1 /* Intel(R) Ethernet Switch Host Interface Driver
2 * Copyright(c) 2013 - 2016 Intel Corporation.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * The full GNU General Public License is included in this distribution in
14 * the file called "COPYING".
16 * Contact Information:
17 * e1000-devel Mailing List <e1000-devel@lists.sourceforge.net>
18 * Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497
21 #include <linux/types.h>
22 #include <linux/module.h>
26 #include <linux/if_macvlan.h>
27 #include <linux/prefetch.h>
31 #define DRV_VERSION "0.21.2-k"
32 #define DRV_SUMMARY "Intel(R) Ethernet Switch Host Interface Driver"
33 const char fm10k_driver_version
[] = DRV_VERSION
;
34 char fm10k_driver_name
[] = "fm10k";
35 static const char fm10k_driver_string
[] = DRV_SUMMARY
;
36 static const char fm10k_copyright
[] =
37 "Copyright (c) 2013 - 2016 Intel Corporation.";
39 MODULE_AUTHOR("Intel Corporation, <linux.nics@intel.com>");
40 MODULE_DESCRIPTION(DRV_SUMMARY
);
41 MODULE_LICENSE("GPL");
42 MODULE_VERSION(DRV_VERSION
);
44 /* single workqueue for entire fm10k driver */
45 struct workqueue_struct
*fm10k_workqueue
;
48 * fm10k_init_module - Driver Registration Routine
50 * fm10k_init_module is the first routine called when the driver is
51 * loaded. All it does is register with the PCI subsystem.
53 static int __init
fm10k_init_module(void)
55 pr_info("%s - version %s\n", fm10k_driver_string
, fm10k_driver_version
);
56 pr_info("%s\n", fm10k_copyright
);
58 /* create driver workqueue */
59 fm10k_workqueue
= alloc_workqueue("fm10k", WQ_MEM_RECLAIM
, 0);
63 return fm10k_register_pci_driver();
65 module_init(fm10k_init_module
);
68 * fm10k_exit_module - Driver Exit Cleanup Routine
70 * fm10k_exit_module is called just before the driver is removed
73 static void __exit
fm10k_exit_module(void)
75 fm10k_unregister_pci_driver();
79 /* destroy driver workqueue */
80 destroy_workqueue(fm10k_workqueue
);
82 module_exit(fm10k_exit_module
);
84 static bool fm10k_alloc_mapped_page(struct fm10k_ring
*rx_ring
,
85 struct fm10k_rx_buffer
*bi
)
87 struct page
*page
= bi
->page
;
90 /* Only page will be NULL if buffer was consumed */
94 /* alloc new page for storage */
95 page
= dev_alloc_page();
96 if (unlikely(!page
)) {
97 rx_ring
->rx_stats
.alloc_failed
++;
101 /* map page for use */
102 dma
= dma_map_page(rx_ring
->dev
, page
, 0, PAGE_SIZE
, DMA_FROM_DEVICE
);
104 /* if mapping failed free memory back to system since
105 * there isn't much point in holding memory we can't use
107 if (dma_mapping_error(rx_ring
->dev
, dma
)) {
110 rx_ring
->rx_stats
.alloc_failed
++;
122 * fm10k_alloc_rx_buffers - Replace used receive buffers
123 * @rx_ring: ring to place buffers on
124 * @cleaned_count: number of buffers to replace
126 void fm10k_alloc_rx_buffers(struct fm10k_ring
*rx_ring
, u16 cleaned_count
)
128 union fm10k_rx_desc
*rx_desc
;
129 struct fm10k_rx_buffer
*bi
;
130 u16 i
= rx_ring
->next_to_use
;
136 rx_desc
= FM10K_RX_DESC(rx_ring
, i
);
137 bi
= &rx_ring
->rx_buffer
[i
];
141 if (!fm10k_alloc_mapped_page(rx_ring
, bi
))
144 /* Refresh the desc even if buffer_addrs didn't change
145 * because each write-back erases this info.
147 rx_desc
->q
.pkt_addr
= cpu_to_le64(bi
->dma
+ bi
->page_offset
);
153 rx_desc
= FM10K_RX_DESC(rx_ring
, 0);
154 bi
= rx_ring
->rx_buffer
;
158 /* clear the status bits for the next_to_use descriptor */
159 rx_desc
->d
.staterr
= 0;
162 } while (cleaned_count
);
166 if (rx_ring
->next_to_use
!= i
) {
167 /* record the next descriptor to use */
168 rx_ring
->next_to_use
= i
;
170 /* update next to alloc since we have filled the ring */
171 rx_ring
->next_to_alloc
= i
;
173 /* Force memory writes to complete before letting h/w
174 * know there are new descriptors to fetch. (Only
175 * applicable for weak-ordered memory model archs,
180 /* notify hardware of new descriptors */
181 writel(i
, rx_ring
->tail
);
186 * fm10k_reuse_rx_page - page flip buffer and store it back on the ring
187 * @rx_ring: rx descriptor ring to store buffers on
188 * @old_buff: donor buffer to have page reused
190 * Synchronizes page for reuse by the interface
192 static void fm10k_reuse_rx_page(struct fm10k_ring
*rx_ring
,
193 struct fm10k_rx_buffer
*old_buff
)
195 struct fm10k_rx_buffer
*new_buff
;
196 u16 nta
= rx_ring
->next_to_alloc
;
198 new_buff
= &rx_ring
->rx_buffer
[nta
];
200 /* update, and store next to alloc */
202 rx_ring
->next_to_alloc
= (nta
< rx_ring
->count
) ? nta
: 0;
204 /* transfer page from old buffer to new buffer */
205 *new_buff
= *old_buff
;
207 /* sync the buffer for use by the device */
208 dma_sync_single_range_for_device(rx_ring
->dev
, old_buff
->dma
,
209 old_buff
->page_offset
,
214 static inline bool fm10k_page_is_reserved(struct page
*page
)
216 return (page_to_nid(page
) != numa_mem_id()) || page_is_pfmemalloc(page
);
219 static bool fm10k_can_reuse_rx_page(struct fm10k_rx_buffer
*rx_buffer
,
221 unsigned int __maybe_unused truesize
)
223 /* avoid re-using remote pages */
224 if (unlikely(fm10k_page_is_reserved(page
)))
227 #if (PAGE_SIZE < 8192)
228 /* if we are only owner of page we can reuse it */
229 if (unlikely(page_count(page
) != 1))
232 /* flip page offset to other buffer */
233 rx_buffer
->page_offset
^= FM10K_RX_BUFSZ
;
235 /* move offset up to the next cache line */
236 rx_buffer
->page_offset
+= truesize
;
238 if (rx_buffer
->page_offset
> (PAGE_SIZE
- FM10K_RX_BUFSZ
))
242 /* Even if we own the page, we are not allowed to use atomic_set()
243 * This would break get_page_unless_zero() users.
251 * fm10k_add_rx_frag - Add contents of Rx buffer to sk_buff
252 * @rx_buffer: buffer containing page to add
253 * @rx_desc: descriptor containing length of buffer written by hardware
254 * @skb: sk_buff to place the data into
256 * This function will add the data contained in rx_buffer->page to the skb.
257 * This is done either through a direct copy if the data in the buffer is
258 * less than the skb header size, otherwise it will just attach the page as
261 * The function will then update the page offset if necessary and return
262 * true if the buffer can be reused by the interface.
264 static bool fm10k_add_rx_frag(struct fm10k_rx_buffer
*rx_buffer
,
265 union fm10k_rx_desc
*rx_desc
,
268 struct page
*page
= rx_buffer
->page
;
269 unsigned char *va
= page_address(page
) + rx_buffer
->page_offset
;
270 unsigned int size
= le16_to_cpu(rx_desc
->w
.length
);
271 #if (PAGE_SIZE < 8192)
272 unsigned int truesize
= FM10K_RX_BUFSZ
;
274 unsigned int truesize
= ALIGN(size
, 512);
276 unsigned int pull_len
;
278 if (unlikely(skb_is_nonlinear(skb
)))
281 if (likely(size
<= FM10K_RX_HDR_LEN
)) {
282 memcpy(__skb_put(skb
, size
), va
, ALIGN(size
, sizeof(long)));
284 /* page is not reserved, we can reuse buffer as-is */
285 if (likely(!fm10k_page_is_reserved(page
)))
288 /* this page cannot be reused so discard it */
293 /* we need the header to contain the greater of either ETH_HLEN or
294 * 60 bytes if the skb->len is less than 60 for skb_pad.
296 pull_len
= eth_get_headlen(va
, FM10K_RX_HDR_LEN
);
298 /* align pull length to size of long to optimize memcpy performance */
299 memcpy(__skb_put(skb
, pull_len
), va
, ALIGN(pull_len
, sizeof(long)));
301 /* update all of the pointers */
306 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
,
307 (unsigned long)va
& ~PAGE_MASK
, size
, truesize
);
309 return fm10k_can_reuse_rx_page(rx_buffer
, page
, truesize
);
312 static struct sk_buff
*fm10k_fetch_rx_buffer(struct fm10k_ring
*rx_ring
,
313 union fm10k_rx_desc
*rx_desc
,
316 struct fm10k_rx_buffer
*rx_buffer
;
319 rx_buffer
= &rx_ring
->rx_buffer
[rx_ring
->next_to_clean
];
320 page
= rx_buffer
->page
;
324 void *page_addr
= page_address(page
) +
325 rx_buffer
->page_offset
;
327 /* prefetch first cache line of first page */
329 #if L1_CACHE_BYTES < 128
330 prefetch(page_addr
+ L1_CACHE_BYTES
);
333 /* allocate a skb to store the frags */
334 skb
= napi_alloc_skb(&rx_ring
->q_vector
->napi
,
336 if (unlikely(!skb
)) {
337 rx_ring
->rx_stats
.alloc_failed
++;
341 /* we will be copying header into skb->data in
342 * pskb_may_pull so it is in our interest to prefetch
343 * it now to avoid a possible cache miss
345 prefetchw(skb
->data
);
348 /* we are reusing so sync this buffer for CPU use */
349 dma_sync_single_range_for_cpu(rx_ring
->dev
,
351 rx_buffer
->page_offset
,
355 /* pull page into skb */
356 if (fm10k_add_rx_frag(rx_buffer
, rx_desc
, skb
)) {
357 /* hand second half of page back to the ring */
358 fm10k_reuse_rx_page(rx_ring
, rx_buffer
);
360 /* we are not reusing the buffer so unmap it */
361 dma_unmap_page(rx_ring
->dev
, rx_buffer
->dma
,
362 PAGE_SIZE
, DMA_FROM_DEVICE
);
365 /* clear contents of rx_buffer */
366 rx_buffer
->page
= NULL
;
371 static inline void fm10k_rx_checksum(struct fm10k_ring
*ring
,
372 union fm10k_rx_desc
*rx_desc
,
375 skb_checksum_none_assert(skb
);
377 /* Rx checksum disabled via ethtool */
378 if (!(ring
->netdev
->features
& NETIF_F_RXCSUM
))
381 /* TCP/UDP checksum error bit is set */
382 if (fm10k_test_staterr(rx_desc
,
383 FM10K_RXD_STATUS_L4E
|
384 FM10K_RXD_STATUS_L4E2
|
385 FM10K_RXD_STATUS_IPE
|
386 FM10K_RXD_STATUS_IPE2
)) {
387 ring
->rx_stats
.csum_err
++;
391 /* It must be a TCP or UDP packet with a valid checksum */
392 if (fm10k_test_staterr(rx_desc
, FM10K_RXD_STATUS_L4CS2
))
393 skb
->encapsulation
= true;
394 else if (!fm10k_test_staterr(rx_desc
, FM10K_RXD_STATUS_L4CS
))
397 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
399 ring
->rx_stats
.csum_good
++;
402 #define FM10K_RSS_L4_TYPES_MASK \
403 (BIT(FM10K_RSSTYPE_IPV4_TCP) | \
404 BIT(FM10K_RSSTYPE_IPV4_UDP) | \
405 BIT(FM10K_RSSTYPE_IPV6_TCP) | \
406 BIT(FM10K_RSSTYPE_IPV6_UDP))
408 static inline void fm10k_rx_hash(struct fm10k_ring
*ring
,
409 union fm10k_rx_desc
*rx_desc
,
414 if (!(ring
->netdev
->features
& NETIF_F_RXHASH
))
417 rss_type
= le16_to_cpu(rx_desc
->w
.pkt_info
) & FM10K_RXD_RSSTYPE_MASK
;
421 skb_set_hash(skb
, le32_to_cpu(rx_desc
->d
.rss
),
422 (BIT(rss_type
) & FM10K_RSS_L4_TYPES_MASK
) ?
423 PKT_HASH_TYPE_L4
: PKT_HASH_TYPE_L3
);
426 static void fm10k_type_trans(struct fm10k_ring
*rx_ring
,
427 union fm10k_rx_desc __maybe_unused
*rx_desc
,
430 struct net_device
*dev
= rx_ring
->netdev
;
431 struct fm10k_l2_accel
*l2_accel
= rcu_dereference_bh(rx_ring
->l2_accel
);
433 /* check to see if DGLORT belongs to a MACVLAN */
435 u16 idx
= le16_to_cpu(FM10K_CB(skb
)->fi
.w
.dglort
) - 1;
437 idx
-= l2_accel
->dglort
;
438 if (idx
< l2_accel
->size
&& l2_accel
->macvlan
[idx
])
439 dev
= l2_accel
->macvlan
[idx
];
444 skb
->protocol
= eth_type_trans(skb
, dev
);
449 /* update MACVLAN statistics */
450 macvlan_count_rx(netdev_priv(dev
), skb
->len
+ ETH_HLEN
, 1,
451 !!(rx_desc
->w
.hdr_info
&
452 cpu_to_le16(FM10K_RXD_HDR_INFO_XC_MASK
)));
456 * fm10k_process_skb_fields - Populate skb header fields from Rx descriptor
457 * @rx_ring: rx descriptor ring packet is being transacted on
458 * @rx_desc: pointer to the EOP Rx descriptor
459 * @skb: pointer to current skb being populated
461 * This function checks the ring, descriptor, and packet information in
462 * order to populate the hash, checksum, VLAN, timestamp, protocol, and
463 * other fields within the skb.
465 static unsigned int fm10k_process_skb_fields(struct fm10k_ring
*rx_ring
,
466 union fm10k_rx_desc
*rx_desc
,
469 unsigned int len
= skb
->len
;
471 fm10k_rx_hash(rx_ring
, rx_desc
, skb
);
473 fm10k_rx_checksum(rx_ring
, rx_desc
, skb
);
475 FM10K_CB(skb
)->fi
.w
.vlan
= rx_desc
->w
.vlan
;
477 skb_record_rx_queue(skb
, rx_ring
->queue_index
);
479 FM10K_CB(skb
)->fi
.d
.glort
= rx_desc
->d
.glort
;
481 if (rx_desc
->w
.vlan
) {
482 u16 vid
= le16_to_cpu(rx_desc
->w
.vlan
);
484 if ((vid
& VLAN_VID_MASK
) != rx_ring
->vid
)
485 __vlan_hwaccel_put_tag(skb
, htons(ETH_P_8021Q
), vid
);
486 else if (vid
& VLAN_PRIO_MASK
)
487 __vlan_hwaccel_put_tag(skb
, htons(ETH_P_8021Q
),
488 vid
& VLAN_PRIO_MASK
);
491 fm10k_type_trans(rx_ring
, rx_desc
, skb
);
497 * fm10k_is_non_eop - process handling of non-EOP buffers
498 * @rx_ring: Rx ring being processed
499 * @rx_desc: Rx descriptor for current buffer
501 * This function updates next to clean. If the buffer is an EOP buffer
502 * this function exits returning false, otherwise it will place the
503 * sk_buff in the next buffer to be chained and return true indicating
504 * that this is in fact a non-EOP buffer.
506 static bool fm10k_is_non_eop(struct fm10k_ring
*rx_ring
,
507 union fm10k_rx_desc
*rx_desc
)
509 u32 ntc
= rx_ring
->next_to_clean
+ 1;
511 /* fetch, update, and store next to clean */
512 ntc
= (ntc
< rx_ring
->count
) ? ntc
: 0;
513 rx_ring
->next_to_clean
= ntc
;
515 prefetch(FM10K_RX_DESC(rx_ring
, ntc
));
517 if (likely(fm10k_test_staterr(rx_desc
, FM10K_RXD_STATUS_EOP
)))
524 * fm10k_cleanup_headers - Correct corrupted or empty headers
525 * @rx_ring: rx descriptor ring packet is being transacted on
526 * @rx_desc: pointer to the EOP Rx descriptor
527 * @skb: pointer to current skb being fixed
529 * Address the case where we are pulling data in on pages only
530 * and as such no data is present in the skb header.
532 * In addition if skb is not at least 60 bytes we need to pad it so that
533 * it is large enough to qualify as a valid Ethernet frame.
535 * Returns true if an error was encountered and skb was freed.
537 static bool fm10k_cleanup_headers(struct fm10k_ring
*rx_ring
,
538 union fm10k_rx_desc
*rx_desc
,
541 if (unlikely((fm10k_test_staterr(rx_desc
,
542 FM10K_RXD_STATUS_RXE
)))) {
543 #define FM10K_TEST_RXD_BIT(rxd, bit) \
544 ((rxd)->w.csum_err & cpu_to_le16(bit))
545 if (FM10K_TEST_RXD_BIT(rx_desc
, FM10K_RXD_ERR_SWITCH_ERROR
))
546 rx_ring
->rx_stats
.switch_errors
++;
547 if (FM10K_TEST_RXD_BIT(rx_desc
, FM10K_RXD_ERR_NO_DESCRIPTOR
))
548 rx_ring
->rx_stats
.drops
++;
549 if (FM10K_TEST_RXD_BIT(rx_desc
, FM10K_RXD_ERR_PP_ERROR
))
550 rx_ring
->rx_stats
.pp_errors
++;
551 if (FM10K_TEST_RXD_BIT(rx_desc
, FM10K_RXD_ERR_SWITCH_READY
))
552 rx_ring
->rx_stats
.link_errors
++;
553 if (FM10K_TEST_RXD_BIT(rx_desc
, FM10K_RXD_ERR_TOO_BIG
))
554 rx_ring
->rx_stats
.length_errors
++;
555 dev_kfree_skb_any(skb
);
556 rx_ring
->rx_stats
.errors
++;
560 /* if eth_skb_pad returns an error the skb was freed */
561 if (eth_skb_pad(skb
))
568 * fm10k_receive_skb - helper function to handle rx indications
569 * @q_vector: structure containing interrupt and ring information
570 * @skb: packet to send up
572 static void fm10k_receive_skb(struct fm10k_q_vector
*q_vector
,
575 napi_gro_receive(&q_vector
->napi
, skb
);
578 static int fm10k_clean_rx_irq(struct fm10k_q_vector
*q_vector
,
579 struct fm10k_ring
*rx_ring
,
582 struct sk_buff
*skb
= rx_ring
->skb
;
583 unsigned int total_bytes
= 0, total_packets
= 0;
584 u16 cleaned_count
= fm10k_desc_unused(rx_ring
);
586 while (likely(total_packets
< budget
)) {
587 union fm10k_rx_desc
*rx_desc
;
589 /* return some buffers to hardware, one at a time is too slow */
590 if (cleaned_count
>= FM10K_RX_BUFFER_WRITE
) {
591 fm10k_alloc_rx_buffers(rx_ring
, cleaned_count
);
595 rx_desc
= FM10K_RX_DESC(rx_ring
, rx_ring
->next_to_clean
);
597 if (!rx_desc
->d
.staterr
)
600 /* This memory barrier is needed to keep us from reading
601 * any other fields out of the rx_desc until we know the
602 * descriptor has been written back
606 /* retrieve a buffer from the ring */
607 skb
= fm10k_fetch_rx_buffer(rx_ring
, rx_desc
, skb
);
609 /* exit if we failed to retrieve a buffer */
615 /* fetch next buffer in frame if non-eop */
616 if (fm10k_is_non_eop(rx_ring
, rx_desc
))
619 /* verify the packet layout is correct */
620 if (fm10k_cleanup_headers(rx_ring
, rx_desc
, skb
)) {
625 /* populate checksum, timestamp, VLAN, and protocol */
626 total_bytes
+= fm10k_process_skb_fields(rx_ring
, rx_desc
, skb
);
628 fm10k_receive_skb(q_vector
, skb
);
630 /* reset skb pointer */
633 /* update budget accounting */
637 /* place incomplete frames back on ring for completion */
640 u64_stats_update_begin(&rx_ring
->syncp
);
641 rx_ring
->stats
.packets
+= total_packets
;
642 rx_ring
->stats
.bytes
+= total_bytes
;
643 u64_stats_update_end(&rx_ring
->syncp
);
644 q_vector
->rx
.total_packets
+= total_packets
;
645 q_vector
->rx
.total_bytes
+= total_bytes
;
647 return total_packets
;
650 #define VXLAN_HLEN (sizeof(struct udphdr) + 8)
651 static struct ethhdr
*fm10k_port_is_vxlan(struct sk_buff
*skb
)
653 struct fm10k_intfc
*interface
= netdev_priv(skb
->dev
);
654 struct fm10k_vxlan_port
*vxlan_port
;
656 /* we can only offload a vxlan if we recognize it as such */
657 vxlan_port
= list_first_entry_or_null(&interface
->vxlan_port
,
658 struct fm10k_vxlan_port
, list
);
662 if (vxlan_port
->port
!= udp_hdr(skb
)->dest
)
665 /* return offset of udp_hdr plus 8 bytes for VXLAN header */
666 return (struct ethhdr
*)(skb_transport_header(skb
) + VXLAN_HLEN
);
669 #define FM10K_NVGRE_RESERVED0_FLAGS htons(0x9FFF)
670 #define NVGRE_TNI htons(0x2000)
671 struct fm10k_nvgre_hdr
{
677 static struct ethhdr
*fm10k_gre_is_nvgre(struct sk_buff
*skb
)
679 struct fm10k_nvgre_hdr
*nvgre_hdr
;
680 int hlen
= ip_hdrlen(skb
);
682 /* currently only IPv4 is supported due to hlen above */
683 if (vlan_get_protocol(skb
) != htons(ETH_P_IP
))
686 /* our transport header should be NVGRE */
687 nvgre_hdr
= (struct fm10k_nvgre_hdr
*)(skb_network_header(skb
) + hlen
);
689 /* verify all reserved flags are 0 */
690 if (nvgre_hdr
->flags
& FM10K_NVGRE_RESERVED0_FLAGS
)
693 /* report start of ethernet header */
694 if (nvgre_hdr
->flags
& NVGRE_TNI
)
695 return (struct ethhdr
*)(nvgre_hdr
+ 1);
697 return (struct ethhdr
*)(&nvgre_hdr
->tni
);
700 __be16
fm10k_tx_encap_offload(struct sk_buff
*skb
)
702 u8 l4_hdr
= 0, inner_l4_hdr
= 0, inner_l4_hlen
;
703 struct ethhdr
*eth_hdr
;
705 if (skb
->inner_protocol_type
!= ENCAP_TYPE_ETHER
||
706 skb
->inner_protocol
!= htons(ETH_P_TEB
))
709 switch (vlan_get_protocol(skb
)) {
710 case htons(ETH_P_IP
):
711 l4_hdr
= ip_hdr(skb
)->protocol
;
713 case htons(ETH_P_IPV6
):
714 l4_hdr
= ipv6_hdr(skb
)->nexthdr
;
722 eth_hdr
= fm10k_port_is_vxlan(skb
);
725 eth_hdr
= fm10k_gre_is_nvgre(skb
);
734 switch (eth_hdr
->h_proto
) {
735 case htons(ETH_P_IP
):
736 inner_l4_hdr
= inner_ip_hdr(skb
)->protocol
;
738 case htons(ETH_P_IPV6
):
739 inner_l4_hdr
= inner_ipv6_hdr(skb
)->nexthdr
;
745 switch (inner_l4_hdr
) {
747 inner_l4_hlen
= inner_tcp_hdrlen(skb
);
756 /* The hardware allows tunnel offloads only if the combined inner and
757 * outer header is 184 bytes or less
759 if (skb_inner_transport_header(skb
) + inner_l4_hlen
-
760 skb_mac_header(skb
) > FM10K_TUNNEL_HEADER_LENGTH
)
763 return eth_hdr
->h_proto
;
766 static int fm10k_tso(struct fm10k_ring
*tx_ring
,
767 struct fm10k_tx_buffer
*first
)
769 struct sk_buff
*skb
= first
->skb
;
770 struct fm10k_tx_desc
*tx_desc
;
774 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
777 if (!skb_is_gso(skb
))
780 /* compute header lengths */
781 if (skb
->encapsulation
) {
782 if (!fm10k_tx_encap_offload(skb
))
784 th
= skb_inner_transport_header(skb
);
786 th
= skb_transport_header(skb
);
789 /* compute offset from SOF to transport header and add header len */
790 hdrlen
= (th
- skb
->data
) + (((struct tcphdr
*)th
)->doff
<< 2);
792 first
->tx_flags
|= FM10K_TX_FLAGS_CSUM
;
794 /* update gso size and bytecount with header size */
795 first
->gso_segs
= skb_shinfo(skb
)->gso_segs
;
796 first
->bytecount
+= (first
->gso_segs
- 1) * hdrlen
;
798 /* populate Tx descriptor header size and mss */
799 tx_desc
= FM10K_TX_DESC(tx_ring
, tx_ring
->next_to_use
);
800 tx_desc
->hdrlen
= hdrlen
;
801 tx_desc
->mss
= cpu_to_le16(skb_shinfo(skb
)->gso_size
);
805 tx_ring
->netdev
->features
&= ~NETIF_F_GSO_UDP_TUNNEL
;
806 if (!net_ratelimit())
807 netdev_err(tx_ring
->netdev
,
808 "TSO requested for unsupported tunnel, disabling offload\n");
812 static void fm10k_tx_csum(struct fm10k_ring
*tx_ring
,
813 struct fm10k_tx_buffer
*first
)
815 struct sk_buff
*skb
= first
->skb
;
816 struct fm10k_tx_desc
*tx_desc
;
819 struct ipv6hdr
*ipv6
;
827 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
830 if (skb
->encapsulation
) {
831 protocol
= fm10k_tx_encap_offload(skb
);
833 if (skb_checksum_help(skb
)) {
834 dev_warn(tx_ring
->dev
,
835 "failed to offload encap csum!\n");
836 tx_ring
->tx_stats
.csum_err
++;
840 network_hdr
.raw
= skb_inner_network_header(skb
);
841 transport_hdr
= skb_inner_transport_header(skb
);
843 protocol
= vlan_get_protocol(skb
);
844 network_hdr
.raw
= skb_network_header(skb
);
845 transport_hdr
= skb_transport_header(skb
);
849 case htons(ETH_P_IP
):
850 l4_hdr
= network_hdr
.ipv4
->protocol
;
852 case htons(ETH_P_IPV6
):
853 l4_hdr
= network_hdr
.ipv6
->nexthdr
;
854 if (likely((transport_hdr
- network_hdr
.raw
) ==
855 sizeof(struct ipv6hdr
)))
857 ipv6_skip_exthdr(skb
, network_hdr
.raw
- skb
->data
+
858 sizeof(struct ipv6hdr
),
860 if (unlikely(frag_off
))
861 l4_hdr
= NEXTHDR_FRAGMENT
;
872 if (skb
->encapsulation
)
875 if (unlikely(net_ratelimit())) {
876 dev_warn(tx_ring
->dev
,
877 "partial checksum, version=%d l4 proto=%x\n",
880 skb_checksum_help(skb
);
881 tx_ring
->tx_stats
.csum_err
++;
885 /* update TX checksum flag */
886 first
->tx_flags
|= FM10K_TX_FLAGS_CSUM
;
887 tx_ring
->tx_stats
.csum_good
++;
890 /* populate Tx descriptor header size and mss */
891 tx_desc
= FM10K_TX_DESC(tx_ring
, tx_ring
->next_to_use
);
896 #define FM10K_SET_FLAG(_input, _flag, _result) \
897 ((_flag <= _result) ? \
898 ((u32)(_input & _flag) * (_result / _flag)) : \
899 ((u32)(_input & _flag) / (_flag / _result)))
901 static u8
fm10k_tx_desc_flags(struct sk_buff
*skb
, u32 tx_flags
)
903 /* set type for advanced descriptor with frame checksum insertion */
906 /* set checksum offload bits */
907 desc_flags
|= FM10K_SET_FLAG(tx_flags
, FM10K_TX_FLAGS_CSUM
,
908 FM10K_TXD_FLAG_CSUM
);
913 static bool fm10k_tx_desc_push(struct fm10k_ring
*tx_ring
,
914 struct fm10k_tx_desc
*tx_desc
, u16 i
,
915 dma_addr_t dma
, unsigned int size
, u8 desc_flags
)
917 /* set RS and INT for last frame in a cache line */
918 if ((++i
& (FM10K_TXD_WB_FIFO_SIZE
- 1)) == 0)
919 desc_flags
|= FM10K_TXD_FLAG_RS
| FM10K_TXD_FLAG_INT
;
921 /* record values to descriptor */
922 tx_desc
->buffer_addr
= cpu_to_le64(dma
);
923 tx_desc
->flags
= desc_flags
;
924 tx_desc
->buflen
= cpu_to_le16(size
);
926 /* return true if we just wrapped the ring */
927 return i
== tx_ring
->count
;
930 static int __fm10k_maybe_stop_tx(struct fm10k_ring
*tx_ring
, u16 size
)
932 netif_stop_subqueue(tx_ring
->netdev
, tx_ring
->queue_index
);
934 /* Memory barrier before checking head and tail */
937 /* Check again in a case another CPU has just made room available */
938 if (likely(fm10k_desc_unused(tx_ring
) < size
))
941 /* A reprieve! - use start_queue because it doesn't call schedule */
942 netif_start_subqueue(tx_ring
->netdev
, tx_ring
->queue_index
);
943 ++tx_ring
->tx_stats
.restart_queue
;
947 static inline int fm10k_maybe_stop_tx(struct fm10k_ring
*tx_ring
, u16 size
)
949 if (likely(fm10k_desc_unused(tx_ring
) >= size
))
951 return __fm10k_maybe_stop_tx(tx_ring
, size
);
954 static void fm10k_tx_map(struct fm10k_ring
*tx_ring
,
955 struct fm10k_tx_buffer
*first
)
957 struct sk_buff
*skb
= first
->skb
;
958 struct fm10k_tx_buffer
*tx_buffer
;
959 struct fm10k_tx_desc
*tx_desc
;
960 struct skb_frag_struct
*frag
;
963 unsigned int data_len
, size
;
964 u32 tx_flags
= first
->tx_flags
;
965 u16 i
= tx_ring
->next_to_use
;
966 u8 flags
= fm10k_tx_desc_flags(skb
, tx_flags
);
968 tx_desc
= FM10K_TX_DESC(tx_ring
, i
);
970 /* add HW VLAN tag */
971 if (skb_vlan_tag_present(skb
))
972 tx_desc
->vlan
= cpu_to_le16(skb_vlan_tag_get(skb
));
976 size
= skb_headlen(skb
);
979 dma
= dma_map_single(tx_ring
->dev
, data
, size
, DMA_TO_DEVICE
);
981 data_len
= skb
->data_len
;
984 for (frag
= &skb_shinfo(skb
)->frags
[0];; frag
++) {
985 if (dma_mapping_error(tx_ring
->dev
, dma
))
988 /* record length, and DMA address */
989 dma_unmap_len_set(tx_buffer
, len
, size
);
990 dma_unmap_addr_set(tx_buffer
, dma
, dma
);
992 while (unlikely(size
> FM10K_MAX_DATA_PER_TXD
)) {
993 if (fm10k_tx_desc_push(tx_ring
, tx_desc
++, i
++, dma
,
994 FM10K_MAX_DATA_PER_TXD
, flags
)) {
995 tx_desc
= FM10K_TX_DESC(tx_ring
, 0);
999 dma
+= FM10K_MAX_DATA_PER_TXD
;
1000 size
-= FM10K_MAX_DATA_PER_TXD
;
1003 if (likely(!data_len
))
1006 if (fm10k_tx_desc_push(tx_ring
, tx_desc
++, i
++,
1007 dma
, size
, flags
)) {
1008 tx_desc
= FM10K_TX_DESC(tx_ring
, 0);
1012 size
= skb_frag_size(frag
);
1015 dma
= skb_frag_dma_map(tx_ring
->dev
, frag
, 0, size
,
1018 tx_buffer
= &tx_ring
->tx_buffer
[i
];
1021 /* write last descriptor with LAST bit set */
1022 flags
|= FM10K_TXD_FLAG_LAST
;
1024 if (fm10k_tx_desc_push(tx_ring
, tx_desc
, i
++, dma
, size
, flags
))
1027 /* record bytecount for BQL */
1028 netdev_tx_sent_queue(txring_txq(tx_ring
), first
->bytecount
);
1030 /* record SW timestamp if HW timestamp is not available */
1031 skb_tx_timestamp(first
->skb
);
1033 /* Force memory writes to complete before letting h/w know there
1034 * are new descriptors to fetch. (Only applicable for weak-ordered
1035 * memory model archs, such as IA-64).
1037 * We also need this memory barrier to make certain all of the
1038 * status bits have been updated before next_to_watch is written.
1042 /* set next_to_watch value indicating a packet is present */
1043 first
->next_to_watch
= tx_desc
;
1045 tx_ring
->next_to_use
= i
;
1047 /* Make sure there is space in the ring for the next send. */
1048 fm10k_maybe_stop_tx(tx_ring
, DESC_NEEDED
);
1050 /* notify HW of packet */
1051 if (netif_xmit_stopped(txring_txq(tx_ring
)) || !skb
->xmit_more
) {
1052 writel(i
, tx_ring
->tail
);
1054 /* we need this if more than one processor can write to our tail
1055 * at a time, it synchronizes IO on IA64/Altix systems
1062 dev_err(tx_ring
->dev
, "TX DMA map failed\n");
1064 /* clear dma mappings for failed tx_buffer map */
1066 tx_buffer
= &tx_ring
->tx_buffer
[i
];
1067 fm10k_unmap_and_free_tx_resource(tx_ring
, tx_buffer
);
1068 if (tx_buffer
== first
)
1075 tx_ring
->next_to_use
= i
;
1078 netdev_tx_t
fm10k_xmit_frame_ring(struct sk_buff
*skb
,
1079 struct fm10k_ring
*tx_ring
)
1081 u16 count
= TXD_USE_COUNT(skb_headlen(skb
));
1082 struct fm10k_tx_buffer
*first
;
1087 /* need: 1 descriptor per page * PAGE_SIZE/FM10K_MAX_DATA_PER_TXD,
1088 * + 1 desc for skb_headlen/FM10K_MAX_DATA_PER_TXD,
1089 * + 2 desc gap to keep tail from touching head
1090 * otherwise try next time
1092 for (f
= 0; f
< skb_shinfo(skb
)->nr_frags
; f
++)
1093 count
+= TXD_USE_COUNT(skb_shinfo(skb
)->frags
[f
].size
);
1095 if (fm10k_maybe_stop_tx(tx_ring
, count
+ 3)) {
1096 tx_ring
->tx_stats
.tx_busy
++;
1097 return NETDEV_TX_BUSY
;
1100 /* record the location of the first descriptor for this packet */
1101 first
= &tx_ring
->tx_buffer
[tx_ring
->next_to_use
];
1103 first
->bytecount
= max_t(unsigned int, skb
->len
, ETH_ZLEN
);
1104 first
->gso_segs
= 1;
1106 /* record initial flags and protocol */
1107 first
->tx_flags
= tx_flags
;
1109 tso
= fm10k_tso(tx_ring
, first
);
1113 fm10k_tx_csum(tx_ring
, first
);
1115 fm10k_tx_map(tx_ring
, first
);
1117 return NETDEV_TX_OK
;
1120 dev_kfree_skb_any(first
->skb
);
1123 return NETDEV_TX_OK
;
1126 static u64
fm10k_get_tx_completed(struct fm10k_ring
*ring
)
1128 return ring
->stats
.packets
;
1131 u64
fm10k_get_tx_pending(struct fm10k_ring
*ring
)
1133 struct fm10k_intfc
*interface
= ring
->q_vector
->interface
;
1134 struct fm10k_hw
*hw
= &interface
->hw
;
1136 u32 head
= fm10k_read_reg(hw
, FM10K_TDH(ring
->reg_idx
));
1137 u32 tail
= fm10k_read_reg(hw
, FM10K_TDT(ring
->reg_idx
));
1139 return ((head
<= tail
) ? tail
: tail
+ ring
->count
) - head
;
1142 bool fm10k_check_tx_hang(struct fm10k_ring
*tx_ring
)
1144 u32 tx_done
= fm10k_get_tx_completed(tx_ring
);
1145 u32 tx_done_old
= tx_ring
->tx_stats
.tx_done_old
;
1146 u32 tx_pending
= fm10k_get_tx_pending(tx_ring
);
1148 clear_check_for_tx_hang(tx_ring
);
1150 /* Check for a hung queue, but be thorough. This verifies
1151 * that a transmit has been completed since the previous
1152 * check AND there is at least one packet pending. By
1153 * requiring this to fail twice we avoid races with
1154 * clearing the ARMED bit and conditions where we
1155 * run the check_tx_hang logic with a transmit completion
1156 * pending but without time to complete it yet.
1158 if (!tx_pending
|| (tx_done_old
!= tx_done
)) {
1159 /* update completed stats and continue */
1160 tx_ring
->tx_stats
.tx_done_old
= tx_done
;
1161 /* reset the countdown */
1162 clear_bit(__FM10K_HANG_CHECK_ARMED
, &tx_ring
->state
);
1167 /* make sure it is true for two checks in a row */
1168 return test_and_set_bit(__FM10K_HANG_CHECK_ARMED
, &tx_ring
->state
);
1172 * fm10k_tx_timeout_reset - initiate reset due to Tx timeout
1173 * @interface: driver private struct
1175 void fm10k_tx_timeout_reset(struct fm10k_intfc
*interface
)
1177 /* Do the reset outside of interrupt context */
1178 if (!test_bit(__FM10K_DOWN
, &interface
->state
)) {
1179 interface
->tx_timeout_count
++;
1180 interface
->flags
|= FM10K_FLAG_RESET_REQUESTED
;
1181 fm10k_service_event_schedule(interface
);
1186 * fm10k_clean_tx_irq - Reclaim resources after transmit completes
1187 * @q_vector: structure containing interrupt and ring information
1188 * @tx_ring: tx ring to clean
1189 * @napi_budget: Used to determine if we are in netpoll
1191 static bool fm10k_clean_tx_irq(struct fm10k_q_vector
*q_vector
,
1192 struct fm10k_ring
*tx_ring
, int napi_budget
)
1194 struct fm10k_intfc
*interface
= q_vector
->interface
;
1195 struct fm10k_tx_buffer
*tx_buffer
;
1196 struct fm10k_tx_desc
*tx_desc
;
1197 unsigned int total_bytes
= 0, total_packets
= 0;
1198 unsigned int budget
= q_vector
->tx
.work_limit
;
1199 unsigned int i
= tx_ring
->next_to_clean
;
1201 if (test_bit(__FM10K_DOWN
, &interface
->state
))
1204 tx_buffer
= &tx_ring
->tx_buffer
[i
];
1205 tx_desc
= FM10K_TX_DESC(tx_ring
, i
);
1206 i
-= tx_ring
->count
;
1209 struct fm10k_tx_desc
*eop_desc
= tx_buffer
->next_to_watch
;
1211 /* if next_to_watch is not set then there is no work pending */
1215 /* prevent any other reads prior to eop_desc */
1216 read_barrier_depends();
1218 /* if DD is not set pending work has not been completed */
1219 if (!(eop_desc
->flags
& FM10K_TXD_FLAG_DONE
))
1222 /* clear next_to_watch to prevent false hangs */
1223 tx_buffer
->next_to_watch
= NULL
;
1225 /* update the statistics for this packet */
1226 total_bytes
+= tx_buffer
->bytecount
;
1227 total_packets
+= tx_buffer
->gso_segs
;
1230 napi_consume_skb(tx_buffer
->skb
, napi_budget
);
1232 /* unmap skb header data */
1233 dma_unmap_single(tx_ring
->dev
,
1234 dma_unmap_addr(tx_buffer
, dma
),
1235 dma_unmap_len(tx_buffer
, len
),
1238 /* clear tx_buffer data */
1239 tx_buffer
->skb
= NULL
;
1240 dma_unmap_len_set(tx_buffer
, len
, 0);
1242 /* unmap remaining buffers */
1243 while (tx_desc
!= eop_desc
) {
1248 i
-= tx_ring
->count
;
1249 tx_buffer
= tx_ring
->tx_buffer
;
1250 tx_desc
= FM10K_TX_DESC(tx_ring
, 0);
1253 /* unmap any remaining paged data */
1254 if (dma_unmap_len(tx_buffer
, len
)) {
1255 dma_unmap_page(tx_ring
->dev
,
1256 dma_unmap_addr(tx_buffer
, dma
),
1257 dma_unmap_len(tx_buffer
, len
),
1259 dma_unmap_len_set(tx_buffer
, len
, 0);
1263 /* move us one more past the eop_desc for start of next pkt */
1268 i
-= tx_ring
->count
;
1269 tx_buffer
= tx_ring
->tx_buffer
;
1270 tx_desc
= FM10K_TX_DESC(tx_ring
, 0);
1273 /* issue prefetch for next Tx descriptor */
1276 /* update budget accounting */
1278 } while (likely(budget
));
1280 i
+= tx_ring
->count
;
1281 tx_ring
->next_to_clean
= i
;
1282 u64_stats_update_begin(&tx_ring
->syncp
);
1283 tx_ring
->stats
.bytes
+= total_bytes
;
1284 tx_ring
->stats
.packets
+= total_packets
;
1285 u64_stats_update_end(&tx_ring
->syncp
);
1286 q_vector
->tx
.total_bytes
+= total_bytes
;
1287 q_vector
->tx
.total_packets
+= total_packets
;
1289 if (check_for_tx_hang(tx_ring
) && fm10k_check_tx_hang(tx_ring
)) {
1290 /* schedule immediate reset if we believe we hung */
1291 struct fm10k_hw
*hw
= &interface
->hw
;
1293 netif_err(interface
, drv
, tx_ring
->netdev
,
1294 "Detected Tx Unit Hang\n"
1296 " TDH, TDT <%x>, <%x>\n"
1297 " next_to_use <%x>\n"
1298 " next_to_clean <%x>\n",
1299 tx_ring
->queue_index
,
1300 fm10k_read_reg(hw
, FM10K_TDH(tx_ring
->reg_idx
)),
1301 fm10k_read_reg(hw
, FM10K_TDT(tx_ring
->reg_idx
)),
1302 tx_ring
->next_to_use
, i
);
1304 netif_stop_subqueue(tx_ring
->netdev
,
1305 tx_ring
->queue_index
);
1307 netif_info(interface
, probe
, tx_ring
->netdev
,
1308 "tx hang %d detected on queue %d, resetting interface\n",
1309 interface
->tx_timeout_count
+ 1,
1310 tx_ring
->queue_index
);
1312 fm10k_tx_timeout_reset(interface
);
1314 /* the netdev is about to reset, no point in enabling stuff */
1318 /* notify netdev of completed buffers */
1319 netdev_tx_completed_queue(txring_txq(tx_ring
),
1320 total_packets
, total_bytes
);
1322 #define TX_WAKE_THRESHOLD min_t(u16, FM10K_MIN_TXD - 1, DESC_NEEDED * 2)
1323 if (unlikely(total_packets
&& netif_carrier_ok(tx_ring
->netdev
) &&
1324 (fm10k_desc_unused(tx_ring
) >= TX_WAKE_THRESHOLD
))) {
1325 /* Make sure that anybody stopping the queue after this
1326 * sees the new next_to_clean.
1329 if (__netif_subqueue_stopped(tx_ring
->netdev
,
1330 tx_ring
->queue_index
) &&
1331 !test_bit(__FM10K_DOWN
, &interface
->state
)) {
1332 netif_wake_subqueue(tx_ring
->netdev
,
1333 tx_ring
->queue_index
);
1334 ++tx_ring
->tx_stats
.restart_queue
;
1342 * fm10k_update_itr - update the dynamic ITR value based on packet size
1344 * Stores a new ITR value based on strictly on packet size. The
1345 * divisors and thresholds used by this function were determined based
1346 * on theoretical maximum wire speed and testing data, in order to
1347 * minimize response time while increasing bulk throughput.
1349 * @ring_container: Container for rings to have ITR updated
1351 static void fm10k_update_itr(struct fm10k_ring_container
*ring_container
)
1353 unsigned int avg_wire_size
, packets
, itr_round
;
1355 /* Only update ITR if we are using adaptive setting */
1356 if (!ITR_IS_ADAPTIVE(ring_container
->itr
))
1359 packets
= ring_container
->total_packets
;
1363 avg_wire_size
= ring_container
->total_bytes
/ packets
;
1365 /* The following is a crude approximation of:
1366 * wmem_default / (size + overhead) = desired_pkts_per_int
1367 * rate / bits_per_byte / (size + ethernet overhead) = pkt_rate
1368 * (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
1370 * Assuming wmem_default is 212992 and overhead is 640 bytes per
1371 * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
1374 * (34 * (size + 24)) / (size + 640) = ITR
1376 * We first do some math on the packet size and then finally bitshift
1377 * by 8 after rounding up. We also have to account for PCIe link speed
1378 * difference as ITR scales based on this.
1380 if (avg_wire_size
<= 360) {
1381 /* Start at 250K ints/sec and gradually drop to 77K ints/sec */
1383 avg_wire_size
+= 376;
1384 } else if (avg_wire_size
<= 1152) {
1385 /* 77K ints/sec to 45K ints/sec */
1387 avg_wire_size
+= 2176;
1388 } else if (avg_wire_size
<= 1920) {
1389 /* 45K ints/sec to 38K ints/sec */
1390 avg_wire_size
+= 4480;
1392 /* plateau at a limit of 38K ints/sec */
1393 avg_wire_size
= 6656;
1396 /* Perform final bitshift for division after rounding up to ensure
1397 * that the calculation will never get below a 1. The bit shift
1398 * accounts for changes in the ITR due to PCIe link speed.
1400 itr_round
= ACCESS_ONCE(ring_container
->itr_scale
) + 8;
1401 avg_wire_size
+= BIT(itr_round
) - 1;
1402 avg_wire_size
>>= itr_round
;
1404 /* write back value and retain adaptive flag */
1405 ring_container
->itr
= avg_wire_size
| FM10K_ITR_ADAPTIVE
;
1408 ring_container
->total_bytes
= 0;
1409 ring_container
->total_packets
= 0;
1412 static void fm10k_qv_enable(struct fm10k_q_vector
*q_vector
)
1414 /* Enable auto-mask and clear the current mask */
1415 u32 itr
= FM10K_ITR_ENABLE
;
1418 fm10k_update_itr(&q_vector
->tx
);
1421 fm10k_update_itr(&q_vector
->rx
);
1423 /* Store Tx itr in timer slot 0 */
1424 itr
|= (q_vector
->tx
.itr
& FM10K_ITR_MAX
);
1426 /* Shift Rx itr to timer slot 1 */
1427 itr
|= (q_vector
->rx
.itr
& FM10K_ITR_MAX
) << FM10K_ITR_INTERVAL1_SHIFT
;
1429 /* Write the final value to the ITR register */
1430 writel(itr
, q_vector
->itr
);
1433 static int fm10k_poll(struct napi_struct
*napi
, int budget
)
1435 struct fm10k_q_vector
*q_vector
=
1436 container_of(napi
, struct fm10k_q_vector
, napi
);
1437 struct fm10k_ring
*ring
;
1438 int per_ring_budget
, work_done
= 0;
1439 bool clean_complete
= true;
1441 fm10k_for_each_ring(ring
, q_vector
->tx
) {
1442 if (!fm10k_clean_tx_irq(q_vector
, ring
, budget
))
1443 clean_complete
= false;
1446 /* Handle case where we are called by netpoll with a budget of 0 */
1450 /* attempt to distribute budget to each queue fairly, but don't
1451 * allow the budget to go below 1 because we'll exit polling
1453 if (q_vector
->rx
.count
> 1)
1454 per_ring_budget
= max(budget
/ q_vector
->rx
.count
, 1);
1456 per_ring_budget
= budget
;
1458 fm10k_for_each_ring(ring
, q_vector
->rx
) {
1459 int work
= fm10k_clean_rx_irq(q_vector
, ring
, per_ring_budget
);
1462 if (work
>= per_ring_budget
)
1463 clean_complete
= false;
1466 /* If all work not completed, return budget and keep polling */
1467 if (!clean_complete
)
1470 /* all work done, exit the polling mode */
1471 napi_complete_done(napi
, work_done
);
1473 /* re-enable the q_vector */
1474 fm10k_qv_enable(q_vector
);
1480 * fm10k_set_qos_queues: Allocate queues for a QOS-enabled device
1481 * @interface: board private structure to initialize
1483 * When QoS (Quality of Service) is enabled, allocate queues for
1484 * each traffic class. If multiqueue isn't available,then abort QoS
1487 * This function handles all combinations of Qos and RSS.
1490 static bool fm10k_set_qos_queues(struct fm10k_intfc
*interface
)
1492 struct net_device
*dev
= interface
->netdev
;
1493 struct fm10k_ring_feature
*f
;
1497 /* Map queue offset and counts onto allocated tx queues */
1498 pcs
= netdev_get_num_tc(dev
);
1503 /* set QoS mask and indices */
1504 f
= &interface
->ring_feature
[RING_F_QOS
];
1506 f
->mask
= BIT(fls(pcs
- 1)) - 1;
1508 /* determine the upper limit for our current DCB mode */
1509 rss_i
= interface
->hw
.mac
.max_queues
/ pcs
;
1510 rss_i
= BIT(fls(rss_i
) - 1);
1512 /* set RSS mask and indices */
1513 f
= &interface
->ring_feature
[RING_F_RSS
];
1514 rss_i
= min_t(u16
, rss_i
, f
->limit
);
1516 f
->mask
= BIT(fls(rss_i
- 1)) - 1;
1518 /* configure pause class to queue mapping */
1519 for (i
= 0; i
< pcs
; i
++)
1520 netdev_set_tc_queue(dev
, i
, rss_i
, rss_i
* i
);
1522 interface
->num_rx_queues
= rss_i
* pcs
;
1523 interface
->num_tx_queues
= rss_i
* pcs
;
1529 * fm10k_set_rss_queues: Allocate queues for RSS
1530 * @interface: board private structure to initialize
1532 * This is our "base" multiqueue mode. RSS (Receive Side Scaling) will try
1533 * to allocate one Rx queue per CPU, and if available, one Tx queue per CPU.
1536 static bool fm10k_set_rss_queues(struct fm10k_intfc
*interface
)
1538 struct fm10k_ring_feature
*f
;
1541 f
= &interface
->ring_feature
[RING_F_RSS
];
1542 rss_i
= min_t(u16
, interface
->hw
.mac
.max_queues
, f
->limit
);
1544 /* record indices and power of 2 mask for RSS */
1546 f
->mask
= BIT(fls(rss_i
- 1)) - 1;
1548 interface
->num_rx_queues
= rss_i
;
1549 interface
->num_tx_queues
= rss_i
;
1555 * fm10k_set_num_queues: Allocate queues for device, feature dependent
1556 * @interface: board private structure to initialize
1558 * This is the top level queue allocation routine. The order here is very
1559 * important, starting with the "most" number of features turned on at once,
1560 * and ending with the smallest set of features. This way large combinations
1561 * can be allocated if they're turned on, and smaller combinations are the
1562 * fallthrough conditions.
1565 static void fm10k_set_num_queues(struct fm10k_intfc
*interface
)
1567 /* Attempt to setup QoS and RSS first */
1568 if (fm10k_set_qos_queues(interface
))
1571 /* If we don't have QoS, just fallback to only RSS. */
1572 fm10k_set_rss_queues(interface
);
1576 * fm10k_reset_num_queues - Reset the number of queues to zero
1577 * @interface: board private structure
1579 * This function should be called whenever we need to reset the number of
1580 * queues after an error condition.
1582 static void fm10k_reset_num_queues(struct fm10k_intfc
*interface
)
1584 interface
->num_tx_queues
= 0;
1585 interface
->num_rx_queues
= 0;
1586 interface
->num_q_vectors
= 0;
1590 * fm10k_alloc_q_vector - Allocate memory for a single interrupt vector
1591 * @interface: board private structure to initialize
1592 * @v_count: q_vectors allocated on interface, used for ring interleaving
1593 * @v_idx: index of vector in interface struct
1594 * @txr_count: total number of Tx rings to allocate
1595 * @txr_idx: index of first Tx ring to allocate
1596 * @rxr_count: total number of Rx rings to allocate
1597 * @rxr_idx: index of first Rx ring to allocate
1599 * We allocate one q_vector. If allocation fails we return -ENOMEM.
1601 static int fm10k_alloc_q_vector(struct fm10k_intfc
*interface
,
1602 unsigned int v_count
, unsigned int v_idx
,
1603 unsigned int txr_count
, unsigned int txr_idx
,
1604 unsigned int rxr_count
, unsigned int rxr_idx
)
1606 struct fm10k_q_vector
*q_vector
;
1607 struct fm10k_ring
*ring
;
1608 int ring_count
, size
;
1610 ring_count
= txr_count
+ rxr_count
;
1611 size
= sizeof(struct fm10k_q_vector
) +
1612 (sizeof(struct fm10k_ring
) * ring_count
);
1614 /* allocate q_vector and rings */
1615 q_vector
= kzalloc(size
, GFP_KERNEL
);
1619 /* initialize NAPI */
1620 netif_napi_add(interface
->netdev
, &q_vector
->napi
,
1621 fm10k_poll
, NAPI_POLL_WEIGHT
);
1623 /* tie q_vector and interface together */
1624 interface
->q_vector
[v_idx
] = q_vector
;
1625 q_vector
->interface
= interface
;
1626 q_vector
->v_idx
= v_idx
;
1628 /* initialize pointer to rings */
1629 ring
= q_vector
->ring
;
1631 /* save Tx ring container info */
1632 q_vector
->tx
.ring
= ring
;
1633 q_vector
->tx
.work_limit
= FM10K_DEFAULT_TX_WORK
;
1634 q_vector
->tx
.itr
= interface
->tx_itr
;
1635 q_vector
->tx
.itr_scale
= interface
->hw
.mac
.itr_scale
;
1636 q_vector
->tx
.count
= txr_count
;
1639 /* assign generic ring traits */
1640 ring
->dev
= &interface
->pdev
->dev
;
1641 ring
->netdev
= interface
->netdev
;
1643 /* configure backlink on ring */
1644 ring
->q_vector
= q_vector
;
1646 /* apply Tx specific ring traits */
1647 ring
->count
= interface
->tx_ring_count
;
1648 ring
->queue_index
= txr_idx
;
1650 /* assign ring to interface */
1651 interface
->tx_ring
[txr_idx
] = ring
;
1653 /* update count and index */
1657 /* push pointer to next ring */
1661 /* save Rx ring container info */
1662 q_vector
->rx
.ring
= ring
;
1663 q_vector
->rx
.itr
= interface
->rx_itr
;
1664 q_vector
->rx
.itr_scale
= interface
->hw
.mac
.itr_scale
;
1665 q_vector
->rx
.count
= rxr_count
;
1668 /* assign generic ring traits */
1669 ring
->dev
= &interface
->pdev
->dev
;
1670 ring
->netdev
= interface
->netdev
;
1671 rcu_assign_pointer(ring
->l2_accel
, interface
->l2_accel
);
1673 /* configure backlink on ring */
1674 ring
->q_vector
= q_vector
;
1676 /* apply Rx specific ring traits */
1677 ring
->count
= interface
->rx_ring_count
;
1678 ring
->queue_index
= rxr_idx
;
1680 /* assign ring to interface */
1681 interface
->rx_ring
[rxr_idx
] = ring
;
1683 /* update count and index */
1687 /* push pointer to next ring */
1691 fm10k_dbg_q_vector_init(q_vector
);
1697 * fm10k_free_q_vector - Free memory allocated for specific interrupt vector
1698 * @interface: board private structure to initialize
1699 * @v_idx: Index of vector to be freed
1701 * This function frees the memory allocated to the q_vector. In addition if
1702 * NAPI is enabled it will delete any references to the NAPI struct prior
1703 * to freeing the q_vector.
1705 static void fm10k_free_q_vector(struct fm10k_intfc
*interface
, int v_idx
)
1707 struct fm10k_q_vector
*q_vector
= interface
->q_vector
[v_idx
];
1708 struct fm10k_ring
*ring
;
1710 fm10k_dbg_q_vector_exit(q_vector
);
1712 fm10k_for_each_ring(ring
, q_vector
->tx
)
1713 interface
->tx_ring
[ring
->queue_index
] = NULL
;
1715 fm10k_for_each_ring(ring
, q_vector
->rx
)
1716 interface
->rx_ring
[ring
->queue_index
] = NULL
;
1718 interface
->q_vector
[v_idx
] = NULL
;
1719 netif_napi_del(&q_vector
->napi
);
1720 kfree_rcu(q_vector
, rcu
);
1724 * fm10k_alloc_q_vectors - Allocate memory for interrupt vectors
1725 * @interface: board private structure to initialize
1727 * We allocate one q_vector per queue interrupt. If allocation fails we
1730 static int fm10k_alloc_q_vectors(struct fm10k_intfc
*interface
)
1732 unsigned int q_vectors
= interface
->num_q_vectors
;
1733 unsigned int rxr_remaining
= interface
->num_rx_queues
;
1734 unsigned int txr_remaining
= interface
->num_tx_queues
;
1735 unsigned int rxr_idx
= 0, txr_idx
= 0, v_idx
= 0;
1738 if (q_vectors
>= (rxr_remaining
+ txr_remaining
)) {
1739 for (; rxr_remaining
; v_idx
++) {
1740 err
= fm10k_alloc_q_vector(interface
, q_vectors
, v_idx
,
1745 /* update counts and index */
1751 for (; v_idx
< q_vectors
; v_idx
++) {
1752 int rqpv
= DIV_ROUND_UP(rxr_remaining
, q_vectors
- v_idx
);
1753 int tqpv
= DIV_ROUND_UP(txr_remaining
, q_vectors
- v_idx
);
1755 err
= fm10k_alloc_q_vector(interface
, q_vectors
, v_idx
,
1762 /* update counts and index */
1763 rxr_remaining
-= rqpv
;
1764 txr_remaining
-= tqpv
;
1772 fm10k_reset_num_queues(interface
);
1775 fm10k_free_q_vector(interface
, v_idx
);
1781 * fm10k_free_q_vectors - Free memory allocated for interrupt vectors
1782 * @interface: board private structure to initialize
1784 * This function frees the memory allocated to the q_vectors. In addition if
1785 * NAPI is enabled it will delete any references to the NAPI struct prior
1786 * to freeing the q_vector.
1788 static void fm10k_free_q_vectors(struct fm10k_intfc
*interface
)
1790 int v_idx
= interface
->num_q_vectors
;
1792 fm10k_reset_num_queues(interface
);
1795 fm10k_free_q_vector(interface
, v_idx
);
1799 * f10k_reset_msix_capability - reset MSI-X capability
1800 * @interface: board private structure to initialize
1802 * Reset the MSI-X capability back to its starting state
1804 static void fm10k_reset_msix_capability(struct fm10k_intfc
*interface
)
1806 pci_disable_msix(interface
->pdev
);
1807 kfree(interface
->msix_entries
);
1808 interface
->msix_entries
= NULL
;
1812 * f10k_init_msix_capability - configure MSI-X capability
1813 * @interface: board private structure to initialize
1815 * Attempt to configure the interrupts using the best available
1816 * capabilities of the hardware and the kernel.
1818 static int fm10k_init_msix_capability(struct fm10k_intfc
*interface
)
1820 struct fm10k_hw
*hw
= &interface
->hw
;
1821 int v_budget
, vector
;
1823 /* It's easy to be greedy for MSI-X vectors, but it really
1824 * doesn't do us much good if we have a lot more vectors
1825 * than CPU's. So let's be conservative and only ask for
1826 * (roughly) the same number of vectors as there are CPU's.
1827 * the default is to use pairs of vectors
1829 v_budget
= max(interface
->num_rx_queues
, interface
->num_tx_queues
);
1830 v_budget
= min_t(u16
, v_budget
, num_online_cpus());
1832 /* account for vectors not related to queues */
1833 v_budget
+= NON_Q_VECTORS(hw
);
1835 /* At the same time, hardware can only support a maximum of
1836 * hw.mac->max_msix_vectors vectors. With features
1837 * such as RSS and VMDq, we can easily surpass the number of Rx and Tx
1838 * descriptor queues supported by our device. Thus, we cap it off in
1839 * those rare cases where the cpu count also exceeds our vector limit.
1841 v_budget
= min_t(int, v_budget
, hw
->mac
.max_msix_vectors
);
1843 /* A failure in MSI-X entry allocation is fatal. */
1844 interface
->msix_entries
= kcalloc(v_budget
, sizeof(struct msix_entry
),
1846 if (!interface
->msix_entries
)
1849 /* populate entry values */
1850 for (vector
= 0; vector
< v_budget
; vector
++)
1851 interface
->msix_entries
[vector
].entry
= vector
;
1853 /* Attempt to enable MSI-X with requested value */
1854 v_budget
= pci_enable_msix_range(interface
->pdev
,
1855 interface
->msix_entries
,
1859 kfree(interface
->msix_entries
);
1860 interface
->msix_entries
= NULL
;
1864 /* record the number of queues available for q_vectors */
1865 interface
->num_q_vectors
= v_budget
- NON_Q_VECTORS(hw
);
1871 * fm10k_cache_ring_qos - Descriptor ring to register mapping for QoS
1872 * @interface: Interface structure continaining rings and devices
1874 * Cache the descriptor ring offsets for Qos
1876 static bool fm10k_cache_ring_qos(struct fm10k_intfc
*interface
)
1878 struct net_device
*dev
= interface
->netdev
;
1879 int pc
, offset
, rss_i
, i
, q_idx
;
1880 u16 pc_stride
= interface
->ring_feature
[RING_F_QOS
].mask
+ 1;
1881 u8 num_pcs
= netdev_get_num_tc(dev
);
1886 rss_i
= interface
->ring_feature
[RING_F_RSS
].indices
;
1888 for (pc
= 0, offset
= 0; pc
< num_pcs
; pc
++, offset
+= rss_i
) {
1890 for (i
= 0; i
< rss_i
; i
++) {
1891 interface
->tx_ring
[offset
+ i
]->reg_idx
= q_idx
;
1892 interface
->tx_ring
[offset
+ i
]->qos_pc
= pc
;
1893 interface
->rx_ring
[offset
+ i
]->reg_idx
= q_idx
;
1894 interface
->rx_ring
[offset
+ i
]->qos_pc
= pc
;
1903 * fm10k_cache_ring_rss - Descriptor ring to register mapping for RSS
1904 * @interface: Interface structure continaining rings and devices
1906 * Cache the descriptor ring offsets for RSS
1908 static void fm10k_cache_ring_rss(struct fm10k_intfc
*interface
)
1912 for (i
= 0; i
< interface
->num_rx_queues
; i
++)
1913 interface
->rx_ring
[i
]->reg_idx
= i
;
1915 for (i
= 0; i
< interface
->num_tx_queues
; i
++)
1916 interface
->tx_ring
[i
]->reg_idx
= i
;
1920 * fm10k_assign_rings - Map rings to network devices
1921 * @interface: Interface structure containing rings and devices
1923 * This function is meant to go though and configure both the network
1924 * devices so that they contain rings, and configure the rings so that
1925 * they function with their network devices.
1927 static void fm10k_assign_rings(struct fm10k_intfc
*interface
)
1929 if (fm10k_cache_ring_qos(interface
))
1932 fm10k_cache_ring_rss(interface
);
1935 static void fm10k_init_reta(struct fm10k_intfc
*interface
)
1937 u16 i
, rss_i
= interface
->ring_feature
[RING_F_RSS
].indices
;
1940 /* If the Rx flow indirection table has been configured manually, we
1941 * need to maintain it when possible.
1943 if (netif_is_rxfh_configured(interface
->netdev
)) {
1944 for (i
= FM10K_RETA_SIZE
; i
--;) {
1945 reta
= interface
->reta
[i
];
1946 if ((((reta
<< 24) >> 24) < rss_i
) &&
1947 (((reta
<< 16) >> 24) < rss_i
) &&
1948 (((reta
<< 8) >> 24) < rss_i
) &&
1949 (((reta
) >> 24) < rss_i
))
1952 /* this should never happen */
1953 dev_err(&interface
->pdev
->dev
,
1954 "RSS indirection table assigned flows out of queue bounds. Reconfiguring.\n");
1955 goto repopulate_reta
;
1958 /* do nothing if all of the elements are in bounds */
1963 fm10k_write_reta(interface
, NULL
);
1967 * fm10k_init_queueing_scheme - Determine proper queueing scheme
1968 * @interface: board private structure to initialize
1970 * We determine which queueing scheme to use based on...
1971 * - Hardware queue count (num_*_queues)
1972 * - defined by miscellaneous hardware support/features (RSS, etc.)
1974 int fm10k_init_queueing_scheme(struct fm10k_intfc
*interface
)
1978 /* Number of supported queues */
1979 fm10k_set_num_queues(interface
);
1981 /* Configure MSI-X capability */
1982 err
= fm10k_init_msix_capability(interface
);
1984 dev_err(&interface
->pdev
->dev
,
1985 "Unable to initialize MSI-X capability\n");
1989 /* Allocate memory for queues */
1990 err
= fm10k_alloc_q_vectors(interface
);
1992 dev_err(&interface
->pdev
->dev
,
1993 "Unable to allocate queue vectors\n");
1994 goto err_alloc_q_vectors
;
1997 /* Map rings to devices, and map devices to physical queues */
1998 fm10k_assign_rings(interface
);
2000 /* Initialize RSS redirection table */
2001 fm10k_init_reta(interface
);
2005 err_alloc_q_vectors
:
2006 fm10k_reset_msix_capability(interface
);
2008 fm10k_reset_num_queues(interface
);
2013 * fm10k_clear_queueing_scheme - Clear the current queueing scheme settings
2014 * @interface: board private structure to clear queueing scheme on
2016 * We go through and clear queueing specific resources and reset the structure
2017 * to pre-load conditions
2019 void fm10k_clear_queueing_scheme(struct fm10k_intfc
*interface
)
2021 fm10k_free_q_vectors(interface
);
2022 fm10k_reset_msix_capability(interface
);