1 /* Intel Ethernet Switch Host Interface Driver
2 * Copyright(c) 2013 - 2014 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 "0.12.2-k"
32 const char fm10k_driver_version
[] = DRV_VERSION
;
33 char fm10k_driver_name
[] = "fm10k";
34 static const char fm10k_driver_string
[] =
35 "Intel(R) Ethernet Switch Host Interface Driver";
36 static const char fm10k_copyright
[] =
37 "Copyright (c) 2013 Intel Corporation.";
39 MODULE_AUTHOR("Intel Corporation, <linux.nics@intel.com>");
40 MODULE_DESCRIPTION("Intel(R) Ethernet Switch Host Interface Driver");
41 MODULE_LICENSE("GPL");
42 MODULE_VERSION(DRV_VERSION
);
45 * fm10k_init_module - Driver Registration Routine
47 * fm10k_init_module is the first routine called when the driver is
48 * loaded. All it does is register with the PCI subsystem.
50 static int __init
fm10k_init_module(void)
52 pr_info("%s - version %s\n", fm10k_driver_string
, fm10k_driver_version
);
53 pr_info("%s\n", fm10k_copyright
);
57 return fm10k_register_pci_driver();
59 module_init(fm10k_init_module
);
62 * fm10k_exit_module - Driver Exit Cleanup Routine
64 * fm10k_exit_module is called just before the driver is removed
67 static void __exit
fm10k_exit_module(void)
69 fm10k_unregister_pci_driver();
73 module_exit(fm10k_exit_module
);
75 static bool fm10k_alloc_mapped_page(struct fm10k_ring
*rx_ring
,
76 struct fm10k_rx_buffer
*bi
)
78 struct page
*page
= bi
->page
;
81 /* Only page will be NULL if buffer was consumed */
85 /* alloc new page for storage */
86 page
= dev_alloc_page();
87 if (unlikely(!page
)) {
88 rx_ring
->rx_stats
.alloc_failed
++;
92 /* map page for use */
93 dma
= dma_map_page(rx_ring
->dev
, page
, 0, PAGE_SIZE
, DMA_FROM_DEVICE
);
95 /* if mapping failed free memory back to system since
96 * there isn't much point in holding memory we can't use
98 if (dma_mapping_error(rx_ring
->dev
, dma
)) {
102 rx_ring
->rx_stats
.alloc_failed
++;
114 * fm10k_alloc_rx_buffers - Replace used receive buffers
115 * @rx_ring: ring to place buffers on
116 * @cleaned_count: number of buffers to replace
118 void fm10k_alloc_rx_buffers(struct fm10k_ring
*rx_ring
, u16 cleaned_count
)
120 union fm10k_rx_desc
*rx_desc
;
121 struct fm10k_rx_buffer
*bi
;
122 u16 i
= rx_ring
->next_to_use
;
128 rx_desc
= FM10K_RX_DESC(rx_ring
, i
);
129 bi
= &rx_ring
->rx_buffer
[i
];
133 if (!fm10k_alloc_mapped_page(rx_ring
, bi
))
136 /* Refresh the desc even if buffer_addrs didn't change
137 * because each write-back erases this info.
139 rx_desc
->q
.pkt_addr
= cpu_to_le64(bi
->dma
+ bi
->page_offset
);
145 rx_desc
= FM10K_RX_DESC(rx_ring
, 0);
146 bi
= rx_ring
->rx_buffer
;
150 /* clear the hdr_addr for the next_to_use descriptor */
151 rx_desc
->q
.hdr_addr
= 0;
154 } while (cleaned_count
);
158 if (rx_ring
->next_to_use
!= i
) {
159 /* record the next descriptor to use */
160 rx_ring
->next_to_use
= i
;
162 /* update next to alloc since we have filled the ring */
163 rx_ring
->next_to_alloc
= i
;
165 /* Force memory writes to complete before letting h/w
166 * know there are new descriptors to fetch. (Only
167 * applicable for weak-ordered memory model archs,
172 /* notify hardware of new descriptors */
173 writel(i
, rx_ring
->tail
);
178 * fm10k_reuse_rx_page - page flip buffer and store it back on the ring
179 * @rx_ring: rx descriptor ring to store buffers on
180 * @old_buff: donor buffer to have page reused
182 * Synchronizes page for reuse by the interface
184 static void fm10k_reuse_rx_page(struct fm10k_ring
*rx_ring
,
185 struct fm10k_rx_buffer
*old_buff
)
187 struct fm10k_rx_buffer
*new_buff
;
188 u16 nta
= rx_ring
->next_to_alloc
;
190 new_buff
= &rx_ring
->rx_buffer
[nta
];
192 /* update, and store next to alloc */
194 rx_ring
->next_to_alloc
= (nta
< rx_ring
->count
) ? nta
: 0;
196 /* transfer page from old buffer to new buffer */
197 memcpy(new_buff
, old_buff
, sizeof(struct fm10k_rx_buffer
));
199 /* sync the buffer for use by the device */
200 dma_sync_single_range_for_device(rx_ring
->dev
, old_buff
->dma
,
201 old_buff
->page_offset
,
206 static bool fm10k_can_reuse_rx_page(struct fm10k_rx_buffer
*rx_buffer
,
208 unsigned int truesize
)
210 /* avoid re-using remote pages */
211 if (unlikely(page_to_nid(page
) != numa_mem_id()))
214 #if (PAGE_SIZE < 8192)
215 /* if we are only owner of page we can reuse it */
216 if (unlikely(page_count(page
) != 1))
219 /* flip page offset to other buffer */
220 rx_buffer
->page_offset
^= FM10K_RX_BUFSZ
;
222 /* Even if we own the page, we are not allowed to use atomic_set()
223 * This would break get_page_unless_zero() users.
225 atomic_inc(&page
->_count
);
227 /* move offset up to the next cache line */
228 rx_buffer
->page_offset
+= truesize
;
230 if (rx_buffer
->page_offset
> (PAGE_SIZE
- FM10K_RX_BUFSZ
))
233 /* bump ref count on page before it is given to the stack */
241 * fm10k_add_rx_frag - Add contents of Rx buffer to sk_buff
242 * @rx_ring: rx descriptor ring to transact packets on
243 * @rx_buffer: buffer containing page to add
244 * @rx_desc: descriptor containing length of buffer written by hardware
245 * @skb: sk_buff to place the data into
247 * This function will add the data contained in rx_buffer->page to the skb.
248 * This is done either through a direct copy if the data in the buffer is
249 * less than the skb header size, otherwise it will just attach the page as
252 * The function will then update the page offset if necessary and return
253 * true if the buffer can be reused by the interface.
255 static bool fm10k_add_rx_frag(struct fm10k_ring
*rx_ring
,
256 struct fm10k_rx_buffer
*rx_buffer
,
257 union fm10k_rx_desc
*rx_desc
,
260 struct page
*page
= rx_buffer
->page
;
261 unsigned int size
= le16_to_cpu(rx_desc
->w
.length
);
262 #if (PAGE_SIZE < 8192)
263 unsigned int truesize
= FM10K_RX_BUFSZ
;
265 unsigned int truesize
= ALIGN(size
, L1_CACHE_BYTES
);
268 if ((size
<= FM10K_RX_HDR_LEN
) && !skb_is_nonlinear(skb
)) {
269 unsigned char *va
= page_address(page
) + rx_buffer
->page_offset
;
271 memcpy(__skb_put(skb
, size
), va
, ALIGN(size
, sizeof(long)));
273 /* we can reuse buffer as-is, just make sure it is local */
274 if (likely(page_to_nid(page
) == numa_mem_id()))
277 /* this page cannot be reused so discard it */
282 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
,
283 rx_buffer
->page_offset
, size
, truesize
);
285 return fm10k_can_reuse_rx_page(rx_buffer
, page
, truesize
);
288 static struct sk_buff
*fm10k_fetch_rx_buffer(struct fm10k_ring
*rx_ring
,
289 union fm10k_rx_desc
*rx_desc
,
292 struct fm10k_rx_buffer
*rx_buffer
;
295 rx_buffer
= &rx_ring
->rx_buffer
[rx_ring
->next_to_clean
];
297 page
= rx_buffer
->page
;
301 void *page_addr
= page_address(page
) +
302 rx_buffer
->page_offset
;
304 /* prefetch first cache line of first page */
306 #if L1_CACHE_BYTES < 128
307 prefetch(page_addr
+ L1_CACHE_BYTES
);
310 /* allocate a skb to store the frags */
311 skb
= napi_alloc_skb(&rx_ring
->q_vector
->napi
,
313 if (unlikely(!skb
)) {
314 rx_ring
->rx_stats
.alloc_failed
++;
318 /* we will be copying header into skb->data in
319 * pskb_may_pull so it is in our interest to prefetch
320 * it now to avoid a possible cache miss
322 prefetchw(skb
->data
);
325 /* we are reusing so sync this buffer for CPU use */
326 dma_sync_single_range_for_cpu(rx_ring
->dev
,
328 rx_buffer
->page_offset
,
332 /* pull page into skb */
333 if (fm10k_add_rx_frag(rx_ring
, rx_buffer
, rx_desc
, skb
)) {
334 /* hand second half of page back to the ring */
335 fm10k_reuse_rx_page(rx_ring
, rx_buffer
);
337 /* we are not reusing the buffer so unmap it */
338 dma_unmap_page(rx_ring
->dev
, rx_buffer
->dma
,
339 PAGE_SIZE
, DMA_FROM_DEVICE
);
342 /* clear contents of rx_buffer */
343 rx_buffer
->page
= NULL
;
348 static inline void fm10k_rx_checksum(struct fm10k_ring
*ring
,
349 union fm10k_rx_desc
*rx_desc
,
352 skb_checksum_none_assert(skb
);
354 /* Rx checksum disabled via ethtool */
355 if (!(ring
->netdev
->features
& NETIF_F_RXCSUM
))
358 /* TCP/UDP checksum error bit is set */
359 if (fm10k_test_staterr(rx_desc
,
360 FM10K_RXD_STATUS_L4E
|
361 FM10K_RXD_STATUS_L4E2
|
362 FM10K_RXD_STATUS_IPE
|
363 FM10K_RXD_STATUS_IPE2
)) {
364 ring
->rx_stats
.csum_err
++;
368 /* It must be a TCP or UDP packet with a valid checksum */
369 if (fm10k_test_staterr(rx_desc
, FM10K_RXD_STATUS_L4CS2
))
370 skb
->encapsulation
= true;
371 else if (!fm10k_test_staterr(rx_desc
, FM10K_RXD_STATUS_L4CS
))
374 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
377 #define FM10K_RSS_L4_TYPES_MASK \
378 ((1ul << FM10K_RSSTYPE_IPV4_TCP) | \
379 (1ul << FM10K_RSSTYPE_IPV4_UDP) | \
380 (1ul << FM10K_RSSTYPE_IPV6_TCP) | \
381 (1ul << FM10K_RSSTYPE_IPV6_UDP))
383 static inline void fm10k_rx_hash(struct fm10k_ring
*ring
,
384 union fm10k_rx_desc
*rx_desc
,
389 if (!(ring
->netdev
->features
& NETIF_F_RXHASH
))
392 rss_type
= le16_to_cpu(rx_desc
->w
.pkt_info
) & FM10K_RXD_RSSTYPE_MASK
;
396 skb_set_hash(skb
, le32_to_cpu(rx_desc
->d
.rss
),
397 (FM10K_RSS_L4_TYPES_MASK
& (1ul << rss_type
)) ?
398 PKT_HASH_TYPE_L4
: PKT_HASH_TYPE_L3
);
401 static void fm10k_rx_hwtstamp(struct fm10k_ring
*rx_ring
,
402 union fm10k_rx_desc
*rx_desc
,
405 struct fm10k_intfc
*interface
= rx_ring
->q_vector
->interface
;
407 FM10K_CB(skb
)->tstamp
= rx_desc
->q
.timestamp
;
409 if (unlikely(interface
->flags
& FM10K_FLAG_RX_TS_ENABLED
))
410 fm10k_systime_to_hwtstamp(interface
, skb_hwtstamps(skb
),
411 le64_to_cpu(rx_desc
->q
.timestamp
));
414 static void fm10k_type_trans(struct fm10k_ring
*rx_ring
,
415 union fm10k_rx_desc
*rx_desc
,
418 struct net_device
*dev
= rx_ring
->netdev
;
419 struct fm10k_l2_accel
*l2_accel
= rcu_dereference_bh(rx_ring
->l2_accel
);
421 /* check to see if DGLORT belongs to a MACVLAN */
423 u16 idx
= le16_to_cpu(FM10K_CB(skb
)->fi
.w
.dglort
) - 1;
425 idx
-= l2_accel
->dglort
;
426 if (idx
< l2_accel
->size
&& l2_accel
->macvlan
[idx
])
427 dev
= l2_accel
->macvlan
[idx
];
432 skb
->protocol
= eth_type_trans(skb
, dev
);
437 /* update MACVLAN statistics */
438 macvlan_count_rx(netdev_priv(dev
), skb
->len
+ ETH_HLEN
, 1,
439 !!(rx_desc
->w
.hdr_info
&
440 cpu_to_le16(FM10K_RXD_HDR_INFO_XC_MASK
)));
444 * fm10k_process_skb_fields - Populate skb header fields from Rx descriptor
445 * @rx_ring: rx descriptor ring packet is being transacted on
446 * @rx_desc: pointer to the EOP Rx descriptor
447 * @skb: pointer to current skb being populated
449 * This function checks the ring, descriptor, and packet information in
450 * order to populate the hash, checksum, VLAN, timestamp, protocol, and
451 * other fields within the skb.
453 static unsigned int fm10k_process_skb_fields(struct fm10k_ring
*rx_ring
,
454 union fm10k_rx_desc
*rx_desc
,
457 unsigned int len
= skb
->len
;
459 fm10k_rx_hash(rx_ring
, rx_desc
, skb
);
461 fm10k_rx_checksum(rx_ring
, rx_desc
, skb
);
463 fm10k_rx_hwtstamp(rx_ring
, rx_desc
, skb
);
465 FM10K_CB(skb
)->fi
.w
.vlan
= rx_desc
->w
.vlan
;
467 skb_record_rx_queue(skb
, rx_ring
->queue_index
);
469 FM10K_CB(skb
)->fi
.d
.glort
= rx_desc
->d
.glort
;
471 if (rx_desc
->w
.vlan
) {
472 u16 vid
= le16_to_cpu(rx_desc
->w
.vlan
);
474 if (vid
!= rx_ring
->vid
)
475 __vlan_hwaccel_put_tag(skb
, htons(ETH_P_8021Q
), vid
);
478 fm10k_type_trans(rx_ring
, rx_desc
, skb
);
484 * fm10k_is_non_eop - process handling of non-EOP buffers
485 * @rx_ring: Rx ring being processed
486 * @rx_desc: Rx descriptor for current buffer
488 * This function updates next to clean. If the buffer is an EOP buffer
489 * this function exits returning false, otherwise it will place the
490 * sk_buff in the next buffer to be chained and return true indicating
491 * that this is in fact a non-EOP buffer.
493 static bool fm10k_is_non_eop(struct fm10k_ring
*rx_ring
,
494 union fm10k_rx_desc
*rx_desc
)
496 u32 ntc
= rx_ring
->next_to_clean
+ 1;
498 /* fetch, update, and store next to clean */
499 ntc
= (ntc
< rx_ring
->count
) ? ntc
: 0;
500 rx_ring
->next_to_clean
= ntc
;
502 prefetch(FM10K_RX_DESC(rx_ring
, ntc
));
504 if (likely(fm10k_test_staterr(rx_desc
, FM10K_RXD_STATUS_EOP
)))
511 * fm10k_pull_tail - fm10k specific version of skb_pull_tail
512 * @rx_ring: rx descriptor ring packet is being transacted on
513 * @rx_desc: pointer to the EOP Rx descriptor
514 * @skb: pointer to current skb being adjusted
516 * This function is an fm10k specific version of __pskb_pull_tail. The
517 * main difference between this version and the original function is that
518 * this function can make several assumptions about the state of things
519 * that allow for significant optimizations versus the standard function.
520 * As a result we can do things like drop a frag and maintain an accurate
521 * truesize for the skb.
523 static void fm10k_pull_tail(struct fm10k_ring
*rx_ring
,
524 union fm10k_rx_desc
*rx_desc
,
527 struct skb_frag_struct
*frag
= &skb_shinfo(skb
)->frags
[0];
529 unsigned int pull_len
;
531 /* it is valid to use page_address instead of kmap since we are
532 * working with pages allocated out of the lomem pool per
533 * alloc_page(GFP_ATOMIC)
535 va
= skb_frag_address(frag
);
537 /* we need the header to contain the greater of either ETH_HLEN or
538 * 60 bytes if the skb->len is less than 60 for skb_pad.
540 pull_len
= eth_get_headlen(va
, FM10K_RX_HDR_LEN
);
542 /* align pull length to size of long to optimize memcpy performance */
543 skb_copy_to_linear_data(skb
, va
, ALIGN(pull_len
, sizeof(long)));
545 /* update all of the pointers */
546 skb_frag_size_sub(frag
, pull_len
);
547 frag
->page_offset
+= pull_len
;
548 skb
->data_len
-= pull_len
;
549 skb
->tail
+= pull_len
;
553 * fm10k_cleanup_headers - Correct corrupted or empty headers
554 * @rx_ring: rx descriptor ring packet is being transacted on
555 * @rx_desc: pointer to the EOP Rx descriptor
556 * @skb: pointer to current skb being fixed
558 * Address the case where we are pulling data in on pages only
559 * and as such no data is present in the skb header.
561 * In addition if skb is not at least 60 bytes we need to pad it so that
562 * it is large enough to qualify as a valid Ethernet frame.
564 * Returns true if an error was encountered and skb was freed.
566 static bool fm10k_cleanup_headers(struct fm10k_ring
*rx_ring
,
567 union fm10k_rx_desc
*rx_desc
,
570 if (unlikely((fm10k_test_staterr(rx_desc
,
571 FM10K_RXD_STATUS_RXE
)))) {
572 dev_kfree_skb_any(skb
);
573 rx_ring
->rx_stats
.errors
++;
577 /* place header in linear portion of buffer */
578 if (skb_is_nonlinear(skb
))
579 fm10k_pull_tail(rx_ring
, rx_desc
, skb
);
581 /* if eth_skb_pad returns an error the skb was freed */
582 if (eth_skb_pad(skb
))
589 * fm10k_receive_skb - helper function to handle rx indications
590 * @q_vector: structure containing interrupt and ring information
591 * @skb: packet to send up
593 static void fm10k_receive_skb(struct fm10k_q_vector
*q_vector
,
596 napi_gro_receive(&q_vector
->napi
, skb
);
599 static bool fm10k_clean_rx_irq(struct fm10k_q_vector
*q_vector
,
600 struct fm10k_ring
*rx_ring
,
603 struct sk_buff
*skb
= rx_ring
->skb
;
604 unsigned int total_bytes
= 0, total_packets
= 0;
605 u16 cleaned_count
= fm10k_desc_unused(rx_ring
);
608 union fm10k_rx_desc
*rx_desc
;
610 /* return some buffers to hardware, one at a time is too slow */
611 if (cleaned_count
>= FM10K_RX_BUFFER_WRITE
) {
612 fm10k_alloc_rx_buffers(rx_ring
, cleaned_count
);
616 rx_desc
= FM10K_RX_DESC(rx_ring
, rx_ring
->next_to_clean
);
618 if (!rx_desc
->d
.staterr
)
621 /* This memory barrier is needed to keep us from reading
622 * any other fields out of the rx_desc until we know the
623 * descriptor has been written back
627 /* retrieve a buffer from the ring */
628 skb
= fm10k_fetch_rx_buffer(rx_ring
, rx_desc
, skb
);
630 /* exit if we failed to retrieve a buffer */
636 /* fetch next buffer in frame if non-eop */
637 if (fm10k_is_non_eop(rx_ring
, rx_desc
))
640 /* verify the packet layout is correct */
641 if (fm10k_cleanup_headers(rx_ring
, rx_desc
, skb
)) {
646 /* populate checksum, timestamp, VLAN, and protocol */
647 total_bytes
+= fm10k_process_skb_fields(rx_ring
, rx_desc
, skb
);
649 fm10k_receive_skb(q_vector
, skb
);
651 /* reset skb pointer */
654 /* update budget accounting */
656 } while (likely(total_packets
< budget
));
658 /* place incomplete frames back on ring for completion */
661 u64_stats_update_begin(&rx_ring
->syncp
);
662 rx_ring
->stats
.packets
+= total_packets
;
663 rx_ring
->stats
.bytes
+= total_bytes
;
664 u64_stats_update_end(&rx_ring
->syncp
);
665 q_vector
->rx
.total_packets
+= total_packets
;
666 q_vector
->rx
.total_bytes
+= total_bytes
;
668 return total_packets
< budget
;
671 #define VXLAN_HLEN (sizeof(struct udphdr) + 8)
672 static struct ethhdr
*fm10k_port_is_vxlan(struct sk_buff
*skb
)
674 struct fm10k_intfc
*interface
= netdev_priv(skb
->dev
);
675 struct fm10k_vxlan_port
*vxlan_port
;
677 /* we can only offload a vxlan if we recognize it as such */
678 vxlan_port
= list_first_entry_or_null(&interface
->vxlan_port
,
679 struct fm10k_vxlan_port
, list
);
683 if (vxlan_port
->port
!= udp_hdr(skb
)->dest
)
686 /* return offset of udp_hdr plus 8 bytes for VXLAN header */
687 return (struct ethhdr
*)(skb_transport_header(skb
) + VXLAN_HLEN
);
690 #define FM10K_NVGRE_RESERVED0_FLAGS htons(0x9FFF)
691 #define NVGRE_TNI htons(0x2000)
692 struct fm10k_nvgre_hdr
{
698 static struct ethhdr
*fm10k_gre_is_nvgre(struct sk_buff
*skb
)
700 struct fm10k_nvgre_hdr
*nvgre_hdr
;
701 int hlen
= ip_hdrlen(skb
);
703 /* currently only IPv4 is supported due to hlen above */
704 if (vlan_get_protocol(skb
) != htons(ETH_P_IP
))
707 /* our transport header should be NVGRE */
708 nvgre_hdr
= (struct fm10k_nvgre_hdr
*)(skb_network_header(skb
) + hlen
);
710 /* verify all reserved flags are 0 */
711 if (nvgre_hdr
->flags
& FM10K_NVGRE_RESERVED0_FLAGS
)
714 /* verify protocol is transparent Ethernet bridging */
715 if (nvgre_hdr
->proto
!= htons(ETH_P_TEB
))
718 /* report start of ethernet header */
719 if (nvgre_hdr
->flags
& NVGRE_TNI
)
720 return (struct ethhdr
*)(nvgre_hdr
+ 1);
722 return (struct ethhdr
*)(&nvgre_hdr
->tni
);
725 static __be16
fm10k_tx_encap_offload(struct sk_buff
*skb
)
727 struct ethhdr
*eth_hdr
;
730 switch (vlan_get_protocol(skb
)) {
731 case htons(ETH_P_IP
):
732 l4_hdr
= ip_hdr(skb
)->protocol
;
734 case htons(ETH_P_IPV6
):
735 l4_hdr
= ipv6_hdr(skb
)->nexthdr
;
743 eth_hdr
= fm10k_port_is_vxlan(skb
);
746 eth_hdr
= fm10k_gre_is_nvgre(skb
);
755 switch (eth_hdr
->h_proto
) {
756 case htons(ETH_P_IP
):
757 case htons(ETH_P_IPV6
):
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
;
825 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
828 if (skb
->encapsulation
) {
829 protocol
= fm10k_tx_encap_offload(skb
);
831 if (skb_checksum_help(skb
)) {
832 dev_warn(tx_ring
->dev
,
833 "failed to offload encap csum!\n");
834 tx_ring
->tx_stats
.csum_err
++;
838 network_hdr
.raw
= skb_inner_network_header(skb
);
840 protocol
= vlan_get_protocol(skb
);
841 network_hdr
.raw
= skb_network_header(skb
);
845 case htons(ETH_P_IP
):
846 l4_hdr
= network_hdr
.ipv4
->protocol
;
848 case htons(ETH_P_IPV6
):
849 l4_hdr
= network_hdr
.ipv6
->nexthdr
;
852 if (unlikely(net_ratelimit())) {
853 dev_warn(tx_ring
->dev
,
854 "partial checksum but ip version=%x!\n",
857 tx_ring
->tx_stats
.csum_err
++;
866 if (skb
->encapsulation
)
869 if (unlikely(net_ratelimit())) {
870 dev_warn(tx_ring
->dev
,
871 "partial checksum but l4 proto=%x!\n",
874 tx_ring
->tx_stats
.csum_err
++;
878 /* update TX checksum flag */
879 first
->tx_flags
|= FM10K_TX_FLAGS_CSUM
;
882 /* populate Tx descriptor header size and mss */
883 tx_desc
= FM10K_TX_DESC(tx_ring
, tx_ring
->next_to_use
);
888 #define FM10K_SET_FLAG(_input, _flag, _result) \
889 ((_flag <= _result) ? \
890 ((u32)(_input & _flag) * (_result / _flag)) : \
891 ((u32)(_input & _flag) / (_flag / _result)))
893 static u8
fm10k_tx_desc_flags(struct sk_buff
*skb
, u32 tx_flags
)
895 /* set type for advanced descriptor with frame checksum insertion */
898 /* set timestamping bits */
899 if (unlikely(skb_shinfo(skb
)->tx_flags
& SKBTX_HW_TSTAMP
) &&
900 likely(skb_shinfo(skb
)->tx_flags
& SKBTX_IN_PROGRESS
))
901 desc_flags
|= FM10K_TXD_FLAG_TIME
;
903 /* set checksum offload bits */
904 desc_flags
|= FM10K_SET_FLAG(tx_flags
, FM10K_TX_FLAGS_CSUM
,
905 FM10K_TXD_FLAG_CSUM
);
910 static bool fm10k_tx_desc_push(struct fm10k_ring
*tx_ring
,
911 struct fm10k_tx_desc
*tx_desc
, u16 i
,
912 dma_addr_t dma
, unsigned int size
, u8 desc_flags
)
914 /* set RS and INT for last frame in a cache line */
915 if ((++i
& (FM10K_TXD_WB_FIFO_SIZE
- 1)) == 0)
916 desc_flags
|= FM10K_TXD_FLAG_RS
| FM10K_TXD_FLAG_INT
;
918 /* record values to descriptor */
919 tx_desc
->buffer_addr
= cpu_to_le64(dma
);
920 tx_desc
->flags
= desc_flags
;
921 tx_desc
->buflen
= cpu_to_le16(size
);
923 /* return true if we just wrapped the ring */
924 return i
== tx_ring
->count
;
927 static int __fm10k_maybe_stop_tx(struct fm10k_ring
*tx_ring
, u16 size
)
929 netif_stop_subqueue(tx_ring
->netdev
, tx_ring
->queue_index
);
933 /* We need to check again in a case another CPU has just
934 * made room available. */
935 if (likely(fm10k_desc_unused(tx_ring
) < size
))
938 /* A reprieve! - use start_queue because it doesn't call schedule */
939 netif_start_subqueue(tx_ring
->netdev
, tx_ring
->queue_index
);
940 ++tx_ring
->tx_stats
.restart_queue
;
944 static inline int fm10k_maybe_stop_tx(struct fm10k_ring
*tx_ring
, u16 size
)
946 if (likely(fm10k_desc_unused(tx_ring
) >= size
))
948 return __fm10k_maybe_stop_tx(tx_ring
, size
);
951 static void fm10k_tx_map(struct fm10k_ring
*tx_ring
,
952 struct fm10k_tx_buffer
*first
)
954 struct sk_buff
*skb
= first
->skb
;
955 struct fm10k_tx_buffer
*tx_buffer
;
956 struct fm10k_tx_desc
*tx_desc
;
957 struct skb_frag_struct
*frag
;
960 unsigned int data_len
, size
;
961 u32 tx_flags
= first
->tx_flags
;
962 u16 i
= tx_ring
->next_to_use
;
963 u8 flags
= fm10k_tx_desc_flags(skb
, tx_flags
);
965 tx_desc
= FM10K_TX_DESC(tx_ring
, i
);
967 /* add HW VLAN tag */
968 if (vlan_tx_tag_present(skb
))
969 tx_desc
->vlan
= cpu_to_le16(vlan_tx_tag_get(skb
));
973 size
= skb_headlen(skb
);
976 dma
= dma_map_single(tx_ring
->dev
, data
, size
, DMA_TO_DEVICE
);
978 data_len
= skb
->data_len
;
981 for (frag
= &skb_shinfo(skb
)->frags
[0];; frag
++) {
982 if (dma_mapping_error(tx_ring
->dev
, dma
))
985 /* record length, and DMA address */
986 dma_unmap_len_set(tx_buffer
, len
, size
);
987 dma_unmap_addr_set(tx_buffer
, dma
, dma
);
989 while (unlikely(size
> FM10K_MAX_DATA_PER_TXD
)) {
990 if (fm10k_tx_desc_push(tx_ring
, tx_desc
++, i
++, dma
,
991 FM10K_MAX_DATA_PER_TXD
, flags
)) {
992 tx_desc
= FM10K_TX_DESC(tx_ring
, 0);
996 dma
+= FM10K_MAX_DATA_PER_TXD
;
997 size
-= FM10K_MAX_DATA_PER_TXD
;
1000 if (likely(!data_len
))
1003 if (fm10k_tx_desc_push(tx_ring
, tx_desc
++, i
++,
1004 dma
, size
, flags
)) {
1005 tx_desc
= FM10K_TX_DESC(tx_ring
, 0);
1009 size
= skb_frag_size(frag
);
1012 dma
= skb_frag_dma_map(tx_ring
->dev
, frag
, 0, size
,
1015 tx_buffer
= &tx_ring
->tx_buffer
[i
];
1018 /* write last descriptor with LAST bit set */
1019 flags
|= FM10K_TXD_FLAG_LAST
;
1021 if (fm10k_tx_desc_push(tx_ring
, tx_desc
, i
++, dma
, size
, flags
))
1024 /* record bytecount for BQL */
1025 netdev_tx_sent_queue(txring_txq(tx_ring
), first
->bytecount
);
1027 /* record SW timestamp if HW timestamp is not available */
1028 skb_tx_timestamp(first
->skb
);
1030 /* Force memory writes to complete before letting h/w know there
1031 * are new descriptors to fetch. (Only applicable for weak-ordered
1032 * memory model archs, such as IA-64).
1034 * We also need this memory barrier to make certain all of the
1035 * status bits have been updated before next_to_watch is written.
1039 /* set next_to_watch value indicating a packet is present */
1040 first
->next_to_watch
= tx_desc
;
1042 tx_ring
->next_to_use
= i
;
1044 /* Make sure there is space in the ring for the next send. */
1045 fm10k_maybe_stop_tx(tx_ring
, DESC_NEEDED
);
1047 /* notify HW of packet */
1048 if (netif_xmit_stopped(txring_txq(tx_ring
)) || !skb
->xmit_more
) {
1049 writel(i
, tx_ring
->tail
);
1051 /* we need this if more than one processor can write to our tail
1052 * at a time, it synchronizes IO on IA64/Altix systems
1059 dev_err(tx_ring
->dev
, "TX DMA map failed\n");
1061 /* clear dma mappings for failed tx_buffer map */
1063 tx_buffer
= &tx_ring
->tx_buffer
[i
];
1064 fm10k_unmap_and_free_tx_resource(tx_ring
, tx_buffer
);
1065 if (tx_buffer
== first
)
1072 tx_ring
->next_to_use
= i
;
1075 netdev_tx_t
fm10k_xmit_frame_ring(struct sk_buff
*skb
,
1076 struct fm10k_ring
*tx_ring
)
1078 struct fm10k_tx_buffer
*first
;
1081 #if PAGE_SIZE > FM10K_MAX_DATA_PER_TXD
1084 u16 count
= TXD_USE_COUNT(skb_headlen(skb
));
1086 /* need: 1 descriptor per page * PAGE_SIZE/FM10K_MAX_DATA_PER_TXD,
1087 * + 1 desc for skb_headlen/FM10K_MAX_DATA_PER_TXD,
1088 * + 2 desc gap to keep tail from touching head
1089 * otherwise try next time
1091 #if PAGE_SIZE > FM10K_MAX_DATA_PER_TXD
1092 for (f
= 0; f
< skb_shinfo(skb
)->nr_frags
; f
++)
1093 count
+= TXD_USE_COUNT(skb_shinfo(skb
)->frags
[f
].size
);
1095 count
+= skb_shinfo(skb
)->nr_frags
;
1097 if (fm10k_maybe_stop_tx(tx_ring
, count
+ 3)) {
1098 tx_ring
->tx_stats
.tx_busy
++;
1099 return NETDEV_TX_BUSY
;
1102 /* record the location of the first descriptor for this packet */
1103 first
= &tx_ring
->tx_buffer
[tx_ring
->next_to_use
];
1105 first
->bytecount
= max_t(unsigned int, skb
->len
, ETH_ZLEN
);
1106 first
->gso_segs
= 1;
1108 /* record initial flags and protocol */
1109 first
->tx_flags
= tx_flags
;
1111 tso
= fm10k_tso(tx_ring
, first
);
1115 fm10k_tx_csum(tx_ring
, first
);
1117 fm10k_tx_map(tx_ring
, first
);
1119 return NETDEV_TX_OK
;
1122 dev_kfree_skb_any(first
->skb
);
1125 return NETDEV_TX_OK
;
1128 static u64
fm10k_get_tx_completed(struct fm10k_ring
*ring
)
1130 return ring
->stats
.packets
;
1133 static u64
fm10k_get_tx_pending(struct fm10k_ring
*ring
)
1135 /* use SW head and tail until we have real hardware */
1136 u32 head
= ring
->next_to_clean
;
1137 u32 tail
= ring
->next_to_use
;
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 netdev_err(interface
->netdev
, "Reset interface\n");
1180 interface
->tx_timeout_count
++;
1181 interface
->flags
|= FM10K_FLAG_RESET_REQUESTED
;
1182 fm10k_service_event_schedule(interface
);
1187 * fm10k_clean_tx_irq - Reclaim resources after transmit completes
1188 * @q_vector: structure containing interrupt and ring information
1189 * @tx_ring: tx ring to clean
1191 static bool fm10k_clean_tx_irq(struct fm10k_q_vector
*q_vector
,
1192 struct fm10k_ring
*tx_ring
)
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 dev_consume_skb_any(tx_buffer
->skb
);
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
;
1355 /* Only update ITR if we are using adaptive setting */
1356 if (!(ring_container
->itr
& FM10K_ITR_ADAPTIVE
))
1359 packets
= ring_container
->total_packets
;
1363 avg_wire_size
= ring_container
->total_bytes
/ packets
;
1365 /* Add 24 bytes to size to account for CRC, preamble, and gap */
1366 avg_wire_size
+= 24;
1368 /* Don't starve jumbo frames */
1369 if (avg_wire_size
> 3000)
1370 avg_wire_size
= 3000;
1372 /* Give a little boost to mid-size frames */
1373 if ((avg_wire_size
> 300) && (avg_wire_size
< 1200))
1378 /* write back value and retain adaptive flag */
1379 ring_container
->itr
= avg_wire_size
| FM10K_ITR_ADAPTIVE
;
1382 ring_container
->total_bytes
= 0;
1383 ring_container
->total_packets
= 0;
1386 static void fm10k_qv_enable(struct fm10k_q_vector
*q_vector
)
1388 /* Enable auto-mask and clear the current mask */
1389 u32 itr
= FM10K_ITR_ENABLE
;
1392 fm10k_update_itr(&q_vector
->tx
);
1395 fm10k_update_itr(&q_vector
->rx
);
1397 /* Store Tx itr in timer slot 0 */
1398 itr
|= (q_vector
->tx
.itr
& FM10K_ITR_MAX
);
1400 /* Shift Rx itr to timer slot 1 */
1401 itr
|= (q_vector
->rx
.itr
& FM10K_ITR_MAX
) << FM10K_ITR_INTERVAL1_SHIFT
;
1403 /* Write the final value to the ITR register */
1404 writel(itr
, q_vector
->itr
);
1407 static int fm10k_poll(struct napi_struct
*napi
, int budget
)
1409 struct fm10k_q_vector
*q_vector
=
1410 container_of(napi
, struct fm10k_q_vector
, napi
);
1411 struct fm10k_ring
*ring
;
1412 int per_ring_budget
;
1413 bool clean_complete
= true;
1415 fm10k_for_each_ring(ring
, q_vector
->tx
)
1416 clean_complete
&= fm10k_clean_tx_irq(q_vector
, ring
);
1418 /* attempt to distribute budget to each queue fairly, but don't
1419 * allow the budget to go below 1 because we'll exit polling
1421 if (q_vector
->rx
.count
> 1)
1422 per_ring_budget
= max(budget
/q_vector
->rx
.count
, 1);
1424 per_ring_budget
= budget
;
1426 fm10k_for_each_ring(ring
, q_vector
->rx
)
1427 clean_complete
&= fm10k_clean_rx_irq(q_vector
, ring
,
1430 /* If all work not completed, return budget and keep polling */
1431 if (!clean_complete
)
1434 /* all work done, exit the polling mode */
1435 napi_complete(napi
);
1437 /* re-enable the q_vector */
1438 fm10k_qv_enable(q_vector
);
1444 * fm10k_set_qos_queues: Allocate queues for a QOS-enabled device
1445 * @interface: board private structure to initialize
1447 * When QoS (Quality of Service) is enabled, allocate queues for
1448 * each traffic class. If multiqueue isn't available,then abort QoS
1451 * This function handles all combinations of Qos and RSS.
1454 static bool fm10k_set_qos_queues(struct fm10k_intfc
*interface
)
1456 struct net_device
*dev
= interface
->netdev
;
1457 struct fm10k_ring_feature
*f
;
1461 /* Map queue offset and counts onto allocated tx queues */
1462 pcs
= netdev_get_num_tc(dev
);
1467 /* set QoS mask and indices */
1468 f
= &interface
->ring_feature
[RING_F_QOS
];
1470 f
->mask
= (1 << fls(pcs
- 1)) - 1;
1472 /* determine the upper limit for our current DCB mode */
1473 rss_i
= interface
->hw
.mac
.max_queues
/ pcs
;
1474 rss_i
= 1 << (fls(rss_i
) - 1);
1476 /* set RSS mask and indices */
1477 f
= &interface
->ring_feature
[RING_F_RSS
];
1478 rss_i
= min_t(u16
, rss_i
, f
->limit
);
1480 f
->mask
= (1 << fls(rss_i
- 1)) - 1;
1482 /* configure pause class to queue mapping */
1483 for (i
= 0; i
< pcs
; i
++)
1484 netdev_set_tc_queue(dev
, i
, rss_i
, rss_i
* i
);
1486 interface
->num_rx_queues
= rss_i
* pcs
;
1487 interface
->num_tx_queues
= rss_i
* pcs
;
1493 * fm10k_set_rss_queues: Allocate queues for RSS
1494 * @interface: board private structure to initialize
1496 * This is our "base" multiqueue mode. RSS (Receive Side Scaling) will try
1497 * to allocate one Rx queue per CPU, and if available, one Tx queue per CPU.
1500 static bool fm10k_set_rss_queues(struct fm10k_intfc
*interface
)
1502 struct fm10k_ring_feature
*f
;
1505 f
= &interface
->ring_feature
[RING_F_RSS
];
1506 rss_i
= min_t(u16
, interface
->hw
.mac
.max_queues
, f
->limit
);
1508 /* record indices and power of 2 mask for RSS */
1510 f
->mask
= (1 << fls(rss_i
- 1)) - 1;
1512 interface
->num_rx_queues
= rss_i
;
1513 interface
->num_tx_queues
= rss_i
;
1519 * fm10k_set_num_queues: Allocate queues for device, feature dependent
1520 * @interface: board private structure to initialize
1522 * This is the top level queue allocation routine. The order here is very
1523 * important, starting with the "most" number of features turned on at once,
1524 * and ending with the smallest set of features. This way large combinations
1525 * can be allocated if they're turned on, and smaller combinations are the
1526 * fallthrough conditions.
1529 static void fm10k_set_num_queues(struct fm10k_intfc
*interface
)
1531 /* Start with base case */
1532 interface
->num_rx_queues
= 1;
1533 interface
->num_tx_queues
= 1;
1535 if (fm10k_set_qos_queues(interface
))
1538 fm10k_set_rss_queues(interface
);
1542 * fm10k_alloc_q_vector - Allocate memory for a single interrupt vector
1543 * @interface: board private structure to initialize
1544 * @v_count: q_vectors allocated on interface, used for ring interleaving
1545 * @v_idx: index of vector in interface struct
1546 * @txr_count: total number of Tx rings to allocate
1547 * @txr_idx: index of first Tx ring to allocate
1548 * @rxr_count: total number of Rx rings to allocate
1549 * @rxr_idx: index of first Rx ring to allocate
1551 * We allocate one q_vector. If allocation fails we return -ENOMEM.
1553 static int fm10k_alloc_q_vector(struct fm10k_intfc
*interface
,
1554 unsigned int v_count
, unsigned int v_idx
,
1555 unsigned int txr_count
, unsigned int txr_idx
,
1556 unsigned int rxr_count
, unsigned int rxr_idx
)
1558 struct fm10k_q_vector
*q_vector
;
1559 struct fm10k_ring
*ring
;
1560 int ring_count
, size
;
1562 ring_count
= txr_count
+ rxr_count
;
1563 size
= sizeof(struct fm10k_q_vector
) +
1564 (sizeof(struct fm10k_ring
) * ring_count
);
1566 /* allocate q_vector and rings */
1567 q_vector
= kzalloc(size
, GFP_KERNEL
);
1571 /* initialize NAPI */
1572 netif_napi_add(interface
->netdev
, &q_vector
->napi
,
1573 fm10k_poll
, NAPI_POLL_WEIGHT
);
1575 /* tie q_vector and interface together */
1576 interface
->q_vector
[v_idx
] = q_vector
;
1577 q_vector
->interface
= interface
;
1578 q_vector
->v_idx
= v_idx
;
1580 /* initialize pointer to rings */
1581 ring
= q_vector
->ring
;
1583 /* save Tx ring container info */
1584 q_vector
->tx
.ring
= ring
;
1585 q_vector
->tx
.work_limit
= FM10K_DEFAULT_TX_WORK
;
1586 q_vector
->tx
.itr
= interface
->tx_itr
;
1587 q_vector
->tx
.count
= txr_count
;
1590 /* assign generic ring traits */
1591 ring
->dev
= &interface
->pdev
->dev
;
1592 ring
->netdev
= interface
->netdev
;
1594 /* configure backlink on ring */
1595 ring
->q_vector
= q_vector
;
1597 /* apply Tx specific ring traits */
1598 ring
->count
= interface
->tx_ring_count
;
1599 ring
->queue_index
= txr_idx
;
1601 /* assign ring to interface */
1602 interface
->tx_ring
[txr_idx
] = ring
;
1604 /* update count and index */
1608 /* push pointer to next ring */
1612 /* save Rx ring container info */
1613 q_vector
->rx
.ring
= ring
;
1614 q_vector
->rx
.itr
= interface
->rx_itr
;
1615 q_vector
->rx
.count
= rxr_count
;
1618 /* assign generic ring traits */
1619 ring
->dev
= &interface
->pdev
->dev
;
1620 ring
->netdev
= interface
->netdev
;
1621 rcu_assign_pointer(ring
->l2_accel
, interface
->l2_accel
);
1623 /* configure backlink on ring */
1624 ring
->q_vector
= q_vector
;
1626 /* apply Rx specific ring traits */
1627 ring
->count
= interface
->rx_ring_count
;
1628 ring
->queue_index
= rxr_idx
;
1630 /* assign ring to interface */
1631 interface
->rx_ring
[rxr_idx
] = ring
;
1633 /* update count and index */
1637 /* push pointer to next ring */
1641 fm10k_dbg_q_vector_init(q_vector
);
1647 * fm10k_free_q_vector - Free memory allocated for specific interrupt vector
1648 * @interface: board private structure to initialize
1649 * @v_idx: Index of vector to be freed
1651 * This function frees the memory allocated to the q_vector. In addition if
1652 * NAPI is enabled it will delete any references to the NAPI struct prior
1653 * to freeing the q_vector.
1655 static void fm10k_free_q_vector(struct fm10k_intfc
*interface
, int v_idx
)
1657 struct fm10k_q_vector
*q_vector
= interface
->q_vector
[v_idx
];
1658 struct fm10k_ring
*ring
;
1660 fm10k_dbg_q_vector_exit(q_vector
);
1662 fm10k_for_each_ring(ring
, q_vector
->tx
)
1663 interface
->tx_ring
[ring
->queue_index
] = NULL
;
1665 fm10k_for_each_ring(ring
, q_vector
->rx
)
1666 interface
->rx_ring
[ring
->queue_index
] = NULL
;
1668 interface
->q_vector
[v_idx
] = NULL
;
1669 netif_napi_del(&q_vector
->napi
);
1670 kfree_rcu(q_vector
, rcu
);
1674 * fm10k_alloc_q_vectors - Allocate memory for interrupt vectors
1675 * @interface: board private structure to initialize
1677 * We allocate one q_vector per queue interrupt. If allocation fails we
1680 static int fm10k_alloc_q_vectors(struct fm10k_intfc
*interface
)
1682 unsigned int q_vectors
= interface
->num_q_vectors
;
1683 unsigned int rxr_remaining
= interface
->num_rx_queues
;
1684 unsigned int txr_remaining
= interface
->num_tx_queues
;
1685 unsigned int rxr_idx
= 0, txr_idx
= 0, v_idx
= 0;
1688 if (q_vectors
>= (rxr_remaining
+ txr_remaining
)) {
1689 for (; rxr_remaining
; v_idx
++) {
1690 err
= fm10k_alloc_q_vector(interface
, q_vectors
, v_idx
,
1695 /* update counts and index */
1701 for (; v_idx
< q_vectors
; v_idx
++) {
1702 int rqpv
= DIV_ROUND_UP(rxr_remaining
, q_vectors
- v_idx
);
1703 int tqpv
= DIV_ROUND_UP(txr_remaining
, q_vectors
- v_idx
);
1705 err
= fm10k_alloc_q_vector(interface
, q_vectors
, v_idx
,
1712 /* update counts and index */
1713 rxr_remaining
-= rqpv
;
1714 txr_remaining
-= tqpv
;
1722 interface
->num_tx_queues
= 0;
1723 interface
->num_rx_queues
= 0;
1724 interface
->num_q_vectors
= 0;
1727 fm10k_free_q_vector(interface
, v_idx
);
1733 * fm10k_free_q_vectors - Free memory allocated for interrupt vectors
1734 * @interface: board private structure to initialize
1736 * This function frees the memory allocated to the q_vectors. In addition if
1737 * NAPI is enabled it will delete any references to the NAPI struct prior
1738 * to freeing the q_vector.
1740 static void fm10k_free_q_vectors(struct fm10k_intfc
*interface
)
1742 int v_idx
= interface
->num_q_vectors
;
1744 interface
->num_tx_queues
= 0;
1745 interface
->num_rx_queues
= 0;
1746 interface
->num_q_vectors
= 0;
1749 fm10k_free_q_vector(interface
, v_idx
);
1753 * f10k_reset_msix_capability - reset MSI-X capability
1754 * @interface: board private structure to initialize
1756 * Reset the MSI-X capability back to its starting state
1758 static void fm10k_reset_msix_capability(struct fm10k_intfc
*interface
)
1760 pci_disable_msix(interface
->pdev
);
1761 kfree(interface
->msix_entries
);
1762 interface
->msix_entries
= NULL
;
1766 * f10k_init_msix_capability - configure MSI-X capability
1767 * @interface: board private structure to initialize
1769 * Attempt to configure the interrupts using the best available
1770 * capabilities of the hardware and the kernel.
1772 static int fm10k_init_msix_capability(struct fm10k_intfc
*interface
)
1774 struct fm10k_hw
*hw
= &interface
->hw
;
1775 int v_budget
, vector
;
1777 /* It's easy to be greedy for MSI-X vectors, but it really
1778 * doesn't do us much good if we have a lot more vectors
1779 * than CPU's. So let's be conservative and only ask for
1780 * (roughly) the same number of vectors as there are CPU's.
1781 * the default is to use pairs of vectors
1783 v_budget
= max(interface
->num_rx_queues
, interface
->num_tx_queues
);
1784 v_budget
= min_t(u16
, v_budget
, num_online_cpus());
1786 /* account for vectors not related to queues */
1787 v_budget
+= NON_Q_VECTORS(hw
);
1789 /* At the same time, hardware can only support a maximum of
1790 * hw.mac->max_msix_vectors vectors. With features
1791 * such as RSS and VMDq, we can easily surpass the number of Rx and Tx
1792 * descriptor queues supported by our device. Thus, we cap it off in
1793 * those rare cases where the cpu count also exceeds our vector limit.
1795 v_budget
= min_t(int, v_budget
, hw
->mac
.max_msix_vectors
);
1797 /* A failure in MSI-X entry allocation is fatal. */
1798 interface
->msix_entries
= kcalloc(v_budget
, sizeof(struct msix_entry
),
1800 if (!interface
->msix_entries
)
1803 /* populate entry values */
1804 for (vector
= 0; vector
< v_budget
; vector
++)
1805 interface
->msix_entries
[vector
].entry
= vector
;
1807 /* Attempt to enable MSI-X with requested value */
1808 v_budget
= pci_enable_msix_range(interface
->pdev
,
1809 interface
->msix_entries
,
1813 kfree(interface
->msix_entries
);
1814 interface
->msix_entries
= NULL
;
1818 /* record the number of queues available for q_vectors */
1819 interface
->num_q_vectors
= v_budget
- NON_Q_VECTORS(hw
);
1825 * fm10k_cache_ring_qos - Descriptor ring to register mapping for QoS
1826 * @interface: Interface structure continaining rings and devices
1828 * Cache the descriptor ring offsets for Qos
1830 static bool fm10k_cache_ring_qos(struct fm10k_intfc
*interface
)
1832 struct net_device
*dev
= interface
->netdev
;
1833 int pc
, offset
, rss_i
, i
, q_idx
;
1834 u16 pc_stride
= interface
->ring_feature
[RING_F_QOS
].mask
+ 1;
1835 u8 num_pcs
= netdev_get_num_tc(dev
);
1840 rss_i
= interface
->ring_feature
[RING_F_RSS
].indices
;
1842 for (pc
= 0, offset
= 0; pc
< num_pcs
; pc
++, offset
+= rss_i
) {
1844 for (i
= 0; i
< rss_i
; i
++) {
1845 interface
->tx_ring
[offset
+ i
]->reg_idx
= q_idx
;
1846 interface
->tx_ring
[offset
+ i
]->qos_pc
= pc
;
1847 interface
->rx_ring
[offset
+ i
]->reg_idx
= q_idx
;
1848 interface
->rx_ring
[offset
+ i
]->qos_pc
= pc
;
1857 * fm10k_cache_ring_rss - Descriptor ring to register mapping for RSS
1858 * @interface: Interface structure continaining rings and devices
1860 * Cache the descriptor ring offsets for RSS
1862 static void fm10k_cache_ring_rss(struct fm10k_intfc
*interface
)
1866 for (i
= 0; i
< interface
->num_rx_queues
; i
++)
1867 interface
->rx_ring
[i
]->reg_idx
= i
;
1869 for (i
= 0; i
< interface
->num_tx_queues
; i
++)
1870 interface
->tx_ring
[i
]->reg_idx
= i
;
1874 * fm10k_assign_rings - Map rings to network devices
1875 * @interface: Interface structure containing rings and devices
1877 * This function is meant to go though and configure both the network
1878 * devices so that they contain rings, and configure the rings so that
1879 * they function with their network devices.
1881 static void fm10k_assign_rings(struct fm10k_intfc
*interface
)
1883 if (fm10k_cache_ring_qos(interface
))
1886 fm10k_cache_ring_rss(interface
);
1889 static void fm10k_init_reta(struct fm10k_intfc
*interface
)
1891 u16 i
, rss_i
= interface
->ring_feature
[RING_F_RSS
].indices
;
1894 /* If the netdev is initialized we have to maintain table if possible */
1895 if (interface
->netdev
->reg_state
) {
1896 for (i
= FM10K_RETA_SIZE
; i
--;) {
1897 reta
= interface
->reta
[i
];
1898 if ((((reta
<< 24) >> 24) < rss_i
) &&
1899 (((reta
<< 16) >> 24) < rss_i
) &&
1900 (((reta
<< 8) >> 24) < rss_i
) &&
1901 (((reta
) >> 24) < rss_i
))
1903 goto repopulate_reta
;
1906 /* do nothing if all of the elements are in bounds */
1911 /* Populate the redirection table 4 entries at a time. To do this
1912 * we are generating the results for n and n+2 and then interleaving
1913 * those with the results with n+1 and n+3.
1915 for (i
= FM10K_RETA_SIZE
; i
--;) {
1916 /* first pass generates n and n+2 */
1917 base
= ((i
* 0x00040004) + 0x00020000) * rss_i
;
1918 reta
= (base
& 0x3F803F80) >> 7;
1920 /* second pass generates n+1 and n+3 */
1921 base
+= 0x00010001 * rss_i
;
1922 reta
|= (base
& 0x3F803F80) << 1;
1924 interface
->reta
[i
] = reta
;
1929 * fm10k_init_queueing_scheme - Determine proper queueing scheme
1930 * @interface: board private structure to initialize
1932 * We determine which queueing scheme to use based on...
1933 * - Hardware queue count (num_*_queues)
1934 * - defined by miscellaneous hardware support/features (RSS, etc.)
1936 int fm10k_init_queueing_scheme(struct fm10k_intfc
*interface
)
1940 /* Number of supported queues */
1941 fm10k_set_num_queues(interface
);
1943 /* Configure MSI-X capability */
1944 err
= fm10k_init_msix_capability(interface
);
1946 dev_err(&interface
->pdev
->dev
,
1947 "Unable to initialize MSI-X capability\n");
1951 /* Allocate memory for queues */
1952 err
= fm10k_alloc_q_vectors(interface
);
1956 /* Map rings to devices, and map devices to physical queues */
1957 fm10k_assign_rings(interface
);
1959 /* Initialize RSS redirection table */
1960 fm10k_init_reta(interface
);
1966 * fm10k_clear_queueing_scheme - Clear the current queueing scheme settings
1967 * @interface: board private structure to clear queueing scheme on
1969 * We go through and clear queueing specific resources and reset the structure
1970 * to pre-load conditions
1972 void fm10k_clear_queueing_scheme(struct fm10k_intfc
*interface
)
1974 fm10k_free_q_vectors(interface
);
1975 fm10k_reset_msix_capability(interface
);