_ring: ring to place buffers on
107 * @cleaned_count: number of buffers to replace
109 void fm10k_alloc_rx_buffers(struct fm10k_ring *rx_ring, u16 cleaned_count)
111 union fm10k_rx_desc *rx_desc;
112 struct fm10k_rx_buffer *bi;
113 u16 i = rx_ring->next_to_use;
119 rx_desc = FM10K_RX_DESC(rx_ring, i);
120 bi = &rx_ring->rx_buffer[i];
124 if (!fm10k_alloc_mapped_page(rx_ring, bi))
127 /* Refresh the desc even if buffer_addrs didn't change
128 * because each write-back erases this info.
130 rx_desc->q.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
136 rx_desc = FM10K_RX_DESC(rx_ring, 0);
137 bi = rx_ring->rx_buffer;
141 /* clear the status bits for the next_to_use descriptor */
142 rx_desc->d.staterr = 0;
145 } while (cleaned_count);
149 if (rx_ring->next_to_use != i) {
150 /* record the next descriptor to use */
151 rx_ring->next_to_use = i;
153 /* update next to alloc since we have filled the ring */
154 rx_ring->next_to_alloc = i;
156 /* Force memory writes to complete before letting h/w
157 * know there are new descriptors to fetch. (Only
158 * applicable for weak-ordered memory model archs,
163 /* notify hardware of new descriptors */
164 writel(i, rx_ring->tail);
169 * fm10k_reuse_rx_page - page flip buffer and store it back on the ring
170 * @rx_ring: rx descriptor ring to store buffers on
171 * @old_buff: donor buffer to have page reused
173 * Synchronizes page for reuse by the interface
175 static void fm10k_reuse_rx_page(struct fm10k_ring *rx_ring,
176 struct fm10k_rx_buffer *old_buff)
178 struct fm10k_rx_buffer *new_buff;
179 u16 nta = rx_ring->next_to_alloc;
181 new_buff = &rx_ring->rx_buffer[nta];
183 /* update, and store next to alloc */
185 rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
187 /* transfer page from old buffer to new buffer */
188 *new_buff = *old_buff;
190 /* sync the buffer for use by the device */
191 dma_sync_single_range_for_device(rx_ring->dev, old_buff->dma,
192 old_buff->page_offset,
197 static inline bool fm10k_page_is_reserved(struct page *page)
199 return (page_to_nid(page) != numa_mem_id()) || page_is_pfmemalloc(page);
202 static bool fm10k_can_reuse_rx_page(struct fm10k_rx_buffer *rx_buffer,
204 unsigned int __maybe_unused truesize)
206 /* avoid re-using remote pages */
207 if (unlikely(fm10k_page_is_reserved(page)))
210 #if (PAGE_SIZE < 8192)
211 /* if we are only owner of page we can reuse it */
212 if (unlikely(page_count(page) != 1))
215 /* flip page offset to other buffer */
216 rx_buffer->page_offset ^= FM10K_RX_BUFSZ;
218 /* move offset up to the next cache line */
219 rx_buffer->page_offset += truesize;
221 if (rx_buffer->page_offset > (PAGE_SIZE - FM10K_RX_BUFSZ))
225 /* Even if we own the page, we are not allowed to use atomic_set()
226 * This would break get_page_unless_zero() users.
234 * fm10k_add_rx_frag - Add contents of Rx buffer to sk_buff
235 * @rx_buffer: buffer containing page to add
236 * @size: packet size from rx_desc
237 * @rx_desc: descriptor containing length of buffer written by hardware
238 * @skb: sk_buff to place the data into
240 * This function will add the data contained in rx_buffer->page to the skb.
241 * This is done either through a direct copy if the data in the buffer is
242 * less than the skb header size, otherwise it will just attach the page as
245 * The function will then update the page offset if necessary and return
246 * true if the buffer can be reused by the interface.
248 static bool fm10k_add_rx_frag(struct fm10k_rx_buffer *rx_buffer,
250 union fm10k_rx_desc *rx_desc,
253 struct page *page = rx_buffer->page;
254 unsigned char *va = page_address(page) + rx_buffer->page_offset;
255 #if (PAGE_SIZE < 8192)
256 unsigned int truesize = FM10K_RX_BUFSZ;
258 unsigned int truesize = ALIGN(size, 512);
260 unsigned int pull_len;
262 if (unlikely(skb_is_nonlinear(skb)))
265 if (likely(size <= FM10K_RX_HDR_LEN)) {
266 memcpy(__skb_put(skb, size), va, ALIGN(size, sizeof(long)));
268 /* page is not reserved, we can reuse buffer as-is */
269 if (likely(!fm10k_page_is_reserved(page)))
272 /* this page cannot be reused so discard it */
277 /* we need the header to contain the greater of either ETH_HLEN or
278 * 60 bytes if the skb->len is less than 60 for skb_pad.
280 pull_len = eth_get_headlen(skb->dev, va, FM10K_RX_HDR_LEN);
282 /* align pull length to size of long to optimize memcpy performance */
283 memcpy(__skb_put(skb, pull_len), va, ALIGN(pull_len, sizeof(long)));
285 /* update all of the pointers */
290 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
291 (unsigned long)va & ~PAGE_MASK, size, truesize);
293 return fm10k_can_reuse_rx_page(rx_buffer, page, truesize);
296 static struct sk_buff *fm10k_fetch_rx_buffer(struct fm10k_ring *rx_ring,
297 union fm10k_rx_desc *rx_desc,
300 unsigned int size = le16_to_cpu(rx_desc->w.length);
301 struct fm10k_rx_buffer *rx_buffer;
304 rx_buffer = &rx_ring->rx_buffer[rx_ring->next_to_clean];
305 page = rx_buffer->page;
309 void *page_addr = page_address(page) +
310 rx_buffer->page_offset;
312 /* prefetch first cache line of first page */
314 #if L1_CACHE_BYTES < 128
315 prefetch((void *)((u8 *)page_addr + L1_CACHE_BYTES));
318 /* allocate a skb to store the frags */
319 skb = napi_alloc_skb(&rx_ring->q_vector->napi,
321 if (unlikely(!skb)) {
322 rx_ring->rx_stats.alloc_failed++;
326 /* we will be copying header into skb->data in
327 * pskb_may_pull so it is in our interest to prefetch
328 * it now to avoid a possible cache miss
330 prefetchw(skb->data);
333 /* we are reusing so sync this buffer for CPU use */
334 dma_sync_single_range_for_cpu(rx_ring->dev,
336 rx_buffer->page_offset,
340 /* pull page into skb */
341 if (fm10k_add_rx_frag(rx_buffer, size, rx_desc, skb)) {
342 /* hand second half of page back to the ring */
343 fm10k_reuse_rx_page(rx_ring, rx_buffer);
345 /* we are not reusing the buffer so unmap it */
346 dma_unmap_page(rx_ring->dev, rx_buffer->dma,
347 PAGE_SIZE, DMA_FROM_DEVICE);
350 /* clear contents of rx_buffer */
351 rx_buffer->page = NULL;
356 static inline void fm10k_rx_checksum(struct fm10k_ring *ring,
357 union fm10k_rx_desc *rx_desc,
360 skb_checksum_none_assert(skb);
362 /* Rx checksum disabled via ethtool */
363 if (!(ring->netdev->features & NETIF_F_RXCSUM))
366 /* TCP/UDP checksum error bit is set */
367 if (fm10k_test_staterr(rx_desc,
368 FM10K_RXD_STATUS_L4E |
369 FM10K_RXD_STATUS_L4E2 |
370 FM10K_RXD_STATUS_IPE |
371 FM10K_RXD_STATUS_IPE2)) {
372 ring->rx_stats.csum_err++;
376 /* It must be a TCP or UDP packet with a valid checksum */
377 if (fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS2))
378 skb->encapsulation = true;
379 else if (!fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS))
382 skb->ip_summed = CHECKSUM_UNNECESSARY;
384 ring->rx_stats.csum_good++;
387 #define FM10K_RSS_L4_TYPES_MASK \
388 (BIT(FM10K_RSSTYPE_IPV4_TCP) | \
389 BIT(FM10K_RSSTYPE_IPV4_UDP) | \
390 BIT(FM10K_RSSTYPE_IPV6_TCP) | \
391 BIT(FM10K_RSSTYPE_IPV6_UDP))
393 static inline void fm10k_rx_hash(struct fm10k_ring *ring,
394 union fm10k_rx_desc *rx_desc,
399 if (!(ring->netdev->features & NETIF_F_RXHASH))
402 rss_type = le16_to_cpu(rx_desc->w.pkt_info) & FM10K_RXD_RSSTYPE_MASK;
406 skb_set_hash(skb, le32_to_cpu(rx_desc->d.rss),
407 (BIT(rss_type) & FM10K_RSS_L4_TYPES_MASK) ?
408 PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3);
411 static void fm10k_type_trans(struct fm10k_ring *rx_ring,
412 union fm10k_rx_desc __maybe_unused *rx_desc,
415 struct net_device *dev = rx_ring->netdev;
416 struct fm10k_l2_accel *l2_accel = rcu_dereference_bh(rx_ring->l2_accel);
418 /* check to see if DGLORT belongs to a MACVLAN */
420 u16 idx = le16_to_cpu(FM10K_CB(skb)->fi.w.dglort) - 1;
422 idx -= l2_accel->dglort;
423 if (idx < l2_accel->size && l2_accel->macvlan[idx])
424 dev = l2_accel->macvlan[idx];
429 /* Record Rx queue, or update macvlan statistics */
431 skb_record_rx_queue(skb, rx_ring->queue_index);
433 macvlan_count_rx(netdev_priv(dev), skb->len + ETH_HLEN, true,
436 skb->protocol = eth_type_trans(skb, dev);
440 * fm10k_process_skb_fields - Populate skb header fields from Rx descriptor
441 * @rx_ring: rx descriptor ring packet is being transacted on
442 * @rx_desc: pointer to the EOP Rx descriptor
443 * @skb: pointer to current skb being populated
445 * This function checks the ring, descriptor, and packet information in
446 * order to populate the hash, checksum, VLAN, timestamp, protocol, and
447 * other fields within the skb.
449 static unsigned int fm10k_process_skb_fields(struct fm10k_ring *rx_ring,
450 union fm10k_rx_desc *rx_desc,
453 unsigned int len = skb->len;
455 fm10k_rx_hash(rx_ring, rx_desc, skb);
457 fm10k_rx_checksum(rx_ring, rx_desc, skb);
459 FM10K_CB(skb)->tstamp = rx_desc->q.timestamp;
461 FM10K_CB(skb)->fi.w.vlan = rx_desc->w.vlan;
463 FM10K_CB(skb)->fi.d.glort = rx_desc->d.glort;
465 if (rx_desc->w.vlan) {
466 u16 vid = le16_to_cpu(rx_desc->w.vlan);
468 if ((vid & VLAN_VID_MASK) != rx_ring->vid)
469 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vid);
470 else if (vid & VLAN_PRIO_MASK)
471 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
472 vid & VLAN_PRIO_MASK);
475 fm10k_type_trans(rx_ring, rx_desc, skb);
481 * fm10k_is_non_eop - process handling of non-EOP buffers
482 * @rx_ring: Rx ring being processed
483 * @rx_desc: Rx descriptor for current buffer
485 * This function updates next to clean. If the buffer is an EOP buffer
486 * this function exits returning false, otherwise it will place the
487 * sk_buff in the next buffer to be chained and return true indicating
488 * that this is in fact a non-EOP buffer.
490 static bool fm10k_is_non_eop(struct fm10k_ring *rx_ring,
491 union fm10k_rx_desc *rx_desc)
493 u32 ntc = rx_ring->next_to_clean + 1;
495 /* fetch, update, and store next to clean */
496 ntc = (ntc < rx_ring->count) ? ntc : 0;
497 rx_ring->next_to_clean = ntc;
499 prefetch(FM10K_RX_DESC(rx_ring, ntc));
501 if (likely(fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_EOP)))
508 * fm10k_cleanup_headers - Correct corrupted or empty headers
509 * @rx_ring: rx descriptor ring packet is being transacted on
510 * @rx_desc: pointer to the EOP Rx descriptor
511 * @skb: pointer to current skb being fixed
513 * Address the case where we are pulling data in on pages only
514 * and as such no data is present in the skb header.
516 * In addition if skb is not at least 60 bytes we need to pad it so that
517 * it is large enough to qualify as a valid Ethernet frame.
519 * Returns true if an error was encountered and skb was freed.
521 static bool fm10k_cleanup_headers(struct fm10k_ring *rx_ring,
522 union fm10k_rx_desc *rx_desc,
525 if (unlikely((fm10k_test_staterr(rx_desc,
526 FM10K_RXD_STATUS_RXE)))) {
527 #define FM10K_TEST_RXD_BIT(rxd, bit) \
528 ((rxd)->w.csum_err & cpu_to_le16(bit))
529 if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_ERROR))
530 rx_ring->rx_stats.switch_errors++;
531 if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_NO_DESCRIPTOR))
532 rx_ring->rx_stats.drops++;
533 if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_PP_ERROR))
534 rx_ring->rx_stats.pp_errors++;
535 if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_READY))
536 rx_ring->rx_stats.link_errors++;
537 if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_TOO_BIG))
538 rx_ring->rx_stats.length_errors++;
539 dev_kfree_skb_any(skb);
540 rx_ring->rx_stats.errors++;
544 /* if eth_skb_pad returns an error the skb was freed */
545 if (eth_skb_pad(skb))
552 * fm10k_receive_skb - helper function to handle rx indications
553 * @q_vector: structure containing interrupt and ring information
554 * @skb: packet to send up
556 static void fm10k_receive_skb(struct fm10k_q_vector *q_vector,
559 napi_gro_receive(&q_vector->napi, skb);
562 static int fm10k_clean_rx_irq(struct fm10k_q_vector *q_vector,
563 struct fm10k_ring *rx_ring,
566 struct sk_buff *skb = rx_ring->skb;
567 unsigned int total_bytes = 0, total_packets = 0;
568 u16 cleaned_count = fm10k_desc_unused(rx_ring);
570 while (likely(total_packets < budget)) {
571 union fm10k_rx_desc *rx_desc;
573 /* return some buffers to hardware, one at a time is too slow */
574 if (cleaned_count >= FM10K_RX_BUFFER_WRITE) {
575 fm10k_alloc_rx_buffers(rx_ring, cleaned_count);
579 rx_desc = FM10K_RX_DESC(rx_ring, rx_ring->next_to_clean);
581 if (!rx_desc->d.staterr)
584 /* This memory barrier is needed to keep us from reading
585 * any other fields out of the rx_desc until we know the
586 * descriptor has been written back
590 /* retrieve a buffer from the ring */
591 skb = fm10k_fetch_rx_buffer(rx_ring, rx_desc, skb);
593 /* exit if we failed to retrieve a buffer */
599 /* fetch next buffer in frame if non-eop */
600 if (fm10k_is_non_eop(rx_ring, rx_desc))
603 /* verify the packet layout is correct */
604 if (fm10k_cleanup_headers(rx_ring, rx_desc, skb)) {
609 /* populate checksum, timestamp, VLAN, and protocol */
610 total_bytes += fm10k_process_skb_fields(rx_ring, rx_desc, skb);
612 fm10k_receive_skb(q_vector, skb);
614 /* reset skb pointer */
617 /* update budget accounting */
621 /* place incomplete frames back on ring for completion */
624 u64_stats_update_begin(&rx_ring->syncp);
625 rx_ring->stats.packets += total_packets;
626 rx_ring->stats.bytes += total_bytes;
627 u64_stats_update_end(&rx_ring->syncp);
628 q_vector->rx.total_packets += total_packets;
629 q_vector->rx.total_bytes += total_bytes;
631 return total_packets;
634 #define VXLAN_HLEN (sizeof(struct udphdr) + 8)
635 static struct ethhdr *fm10k_port_is_vxlan(struct sk_buff *skb)
637 struct fm10k_intfc *interface = netdev_priv(skb->dev);
639 if (interface->vxlan_port != udp_hdr(skb)->dest)
642 /* return offset of udp_hdr plus 8 bytes for VXLAN header */
643 return (struct ethhdr *)(skb_transport_header(skb) + VXLAN_HLEN);
646 #define FM10K_NVGRE_RESERVED0_FLAGS htons(0x9FFF)
647 #define NVGRE_TNI htons(0x2000)
648 struct fm10k_nvgre_hdr {
654 static struct ethhdr *fm10k_gre_is_nvgre(struct sk_buff *skb)
656 struct fm10k_nvgre_hdr *nvgre_hdr;
657 int hlen = ip_hdrlen(skb);
659 /* currently only IPv4 is supported due to hlen above */
660 if (vlan_get_protocol(skb) != htons(ETH_P_IP))
663 /* our transport header should be NVGRE */
664 nvgre_hdr = (struct fm10k_nvgre_hdr *)(skb_network_header(skb) + hlen);
666 /* verify all reserved flags are 0 */
667 if (nvgre_hdr->flags & FM10K_NVGRE_RESERVED0_FLAGS)
670 /* report start of ethernet header */
671 if (nvgre_hdr->flags & NVGRE_TNI)
672 return (struct ethhdr *)(nvgre_hdr + 1);
674 return (struct ethhdr *)(&nvgre_hdr->tni);
677 __be16 fm10k_tx_encap_offload(struct sk_buff *skb)
679 u8 l4_hdr = 0, inner_l4_hdr = 0, inner_l4_hlen;
680 struct ethhdr *eth_hdr;
682 if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
683 skb->inner_protocol != htons(ETH_P_TEB))
686 switch (vlan_get_protocol(skb)) {
687 case htons(ETH_P_IP):
688 l4_hdr = ip_hdr(skb)->protocol;
690 case htons(ETH_P_IPV6):
691 l4_hdr = ipv6_hdr(skb)->nexthdr;
699 eth_hdr = fm10k_port_is_vxlan(skb);
702 eth_hdr = fm10k_gre_is_nvgre(skb);
711 switch (eth_hdr->h_proto) {
712 case htons(ETH_P_IP):
713 inner_l4_hdr = inner_ip_hdr(skb)->protocol;
715 case htons(ETH_P_IPV6):
716 inner_l4_hdr = inner_ipv6_hdr(skb)->nexthdr;
722 switch (inner_l4_hdr) {
724 inner_l4_hlen = inner_tcp_hdrlen(skb);
733 /* The hardware allows tunnel offloads only if the combined inner and
734 * outer header is 184 bytes or less
736 if (skb_inner_transport_header(skb) + inner_l4_hlen -
737 skb_mac_header(skb) > FM10K_TUNNEL_HEADER_LENGTH)
740 return eth_hdr->h_proto;
743 static int fm10k_tso(struct fm10k_ring *tx_ring,
744 struct fm10k_tx_buffer *first)
746 struct sk_buff *skb = first->skb;
747 struct fm10k_tx_desc *tx_desc;
751 if (skb->ip_summed != CHECKSUM_PARTIAL)
754 if (!skb_is_gso(skb))
757 /* compute header lengths */
758 if (skb->encapsulation) {
759 if (!fm10k_tx_encap_offload(skb))
761 th = skb_inner_transport_header(skb);
763 th = skb_transport_header(skb);
766 /* compute offset from SOF to transport header and add header len */
767 hdrlen = (th - skb->data) + (((struct tcphdr *)th)->doff << 2);
769 first->tx_flags |= FM10K_TX_FLAGS_CSUM;
771 /* update gso size and bytecount with header size */
772 first->gso_segs = skb_shinfo(skb)->gso_segs;
773 first->bytecount += (first->gso_segs - 1) * hdrlen;
775 /* populate Tx descriptor header size and mss */
776 tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use);
777 tx_desc->hdrlen = hdrlen;
778 tx_desc->mss = cpu_to_le16(skb_shinfo(skb)->gso_size);
783 tx_ring->netdev->features &= ~NETIF_F_GSO_UDP_TUNNEL;
785 netdev_err(tx_ring->netdev,
786 "TSO requested for unsupported tunnel, disabling offload\n");
790 static void fm10k_tx_csum(struct fm10k_ring *tx_ring,
791 struct fm10k_tx_buffer *first)
793 struct sk_buff *skb = first->skb;
794 struct fm10k_tx_desc *tx_desc;
797 struct ipv6hdr *ipv6;
805 if (skb->ip_summed != CHECKSUM_PARTIAL)
808 if (skb->encapsulation) {
809 protocol = fm10k_tx_encap_offload(skb);
811 if (skb_checksum_help(skb)) {
812 dev_warn(tx_ring->dev,
813 "failed to offload encap csum!\n");
814 tx_ring->tx_stats.csum_err++;
818 network_hdr.raw = skb_inner_network_header(skb);
819 transport_hdr = skb_inner_transport_header(skb);
821 protocol = vlan_get_protocol(skb);
822 network_hdr.raw = skb_network_header(skb);
823 transport_hdr = skb_transport_header(skb);
827 case htons(ETH_P_IP):
828 l4_hdr = network_hdr.ipv4->protocol;
830 case htons(ETH_P_IPV6):
831 l4_hdr = network_hdr.ipv6->nexthdr;
832 if (likely((transport_hdr - network_hdr.raw) ==
833 sizeof(struct ipv6hdr)))
835 ipv6_skip_exthdr(skb, network_hdr.raw - skb->data +
836 sizeof(struct ipv6hdr),
838 if (unlikely(frag_off))
839 l4_hdr = NEXTHDR_FRAGMENT;
850 if (skb->encapsulation)
854 if (unlikely(net_ratelimit())) {
855 dev_warn(tx_ring->dev,
856 "partial checksum, version=%d l4 proto=%x\n",
859 skb_checksum_help(skb);
860 tx_ring->tx_stats.csum_err++;
864 /* update TX checksum flag */
865 first->tx_flags |= FM10K_TX_FLAGS_CSUM;
866 tx_ring->tx_stats.csum_good++;
869 /* populate Tx descriptor header size and mss */
870 tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use);
875 #define FM10K_SET_FLAG(_input, _flag, _result) \
876 ((_flag <= _result) ? \
877 ((u32)(_input & _flag) * (_result / _flag)) : \
878 ((u32)(_input & _flag) / (_flag / _result)))
880 static u8 fm10k_tx_desc_flags(struct sk_buff *skb, u32 tx_flags)
882 /* set type for advanced descriptor with frame checksum insertion */
885 /* set checksum offload bits */
886 desc_flags |= FM10K_SET_FLAG(tx_flags, FM10K_TX_FLAGS_CSUM,
887 FM10K_TXD_FLAG_CSUM);
892 static bool fm10k_tx_desc_push(struct fm10k_ring *tx_ring,
893 struct fm10k_tx_desc *tx_desc, u16 i,
894 dma_addr_t dma, unsigned int size, u8 desc_flags)
896 /* set RS and INT for last frame in a cache line */
897 if ((++i & (FM10K_TXD_WB_FIFO_SIZE - 1)) == 0)
898 desc_flags |= FM10K_TXD_FLAG_RS | FM10K_TXD_FLAG_INT;
900 /* record values to descriptor */
901 tx_desc->buffer_addr = cpu_to_le64(dma);
902 tx_desc->flags = desc_flags;
903 tx_desc->buflen = cpu_to_le16(size);
905 /* return true if we just wrapped the ring */
906 return i == tx_ring->count;
909 static int __fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size)
911 netif_stop_subqueue(tx_ring->netdev, tx_ring->queue_index);
913 /* Memory barrier before checking head and tail */
916 /* Check again in a case another CPU has just made room available */
917 if (likely(fm10k_desc_unused(tx_ring) < size))
920 /* A reprieve! - use start_queue because it doesn't call schedule */
921 netif_start_subqueue(tx_ring->netdev, tx_ring->queue_index);
922 ++tx_ring->tx_stats.restart_queue;
926 static inline int fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size)
928 if (likely(fm10k_desc_unused(tx_ring) >= size))
930 return __fm10k_maybe_stop_tx(tx_ring, size);
933 static void fm10k_tx_map(struct fm10k_ring *tx_ring,
934 struct fm10k_tx_buffer *first)
936 struct sk_buff *skb = first->skb;
937 struct fm10k_tx_buffer *tx_buffer;
938 struct fm10k_tx_desc *tx_desc;
942 unsigned int data_len, size;
943 u32 tx_flags = first->tx_flags;
944 u16 i = tx_ring->next_to_use;
945 u8 flags = fm10k_tx_desc_flags(skb, tx_flags);
947 tx_desc = FM10K_TX_DESC(tx_ring, i);
949 /* add HW VLAN tag */
950 if (skb_vlan_tag_present(skb))
951 tx_desc->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
955 size = skb_headlen(skb);
958 dma = dma_map_single(tx_ring->dev, data, size, DMA_TO_DEVICE);
960 data_len = skb->data_len;
963 for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
964 if (dma_mapping_error(tx_ring->dev, dma))
967 /* record length, and DMA address */
968 dma_unmap_len_set(tx_buffer, len, size);
969 dma_unmap_addr_set(tx_buffer, dma, dma);
971 while (unlikely(size > FM10K_MAX_DATA_PER_TXD)) {
972 if (fm10k_tx_desc_push(tx_ring, tx_desc++, i++, dma,
973 FM10K_MAX_DATA_PER_TXD, flags)) {
974 tx_desc = FM10K_TX_DESC(tx_ring, 0);
978 dma += FM10K_MAX_DATA_PER_TXD;
979 size -= FM10K_MAX_DATA_PER_TXD;
982 if (likely(!data_len))
985 if (fm10k_tx_desc_push(tx_ring, tx_desc++, i++,
987 tx_desc = FM10K_TX_DESC(tx_ring, 0);
991 size = skb_frag_size(frag);
994 dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
997 tx_buffer = &tx_ring->tx_buffer[i];
1000 /* write last descriptor with LAST bit set */
1001 flags |= FM10K_TXD_FLAG_LAST;
1003 if (fm10k_tx_desc_push(tx_ring, tx_desc, i++, dma, size, flags))
1006 /* record bytecount for BQL */
1007 netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount);
1009 /* record SW timestamp if HW timestamp is not available */
1010 skb_tx_timestamp(first->skb);
1012 /* Force memory writes to complete before letting h/w know there
1013 * are new descriptors to fetch. (Only applicable for weak-ordered
1014 * memory model archs, such as IA-64).
1016 * We also need this memory barrier to make certain all of the
1017 * status bits have been updated before next_to_watch is written.
1021 /* set next_to_watch value indicating a packet is present */
1022 first->next_to_watch = tx_desc;
1024 tx_ring->next_to_use = i;
1026 /* Make sure there is space in the ring for the next send. */
1027 fm10k_maybe_stop_tx(tx_ring, DESC_NEEDED);
1029 /* notify HW of packet */
1030 if (netif_xmit_stopped(txring_txq(tx_ring)) || !netdev_xmit_more()) {
1031 writel(i, tx_ring->tail);
1036 dev_err(tx_ring->dev, "TX DMA map failed\n");
1038 /* clear dma mappings for failed tx_buffer map */
1040 tx_buffer = &tx_ring->tx_buffer[i];
1041 fm10k_unmap_and_free_tx_resource(tx_ring, tx_buffer);
1042 if (tx_buffer == first)
1049 tx_ring->next_to_use = i;
1052 netdev_tx_t fm10k_xmit_frame_ring(struct sk_buff *skb,
1053 struct fm10k_ring *tx_ring)
1055 u16 count = TXD_USE_COUNT(skb_headlen(skb));
1056 struct fm10k_tx_buffer *first;
1061 /* need: 1 descriptor per page * PAGE_SIZE/FM10K_MAX_DATA_PER_TXD,
1062 * + 1 desc for skb_headlen/FM10K_MAX_DATA_PER_TXD,
1063 * + 2 desc gap to keep tail from touching head
1064 * otherwise try next time
1066 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++) {
1067 skb_frag_t *frag = &skb_shinfo(skb)->frags[f];
1069 count += TXD_USE_COUNT(skb_frag_size(frag));
1072 if (fm10k_maybe_stop_tx(tx_ring, count + 3)) {
1073 tx_ring->tx_stats.tx_busy++;
1074 return NETDEV_TX_BUSY;
1077 /* record the location of the first descriptor for this packet */
1078 first = &tx_ring->tx_buffer[tx_ring->next_to_use];
1080 first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN);
1081 first->gso_segs = 1;
1083 /* record initial flags and protocol */
1084 first->tx_flags = tx_flags;
1086 tso = fm10k_tso(tx_ring, first);
1090 fm10k_tx_csum(tx_ring, first);
1092 fm10k_tx_map(tx_ring, first);
1094 return NETDEV_TX_OK;
1097 dev_kfree_skb_any(first->skb);
1100 return NETDEV_TX_OK;
1103 static u64 fm10k_get_tx_completed(struct fm10k_ring *ring)
1105 return ring->stats.packets;
1109 * fm10k_get_tx_pending - how many Tx descriptors not processed
1110 * @ring: the ring structure
1111 * @in_sw: is tx_pending being checked in SW or in HW?
1113 u64 fm10k_get_tx_pending(struct fm10k_ring *ring, bool in_sw)
1115 struct fm10k_intfc *interface = ring->q_vector->interface;
1116 struct fm10k_hw *hw = &interface->hw;
1119 if (likely(in_sw)) {
1120 head = ring->next_to_clean;
1121 tail = ring->next_to_use;
1123 head = fm10k_read_reg(hw, FM10K_TDH(ring->reg_idx));
1124 tail = fm10k_read_reg(hw, FM10K_TDT(ring->reg_idx));
1127 return ((head <= tail) ? tail : tail + ring->count) - head;
1130 bool fm10k_check_tx_hang(struct fm10k_ring *tx_ring)
1132 u32 tx_done = fm10k_get_tx_completed(tx_ring);
1133 u32 tx_done_old = tx_ring->tx_stats.tx_done_old;
1134 u32 tx_pending = fm10k_get_tx_pending(tx_ring, true);
1136 clear_check_for_tx_hang(tx_ring);
1138 /* Check for a hung queue, but be thorough. This verifies
1139 * that a transmit has been completed since the previous
1140 * check AND there is at least one packet pending. By
1141 * requiring this to fail twice we avoid races with
1142 * clearing the ARMED bit and conditions where we
1143 * run the check_tx_hang logic with a transmit completion
1144 * pending but without time to complete it yet.
1146 if (!tx_pending || (tx_done_old != tx_done)) {
1147 /* update completed stats and continue */
1148 tx_ring->tx_stats.tx_done_old = tx_done;
1149 /* reset the countdown */
1150 clear_bit(__FM10K_HANG_CHECK_ARMED, tx_ring->state);
1155 /* make sure it is true for two checks in a row */
1156 return test_and_set_bit(__FM10K_HANG_CHECK_ARMED, tx_ring->state);
1160 * fm10k_tx_timeout_reset - initiate reset due to Tx timeout
1161 * @interface: driver private struct
1163 void fm10k_tx_timeout_reset(struct fm10k_intfc *interface)
1165 /* Do the reset outside of interrupt context */
1166 if (!test_bit(__FM10K_DOWN, interface->state)) {
1167 interface->tx_timeout_count++;
1168 set_bit(FM10K_FLAG_RESET_REQUESTED, interface->flags);
1169 fm10k_service_event_schedule(interface);
1174 * fm10k_clean_tx_irq - Reclaim resources after transmit completes
1175 * @q_vector: structure containing interrupt and ring information
1176 * @tx_ring: tx ring to clean
1177 * @napi_budget: Used to determine if we are in netpoll
1179 static bool fm10k_clean_tx_irq(struct fm10k_q_vector *q_vector,
1180 struct fm10k_ring *tx_ring, int napi_budget)
1182 struct fm10k_intfc *interface = q_vector->interface;
1183 struct fm10k_tx_buffer *tx_buffer;
1184 struct fm10k_tx_desc *tx_desc;
1185 unsigned int total_bytes = 0, total_packets = 0;
1186 unsigned int budget = q_vector->tx.work_limit;
1187 unsigned int i = tx_ring->next_to_clean;
1189 if (test_bit(__FM10K_DOWN, interface->state))
1192 tx_buffer = &tx_ring->tx_buffer[i];
1193 tx_desc = FM10K_TX_DESC(tx_ring, i);
1194 i -= tx_ring->count;
1197 struct fm10k_tx_desc *eop_desc = tx_buffer->next_to_watch;
1199 /* if next_to_watch is not set then there is no work pending */
1203 /* prevent any other reads prior to eop_desc */
1206 /* if DD is not set pending work has not been completed */
1207 if (!(eop_desc->flags & FM10K_TXD_FLAG_DONE))
1210 /* clear next_to_watch to prevent false hangs */
1211 tx_buffer->next_to_watch = NULL;
1213 /* update the statistics for this packet */
1214 total_bytes += tx_buffer->bytecount;
1215 total_packets += tx_buffer->gso_segs;
1218 napi_consume_skb(tx_buffer->skb, napi_budget);
1220 /* unmap skb header data */
1221 dma_unmap_single(tx_ring->dev,
1222 dma_unmap_addr(tx_buffer, dma),
1223 dma_unmap_len(tx_buffer, len),
1226 /* clear tx_buffer data */
1227 tx_buffer->skb = NULL;
1228 dma_unmap_len_set(tx_buffer, len, 0);
1230 /* unmap remaining buffers */
1231 while (tx_desc != eop_desc) {
1236 i -= tx_ring->count;
1237 tx_buffer = tx_ring->tx_buffer;
1238 tx_desc = FM10K_TX_DESC(tx_ring, 0);
1241 /* unmap any remaining paged data */
1242 if (dma_unmap_len(tx_buffer, len)) {
1243 dma_unmap_page(tx_ring->dev,
1244 dma_unmap_addr(tx_buffer, dma),
1245 dma_unmap_len(tx_buffer, len),
1247 dma_unmap_len_set(tx_buffer, len, 0);
1251 /* move us one more past the eop_desc for start of next pkt */
1256 i -= tx_ring->count;
1257 tx_buffer = tx_ring->tx_buffer;
1258 tx_desc = FM10K_TX_DESC(tx_ring, 0);
1261 /* issue prefetch for next Tx descriptor */
1264 /* update budget accounting */
1266 } while (likely(budget));
1268 i += tx_ring->count;
1269 tx_ring->next_to_clean = i;
1270 u64_stats_update_begin(&tx_ring->syncp);
1271 tx_ring->stats.bytes += total_bytes;
1272 tx_ring->stats.packets += total_packets;
1273 u64_stats_update_end(&tx_ring->syncp);
1274 q_vector->tx.total_bytes += total_bytes;
1275 q_vector->tx.total_packets += total_packets;
1277 if (check_for_tx_hang(tx_ring) && fm10k_check_tx_hang(tx_ring)) {
1278 /* schedule immediate reset if we believe we hung */
1279 struct fm10k_hw *hw = &interface->hw;
1281 netif_err(interface, drv, tx_ring->netdev,
1282 "Detected Tx Unit Hang\n"
1284 " TDH, TDT <%x>, <%x>\n"
1285 " next_to_use <%x>\n"
1286 " next_to_clean <%x>\n",
1287 tx_ring->queue_index,
1288 fm10k_read_reg(hw, FM10K_TDH(tx_ring->reg_idx)),
1289 fm10k_read_reg(hw, FM10K_TDT(tx_ring->reg_idx)),
1290 tx_ring->next_to_use, i);
1292 netif_stop_subqueue(tx_ring->netdev,
1293 tx_ring->queue_index);
1295 netif_info(interface, probe, tx_ring->netdev,
1296 "tx hang %d detected on queue %d, resetting interface\n",
1297 interface->tx_timeout_count + 1,
1298 tx_ring->queue_index);
1300 fm10k_tx_timeout_reset(interface);
1302 /* the netdev is about to reset, no point in enabling stuff */
1306 /* notify netdev of completed buffers */
1307 netdev_tx_completed_queue(txring_txq(tx_ring),
1308 total_packets, total_bytes);
1310 #define TX_WAKE_THRESHOLD min_t(u16, FM10K_MIN_TXD - 1, DESC_NEEDED * 2)
1311 if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) &&
1312 (fm10k_desc_unused(tx_ring) >= TX_WAKE_THRESHOLD))) {
1313 /* Make sure that anybody stopping the queue after this
1314 * sees the new next_to_clean.
1317 if (__netif_subqueue_stopped(tx_ring->netdev,
1318 tx_ring->queue_index) &&
1319 !test_bit(__FM10K_DOWN, interface->state)) {
1320 netif_wake_subqueue(tx_ring->netdev,
1321 tx_ring->queue_index);
1322 ++tx_ring->tx_stats.restart_queue;
1330 * fm10k_update_itr - update the dynamic ITR value based on packet size
1332 * Stores a new ITR value based on strictly on packet size. The
1333 * divisors and thresholds used by this function were determined based
1334 * on theoretical maximum wire speed and testing data, in order to
1335 * minimize response time while increasing bulk throughput.
1337 * @ring_container: Container for rings to have ITR updated
1339 static void fm10k_update_itr(struct fm10k_ring_container *ring_container)
1341 unsigned int avg_wire_size, packets, itr_round;
1343 /* Only update ITR if we are using adaptive setting */
1344 if (!ITR_IS_ADAPTIVE(ring_container->itr))
1347 packets = ring_container->total_packets;
1351 avg_wire_size = ring_container->total_bytes / packets;
1353 /* The following is a crude approximation of:
1354 * wmem_default / (size + overhead) = desired_pkts_per_int
1355 * rate / bits_per_byte / (size + ethernet overhead) = pkt_rate
1356 * (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
1358 * Assuming wmem_default is 212992 and overhead is 640 bytes per
1359 * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
1362 * (34 * (size + 24)) / (size + 640) = ITR
1364 * We first do some math on the packet size and then finally bitshift
1365 * by 8 after rounding up. We also have to account for PCIe link speed
1366 * difference as ITR scales based on this.
1368 if (avg_wire_size <= 360) {
1369 /* Start at 250K ints/sec and gradually drop to 77K ints/sec */
1371 avg_wire_size += 376;
1372 } else if (avg_wire_size <= 1152) {
1373 /* 77K ints/sec to 45K ints/sec */
1375 avg_wire_size += 2176;
1376 } else if (avg_wire_size <= 1920) {
1377 /* 45K ints/sec to 38K ints/sec */
1378 avg_wire_size += 4480;
1380 /* plateau at a limit of 38K ints/sec */
1381 avg_wire_size = 6656;
1384 /* Perform final bitshift for division after rounding up to ensure
1385 * that the calculation will never get below a 1. The bit shift
1386 * accounts for changes in the ITR due to PCIe link speed.
1388 itr_round = READ_ONCE(ring_container->itr_scale) + 8;
1389 avg_wire_size += BIT(itr_round) - 1;
1390 avg_wire_size >>= itr_round;
1392 /* write back value and retain adaptive flag */
1393 ring_container->itr = avg_wire_size | FM10K_ITR_ADAPTIVE;
1396 ring_container->total_bytes = 0;
1397 ring_container->total_packets = 0;
1400 static void fm10k_qv_enable(struct fm10k_q_vector *q_vector)
1402 /* Enable auto-mask and clear the current mask */
1403 u32 itr = FM10K_ITR_ENABLE;
1406 fm10k_update_itr(&q_vector->tx);
1409 fm10k_update_itr(&q_vector->rx);
1411 /* Store Tx itr in timer slot 0 */
1412 itr |= (q_vector->tx.itr & FM10K_ITR_MAX);
1414 /* Shift Rx itr to timer slot 1 */
1415 itr |= (q_vector->rx.itr & FM10K_ITR_MAX) << FM10K_ITR_INTERVAL1_SHIFT;
1417 /* Write the final value to the ITR register */
1418 writel(itr, q_vector->itr);
1421 static int fm10k_poll(struct napi_struct *napi, int budget)
1423 struct fm10k_q_vector *q_vector =
1424 container_of(napi, struct fm10k_q_vector, napi);
1425 struct fm10k_ring *ring;
1426 int per_ring_budget, work_done = 0;
1427 bool clean_complete = true;
1429 fm10k_for_each_ring(ring, q_vector->tx) {
1430 if (!fm10k_clean_tx_irq(q_vector, ring, budget))
1431 clean_complete = false;
1434 /* Handle case where we are called by netpoll with a budget of 0 */
1438 /* attempt to distribute budget to each queue fairly, but don't
1439 * allow the budget to go below 1 because we'll exit polling
1441 if (q_vector->rx.count > 1)
1442 per_ring_budget = max(budget / q_vector->rx.count, 1);
1444 per_ring_budget = budget;
1446 fm10k_for_each_ring(ring, q_vector->rx) {
1447 int work = fm10k_clean_rx_irq(q_vector, ring, per_ring_budget);
1450 if (work >= per_ring_budget)
1451 clean_complete = false;
1454 /* If all work not completed, return budget and keep polling */
1455 if (!clean_complete)
1458 /* Exit the polling mode, but don't re-enable interrupts if stack might
1459 * poll us due to busy-polling
1461 if (likely(napi_complete_done(napi, work_done)))
1462 fm10k_qv_enable(q_vector);
1464 return min(work_done, budget - 1);
1468 * fm10k_set_qos_queues: Allocate queues for a QOS-enabled device
1469 * @interface: board private structure to initialize
1471 * When QoS (Quality of Service) is enabled, allocate queues for
1472 * each traffic class. If multiqueue isn't available,then abort QoS
1475 * This function handles all combinations of Qos and RSS.
1478 static bool fm10k_set_qos_queues(struct fm10k_intfc *interface)
1480 struct net_device *dev = interface->netdev;
1481 struct fm10k_ring_feature *f;
1485 /* Map queue offset and counts onto allocated tx queues */
1486 pcs = netdev_get_num_tc(dev);
1491 /* set QoS mask and indices */
1492 f = &interface->ring_feature[RING_F_QOS];
1494 f->mask = BIT(fls(pcs - 1)) - 1;
1496 /* determine the upper limit for our current DCB mode */
1497 rss_i = interface->hw.mac.max_queues / pcs;
1498 rss_i = BIT(fls(rss_i) - 1);
1500 /* set RSS mask and indices */
1501 f = &interface->ring_feature[RING_F_RSS];
1502 rss_i = min_t(u16, rss_i, f->limit);
1504 f->mask = BIT(fls(rss_i - 1)) - 1;
1506 /* configure pause class to queue mapping */
1507 for (i = 0; i < pcs; i++)
1508 netdev_set_tc_queue(dev, i, rss_i, rss_i * i);
1510 interface->num_rx_queues = rss_i * pcs;
1511 interface->num_tx_queues = rss_i * pcs;
1517 * fm10k_set_rss_queues: Allocate queues for RSS
1518 * @interface: board private structure to initialize
1520 * This is our "base" multiqueue mode. RSS (Receive Side Scaling) will try
1521 * to allocate one Rx queue per CPU, and if available, one Tx queue per CPU.
1524 static bool fm10k_set_rss_queues(struct fm10k_intfc *interface)
1526 struct fm10k_ring_feature *f;
1529 f = &interface->ring_feature[RING_F_RSS];
1530 rss_i = min_t(u16, interface->hw.mac.max_queues, f->limit);
1532 /* record indices and power of 2 mask for RSS */
1534 f->mask = BIT(fls(rss_i - 1)) - 1;
1536 interface->num_rx_queues = rss_i;
1537 interface->num_tx_queues = rss_i;
1543 * fm10k_set_num_queues: Allocate queues for device, feature dependent
1544 * @interface: board private structure to initialize
1546 * This is the top level queue allocation routine. The order here is very
1547 * important, starting with the "most" number of features turned on at once,
1548 * and ending with the smallest set of features. This way large combinations
1549 * can be allocated if they're turned on, and smaller combinations are the
1550 * fall through conditions.
1553 static void fm10k_set_num_queues(struct fm10k_intfc *interface)
1555 /* Attempt to setup QoS and RSS first */
1556 if (fm10k_set_qos_queues(interface))
1559 /* If we don't have QoS, just fallback to only RSS. */
1560 fm10k_set_rss_queues(interface);
1564 * fm10k_reset_num_queues - Reset the number of queues to zero
1565 * @interface: board private structure
1567 * This function should be called whenever we need to reset the number of
1568 * queues after an error condition.
1570 static void fm10k_reset_num_queues(struct fm10k_intfc *interface)
1572 interface->num_tx_queues = 0;
1573 interface->num_rx_queues = 0;
1574 interface->num_q_vectors = 0;
1578 * fm10k_alloc_q_vector - Allocate memory for a single interrupt vector
1579 * @interface: board private structure to initialize
1580 * @v_count: q_vectors allocated on interface, used for ring interleaving
1581 * @v_idx: index of vector in interface struct
1582 * @txr_count: total number of Tx rings to allocate
1583 * @txr_idx: index of first Tx ring to allocate
1584 * @rxr_count: total number of Rx rings to allocate
1585 * @rxr_idx: index of first Rx ring to allocate
1587 * We allocate one q_vector. If allocation fails we return -ENOMEM.
1589 static int fm10k_alloc_q_vector(struct fm10k_intfc *interface,
1590 unsigned int v_count, unsigned int v_idx,
1591 unsigned int txr_count, unsigned int txr_idx,
1592 unsigned int rxr_count, unsigned int rxr_idx)
1594 struct fm10k_q_vector *q_vector;
1595 struct fm10k_ring *ring;
1598 ring_count = txr_count + rxr_count;
1600 /* allocate q_vector and rings */
1601 q_vector = kzalloc(struct_size(q_vector, ring, ring_count), GFP_KERNEL);
1605 /* initialize NAPI */
1606 netif_napi_add(interface->netdev, &q_vector->napi,
1607 fm10k_poll, NAPI_POLL_WEIGHT);
1609 /* tie q_vector and interface together */
1610 interface->q_vector[v_idx] = q_vector;
1611 q_vector->interface = interface;
1612 q_vector->v_idx = v_idx;
1614 /* initialize pointer to rings */
1615 ring = q_vector->ring;
1617 /* save Tx ring container info */
1618 q_vector->tx.ring = ring;
1619 q_vector->tx.work_limit = FM10K_DEFAULT_TX_WORK;
1620 q_vector->tx.itr = interface->tx_itr;
1621 q_vector->tx.itr_scale = interface->hw.mac.itr_scale;
1622 q_vector->tx.count = txr_count;
1625 /* assign generic ring traits */
1626 ring->dev = &interface->pdev->dev;
1627 ring->netdev = interface->netdev;
1629 /* configure backlink on ring */
1630 ring->q_vector = q_vector;
1632 /* apply Tx specific ring traits */
1633 ring->count = interface->tx_ring_count;
1634 ring->queue_index = txr_idx;
1636 /* assign ring to interface */
1637 interface->tx_ring[txr_idx] = ring;
1639 /* update count and index */
1643 /* push pointer to next ring */
1647 /* save Rx ring container info */
1648 q_vector->rx.ring = ring;
1649 q_vector->rx.itr = interface->rx_itr;
1650 q_vector->rx.itr_scale = interface->hw.mac.itr_scale;
1651 q_vector->rx.count = rxr_count;
1654 /* assign generic ring traits */
1655 ring->dev = &interface->pdev->dev;
1656 ring->netdev = interface->netdev;
1657 rcu_assign_pointer(ring->l2_accel, interface->l2_accel);
1659 /* configure backlink on ring */
1660 ring->q_vector = q_vector;
1662 /* apply Rx specific ring traits */
1663 ring->count = interface->rx_ring_count;
1664 ring->queue_index = rxr_idx;
1666 /* assign ring to interface */
1667 interface->rx_ring[rxr_idx] = ring;
1669 /* update count and index */
1673 /* push pointer to next ring */
1677 fm10k_dbg_q_vector_init(q_vector);
1683 * fm10k_free_q_vector - Free memory allocated for specific interrupt vector
1684 * @interface: board private structure to initialize
1685 * @v_idx: Index of vector to be freed
1687 * This function frees the memory allocated to the q_vector. In addition if
1688 * NAPI is enabled it will delete any references to the NAPI struct prior
1689 * to freeing the q_vector.
1691 static void fm10k_free_q_vector(struct fm10k_intfc *interface, int v_idx)
1693 struct fm10k_q_vector *q_vector = interface->q_vector[v_idx];
1694 struct fm10k_ring *ring;
1696 fm10k_dbg_q_vector_exit(q_vector);
1698 fm10k_for_each_ring(ring, q_vector->tx)
1699 interface->tx_ring[ring->queue_index] = NULL;
1701 fm10k_for_each_ring(ring, q_vector->rx)
1702 interface->rx_ring[ring->queue_index] = NULL;
1704 interface->q_vector[v_idx] = NULL;
1705 netif_napi_del(&q_vector->napi);
1706 kfree_rcu(q_vector, rcu);
1710 * fm10k_alloc_q_vectors - Allocate memory for interrupt vectors
1711 * @interface: board private structure to initialize
1713 * We allocate one q_vector per queue interrupt. If allocation fails we
1716 static int fm10k_alloc_q_vectors(struct fm10k_intfc *interface)
1718 unsigned int q_vectors = interface->num_q_vectors;
1719 unsigned int rxr_remaining = interface->num_rx_queues;
1720 unsigned int txr_remaining = interface->num_tx_queues;
1721 unsigned int rxr_idx = 0, txr_idx = 0, v_idx = 0;
1724 if (q_vectors >= (rxr_remaining + txr_remaining)) {
1725 for (; rxr_remaining; v_idx++) {
1726 err = fm10k_alloc_q_vector(interface, q_vectors, v_idx,
1731 /* update counts and index */
1737 for (; v_idx < q_vectors; v_idx++) {
1738 int rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors - v_idx);
1739 int tqpv = DIV_ROUND_UP(txr_remaining, q_vectors - v_idx);
1741 err = fm10k_alloc_q_vector(interface, q_vectors, v_idx,
1748 /* update counts and index */
1749 rxr_remaining -= rqpv;
1750 txr_remaining -= tqpv;
1758 fm10k_reset_num_queues(interface);
1761 fm10k_free_q_vector(interface, v_idx);
1767 * fm10k_free_q_vectors - Free memory allocated for interrupt vectors
1768 * @interface: board private structure to initialize
1770 * This function frees the memory allocated to the q_vectors. In addition if
1771 * NAPI is enabled it will delete any references to the NAPI struct prior
1772 * to freeing the q_vector.
1774 static void fm10k_free_q_vectors(struct fm10k_intfc *interface)
1776 int v_idx = interface->num_q_vectors;
1778 fm10k_reset_num_queues(interface);
1781 fm10k_free_q_vector(interface, v_idx);
1785 * f10k_reset_msix_capability - reset MSI-X capability
1786 * @interface: board private structure to initialize
1788 * Reset the MSI-X capability back to its starting state
1790 static void fm10k_reset_msix_capability(struct fm10k_intfc *interface)
1792 pci_disable_msix(interface->pdev);
1793 kfree(interface->msix_entries);
1794 interface->msix_entries = NULL;
1798 * f10k_init_msix_capability - configure MSI-X capability
1799 * @interface: board private structure to initialize
1801 * Attempt to configure the interrupts using the best available
1802 * capabilities of the hardware and the kernel.
1804 static int fm10k_init_msix_capability(struct fm10k_intfc *interface)
1806 struct fm10k_hw *hw = &interface->hw;
1807 int v_budget, vector;
1809 /* It's easy to be greedy for MSI-X vectors, but it really
1810 * doesn't do us much good if we have a lot more vectors
1811 * than CPU's. So let's be conservative and only ask for
1812 * (roughly) the same number of vectors as there are CPU's.
1813 * the default is to use pairs of vectors
1815 v_budget = max(interface->num_rx_queues, interface->num_tx_queues);
1816 v_budget = min_t(u16, v_budget, num_online_cpus());
1818 /* account for vectors not related to queues */
1819 v_budget += NON_Q_VECTORS;
1821 /* At the same time, hardware can only support a maximum of
1822 * hw.mac->max_msix_vectors vectors. With features
1823 * such as RSS and VMDq, we can easily surpass the number of Rx and Tx
1824 * descriptor queues supported by our device. Thus, we cap it off in
1825 * those rare cases where the cpu count also exceeds our vector limit.
1827 v_budget = min_t(int, v_budget, hw->mac.max_msix_vectors);
1829 /* A failure in MSI-X entry allocation is fatal. */
1830 interface->msix_entries = kcalloc(v_budget, sizeof(struct msix_entry),
1832 if (!interface->msix_entries)
1835 /* populate entry values */
1836 for (vector = 0; vector < v_budget; vector++)
1837 interface->msix_entries[vector].entry = vector;
1839 /* Attempt to enable MSI-X with requested value */
1840 v_budget = pci_enable_msix_range(interface->pdev,
1841 interface->msix_entries,
1845 kfree(interface->msix_entries);
1846 interface->msix_entries = NULL;
1850 /* record the number of queues available for q_vectors */
1851 interface->num_q_vectors = v_budget - NON_Q_VECTORS;
1857 * fm10k_cache_ring_qos - Descriptor ring to register mapping for QoS
1858 * @interface: Interface structure continaining rings and devices
1860 * Cache the descriptor ring offsets for Qos
1862 static bool fm10k_cache_ring_qos(struct fm10k_intfc *interface)
1864 struct net_device *dev = interface->netdev;
1865 int pc, offset, rss_i, i;
1866 u16 pc_stride = interface->ring_feature[RING_F_QOS].mask + 1;
1867 u8 num_pcs = netdev_get_num_tc(dev);
1872 rss_i = interface->ring_feature[RING_F_RSS].indices;
1874 for (pc = 0, offset = 0; pc < num_pcs; pc++, offset += rss_i) {
1877 for (i = 0; i < rss_i; i++) {
1878 interface->tx_ring[offset + i]->reg_idx = q_idx;
1879 interface->tx_ring[offset + i]->qos_pc = pc;
1880 interface->rx_ring[offset + i]->reg_idx = q_idx;
1881 interface->rx_ring[offset + i]->qos_pc = pc;
1890 * fm10k_cache_ring_rss - Descriptor ring to register mapping for RSS
1891 * @interface: Interface structure continaining rings and devices
1893 * Cache the descriptor ring offsets for RSS
1895 static void fm10k_cache_ring_rss(struct fm10k_intfc *interface)
1899 for (i = 0; i < interface->num_rx_queues; i++)
1900 interface->rx_ring[i]->reg_idx = i;
1902 for (i = 0; i < interface->num_tx_queues; i++)
1903 interface->tx_ring[i]->reg_idx = i;
1907 * fm10k_assign_rings - Map rings to network devices
1908 * @interface: Interface structure containing rings and devices
1910 * This function is meant to go though and configure both the network
1911 * devices so that they contain rings, and configure the rings so that
1912 * they function with their network devices.
1914 static void fm10k_assign_rings(struct fm10k_intfc *interface)
1916 if (fm10k_cache_ring_qos(interface))
1919 fm10k_cache_ring_rss(interface);
1922 static void fm10k_init_reta(struct fm10k_intfc *interface)
1924 u16 i, rss_i = interface->ring_feature[RING_F_RSS].indices;
1927 /* If the Rx flow indirection table has been configured manually, we
1928 * need to maintain it when possible.
1930 if (netif_is_rxfh_configured(interface->netdev)) {
1931 for (i = FM10K_RETA_SIZE; i--;) {
1932 reta = interface->reta[i];
1933 if ((((reta << 24) >> 24) < rss_i) &&
1934 (((reta << 16) >> 24) < rss_i) &&
1935 (((reta << 8) >> 24) < rss_i) &&
1936 (((reta) >> 24) < rss_i))
1939 /* this should never happen */
1940 dev_err(&interface->pdev->dev,
1941 "RSS indirection table assigned flows out of queue bounds. Reconfiguring.\n");
1942 goto repopulate_reta;
1945 /* do nothing if all of the elements are in bounds */
1950 fm10k_write_reta(interface, NULL);
1954 * fm10k_init_queueing_scheme - Determine proper queueing scheme
1955 * @interface: board private structure to initialize
1957 * We determine which queueing scheme to use based on...
1958 * - Hardware queue count (num_*_queues)
1959 * - defined by miscellaneous hardware support/features (RSS, etc.)
1961 int fm10k_init_queueing_scheme(struct fm10k_intfc *interface)
1965 /* Number of supported queues */
1966 fm10k_set_num_queues(interface);
1968 /* Configure MSI-X capability */
1969 err = fm10k_init_msix_capability(interface);
1971 dev_err(&interface->pdev->dev,
1972 "Unable to initialize MSI-X capability\n");
1976 /* Allocate memory for queues */
1977 err = fm10k_alloc_q_vectors(interface);
1979 dev_err(&interface->pdev->dev,
1980 "Unable to allocate queue vectors\n");
1981 goto err_alloc_q_vectors;
1984 /* Map rings to devices, and map devices to physical queues */
1985 fm10k_assign_rings(interface);
1987 /* Initialize RSS redirection table */
1988 fm10k_init_reta(interface);
1992 err_alloc_q_vectors:
1993 fm10k_reset_msix_capability(interface);
1995 fm10k_reset_num_queues(interface);
2000 * fm10k_clear_queueing_scheme - Clear the current queueing scheme settings
2001 * @interface: board private structure to clear queueing scheme on
2003 * We go through and clear queueing specific resources and reset the structure
2004 * to pre-load conditions
2006 void fm10k_clear_queueing_scheme(struct fm10k_intfc *interface)
2008 fm10k_free_q_vectors(interface);
2009 fm10k_reset_msix_capability(interface);