1 /*******************************************************************************
3 * Intel Ethernet Controller XL710 Family Linux Virtual Function Driver
4 * Copyright(c) 2013 - 2016 Intel Corporation.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms and conditions of the GNU General Public License,
8 * version 2, as published by the Free Software Foundation.
10 * This program is distributed in the hope it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program. If not, see <http://www.gnu.org/licenses/>.
18 * The full GNU General Public License is included in this distribution in
19 * the file called "COPYING".
21 * Contact Information:
22 * e1000-devel Mailing List <e1000-devel@lists.sourceforge.net>
23 * Intel Corporation, 5200 N.E. Elam Young Parkway, Hillsboro, OR 97124-6497
25 ******************************************************************************/
27 #include <linux/prefetch.h>
28 #include <net/busy_poll.h>
31 #include "i40e_prototype.h"
33 static inline __le64
build_ctob(u32 td_cmd
, u32 td_offset
, unsigned int size
,
36 return cpu_to_le64(I40E_TX_DESC_DTYPE_DATA
|
37 ((u64
)td_cmd
<< I40E_TXD_QW1_CMD_SHIFT
) |
38 ((u64
)td_offset
<< I40E_TXD_QW1_OFFSET_SHIFT
) |
39 ((u64
)size
<< I40E_TXD_QW1_TX_BUF_SZ_SHIFT
) |
40 ((u64
)td_tag
<< I40E_TXD_QW1_L2TAG1_SHIFT
));
43 #define I40E_TXD_CMD (I40E_TX_DESC_CMD_EOP | I40E_TX_DESC_CMD_RS)
46 * i40e_unmap_and_free_tx_resource - Release a Tx buffer
47 * @ring: the ring that owns the buffer
48 * @tx_buffer: the buffer to free
50 static void i40e_unmap_and_free_tx_resource(struct i40e_ring
*ring
,
51 struct i40e_tx_buffer
*tx_buffer
)
54 dev_kfree_skb_any(tx_buffer
->skb
);
55 if (dma_unmap_len(tx_buffer
, len
))
56 dma_unmap_single(ring
->dev
,
57 dma_unmap_addr(tx_buffer
, dma
),
58 dma_unmap_len(tx_buffer
, len
),
60 } else if (dma_unmap_len(tx_buffer
, len
)) {
61 dma_unmap_page(ring
->dev
,
62 dma_unmap_addr(tx_buffer
, dma
),
63 dma_unmap_len(tx_buffer
, len
),
67 if (tx_buffer
->tx_flags
& I40E_TX_FLAGS_FD_SB
)
68 kfree(tx_buffer
->raw_buf
);
70 tx_buffer
->next_to_watch
= NULL
;
71 tx_buffer
->skb
= NULL
;
72 dma_unmap_len_set(tx_buffer
, len
, 0);
73 /* tx_buffer must be completely set up in the transmit path */
77 * i40evf_clean_tx_ring - Free any empty Tx buffers
78 * @tx_ring: ring to be cleaned
80 void i40evf_clean_tx_ring(struct i40e_ring
*tx_ring
)
82 unsigned long bi_size
;
85 /* ring already cleared, nothing to do */
89 /* Free all the Tx ring sk_buffs */
90 for (i
= 0; i
< tx_ring
->count
; i
++)
91 i40e_unmap_and_free_tx_resource(tx_ring
, &tx_ring
->tx_bi
[i
]);
93 bi_size
= sizeof(struct i40e_tx_buffer
) * tx_ring
->count
;
94 memset(tx_ring
->tx_bi
, 0, bi_size
);
96 /* Zero out the descriptor ring */
97 memset(tx_ring
->desc
, 0, tx_ring
->size
);
99 tx_ring
->next_to_use
= 0;
100 tx_ring
->next_to_clean
= 0;
102 if (!tx_ring
->netdev
)
105 /* cleanup Tx queue statistics */
106 netdev_tx_reset_queue(netdev_get_tx_queue(tx_ring
->netdev
,
107 tx_ring
->queue_index
));
111 * i40evf_free_tx_resources - Free Tx resources per queue
112 * @tx_ring: Tx descriptor ring for a specific queue
114 * Free all transmit software resources
116 void i40evf_free_tx_resources(struct i40e_ring
*tx_ring
)
118 i40evf_clean_tx_ring(tx_ring
);
119 kfree(tx_ring
->tx_bi
);
120 tx_ring
->tx_bi
= NULL
;
123 dma_free_coherent(tx_ring
->dev
, tx_ring
->size
,
124 tx_ring
->desc
, tx_ring
->dma
);
125 tx_ring
->desc
= NULL
;
130 * i40evf_get_tx_pending - how many Tx descriptors not processed
131 * @tx_ring: the ring of descriptors
132 * @in_sw: is tx_pending being checked in SW or HW
134 * Since there is no access to the ring head register
135 * in XL710, we need to use our local copies
137 u32
i40evf_get_tx_pending(struct i40e_ring
*ring
, bool in_sw
)
142 head
= i40e_get_head(ring
);
144 head
= ring
->next_to_clean
;
145 tail
= readl(ring
->tail
);
148 return (head
< tail
) ?
149 tail
- head
: (tail
+ ring
->count
- head
);
154 #define WB_STRIDE 0x3
157 * i40e_clean_tx_irq - Reclaim resources after transmit completes
158 * @vsi: the VSI we care about
159 * @tx_ring: Tx ring to clean
160 * @napi_budget: Used to determine if we are in netpoll
162 * Returns true if there's any budget left (e.g. the clean is finished)
164 static bool i40e_clean_tx_irq(struct i40e_vsi
*vsi
,
165 struct i40e_ring
*tx_ring
, int napi_budget
)
167 u16 i
= tx_ring
->next_to_clean
;
168 struct i40e_tx_buffer
*tx_buf
;
169 struct i40e_tx_desc
*tx_head
;
170 struct i40e_tx_desc
*tx_desc
;
171 unsigned int total_bytes
= 0, total_packets
= 0;
172 unsigned int budget
= vsi
->work_limit
;
174 tx_buf
= &tx_ring
->tx_bi
[i
];
175 tx_desc
= I40E_TX_DESC(tx_ring
, i
);
178 tx_head
= I40E_TX_DESC(tx_ring
, i40e_get_head(tx_ring
));
181 struct i40e_tx_desc
*eop_desc
= tx_buf
->next_to_watch
;
183 /* if next_to_watch is not set then there is no work pending */
187 /* prevent any other reads prior to eop_desc */
188 read_barrier_depends();
190 /* we have caught up to head, no work left to do */
191 if (tx_head
== tx_desc
)
194 /* clear next_to_watch to prevent false hangs */
195 tx_buf
->next_to_watch
= NULL
;
197 /* update the statistics for this packet */
198 total_bytes
+= tx_buf
->bytecount
;
199 total_packets
+= tx_buf
->gso_segs
;
202 napi_consume_skb(tx_buf
->skb
, napi_budget
);
204 /* unmap skb header data */
205 dma_unmap_single(tx_ring
->dev
,
206 dma_unmap_addr(tx_buf
, dma
),
207 dma_unmap_len(tx_buf
, len
),
210 /* clear tx_buffer data */
212 dma_unmap_len_set(tx_buf
, len
, 0);
214 /* unmap remaining buffers */
215 while (tx_desc
!= eop_desc
) {
222 tx_buf
= tx_ring
->tx_bi
;
223 tx_desc
= I40E_TX_DESC(tx_ring
, 0);
226 /* unmap any remaining paged data */
227 if (dma_unmap_len(tx_buf
, len
)) {
228 dma_unmap_page(tx_ring
->dev
,
229 dma_unmap_addr(tx_buf
, dma
),
230 dma_unmap_len(tx_buf
, len
),
232 dma_unmap_len_set(tx_buf
, len
, 0);
236 /* move us one more past the eop_desc for start of next pkt */
242 tx_buf
= tx_ring
->tx_bi
;
243 tx_desc
= I40E_TX_DESC(tx_ring
, 0);
248 /* update budget accounting */
250 } while (likely(budget
));
253 tx_ring
->next_to_clean
= i
;
254 u64_stats_update_begin(&tx_ring
->syncp
);
255 tx_ring
->stats
.bytes
+= total_bytes
;
256 tx_ring
->stats
.packets
+= total_packets
;
257 u64_stats_update_end(&tx_ring
->syncp
);
258 tx_ring
->q_vector
->tx
.total_bytes
+= total_bytes
;
259 tx_ring
->q_vector
->tx
.total_packets
+= total_packets
;
261 if (tx_ring
->flags
& I40E_TXR_FLAGS_WB_ON_ITR
) {
263 /* check to see if there are < 4 descriptors
264 * waiting to be written back, then kick the hardware to force
265 * them to be written back in case we stay in NAPI.
266 * In this mode on X722 we do not enable Interrupt.
268 j
= i40evf_get_tx_pending(tx_ring
, false);
271 ((j
/ (WB_STRIDE
+ 1)) == 0) && (j
> 0) &&
272 !test_bit(__I40E_DOWN
, &vsi
->state
) &&
273 (I40E_DESC_UNUSED(tx_ring
) != tx_ring
->count
))
274 tx_ring
->arm_wb
= true;
277 netdev_tx_completed_queue(netdev_get_tx_queue(tx_ring
->netdev
,
278 tx_ring
->queue_index
),
279 total_packets
, total_bytes
);
281 #define TX_WAKE_THRESHOLD (DESC_NEEDED * 2)
282 if (unlikely(total_packets
&& netif_carrier_ok(tx_ring
->netdev
) &&
283 (I40E_DESC_UNUSED(tx_ring
) >= TX_WAKE_THRESHOLD
))) {
284 /* Make sure that anybody stopping the queue after this
285 * sees the new next_to_clean.
288 if (__netif_subqueue_stopped(tx_ring
->netdev
,
289 tx_ring
->queue_index
) &&
290 !test_bit(__I40E_DOWN
, &vsi
->state
)) {
291 netif_wake_subqueue(tx_ring
->netdev
,
292 tx_ring
->queue_index
);
293 ++tx_ring
->tx_stats
.restart_queue
;
301 * i40evf_enable_wb_on_itr - Arm hardware to do a wb, interrupts are not enabled
302 * @vsi: the VSI we care about
303 * @q_vector: the vector on which to enable writeback
306 static void i40e_enable_wb_on_itr(struct i40e_vsi
*vsi
,
307 struct i40e_q_vector
*q_vector
)
309 u16 flags
= q_vector
->tx
.ring
[0].flags
;
312 if (!(flags
& I40E_TXR_FLAGS_WB_ON_ITR
))
315 if (q_vector
->arm_wb_state
)
318 val
= I40E_VFINT_DYN_CTLN1_WB_ON_ITR_MASK
|
319 I40E_VFINT_DYN_CTLN1_ITR_INDX_MASK
; /* set noitr */
322 I40E_VFINT_DYN_CTLN1(q_vector
->v_idx
+
323 vsi
->base_vector
- 1), val
);
324 q_vector
->arm_wb_state
= true;
328 * i40evf_force_wb - Issue SW Interrupt so HW does a wb
329 * @vsi: the VSI we care about
330 * @q_vector: the vector on which to force writeback
333 void i40evf_force_wb(struct i40e_vsi
*vsi
, struct i40e_q_vector
*q_vector
)
335 u32 val
= I40E_VFINT_DYN_CTLN1_INTENA_MASK
|
336 I40E_VFINT_DYN_CTLN1_ITR_INDX_MASK
| /* set noitr */
337 I40E_VFINT_DYN_CTLN1_SWINT_TRIG_MASK
|
338 I40E_VFINT_DYN_CTLN1_SW_ITR_INDX_ENA_MASK
339 /* allow 00 to be written to the index */;
342 I40E_VFINT_DYN_CTLN1(q_vector
->v_idx
+ vsi
->base_vector
- 1),
347 * i40e_set_new_dynamic_itr - Find new ITR level
348 * @rc: structure containing ring performance data
350 * Returns true if ITR changed, false if not
352 * Stores a new ITR value based on packets and byte counts during
353 * the last interrupt. The advantage of per interrupt computation
354 * is faster updates and more accurate ITR for the current traffic
355 * pattern. Constants in this function were computed based on
356 * theoretical maximum wire speed and thresholds were set based on
357 * testing data as well as attempting to minimize response time
358 * while increasing bulk throughput.
360 static bool i40e_set_new_dynamic_itr(struct i40e_ring_container
*rc
)
362 enum i40e_latency_range new_latency_range
= rc
->latency_range
;
363 struct i40e_q_vector
*qv
= rc
->ring
->q_vector
;
364 u32 new_itr
= rc
->itr
;
368 if (rc
->total_packets
== 0 || !rc
->itr
)
371 /* simple throttlerate management
372 * 0-10MB/s lowest (50000 ints/s)
373 * 10-20MB/s low (20000 ints/s)
374 * 20-1249MB/s bulk (18000 ints/s)
375 * > 40000 Rx packets per second (8000 ints/s)
377 * The math works out because the divisor is in 10^(-6) which
378 * turns the bytes/us input value into MB/s values, but
379 * make sure to use usecs, as the register values written
380 * are in 2 usec increments in the ITR registers, and make sure
381 * to use the smoothed values that the countdown timer gives us.
383 usecs
= (rc
->itr
<< 1) * ITR_COUNTDOWN_START
;
384 bytes_per_int
= rc
->total_bytes
/ usecs
;
386 switch (new_latency_range
) {
387 case I40E_LOWEST_LATENCY
:
388 if (bytes_per_int
> 10)
389 new_latency_range
= I40E_LOW_LATENCY
;
391 case I40E_LOW_LATENCY
:
392 if (bytes_per_int
> 20)
393 new_latency_range
= I40E_BULK_LATENCY
;
394 else if (bytes_per_int
<= 10)
395 new_latency_range
= I40E_LOWEST_LATENCY
;
397 case I40E_BULK_LATENCY
:
398 case I40E_ULTRA_LATENCY
:
400 if (bytes_per_int
<= 20)
401 new_latency_range
= I40E_LOW_LATENCY
;
405 /* this is to adjust RX more aggressively when streaming small
406 * packets. The value of 40000 was picked as it is just beyond
407 * what the hardware can receive per second if in low latency
410 #define RX_ULTRA_PACKET_RATE 40000
412 if ((((rc
->total_packets
* 1000000) / usecs
) > RX_ULTRA_PACKET_RATE
) &&
414 new_latency_range
= I40E_ULTRA_LATENCY
;
416 rc
->latency_range
= new_latency_range
;
418 switch (new_latency_range
) {
419 case I40E_LOWEST_LATENCY
:
420 new_itr
= I40E_ITR_50K
;
422 case I40E_LOW_LATENCY
:
423 new_itr
= I40E_ITR_20K
;
425 case I40E_BULK_LATENCY
:
426 new_itr
= I40E_ITR_18K
;
428 case I40E_ULTRA_LATENCY
:
429 new_itr
= I40E_ITR_8K
;
436 rc
->total_packets
= 0;
438 if (new_itr
!= rc
->itr
) {
447 * i40evf_setup_tx_descriptors - Allocate the Tx descriptors
448 * @tx_ring: the tx ring to set up
450 * Return 0 on success, negative on error
452 int i40evf_setup_tx_descriptors(struct i40e_ring
*tx_ring
)
454 struct device
*dev
= tx_ring
->dev
;
460 /* warn if we are about to overwrite the pointer */
461 WARN_ON(tx_ring
->tx_bi
);
462 bi_size
= sizeof(struct i40e_tx_buffer
) * tx_ring
->count
;
463 tx_ring
->tx_bi
= kzalloc(bi_size
, GFP_KERNEL
);
467 /* round up to nearest 4K */
468 tx_ring
->size
= tx_ring
->count
* sizeof(struct i40e_tx_desc
);
469 /* add u32 for head writeback, align after this takes care of
470 * guaranteeing this is at least one cache line in size
472 tx_ring
->size
+= sizeof(u32
);
473 tx_ring
->size
= ALIGN(tx_ring
->size
, 4096);
474 tx_ring
->desc
= dma_alloc_coherent(dev
, tx_ring
->size
,
475 &tx_ring
->dma
, GFP_KERNEL
);
476 if (!tx_ring
->desc
) {
477 dev_info(dev
, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
482 tx_ring
->next_to_use
= 0;
483 tx_ring
->next_to_clean
= 0;
487 kfree(tx_ring
->tx_bi
);
488 tx_ring
->tx_bi
= NULL
;
493 * i40evf_clean_rx_ring - Free Rx buffers
494 * @rx_ring: ring to be cleaned
496 void i40evf_clean_rx_ring(struct i40e_ring
*rx_ring
)
498 struct device
*dev
= rx_ring
->dev
;
499 struct i40e_rx_buffer
*rx_bi
;
500 unsigned long bi_size
;
503 /* ring already cleared, nothing to do */
507 if (ring_is_ps_enabled(rx_ring
)) {
508 int bufsz
= ALIGN(rx_ring
->rx_hdr_len
, 256) * rx_ring
->count
;
510 rx_bi
= &rx_ring
->rx_bi
[0];
511 if (rx_bi
->hdr_buf
) {
512 dma_free_coherent(dev
,
516 for (i
= 0; i
< rx_ring
->count
; i
++) {
517 rx_bi
= &rx_ring
->rx_bi
[i
];
519 rx_bi
->hdr_buf
= NULL
;
523 /* Free all the Rx ring sk_buffs */
524 for (i
= 0; i
< rx_ring
->count
; i
++) {
525 rx_bi
= &rx_ring
->rx_bi
[i
];
527 dma_unmap_single(dev
,
534 dev_kfree_skb(rx_bi
->skb
);
538 if (rx_bi
->page_dma
) {
545 __free_page(rx_bi
->page
);
547 rx_bi
->page_offset
= 0;
551 bi_size
= sizeof(struct i40e_rx_buffer
) * rx_ring
->count
;
552 memset(rx_ring
->rx_bi
, 0, bi_size
);
554 /* Zero out the descriptor ring */
555 memset(rx_ring
->desc
, 0, rx_ring
->size
);
557 rx_ring
->next_to_clean
= 0;
558 rx_ring
->next_to_use
= 0;
562 * i40evf_free_rx_resources - Free Rx resources
563 * @rx_ring: ring to clean the resources from
565 * Free all receive software resources
567 void i40evf_free_rx_resources(struct i40e_ring
*rx_ring
)
569 i40evf_clean_rx_ring(rx_ring
);
570 kfree(rx_ring
->rx_bi
);
571 rx_ring
->rx_bi
= NULL
;
574 dma_free_coherent(rx_ring
->dev
, rx_ring
->size
,
575 rx_ring
->desc
, rx_ring
->dma
);
576 rx_ring
->desc
= NULL
;
581 * i40evf_alloc_rx_headers - allocate rx header buffers
582 * @rx_ring: ring to alloc buffers
584 * Allocate rx header buffers for the entire ring. As these are static,
585 * this is only called when setting up a new ring.
587 void i40evf_alloc_rx_headers(struct i40e_ring
*rx_ring
)
589 struct device
*dev
= rx_ring
->dev
;
590 struct i40e_rx_buffer
*rx_bi
;
596 if (rx_ring
->rx_bi
[0].hdr_buf
)
598 /* Make sure the buffers don't cross cache line boundaries. */
599 buf_size
= ALIGN(rx_ring
->rx_hdr_len
, 256);
600 buffer
= dma_alloc_coherent(dev
, buf_size
* rx_ring
->count
,
604 for (i
= 0; i
< rx_ring
->count
; i
++) {
605 rx_bi
= &rx_ring
->rx_bi
[i
];
606 rx_bi
->dma
= dma
+ (i
* buf_size
);
607 rx_bi
->hdr_buf
= buffer
+ (i
* buf_size
);
612 * i40evf_setup_rx_descriptors - Allocate Rx descriptors
613 * @rx_ring: Rx descriptor ring (for a specific queue) to setup
615 * Returns 0 on success, negative on failure
617 int i40evf_setup_rx_descriptors(struct i40e_ring
*rx_ring
)
619 struct device
*dev
= rx_ring
->dev
;
622 /* warn if we are about to overwrite the pointer */
623 WARN_ON(rx_ring
->rx_bi
);
624 bi_size
= sizeof(struct i40e_rx_buffer
) * rx_ring
->count
;
625 rx_ring
->rx_bi
= kzalloc(bi_size
, GFP_KERNEL
);
629 u64_stats_init(&rx_ring
->syncp
);
631 /* Round up to nearest 4K */
632 rx_ring
->size
= ring_is_16byte_desc_enabled(rx_ring
)
633 ? rx_ring
->count
* sizeof(union i40e_16byte_rx_desc
)
634 : rx_ring
->count
* sizeof(union i40e_32byte_rx_desc
);
635 rx_ring
->size
= ALIGN(rx_ring
->size
, 4096);
636 rx_ring
->desc
= dma_alloc_coherent(dev
, rx_ring
->size
,
637 &rx_ring
->dma
, GFP_KERNEL
);
639 if (!rx_ring
->desc
) {
640 dev_info(dev
, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
645 rx_ring
->next_to_clean
= 0;
646 rx_ring
->next_to_use
= 0;
650 kfree(rx_ring
->rx_bi
);
651 rx_ring
->rx_bi
= NULL
;
656 * i40e_release_rx_desc - Store the new tail and head values
657 * @rx_ring: ring to bump
658 * @val: new head index
660 static inline void i40e_release_rx_desc(struct i40e_ring
*rx_ring
, u32 val
)
662 rx_ring
->next_to_use
= val
;
663 /* Force memory writes to complete before letting h/w
664 * know there are new descriptors to fetch. (Only
665 * applicable for weak-ordered memory model archs,
669 writel(val
, rx_ring
->tail
);
673 * i40evf_alloc_rx_buffers_ps - Replace used receive buffers; packet split
674 * @rx_ring: ring to place buffers on
675 * @cleaned_count: number of buffers to replace
677 * Returns true if any errors on allocation
679 bool i40evf_alloc_rx_buffers_ps(struct i40e_ring
*rx_ring
, u16 cleaned_count
)
681 u16 i
= rx_ring
->next_to_use
;
682 union i40e_rx_desc
*rx_desc
;
683 struct i40e_rx_buffer
*bi
;
684 const int current_node
= numa_node_id();
686 /* do nothing if no valid netdev defined */
687 if (!rx_ring
->netdev
|| !cleaned_count
)
690 while (cleaned_count
--) {
691 rx_desc
= I40E_RX_DESC(rx_ring
, i
);
692 bi
= &rx_ring
->rx_bi
[i
];
694 if (bi
->skb
) /* desc is in use */
697 /* If we've been moved to a different NUMA node, release the
698 * page so we can get a new one on the current node.
700 if (bi
->page
&& page_to_nid(bi
->page
) != current_node
) {
701 dma_unmap_page(rx_ring
->dev
,
705 __free_page(bi
->page
);
708 rx_ring
->rx_stats
.realloc_count
++;
709 } else if (bi
->page
) {
710 rx_ring
->rx_stats
.page_reuse_count
++;
714 bi
->page
= alloc_page(GFP_ATOMIC
);
716 rx_ring
->rx_stats
.alloc_page_failed
++;
719 bi
->page_dma
= dma_map_page(rx_ring
->dev
,
724 if (dma_mapping_error(rx_ring
->dev
, bi
->page_dma
)) {
725 rx_ring
->rx_stats
.alloc_page_failed
++;
726 __free_page(bi
->page
);
735 /* Refresh the desc even if buffer_addrs didn't change
736 * because each write-back erases this info.
738 rx_desc
->read
.pkt_addr
=
739 cpu_to_le64(bi
->page_dma
+ bi
->page_offset
);
740 rx_desc
->read
.hdr_addr
= cpu_to_le64(bi
->dma
);
742 if (i
== rx_ring
->count
)
746 if (rx_ring
->next_to_use
!= i
)
747 i40e_release_rx_desc(rx_ring
, i
);
752 if (rx_ring
->next_to_use
!= i
)
753 i40e_release_rx_desc(rx_ring
, i
);
755 /* make sure to come back via polling to try again after
762 * i40evf_alloc_rx_buffers_1buf - Replace used receive buffers; single buffer
763 * @rx_ring: ring to place buffers on
764 * @cleaned_count: number of buffers to replace
766 * Returns true if any errors on allocation
768 bool i40evf_alloc_rx_buffers_1buf(struct i40e_ring
*rx_ring
, u16 cleaned_count
)
770 u16 i
= rx_ring
->next_to_use
;
771 union i40e_rx_desc
*rx_desc
;
772 struct i40e_rx_buffer
*bi
;
775 /* do nothing if no valid netdev defined */
776 if (!rx_ring
->netdev
|| !cleaned_count
)
779 while (cleaned_count
--) {
780 rx_desc
= I40E_RX_DESC(rx_ring
, i
);
781 bi
= &rx_ring
->rx_bi
[i
];
785 skb
= __netdev_alloc_skb_ip_align(rx_ring
->netdev
,
790 rx_ring
->rx_stats
.alloc_buff_failed
++;
793 /* initialize queue mapping */
794 skb_record_rx_queue(skb
, rx_ring
->queue_index
);
799 bi
->dma
= dma_map_single(rx_ring
->dev
,
803 if (dma_mapping_error(rx_ring
->dev
, bi
->dma
)) {
804 rx_ring
->rx_stats
.alloc_buff_failed
++;
806 dev_kfree_skb(bi
->skb
);
812 rx_desc
->read
.pkt_addr
= cpu_to_le64(bi
->dma
);
813 rx_desc
->read
.hdr_addr
= 0;
815 if (i
== rx_ring
->count
)
819 if (rx_ring
->next_to_use
!= i
)
820 i40e_release_rx_desc(rx_ring
, i
);
825 if (rx_ring
->next_to_use
!= i
)
826 i40e_release_rx_desc(rx_ring
, i
);
828 /* make sure to come back via polling to try again after
835 * i40e_receive_skb - Send a completed packet up the stack
836 * @rx_ring: rx ring in play
837 * @skb: packet to send up
838 * @vlan_tag: vlan tag for packet
840 static void i40e_receive_skb(struct i40e_ring
*rx_ring
,
841 struct sk_buff
*skb
, u16 vlan_tag
)
843 struct i40e_q_vector
*q_vector
= rx_ring
->q_vector
;
845 if ((rx_ring
->netdev
->features
& NETIF_F_HW_VLAN_CTAG_RX
) &&
846 (vlan_tag
& VLAN_VID_MASK
))
847 __vlan_hwaccel_put_tag(skb
, htons(ETH_P_8021Q
), vlan_tag
);
849 napi_gro_receive(&q_vector
->napi
, skb
);
853 * i40e_rx_checksum - Indicate in skb if hw indicated a good cksum
854 * @vsi: the VSI we care about
855 * @skb: skb currently being received and modified
856 * @rx_status: status value of last descriptor in packet
857 * @rx_error: error value of last descriptor in packet
858 * @rx_ptype: ptype value of last descriptor in packet
860 static inline void i40e_rx_checksum(struct i40e_vsi
*vsi
,
866 struct i40e_rx_ptype_decoded decoded
= decode_rx_desc_ptype(rx_ptype
);
867 bool ipv4
, ipv6
, ipv4_tunnel
, ipv6_tunnel
;
869 skb
->ip_summed
= CHECKSUM_NONE
;
871 /* Rx csum enabled and ip headers found? */
872 if (!(vsi
->netdev
->features
& NETIF_F_RXCSUM
))
875 /* did the hardware decode the packet and checksum? */
876 if (!(rx_status
& BIT(I40E_RX_DESC_STATUS_L3L4P_SHIFT
)))
879 /* both known and outer_ip must be set for the below code to work */
880 if (!(decoded
.known
&& decoded
.outer_ip
))
883 ipv4
= (decoded
.outer_ip
== I40E_RX_PTYPE_OUTER_IP
) &&
884 (decoded
.outer_ip_ver
== I40E_RX_PTYPE_OUTER_IPV4
);
885 ipv6
= (decoded
.outer_ip
== I40E_RX_PTYPE_OUTER_IP
) &&
886 (decoded
.outer_ip_ver
== I40E_RX_PTYPE_OUTER_IPV6
);
889 (rx_error
& (BIT(I40E_RX_DESC_ERROR_IPE_SHIFT
) |
890 BIT(I40E_RX_DESC_ERROR_EIPE_SHIFT
))))
893 /* likely incorrect csum if alternate IP extension headers found */
895 rx_status
& BIT(I40E_RX_DESC_STATUS_IPV6EXADD_SHIFT
))
896 /* don't increment checksum err here, non-fatal err */
899 /* there was some L4 error, count error and punt packet to the stack */
900 if (rx_error
& BIT(I40E_RX_DESC_ERROR_L4E_SHIFT
))
903 /* handle packets that were not able to be checksummed due
904 * to arrival speed, in this case the stack can compute
907 if (rx_error
& BIT(I40E_RX_DESC_ERROR_PPRS_SHIFT
))
910 /* The hardware supported by this driver does not validate outer
911 * checksums for tunneled VXLAN or GENEVE frames. I don't agree
912 * with it but the specification states that you "MAY validate", it
913 * doesn't make it a hard requirement so if we have validated the
914 * inner checksum report CHECKSUM_UNNECESSARY.
917 ipv4_tunnel
= (rx_ptype
>= I40E_RX_PTYPE_GRENAT4_MAC_PAY3
) &&
918 (rx_ptype
<= I40E_RX_PTYPE_GRENAT4_MACVLAN_IPV6_ICMP_PAY4
);
919 ipv6_tunnel
= (rx_ptype
>= I40E_RX_PTYPE_GRENAT6_MAC_PAY3
) &&
920 (rx_ptype
<= I40E_RX_PTYPE_GRENAT6_MACVLAN_IPV6_ICMP_PAY4
);
922 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
923 skb
->csum_level
= ipv4_tunnel
|| ipv6_tunnel
;
928 vsi
->back
->hw_csum_rx_error
++;
932 * i40e_ptype_to_htype - get a hash type
933 * @ptype: the ptype value from the descriptor
935 * Returns a hash type to be used by skb_set_hash
937 static inline enum pkt_hash_types
i40e_ptype_to_htype(u8 ptype
)
939 struct i40e_rx_ptype_decoded decoded
= decode_rx_desc_ptype(ptype
);
942 return PKT_HASH_TYPE_NONE
;
944 if (decoded
.outer_ip
== I40E_RX_PTYPE_OUTER_IP
&&
945 decoded
.payload_layer
== I40E_RX_PTYPE_PAYLOAD_LAYER_PAY4
)
946 return PKT_HASH_TYPE_L4
;
947 else if (decoded
.outer_ip
== I40E_RX_PTYPE_OUTER_IP
&&
948 decoded
.payload_layer
== I40E_RX_PTYPE_PAYLOAD_LAYER_PAY3
)
949 return PKT_HASH_TYPE_L3
;
951 return PKT_HASH_TYPE_L2
;
955 * i40e_rx_hash - set the hash value in the skb
956 * @ring: descriptor ring
957 * @rx_desc: specific descriptor
959 static inline void i40e_rx_hash(struct i40e_ring
*ring
,
960 union i40e_rx_desc
*rx_desc
,
965 const __le64 rss_mask
=
966 cpu_to_le64((u64
)I40E_RX_DESC_FLTSTAT_RSS_HASH
<<
967 I40E_RX_DESC_STATUS_FLTSTAT_SHIFT
);
969 if (ring
->netdev
->features
& NETIF_F_RXHASH
)
972 if ((rx_desc
->wb
.qword1
.status_error_len
& rss_mask
) == rss_mask
) {
973 hash
= le32_to_cpu(rx_desc
->wb
.qword0
.hi_dword
.rss
);
974 skb_set_hash(skb
, hash
, i40e_ptype_to_htype(rx_ptype
));
979 * i40e_clean_rx_irq_ps - Reclaim resources after receive; packet split
980 * @rx_ring: rx ring to clean
981 * @budget: how many cleans we're allowed
983 * Returns true if there's any budget left (e.g. the clean is finished)
985 static int i40e_clean_rx_irq_ps(struct i40e_ring
*rx_ring
, const int budget
)
987 unsigned int total_rx_bytes
= 0, total_rx_packets
= 0;
988 u16 rx_packet_len
, rx_header_len
, rx_sph
, rx_hbo
;
989 u16 cleaned_count
= I40E_DESC_UNUSED(rx_ring
);
990 struct i40e_vsi
*vsi
= rx_ring
->vsi
;
991 u16 i
= rx_ring
->next_to_clean
;
992 union i40e_rx_desc
*rx_desc
;
993 u32 rx_error
, rx_status
;
994 bool failure
= false;
1000 struct i40e_rx_buffer
*rx_bi
;
1001 struct sk_buff
*skb
;
1003 /* return some buffers to hardware, one at a time is too slow */
1004 if (cleaned_count
>= I40E_RX_BUFFER_WRITE
) {
1005 failure
= failure
||
1006 i40evf_alloc_rx_buffers_ps(rx_ring
,
1011 i
= rx_ring
->next_to_clean
;
1012 rx_desc
= I40E_RX_DESC(rx_ring
, i
);
1013 qword
= le64_to_cpu(rx_desc
->wb
.qword1
.status_error_len
);
1014 rx_status
= (qword
& I40E_RXD_QW1_STATUS_MASK
) >>
1015 I40E_RXD_QW1_STATUS_SHIFT
;
1017 if (!(rx_status
& BIT(I40E_RX_DESC_STATUS_DD_SHIFT
)))
1020 /* This memory barrier is needed to keep us from reading
1021 * any other fields out of the rx_desc until we know the
1025 /* sync header buffer for reading */
1026 dma_sync_single_range_for_cpu(rx_ring
->dev
,
1027 rx_ring
->rx_bi
[0].dma
,
1028 i
* rx_ring
->rx_hdr_len
,
1029 rx_ring
->rx_hdr_len
,
1031 rx_bi
= &rx_ring
->rx_bi
[i
];
1034 skb
= __netdev_alloc_skb_ip_align(rx_ring
->netdev
,
1035 rx_ring
->rx_hdr_len
,
1039 rx_ring
->rx_stats
.alloc_buff_failed
++;
1044 /* initialize queue mapping */
1045 skb_record_rx_queue(skb
, rx_ring
->queue_index
);
1046 /* we are reusing so sync this buffer for CPU use */
1047 dma_sync_single_range_for_cpu(rx_ring
->dev
,
1048 rx_ring
->rx_bi
[0].dma
,
1049 i
* rx_ring
->rx_hdr_len
,
1050 rx_ring
->rx_hdr_len
,
1053 rx_packet_len
= (qword
& I40E_RXD_QW1_LENGTH_PBUF_MASK
) >>
1054 I40E_RXD_QW1_LENGTH_PBUF_SHIFT
;
1055 rx_header_len
= (qword
& I40E_RXD_QW1_LENGTH_HBUF_MASK
) >>
1056 I40E_RXD_QW1_LENGTH_HBUF_SHIFT
;
1057 rx_sph
= (qword
& I40E_RXD_QW1_LENGTH_SPH_MASK
) >>
1058 I40E_RXD_QW1_LENGTH_SPH_SHIFT
;
1060 rx_error
= (qword
& I40E_RXD_QW1_ERROR_MASK
) >>
1061 I40E_RXD_QW1_ERROR_SHIFT
;
1062 rx_hbo
= rx_error
& BIT(I40E_RX_DESC_ERROR_HBO_SHIFT
);
1063 rx_error
&= ~BIT(I40E_RX_DESC_ERROR_HBO_SHIFT
);
1065 rx_ptype
= (qword
& I40E_RXD_QW1_PTYPE_MASK
) >>
1066 I40E_RXD_QW1_PTYPE_SHIFT
;
1067 /* sync half-page for reading */
1068 dma_sync_single_range_for_cpu(rx_ring
->dev
,
1073 prefetch(page_address(rx_bi
->page
) + rx_bi
->page_offset
);
1077 if (rx_hbo
|| rx_sph
) {
1081 len
= I40E_RX_HDR_SIZE
;
1083 len
= rx_header_len
;
1084 memcpy(__skb_put(skb
, len
), rx_bi
->hdr_buf
, len
);
1085 } else if (skb
->len
== 0) {
1087 unsigned char *va
= page_address(rx_bi
->page
) +
1090 len
= min(rx_packet_len
, rx_ring
->rx_hdr_len
);
1091 memcpy(__skb_put(skb
, len
), va
, len
);
1093 rx_packet_len
-= len
;
1095 /* Get the rest of the data if this was a header split */
1096 if (rx_packet_len
) {
1097 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
,
1099 rx_bi
->page_offset
+ copysize
,
1100 rx_packet_len
, I40E_RXBUFFER_2048
);
1102 /* If the page count is more than 2, then both halves
1103 * of the page are used and we need to free it. Do it
1104 * here instead of in the alloc code. Otherwise one
1105 * of the half-pages might be released between now and
1106 * then, and we wouldn't know which one to use.
1107 * Don't call get_page and free_page since those are
1108 * both expensive atomic operations that just change
1109 * the refcount in opposite directions. Just give the
1110 * page to the stack; he can have our refcount.
1112 if (page_count(rx_bi
->page
) > 2) {
1113 dma_unmap_page(rx_ring
->dev
,
1118 rx_bi
->page_dma
= 0;
1119 rx_ring
->rx_stats
.realloc_count
++;
1121 get_page(rx_bi
->page
);
1122 /* switch to the other half-page here; the
1123 * allocation code programs the right addr
1124 * into HW. If we haven't used this half-page,
1125 * the address won't be changed, and HW can
1126 * just use it next time through.
1128 rx_bi
->page_offset
^= PAGE_SIZE
/ 2;
1132 I40E_RX_INCREMENT(rx_ring
, i
);
1135 !(rx_status
& BIT(I40E_RX_DESC_STATUS_EOF_SHIFT
)))) {
1136 struct i40e_rx_buffer
*next_buffer
;
1138 next_buffer
= &rx_ring
->rx_bi
[i
];
1139 next_buffer
->skb
= skb
;
1140 rx_ring
->rx_stats
.non_eop_descs
++;
1144 /* ERR_MASK will only have valid bits if EOP set */
1145 if (unlikely(rx_error
& BIT(I40E_RX_DESC_ERROR_RXE_SHIFT
))) {
1146 dev_kfree_skb_any(skb
);
1150 i40e_rx_hash(rx_ring
, rx_desc
, skb
, rx_ptype
);
1152 /* probably a little skewed due to removing CRC */
1153 total_rx_bytes
+= skb
->len
;
1156 skb
->protocol
= eth_type_trans(skb
, rx_ring
->netdev
);
1158 i40e_rx_checksum(vsi
, skb
, rx_status
, rx_error
, rx_ptype
);
1160 vlan_tag
= rx_status
& BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT
)
1161 ? le16_to_cpu(rx_desc
->wb
.qword0
.lo_dword
.l2tag1
)
1165 i40e_rx_is_fcoe(rx_ptype
) &&
1166 !i40e_fcoe_handle_offload(rx_ring
, rx_desc
, skb
))) {
1167 dev_kfree_skb_any(skb
);
1171 i40e_receive_skb(rx_ring
, skb
, vlan_tag
);
1173 rx_desc
->wb
.qword1
.status_error_len
= 0;
1175 } while (likely(total_rx_packets
< budget
));
1177 u64_stats_update_begin(&rx_ring
->syncp
);
1178 rx_ring
->stats
.packets
+= total_rx_packets
;
1179 rx_ring
->stats
.bytes
+= total_rx_bytes
;
1180 u64_stats_update_end(&rx_ring
->syncp
);
1181 rx_ring
->q_vector
->rx
.total_packets
+= total_rx_packets
;
1182 rx_ring
->q_vector
->rx
.total_bytes
+= total_rx_bytes
;
1184 return failure
? budget
: total_rx_packets
;
1188 * i40e_clean_rx_irq_1buf - Reclaim resources after receive; single buffer
1189 * @rx_ring: rx ring to clean
1190 * @budget: how many cleans we're allowed
1192 * Returns number of packets cleaned
1194 static int i40e_clean_rx_irq_1buf(struct i40e_ring
*rx_ring
, int budget
)
1196 unsigned int total_rx_bytes
= 0, total_rx_packets
= 0;
1197 u16 cleaned_count
= I40E_DESC_UNUSED(rx_ring
);
1198 struct i40e_vsi
*vsi
= rx_ring
->vsi
;
1199 union i40e_rx_desc
*rx_desc
;
1200 u32 rx_error
, rx_status
;
1202 bool failure
= false;
1208 struct i40e_rx_buffer
*rx_bi
;
1209 struct sk_buff
*skb
;
1211 /* return some buffers to hardware, one at a time is too slow */
1212 if (cleaned_count
>= I40E_RX_BUFFER_WRITE
) {
1213 failure
= failure
||
1214 i40evf_alloc_rx_buffers_1buf(rx_ring
,
1219 i
= rx_ring
->next_to_clean
;
1220 rx_desc
= I40E_RX_DESC(rx_ring
, i
);
1221 qword
= le64_to_cpu(rx_desc
->wb
.qword1
.status_error_len
);
1222 rx_status
= (qword
& I40E_RXD_QW1_STATUS_MASK
) >>
1223 I40E_RXD_QW1_STATUS_SHIFT
;
1225 if (!(rx_status
& BIT(I40E_RX_DESC_STATUS_DD_SHIFT
)))
1228 /* This memory barrier is needed to keep us from reading
1229 * any other fields out of the rx_desc until we know the
1234 rx_bi
= &rx_ring
->rx_bi
[i
];
1236 prefetch(skb
->data
);
1238 rx_packet_len
= (qword
& I40E_RXD_QW1_LENGTH_PBUF_MASK
) >>
1239 I40E_RXD_QW1_LENGTH_PBUF_SHIFT
;
1241 rx_error
= (qword
& I40E_RXD_QW1_ERROR_MASK
) >>
1242 I40E_RXD_QW1_ERROR_SHIFT
;
1243 rx_error
&= ~BIT(I40E_RX_DESC_ERROR_HBO_SHIFT
);
1245 rx_ptype
= (qword
& I40E_RXD_QW1_PTYPE_MASK
) >>
1246 I40E_RXD_QW1_PTYPE_SHIFT
;
1250 /* Get the header and possibly the whole packet
1251 * If this is an skb from previous receive dma will be 0
1253 skb_put(skb
, rx_packet_len
);
1254 dma_unmap_single(rx_ring
->dev
, rx_bi
->dma
, rx_ring
->rx_buf_len
,
1258 I40E_RX_INCREMENT(rx_ring
, i
);
1261 !(rx_status
& BIT(I40E_RX_DESC_STATUS_EOF_SHIFT
)))) {
1262 rx_ring
->rx_stats
.non_eop_descs
++;
1266 /* ERR_MASK will only have valid bits if EOP set */
1267 if (unlikely(rx_error
& BIT(I40E_RX_DESC_ERROR_RXE_SHIFT
))) {
1268 dev_kfree_skb_any(skb
);
1272 i40e_rx_hash(rx_ring
, rx_desc
, skb
, rx_ptype
);
1273 /* probably a little skewed due to removing CRC */
1274 total_rx_bytes
+= skb
->len
;
1277 skb
->protocol
= eth_type_trans(skb
, rx_ring
->netdev
);
1279 i40e_rx_checksum(vsi
, skb
, rx_status
, rx_error
, rx_ptype
);
1281 vlan_tag
= rx_status
& BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT
)
1282 ? le16_to_cpu(rx_desc
->wb
.qword0
.lo_dword
.l2tag1
)
1284 i40e_receive_skb(rx_ring
, skb
, vlan_tag
);
1286 rx_desc
->wb
.qword1
.status_error_len
= 0;
1287 } while (likely(total_rx_packets
< budget
));
1289 u64_stats_update_begin(&rx_ring
->syncp
);
1290 rx_ring
->stats
.packets
+= total_rx_packets
;
1291 rx_ring
->stats
.bytes
+= total_rx_bytes
;
1292 u64_stats_update_end(&rx_ring
->syncp
);
1293 rx_ring
->q_vector
->rx
.total_packets
+= total_rx_packets
;
1294 rx_ring
->q_vector
->rx
.total_bytes
+= total_rx_bytes
;
1296 return failure
? budget
: total_rx_packets
;
1299 static u32
i40e_buildreg_itr(const int type
, const u16 itr
)
1303 val
= I40E_VFINT_DYN_CTLN1_INTENA_MASK
|
1304 /* Don't clear PBA because that can cause lost interrupts that
1305 * came in while we were cleaning/polling
1307 (type
<< I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT
) |
1308 (itr
<< I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT
);
1313 /* a small macro to shorten up some long lines */
1314 #define INTREG I40E_VFINT_DYN_CTLN1
1317 * i40e_update_enable_itr - Update itr and re-enable MSIX interrupt
1318 * @vsi: the VSI we care about
1319 * @q_vector: q_vector for which itr is being updated and interrupt enabled
1322 static inline void i40e_update_enable_itr(struct i40e_vsi
*vsi
,
1323 struct i40e_q_vector
*q_vector
)
1325 struct i40e_hw
*hw
= &vsi
->back
->hw
;
1326 bool rx
= false, tx
= false;
1330 vector
= (q_vector
->v_idx
+ vsi
->base_vector
);
1332 /* avoid dynamic calculation if in countdown mode OR if
1333 * all dynamic is disabled
1335 rxval
= txval
= i40e_buildreg_itr(I40E_ITR_NONE
, 0);
1337 if (q_vector
->itr_countdown
> 0 ||
1338 (!ITR_IS_DYNAMIC(vsi
->rx_itr_setting
) &&
1339 !ITR_IS_DYNAMIC(vsi
->tx_itr_setting
))) {
1343 if (ITR_IS_DYNAMIC(vsi
->rx_itr_setting
)) {
1344 rx
= i40e_set_new_dynamic_itr(&q_vector
->rx
);
1345 rxval
= i40e_buildreg_itr(I40E_RX_ITR
, q_vector
->rx
.itr
);
1348 if (ITR_IS_DYNAMIC(vsi
->tx_itr_setting
)) {
1349 tx
= i40e_set_new_dynamic_itr(&q_vector
->tx
);
1350 txval
= i40e_buildreg_itr(I40E_TX_ITR
, q_vector
->tx
.itr
);
1354 /* get the higher of the two ITR adjustments and
1355 * use the same value for both ITR registers
1356 * when in adaptive mode (Rx and/or Tx)
1358 u16 itr
= max(q_vector
->tx
.itr
, q_vector
->rx
.itr
);
1360 q_vector
->tx
.itr
= q_vector
->rx
.itr
= itr
;
1361 txval
= i40e_buildreg_itr(I40E_TX_ITR
, itr
);
1363 rxval
= i40e_buildreg_itr(I40E_RX_ITR
, itr
);
1367 /* only need to enable the interrupt once, but need
1368 * to possibly update both ITR values
1371 /* set the INTENA_MSK_MASK so that this first write
1372 * won't actually enable the interrupt, instead just
1373 * updating the ITR (it's bit 31 PF and VF)
1376 /* don't check _DOWN because interrupt isn't being enabled */
1377 wr32(hw
, INTREG(vector
- 1), rxval
);
1381 if (!test_bit(__I40E_DOWN
, &vsi
->state
))
1382 wr32(hw
, INTREG(vector
- 1), txval
);
1384 if (q_vector
->itr_countdown
)
1385 q_vector
->itr_countdown
--;
1387 q_vector
->itr_countdown
= ITR_COUNTDOWN_START
;
1391 * i40evf_napi_poll - NAPI polling Rx/Tx cleanup routine
1392 * @napi: napi struct with our devices info in it
1393 * @budget: amount of work driver is allowed to do this pass, in packets
1395 * This function will clean all queues associated with a q_vector.
1397 * Returns the amount of work done
1399 int i40evf_napi_poll(struct napi_struct
*napi
, int budget
)
1401 struct i40e_q_vector
*q_vector
=
1402 container_of(napi
, struct i40e_q_vector
, napi
);
1403 struct i40e_vsi
*vsi
= q_vector
->vsi
;
1404 struct i40e_ring
*ring
;
1405 bool clean_complete
= true;
1406 bool arm_wb
= false;
1407 int budget_per_ring
;
1410 if (test_bit(__I40E_DOWN
, &vsi
->state
)) {
1411 napi_complete(napi
);
1415 /* Since the actual Tx work is minimal, we can give the Tx a larger
1416 * budget and be more aggressive about cleaning up the Tx descriptors.
1418 i40e_for_each_ring(ring
, q_vector
->tx
) {
1419 if (!i40e_clean_tx_irq(vsi
, ring
, budget
)) {
1420 clean_complete
= false;
1423 arm_wb
|= ring
->arm_wb
;
1424 ring
->arm_wb
= false;
1427 /* Handle case where we are called by netpoll with a budget of 0 */
1431 /* We attempt to distribute budget to each Rx queue fairly, but don't
1432 * allow the budget to go below 1 because that would exit polling early.
1434 budget_per_ring
= max(budget
/q_vector
->num_ringpairs
, 1);
1436 i40e_for_each_ring(ring
, q_vector
->rx
) {
1439 if (ring_is_ps_enabled(ring
))
1440 cleaned
= i40e_clean_rx_irq_ps(ring
, budget_per_ring
);
1442 cleaned
= i40e_clean_rx_irq_1buf(ring
, budget_per_ring
);
1444 work_done
+= cleaned
;
1445 /* if we clean as many as budgeted, we must not be done */
1446 if (cleaned
>= budget_per_ring
)
1447 clean_complete
= false;
1450 /* If work not completed, return budget and polling will return */
1451 if (!clean_complete
) {
1454 q_vector
->tx
.ring
[0].tx_stats
.tx_force_wb
++;
1455 i40e_enable_wb_on_itr(vsi
, q_vector
);
1460 if (vsi
->back
->flags
& I40E_TXR_FLAGS_WB_ON_ITR
)
1461 q_vector
->arm_wb_state
= false;
1463 /* Work is done so exit the polling mode and re-enable the interrupt */
1464 napi_complete_done(napi
, work_done
);
1465 i40e_update_enable_itr(vsi
, q_vector
);
1470 * i40evf_tx_prepare_vlan_flags - prepare generic TX VLAN tagging flags for HW
1472 * @tx_ring: ring to send buffer on
1473 * @flags: the tx flags to be set
1475 * Checks the skb and set up correspondingly several generic transmit flags
1476 * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
1478 * Returns error code indicate the frame should be dropped upon error and the
1479 * otherwise returns 0 to indicate the flags has been set properly.
1481 static inline int i40evf_tx_prepare_vlan_flags(struct sk_buff
*skb
,
1482 struct i40e_ring
*tx_ring
,
1485 __be16 protocol
= skb
->protocol
;
1488 if (protocol
== htons(ETH_P_8021Q
) &&
1489 !(tx_ring
->netdev
->features
& NETIF_F_HW_VLAN_CTAG_TX
)) {
1490 /* When HW VLAN acceleration is turned off by the user the
1491 * stack sets the protocol to 8021q so that the driver
1492 * can take any steps required to support the SW only
1493 * VLAN handling. In our case the driver doesn't need
1494 * to take any further steps so just set the protocol
1495 * to the encapsulated ethertype.
1497 skb
->protocol
= vlan_get_protocol(skb
);
1501 /* if we have a HW VLAN tag being added, default to the HW one */
1502 if (skb_vlan_tag_present(skb
)) {
1503 tx_flags
|= skb_vlan_tag_get(skb
) << I40E_TX_FLAGS_VLAN_SHIFT
;
1504 tx_flags
|= I40E_TX_FLAGS_HW_VLAN
;
1505 /* else if it is a SW VLAN, check the next protocol and store the tag */
1506 } else if (protocol
== htons(ETH_P_8021Q
)) {
1507 struct vlan_hdr
*vhdr
, _vhdr
;
1509 vhdr
= skb_header_pointer(skb
, ETH_HLEN
, sizeof(_vhdr
), &_vhdr
);
1513 protocol
= vhdr
->h_vlan_encapsulated_proto
;
1514 tx_flags
|= ntohs(vhdr
->h_vlan_TCI
) << I40E_TX_FLAGS_VLAN_SHIFT
;
1515 tx_flags
|= I40E_TX_FLAGS_SW_VLAN
;
1524 * i40e_tso - set up the tso context descriptor
1525 * @skb: ptr to the skb we're sending
1526 * @hdr_len: ptr to the size of the packet header
1527 * @cd_type_cmd_tso_mss: Quad Word 1
1529 * Returns 0 if no TSO can happen, 1 if tso is going, or error
1531 static int i40e_tso(struct sk_buff
*skb
, u8
*hdr_len
, u64
*cd_type_cmd_tso_mss
)
1533 u64 cd_cmd
, cd_tso_len
, cd_mss
;
1544 u32 paylen
, l4_offset
;
1547 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1550 if (!skb_is_gso(skb
))
1553 err
= skb_cow_head(skb
, 0);
1557 ip
.hdr
= skb_network_header(skb
);
1558 l4
.hdr
= skb_transport_header(skb
);
1560 /* initialize outer IP header fields */
1561 if (ip
.v4
->version
== 4) {
1565 ip
.v6
->payload_len
= 0;
1568 if (skb_shinfo(skb
)->gso_type
& (SKB_GSO_GRE
|
1571 SKB_GSO_UDP_TUNNEL
|
1572 SKB_GSO_UDP_TUNNEL_CSUM
)) {
1573 if (skb_shinfo(skb
)->gso_type
& SKB_GSO_UDP_TUNNEL_CSUM
) {
1574 /* determine offset of outer transport header */
1575 l4_offset
= l4
.hdr
- skb
->data
;
1577 /* remove payload length from outer checksum */
1578 paylen
= skb
->len
- l4_offset
;
1579 csum_replace_by_diff(&l4
.udp
->check
, htonl(paylen
));
1582 /* reset pointers to inner headers */
1583 ip
.hdr
= skb_inner_network_header(skb
);
1584 l4
.hdr
= skb_inner_transport_header(skb
);
1586 /* initialize inner IP header fields */
1587 if (ip
.v4
->version
== 4) {
1591 ip
.v6
->payload_len
= 0;
1595 /* determine offset of inner transport header */
1596 l4_offset
= l4
.hdr
- skb
->data
;
1598 /* remove payload length from inner checksum */
1599 paylen
= skb
->len
- l4_offset
;
1600 csum_replace_by_diff(&l4
.tcp
->check
, htonl(paylen
));
1602 /* compute length of segmentation header */
1603 *hdr_len
= (l4
.tcp
->doff
* 4) + l4_offset
;
1605 /* find the field values */
1606 cd_cmd
= I40E_TX_CTX_DESC_TSO
;
1607 cd_tso_len
= skb
->len
- *hdr_len
;
1608 cd_mss
= skb_shinfo(skb
)->gso_size
;
1609 *cd_type_cmd_tso_mss
|= (cd_cmd
<< I40E_TXD_CTX_QW1_CMD_SHIFT
) |
1610 (cd_tso_len
<< I40E_TXD_CTX_QW1_TSO_LEN_SHIFT
) |
1611 (cd_mss
<< I40E_TXD_CTX_QW1_MSS_SHIFT
);
1616 * i40e_tx_enable_csum - Enable Tx checksum offloads
1618 * @tx_flags: pointer to Tx flags currently set
1619 * @td_cmd: Tx descriptor command bits to set
1620 * @td_offset: Tx descriptor header offsets to set
1621 * @tx_ring: Tx descriptor ring
1622 * @cd_tunneling: ptr to context desc bits
1624 static int i40e_tx_enable_csum(struct sk_buff
*skb
, u32
*tx_flags
,
1625 u32
*td_cmd
, u32
*td_offset
,
1626 struct i40e_ring
*tx_ring
,
1639 unsigned char *exthdr
;
1640 u32 offset
, cmd
= 0;
1644 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1647 ip
.hdr
= skb_network_header(skb
);
1648 l4
.hdr
= skb_transport_header(skb
);
1650 /* compute outer L2 header size */
1651 offset
= ((ip
.hdr
- skb
->data
) / 2) << I40E_TX_DESC_LENGTH_MACLEN_SHIFT
;
1653 if (skb
->encapsulation
) {
1655 /* define outer network header type */
1656 if (*tx_flags
& I40E_TX_FLAGS_IPV4
) {
1657 tunnel
|= (*tx_flags
& I40E_TX_FLAGS_TSO
) ?
1658 I40E_TX_CTX_EXT_IP_IPV4
:
1659 I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM
;
1661 l4_proto
= ip
.v4
->protocol
;
1662 } else if (*tx_flags
& I40E_TX_FLAGS_IPV6
) {
1663 tunnel
|= I40E_TX_CTX_EXT_IP_IPV6
;
1665 exthdr
= ip
.hdr
+ sizeof(*ip
.v6
);
1666 l4_proto
= ip
.v6
->nexthdr
;
1667 if (l4
.hdr
!= exthdr
)
1668 ipv6_skip_exthdr(skb
, exthdr
- skb
->data
,
1669 &l4_proto
, &frag_off
);
1672 /* define outer transport */
1675 tunnel
|= I40E_TXD_CTX_UDP_TUNNELING
;
1676 *tx_flags
|= I40E_TX_FLAGS_VXLAN_TUNNEL
;
1679 tunnel
|= I40E_TXD_CTX_GRE_TUNNELING
;
1680 *tx_flags
|= I40E_TX_FLAGS_VXLAN_TUNNEL
;
1684 *tx_flags
|= I40E_TX_FLAGS_VXLAN_TUNNEL
;
1685 l4
.hdr
= skb_inner_network_header(skb
);
1688 if (*tx_flags
& I40E_TX_FLAGS_TSO
)
1691 skb_checksum_help(skb
);
1695 /* compute outer L3 header size */
1696 tunnel
|= ((l4
.hdr
- ip
.hdr
) / 4) <<
1697 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT
;
1699 /* switch IP header pointer from outer to inner header */
1700 ip
.hdr
= skb_inner_network_header(skb
);
1702 /* compute tunnel header size */
1703 tunnel
|= ((ip
.hdr
- l4
.hdr
) / 2) <<
1704 I40E_TXD_CTX_QW0_NATLEN_SHIFT
;
1706 /* indicate if we need to offload outer UDP header */
1707 if ((*tx_flags
& I40E_TX_FLAGS_TSO
) &&
1708 (skb_shinfo(skb
)->gso_type
& SKB_GSO_UDP_TUNNEL_CSUM
))
1709 tunnel
|= I40E_TXD_CTX_QW0_L4T_CS_MASK
;
1711 /* record tunnel offload values */
1712 *cd_tunneling
|= tunnel
;
1714 /* switch L4 header pointer from outer to inner */
1715 l4
.hdr
= skb_inner_transport_header(skb
);
1718 /* reset type as we transition from outer to inner headers */
1719 *tx_flags
&= ~(I40E_TX_FLAGS_IPV4
| I40E_TX_FLAGS_IPV6
);
1720 if (ip
.v4
->version
== 4)
1721 *tx_flags
|= I40E_TX_FLAGS_IPV4
;
1722 if (ip
.v6
->version
== 6)
1723 *tx_flags
|= I40E_TX_FLAGS_IPV6
;
1726 /* Enable IP checksum offloads */
1727 if (*tx_flags
& I40E_TX_FLAGS_IPV4
) {
1728 l4_proto
= ip
.v4
->protocol
;
1729 /* the stack computes the IP header already, the only time we
1730 * need the hardware to recompute it is in the case of TSO.
1732 cmd
|= (*tx_flags
& I40E_TX_FLAGS_TSO
) ?
1733 I40E_TX_DESC_CMD_IIPT_IPV4_CSUM
:
1734 I40E_TX_DESC_CMD_IIPT_IPV4
;
1735 } else if (*tx_flags
& I40E_TX_FLAGS_IPV6
) {
1736 cmd
|= I40E_TX_DESC_CMD_IIPT_IPV6
;
1738 exthdr
= ip
.hdr
+ sizeof(*ip
.v6
);
1739 l4_proto
= ip
.v6
->nexthdr
;
1740 if (l4
.hdr
!= exthdr
)
1741 ipv6_skip_exthdr(skb
, exthdr
- skb
->data
,
1742 &l4_proto
, &frag_off
);
1745 /* compute inner L3 header size */
1746 offset
|= ((l4
.hdr
- ip
.hdr
) / 4) << I40E_TX_DESC_LENGTH_IPLEN_SHIFT
;
1748 /* Enable L4 checksum offloads */
1751 /* enable checksum offloads */
1752 cmd
|= I40E_TX_DESC_CMD_L4T_EOFT_TCP
;
1753 offset
|= l4
.tcp
->doff
<< I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT
;
1756 /* enable SCTP checksum offload */
1757 cmd
|= I40E_TX_DESC_CMD_L4T_EOFT_SCTP
;
1758 offset
|= (sizeof(struct sctphdr
) >> 2) <<
1759 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT
;
1762 /* enable UDP checksum offload */
1763 cmd
|= I40E_TX_DESC_CMD_L4T_EOFT_UDP
;
1764 offset
|= (sizeof(struct udphdr
) >> 2) <<
1765 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT
;
1768 if (*tx_flags
& I40E_TX_FLAGS_TSO
)
1770 skb_checksum_help(skb
);
1775 *td_offset
|= offset
;
1781 * i40e_create_tx_ctx Build the Tx context descriptor
1782 * @tx_ring: ring to create the descriptor on
1783 * @cd_type_cmd_tso_mss: Quad Word 1
1784 * @cd_tunneling: Quad Word 0 - bits 0-31
1785 * @cd_l2tag2: Quad Word 0 - bits 32-63
1787 static void i40e_create_tx_ctx(struct i40e_ring
*tx_ring
,
1788 const u64 cd_type_cmd_tso_mss
,
1789 const u32 cd_tunneling
, const u32 cd_l2tag2
)
1791 struct i40e_tx_context_desc
*context_desc
;
1792 int i
= tx_ring
->next_to_use
;
1794 if ((cd_type_cmd_tso_mss
== I40E_TX_DESC_DTYPE_CONTEXT
) &&
1795 !cd_tunneling
&& !cd_l2tag2
)
1798 /* grab the next descriptor */
1799 context_desc
= I40E_TX_CTXTDESC(tx_ring
, i
);
1802 tx_ring
->next_to_use
= (i
< tx_ring
->count
) ? i
: 0;
1804 /* cpu_to_le32 and assign to struct fields */
1805 context_desc
->tunneling_params
= cpu_to_le32(cd_tunneling
);
1806 context_desc
->l2tag2
= cpu_to_le16(cd_l2tag2
);
1807 context_desc
->rsvd
= cpu_to_le16(0);
1808 context_desc
->type_cmd_tso_mss
= cpu_to_le64(cd_type_cmd_tso_mss
);
1812 * __i40evf_chk_linearize - Check if there are more than 8 buffers per packet
1815 * Note: Our HW can't DMA more than 8 buffers to build a packet on the wire
1816 * and so we need to figure out the cases where we need to linearize the skb.
1818 * For TSO we need to count the TSO header and segment payload separately.
1819 * As such we need to check cases where we have 7 fragments or more as we
1820 * can potentially require 9 DMA transactions, 1 for the TSO header, 1 for
1821 * the segment payload in the first descriptor, and another 7 for the
1824 bool __i40evf_chk_linearize(struct sk_buff
*skb
)
1826 const struct skb_frag_struct
*frag
, *stale
;
1829 /* no need to check if number of frags is less than 7 */
1830 nr_frags
= skb_shinfo(skb
)->nr_frags
;
1831 if (nr_frags
< (I40E_MAX_BUFFER_TXD
- 1))
1834 /* We need to walk through the list and validate that each group
1835 * of 6 fragments totals at least gso_size. However we don't need
1836 * to perform such validation on the last 6 since the last 6 cannot
1837 * inherit any data from a descriptor after them.
1839 nr_frags
-= I40E_MAX_BUFFER_TXD
- 2;
1840 frag
= &skb_shinfo(skb
)->frags
[0];
1842 /* Initialize size to the negative value of gso_size minus 1. We
1843 * use this as the worst case scenerio in which the frag ahead
1844 * of us only provides one byte which is why we are limited to 6
1845 * descriptors for a single transmit as the header and previous
1846 * fragment are already consuming 2 descriptors.
1848 sum
= 1 - skb_shinfo(skb
)->gso_size
;
1850 /* Add size of frags 0 through 4 to create our initial sum */
1851 sum
+= skb_frag_size(frag
++);
1852 sum
+= skb_frag_size(frag
++);
1853 sum
+= skb_frag_size(frag
++);
1854 sum
+= skb_frag_size(frag
++);
1855 sum
+= skb_frag_size(frag
++);
1857 /* Walk through fragments adding latest fragment, testing it, and
1858 * then removing stale fragments from the sum.
1860 stale
= &skb_shinfo(skb
)->frags
[0];
1862 sum
+= skb_frag_size(frag
++);
1864 /* if sum is negative we failed to make sufficient progress */
1868 /* use pre-decrement to avoid processing last fragment */
1872 sum
-= skb_frag_size(stale
++);
1879 * __i40evf_maybe_stop_tx - 2nd level check for tx stop conditions
1880 * @tx_ring: the ring to be checked
1881 * @size: the size buffer we want to assure is available
1883 * Returns -EBUSY if a stop is needed, else 0
1885 int __i40evf_maybe_stop_tx(struct i40e_ring
*tx_ring
, int size
)
1887 netif_stop_subqueue(tx_ring
->netdev
, tx_ring
->queue_index
);
1888 /* Memory barrier before checking head and tail */
1891 /* Check again in a case another CPU has just made room available. */
1892 if (likely(I40E_DESC_UNUSED(tx_ring
) < size
))
1895 /* A reprieve! - use start_queue because it doesn't call schedule */
1896 netif_start_subqueue(tx_ring
->netdev
, tx_ring
->queue_index
);
1897 ++tx_ring
->tx_stats
.restart_queue
;
1902 * i40evf_tx_map - Build the Tx descriptor
1903 * @tx_ring: ring to send buffer on
1905 * @first: first buffer info buffer to use
1906 * @tx_flags: collected send information
1907 * @hdr_len: size of the packet header
1908 * @td_cmd: the command field in the descriptor
1909 * @td_offset: offset for checksum or crc
1911 static inline void i40evf_tx_map(struct i40e_ring
*tx_ring
, struct sk_buff
*skb
,
1912 struct i40e_tx_buffer
*first
, u32 tx_flags
,
1913 const u8 hdr_len
, u32 td_cmd
, u32 td_offset
)
1915 unsigned int data_len
= skb
->data_len
;
1916 unsigned int size
= skb_headlen(skb
);
1917 struct skb_frag_struct
*frag
;
1918 struct i40e_tx_buffer
*tx_bi
;
1919 struct i40e_tx_desc
*tx_desc
;
1920 u16 i
= tx_ring
->next_to_use
;
1925 bool tail_bump
= true;
1928 if (tx_flags
& I40E_TX_FLAGS_HW_VLAN
) {
1929 td_cmd
|= I40E_TX_DESC_CMD_IL2TAG1
;
1930 td_tag
= (tx_flags
& I40E_TX_FLAGS_VLAN_MASK
) >>
1931 I40E_TX_FLAGS_VLAN_SHIFT
;
1934 if (tx_flags
& (I40E_TX_FLAGS_TSO
| I40E_TX_FLAGS_FSO
))
1935 gso_segs
= skb_shinfo(skb
)->gso_segs
;
1939 /* multiply data chunks by size of headers */
1940 first
->bytecount
= skb
->len
- hdr_len
+ (gso_segs
* hdr_len
);
1941 first
->gso_segs
= gso_segs
;
1943 first
->tx_flags
= tx_flags
;
1945 dma
= dma_map_single(tx_ring
->dev
, skb
->data
, size
, DMA_TO_DEVICE
);
1947 tx_desc
= I40E_TX_DESC(tx_ring
, i
);
1950 for (frag
= &skb_shinfo(skb
)->frags
[0];; frag
++) {
1951 unsigned int max_data
= I40E_MAX_DATA_PER_TXD_ALIGNED
;
1953 if (dma_mapping_error(tx_ring
->dev
, dma
))
1956 /* record length, and DMA address */
1957 dma_unmap_len_set(tx_bi
, len
, size
);
1958 dma_unmap_addr_set(tx_bi
, dma
, dma
);
1960 /* align size to end of page */
1961 max_data
+= -dma
& (I40E_MAX_READ_REQ_SIZE
- 1);
1962 tx_desc
->buffer_addr
= cpu_to_le64(dma
);
1964 while (unlikely(size
> I40E_MAX_DATA_PER_TXD
)) {
1965 tx_desc
->cmd_type_offset_bsz
=
1966 build_ctob(td_cmd
, td_offset
,
1973 if (i
== tx_ring
->count
) {
1974 tx_desc
= I40E_TX_DESC(tx_ring
, 0);
1981 max_data
= I40E_MAX_DATA_PER_TXD_ALIGNED
;
1982 tx_desc
->buffer_addr
= cpu_to_le64(dma
);
1985 if (likely(!data_len
))
1988 tx_desc
->cmd_type_offset_bsz
= build_ctob(td_cmd
, td_offset
,
1995 if (i
== tx_ring
->count
) {
1996 tx_desc
= I40E_TX_DESC(tx_ring
, 0);
2000 size
= skb_frag_size(frag
);
2003 dma
= skb_frag_dma_map(tx_ring
->dev
, frag
, 0, size
,
2006 tx_bi
= &tx_ring
->tx_bi
[i
];
2009 /* set next_to_watch value indicating a packet is present */
2010 first
->next_to_watch
= tx_desc
;
2013 if (i
== tx_ring
->count
)
2016 tx_ring
->next_to_use
= i
;
2018 netdev_tx_sent_queue(netdev_get_tx_queue(tx_ring
->netdev
,
2019 tx_ring
->queue_index
),
2021 i40e_maybe_stop_tx(tx_ring
, DESC_NEEDED
);
2023 /* Algorithm to optimize tail and RS bit setting:
2024 * if xmit_more is supported
2025 * if xmit_more is true
2026 * do not update tail and do not mark RS bit.
2027 * if xmit_more is false and last xmit_more was false
2028 * if every packet spanned less than 4 desc
2029 * then set RS bit on 4th packet and update tail
2032 * update tail and set RS bit on every packet.
2033 * if xmit_more is false and last_xmit_more was true
2034 * update tail and set RS bit.
2036 * Optimization: wmb to be issued only in case of tail update.
2037 * Also optimize the Descriptor WB path for RS bit with the same
2040 * Note: If there are less than 4 packets
2041 * pending and interrupts were disabled the service task will
2042 * trigger a force WB.
2044 if (skb
->xmit_more
&&
2045 !netif_xmit_stopped(netdev_get_tx_queue(tx_ring
->netdev
,
2046 tx_ring
->queue_index
))) {
2047 tx_ring
->flags
|= I40E_TXR_FLAGS_LAST_XMIT_MORE_SET
;
2049 } else if (!skb
->xmit_more
&&
2050 !netif_xmit_stopped(netdev_get_tx_queue(tx_ring
->netdev
,
2051 tx_ring
->queue_index
)) &&
2052 (!(tx_ring
->flags
& I40E_TXR_FLAGS_LAST_XMIT_MORE_SET
)) &&
2053 (tx_ring
->packet_stride
< WB_STRIDE
) &&
2054 (desc_count
< WB_STRIDE
)) {
2055 tx_ring
->packet_stride
++;
2057 tx_ring
->packet_stride
= 0;
2058 tx_ring
->flags
&= ~I40E_TXR_FLAGS_LAST_XMIT_MORE_SET
;
2062 tx_ring
->packet_stride
= 0;
2064 tx_desc
->cmd_type_offset_bsz
=
2065 build_ctob(td_cmd
, td_offset
, size
, td_tag
) |
2066 cpu_to_le64((u64
)(do_rs
? I40E_TXD_CMD
:
2067 I40E_TX_DESC_CMD_EOP
) <<
2068 I40E_TXD_QW1_CMD_SHIFT
);
2070 /* notify HW of packet */
2072 prefetchw(tx_desc
+ 1);
2075 /* Force memory writes to complete before letting h/w
2076 * know there are new descriptors to fetch. (Only
2077 * applicable for weak-ordered memory model archs,
2081 writel(i
, tx_ring
->tail
);
2087 dev_info(tx_ring
->dev
, "TX DMA map failed\n");
2089 /* clear dma mappings for failed tx_bi map */
2091 tx_bi
= &tx_ring
->tx_bi
[i
];
2092 i40e_unmap_and_free_tx_resource(tx_ring
, tx_bi
);
2100 tx_ring
->next_to_use
= i
;
2104 * i40e_xmit_frame_ring - Sends buffer on Tx ring
2106 * @tx_ring: ring to send buffer on
2108 * Returns NETDEV_TX_OK if sent, else an error code
2110 static netdev_tx_t
i40e_xmit_frame_ring(struct sk_buff
*skb
,
2111 struct i40e_ring
*tx_ring
)
2113 u64 cd_type_cmd_tso_mss
= I40E_TX_DESC_DTYPE_CONTEXT
;
2114 u32 cd_tunneling
= 0, cd_l2tag2
= 0;
2115 struct i40e_tx_buffer
*first
;
2123 /* prefetch the data, we'll need it later */
2124 prefetch(skb
->data
);
2126 count
= i40e_xmit_descriptor_count(skb
);
2127 if (i40e_chk_linearize(skb
, count
)) {
2128 if (__skb_linearize(skb
))
2130 count
= i40e_txd_use_count(skb
->len
);
2131 tx_ring
->tx_stats
.tx_linearize
++;
2134 /* need: 1 descriptor per page * PAGE_SIZE/I40E_MAX_DATA_PER_TXD,
2135 * + 1 desc for skb_head_len/I40E_MAX_DATA_PER_TXD,
2136 * + 4 desc gap to avoid the cache line where head is,
2137 * + 1 desc for context descriptor,
2138 * otherwise try next time
2140 if (i40e_maybe_stop_tx(tx_ring
, count
+ 4 + 1)) {
2141 tx_ring
->tx_stats
.tx_busy
++;
2142 return NETDEV_TX_BUSY
;
2145 /* prepare the xmit flags */
2146 if (i40evf_tx_prepare_vlan_flags(skb
, tx_ring
, &tx_flags
))
2149 /* obtain protocol of skb */
2150 protocol
= vlan_get_protocol(skb
);
2152 /* record the location of the first descriptor for this packet */
2153 first
= &tx_ring
->tx_bi
[tx_ring
->next_to_use
];
2155 /* setup IPv4/IPv6 offloads */
2156 if (protocol
== htons(ETH_P_IP
))
2157 tx_flags
|= I40E_TX_FLAGS_IPV4
;
2158 else if (protocol
== htons(ETH_P_IPV6
))
2159 tx_flags
|= I40E_TX_FLAGS_IPV6
;
2161 tso
= i40e_tso(skb
, &hdr_len
, &cd_type_cmd_tso_mss
);
2166 tx_flags
|= I40E_TX_FLAGS_TSO
;
2168 /* Always offload the checksum, since it's in the data descriptor */
2169 tso
= i40e_tx_enable_csum(skb
, &tx_flags
, &td_cmd
, &td_offset
,
2170 tx_ring
, &cd_tunneling
);
2174 skb_tx_timestamp(skb
);
2176 /* always enable CRC insertion offload */
2177 td_cmd
|= I40E_TX_DESC_CMD_ICRC
;
2179 i40e_create_tx_ctx(tx_ring
, cd_type_cmd_tso_mss
,
2180 cd_tunneling
, cd_l2tag2
);
2182 i40evf_tx_map(tx_ring
, skb
, first
, tx_flags
, hdr_len
,
2185 return NETDEV_TX_OK
;
2188 dev_kfree_skb_any(skb
);
2189 return NETDEV_TX_OK
;
2193 * i40evf_xmit_frame - Selects the correct VSI and Tx queue to send buffer
2195 * @netdev: network interface device structure
2197 * Returns NETDEV_TX_OK if sent, else an error code
2199 netdev_tx_t
i40evf_xmit_frame(struct sk_buff
*skb
, struct net_device
*netdev
)
2201 struct i40evf_adapter
*adapter
= netdev_priv(netdev
);
2202 struct i40e_ring
*tx_ring
= &adapter
->tx_rings
[skb
->queue_mapping
];
2204 /* hardware can't handle really short frames, hardware padding works
2207 if (unlikely(skb
->len
< I40E_MIN_TX_LEN
)) {
2208 if (skb_pad(skb
, I40E_MIN_TX_LEN
- skb
->len
))
2209 return NETDEV_TX_OK
;
2210 skb
->len
= I40E_MIN_TX_LEN
;
2211 skb_set_tail_pointer(skb
, I40E_MIN_TX_LEN
);
2214 return i40e_xmit_frame_ring(skb
, tx_ring
);