1 /*******************************************************************************
3 * Intel Ethernet Controller XL710 Family Linux Virtual Function Driver
4 * Copyright(c) 2013 - 2014 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 if (tx_buffer->tx_flags & I40E_TX_FLAGS_FD_SB)
55 kfree(tx_buffer->raw_buf);
57 dev_kfree_skb_any(tx_buffer->skb);
59 if (dma_unmap_len(tx_buffer, len))
60 dma_unmap_single(ring->dev,
61 dma_unmap_addr(tx_buffer, dma),
62 dma_unmap_len(tx_buffer, len),
64 } else if (dma_unmap_len(tx_buffer, len)) {
65 dma_unmap_page(ring->dev,
66 dma_unmap_addr(tx_buffer, dma),
67 dma_unmap_len(tx_buffer, len),
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 * i40e_get_head - Retrieve head from head writeback
131 * @tx_ring: tx ring to fetch head of
133 * Returns value of Tx ring head based on value stored
134 * in head write-back location
136 static inline u32 i40e_get_head(struct i40e_ring *tx_ring)
138 void *head = (struct i40e_tx_desc *)tx_ring->desc + tx_ring->count;
140 return le32_to_cpu(*(volatile __le32 *)head);
143 #define WB_STRIDE 0x3
146 * i40e_clean_tx_irq - Reclaim resources after transmit completes
147 * @tx_ring: tx ring to clean
148 * @budget: how many cleans we're allowed
150 * Returns true if there's any budget left (e.g. the clean is finished)
152 static bool i40e_clean_tx_irq(struct i40e_ring *tx_ring, int budget)
154 u16 i = tx_ring->next_to_clean;
155 struct i40e_tx_buffer *tx_buf;
156 struct i40e_tx_desc *tx_head;
157 struct i40e_tx_desc *tx_desc;
158 unsigned int total_packets = 0;
159 unsigned int total_bytes = 0;
161 tx_buf = &tx_ring->tx_bi[i];
162 tx_desc = I40E_TX_DESC(tx_ring, i);
165 tx_head = I40E_TX_DESC(tx_ring, i40e_get_head(tx_ring));
168 struct i40e_tx_desc *eop_desc = tx_buf->next_to_watch;
170 /* if next_to_watch is not set then there is no work pending */
174 /* prevent any other reads prior to eop_desc */
175 read_barrier_depends();
177 /* we have caught up to head, no work left to do */
178 if (tx_head == tx_desc)
181 /* clear next_to_watch to prevent false hangs */
182 tx_buf->next_to_watch = NULL;
184 /* update the statistics for this packet */
185 total_bytes += tx_buf->bytecount;
186 total_packets += tx_buf->gso_segs;
189 dev_kfree_skb_any(tx_buf->skb);
191 /* unmap skb header data */
192 dma_unmap_single(tx_ring->dev,
193 dma_unmap_addr(tx_buf, dma),
194 dma_unmap_len(tx_buf, len),
197 /* clear tx_buffer data */
199 dma_unmap_len_set(tx_buf, len, 0);
201 /* unmap remaining buffers */
202 while (tx_desc != eop_desc) {
209 tx_buf = tx_ring->tx_bi;
210 tx_desc = I40E_TX_DESC(tx_ring, 0);
213 /* unmap any remaining paged data */
214 if (dma_unmap_len(tx_buf, len)) {
215 dma_unmap_page(tx_ring->dev,
216 dma_unmap_addr(tx_buf, dma),
217 dma_unmap_len(tx_buf, len),
219 dma_unmap_len_set(tx_buf, len, 0);
223 /* move us one more past the eop_desc for start of next pkt */
229 tx_buf = tx_ring->tx_bi;
230 tx_desc = I40E_TX_DESC(tx_ring, 0);
235 /* update budget accounting */
237 } while (likely(budget));
240 tx_ring->next_to_clean = i;
241 u64_stats_update_begin(&tx_ring->syncp);
242 tx_ring->stats.bytes += total_bytes;
243 tx_ring->stats.packets += total_packets;
244 u64_stats_update_end(&tx_ring->syncp);
245 tx_ring->q_vector->tx.total_bytes += total_bytes;
246 tx_ring->q_vector->tx.total_packets += total_packets;
248 /* check to see if there are any non-cache aligned descriptors
249 * waiting to be written back, and kick the hardware to force
250 * them to be written back in case of napi polling
253 !((i & WB_STRIDE) == WB_STRIDE) &&
254 !test_bit(__I40E_DOWN, &tx_ring->vsi->state) &&
255 (I40E_DESC_UNUSED(tx_ring) != tx_ring->count))
256 tx_ring->arm_wb = true;
258 netdev_tx_completed_queue(netdev_get_tx_queue(tx_ring->netdev,
259 tx_ring->queue_index),
260 total_packets, total_bytes);
262 #define TX_WAKE_THRESHOLD (DESC_NEEDED * 2)
263 if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) &&
264 (I40E_DESC_UNUSED(tx_ring) >= TX_WAKE_THRESHOLD))) {
265 /* Make sure that anybody stopping the queue after this
266 * sees the new next_to_clean.
269 if (__netif_subqueue_stopped(tx_ring->netdev,
270 tx_ring->queue_index) &&
271 !test_bit(__I40E_DOWN, &tx_ring->vsi->state)) {
272 netif_wake_subqueue(tx_ring->netdev,
273 tx_ring->queue_index);
274 ++tx_ring->tx_stats.restart_queue;
282 * i40evf_force_wb -Arm hardware to do a wb on noncache aligned descriptors
283 * @vsi: the VSI we care about
284 * @q_vector: the vector on which to force writeback
287 static void i40evf_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector)
289 u16 flags = q_vector->tx.ring[0].flags;
291 if (flags & I40E_TXR_FLAGS_WB_ON_ITR) {
294 if (q_vector->arm_wb_state)
297 val = I40E_VFINT_DYN_CTLN1_WB_ON_ITR_MASK;
300 I40E_VFINT_DYN_CTLN1(q_vector->v_idx +
301 vsi->base_vector - 1),
303 q_vector->arm_wb_state = true;
305 u32 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
306 I40E_VFINT_DYN_CTLN1_ITR_INDX_MASK | /* set noitr */
307 I40E_VFINT_DYN_CTLN1_SWINT_TRIG_MASK |
308 I40E_VFINT_DYN_CTLN1_SW_ITR_INDX_ENA_MASK;
309 /* allow 00 to be written to the index */
312 I40E_VFINT_DYN_CTLN1(q_vector->v_idx +
313 vsi->base_vector - 1), val);
318 * i40e_set_new_dynamic_itr - Find new ITR level
319 * @rc: structure containing ring performance data
321 * Stores a new ITR value based on packets and byte counts during
322 * the last interrupt. The advantage of per interrupt computation
323 * is faster updates and more accurate ITR for the current traffic
324 * pattern. Constants in this function were computed based on
325 * theoretical maximum wire speed and thresholds were set based on
326 * testing data as well as attempting to minimize response time
327 * while increasing bulk throughput.
329 static void i40e_set_new_dynamic_itr(struct i40e_ring_container *rc)
331 enum i40e_latency_range new_latency_range = rc->latency_range;
332 u32 new_itr = rc->itr;
335 if (rc->total_packets == 0 || !rc->itr)
338 /* simple throttlerate management
339 * 0-10MB/s lowest (100000 ints/s)
340 * 10-20MB/s low (20000 ints/s)
341 * 20-1249MB/s bulk (8000 ints/s)
343 bytes_per_int = rc->total_bytes / rc->itr;
344 switch (new_latency_range) {
345 case I40E_LOWEST_LATENCY:
346 if (bytes_per_int > 10)
347 new_latency_range = I40E_LOW_LATENCY;
349 case I40E_LOW_LATENCY:
350 if (bytes_per_int > 20)
351 new_latency_range = I40E_BULK_LATENCY;
352 else if (bytes_per_int <= 10)
353 new_latency_range = I40E_LOWEST_LATENCY;
355 case I40E_BULK_LATENCY:
356 if (bytes_per_int <= 20)
357 new_latency_range = I40E_LOW_LATENCY;
360 if (bytes_per_int <= 20)
361 new_latency_range = I40E_LOW_LATENCY;
364 rc->latency_range = new_latency_range;
366 switch (new_latency_range) {
367 case I40E_LOWEST_LATENCY:
368 new_itr = I40E_ITR_100K;
370 case I40E_LOW_LATENCY:
371 new_itr = I40E_ITR_20K;
373 case I40E_BULK_LATENCY:
374 new_itr = I40E_ITR_8K;
380 if (new_itr != rc->itr)
384 rc->total_packets = 0;
388 * i40evf_setup_tx_descriptors - Allocate the Tx descriptors
389 * @tx_ring: the tx ring to set up
391 * Return 0 on success, negative on error
393 int i40evf_setup_tx_descriptors(struct i40e_ring *tx_ring)
395 struct device *dev = tx_ring->dev;
401 /* warn if we are about to overwrite the pointer */
402 WARN_ON(tx_ring->tx_bi);
403 bi_size = sizeof(struct i40e_tx_buffer) * tx_ring->count;
404 tx_ring->tx_bi = kzalloc(bi_size, GFP_KERNEL);
408 /* round up to nearest 4K */
409 tx_ring->size = tx_ring->count * sizeof(struct i40e_tx_desc);
410 /* add u32 for head writeback, align after this takes care of
411 * guaranteeing this is at least one cache line in size
413 tx_ring->size += sizeof(u32);
414 tx_ring->size = ALIGN(tx_ring->size, 4096);
415 tx_ring->desc = dma_alloc_coherent(dev, tx_ring->size,
416 &tx_ring->dma, GFP_KERNEL);
417 if (!tx_ring->desc) {
418 dev_info(dev, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
423 tx_ring->next_to_use = 0;
424 tx_ring->next_to_clean = 0;
428 kfree(tx_ring->tx_bi);
429 tx_ring->tx_bi = NULL;
434 * i40evf_clean_rx_ring - Free Rx buffers
435 * @rx_ring: ring to be cleaned
437 void i40evf_clean_rx_ring(struct i40e_ring *rx_ring)
439 struct device *dev = rx_ring->dev;
440 struct i40e_rx_buffer *rx_bi;
441 unsigned long bi_size;
444 /* ring already cleared, nothing to do */
448 if (ring_is_ps_enabled(rx_ring)) {
449 int bufsz = ALIGN(rx_ring->rx_hdr_len, 256) * rx_ring->count;
451 rx_bi = &rx_ring->rx_bi[0];
452 if (rx_bi->hdr_buf) {
453 dma_free_coherent(dev,
457 for (i = 0; i < rx_ring->count; i++) {
458 rx_bi = &rx_ring->rx_bi[i];
460 rx_bi->hdr_buf = NULL;
464 /* Free all the Rx ring sk_buffs */
465 for (i = 0; i < rx_ring->count; i++) {
466 rx_bi = &rx_ring->rx_bi[i];
468 dma_unmap_single(dev,
475 dev_kfree_skb(rx_bi->skb);
479 if (rx_bi->page_dma) {
486 __free_page(rx_bi->page);
488 rx_bi->page_offset = 0;
492 bi_size = sizeof(struct i40e_rx_buffer) * rx_ring->count;
493 memset(rx_ring->rx_bi, 0, bi_size);
495 /* Zero out the descriptor ring */
496 memset(rx_ring->desc, 0, rx_ring->size);
498 rx_ring->next_to_clean = 0;
499 rx_ring->next_to_use = 0;
503 * i40evf_free_rx_resources - Free Rx resources
504 * @rx_ring: ring to clean the resources from
506 * Free all receive software resources
508 void i40evf_free_rx_resources(struct i40e_ring *rx_ring)
510 i40evf_clean_rx_ring(rx_ring);
511 kfree(rx_ring->rx_bi);
512 rx_ring->rx_bi = NULL;
515 dma_free_coherent(rx_ring->dev, rx_ring->size,
516 rx_ring->desc, rx_ring->dma);
517 rx_ring->desc = NULL;
522 * i40evf_alloc_rx_headers - allocate rx header buffers
523 * @rx_ring: ring to alloc buffers
525 * Allocate rx header buffers for the entire ring. As these are static,
526 * this is only called when setting up a new ring.
528 void i40evf_alloc_rx_headers(struct i40e_ring *rx_ring)
530 struct device *dev = rx_ring->dev;
531 struct i40e_rx_buffer *rx_bi;
537 if (rx_ring->rx_bi[0].hdr_buf)
539 /* Make sure the buffers don't cross cache line boundaries. */
540 buf_size = ALIGN(rx_ring->rx_hdr_len, 256);
541 buffer = dma_alloc_coherent(dev, buf_size * rx_ring->count,
545 for (i = 0; i < rx_ring->count; i++) {
546 rx_bi = &rx_ring->rx_bi[i];
547 rx_bi->dma = dma + (i * buf_size);
548 rx_bi->hdr_buf = buffer + (i * buf_size);
553 * i40evf_setup_rx_descriptors - Allocate Rx descriptors
554 * @rx_ring: Rx descriptor ring (for a specific queue) to setup
556 * Returns 0 on success, negative on failure
558 int i40evf_setup_rx_descriptors(struct i40e_ring *rx_ring)
560 struct device *dev = rx_ring->dev;
563 /* warn if we are about to overwrite the pointer */
564 WARN_ON(rx_ring->rx_bi);
565 bi_size = sizeof(struct i40e_rx_buffer) * rx_ring->count;
566 rx_ring->rx_bi = kzalloc(bi_size, GFP_KERNEL);
570 u64_stats_init(&rx_ring->syncp);
572 /* Round up to nearest 4K */
573 rx_ring->size = ring_is_16byte_desc_enabled(rx_ring)
574 ? rx_ring->count * sizeof(union i40e_16byte_rx_desc)
575 : rx_ring->count * sizeof(union i40e_32byte_rx_desc);
576 rx_ring->size = ALIGN(rx_ring->size, 4096);
577 rx_ring->desc = dma_alloc_coherent(dev, rx_ring->size,
578 &rx_ring->dma, GFP_KERNEL);
580 if (!rx_ring->desc) {
581 dev_info(dev, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
586 rx_ring->next_to_clean = 0;
587 rx_ring->next_to_use = 0;
591 kfree(rx_ring->rx_bi);
592 rx_ring->rx_bi = NULL;
597 * i40e_release_rx_desc - Store the new tail and head values
598 * @rx_ring: ring to bump
599 * @val: new head index
601 static inline void i40e_release_rx_desc(struct i40e_ring *rx_ring, u32 val)
603 rx_ring->next_to_use = val;
604 /* Force memory writes to complete before letting h/w
605 * know there are new descriptors to fetch. (Only
606 * applicable for weak-ordered memory model archs,
610 writel(val, rx_ring->tail);
614 * i40evf_alloc_rx_buffers_ps - Replace used receive buffers; packet split
615 * @rx_ring: ring to place buffers on
616 * @cleaned_count: number of buffers to replace
618 void i40evf_alloc_rx_buffers_ps(struct i40e_ring *rx_ring, u16 cleaned_count)
620 u16 i = rx_ring->next_to_use;
621 union i40e_rx_desc *rx_desc;
622 struct i40e_rx_buffer *bi;
624 /* do nothing if no valid netdev defined */
625 if (!rx_ring->netdev || !cleaned_count)
628 while (cleaned_count--) {
629 rx_desc = I40E_RX_DESC(rx_ring, i);
630 bi = &rx_ring->rx_bi[i];
632 if (bi->skb) /* desc is in use */
635 bi->page = alloc_page(GFP_ATOMIC);
637 rx_ring->rx_stats.alloc_page_failed++;
643 /* use a half page if we're re-using */
644 bi->page_offset ^= PAGE_SIZE / 2;
645 bi->page_dma = dma_map_page(rx_ring->dev,
650 if (dma_mapping_error(rx_ring->dev,
652 rx_ring->rx_stats.alloc_page_failed++;
658 dma_sync_single_range_for_device(rx_ring->dev,
663 /* Refresh the desc even if buffer_addrs didn't change
664 * because each write-back erases this info.
666 rx_desc->read.pkt_addr = cpu_to_le64(bi->page_dma);
667 rx_desc->read.hdr_addr = cpu_to_le64(bi->dma);
669 if (i == rx_ring->count)
674 if (rx_ring->next_to_use != i)
675 i40e_release_rx_desc(rx_ring, i);
679 * i40evf_alloc_rx_buffers_1buf - Replace used receive buffers; single buffer
680 * @rx_ring: ring to place buffers on
681 * @cleaned_count: number of buffers to replace
683 void i40evf_alloc_rx_buffers_1buf(struct i40e_ring *rx_ring, u16 cleaned_count)
685 u16 i = rx_ring->next_to_use;
686 union i40e_rx_desc *rx_desc;
687 struct i40e_rx_buffer *bi;
690 /* do nothing if no valid netdev defined */
691 if (!rx_ring->netdev || !cleaned_count)
694 while (cleaned_count--) {
695 rx_desc = I40E_RX_DESC(rx_ring, i);
696 bi = &rx_ring->rx_bi[i];
700 skb = netdev_alloc_skb_ip_align(rx_ring->netdev,
701 rx_ring->rx_buf_len);
703 rx_ring->rx_stats.alloc_buff_failed++;
706 /* initialize queue mapping */
707 skb_record_rx_queue(skb, rx_ring->queue_index);
712 bi->dma = dma_map_single(rx_ring->dev,
716 if (dma_mapping_error(rx_ring->dev, bi->dma)) {
717 rx_ring->rx_stats.alloc_buff_failed++;
723 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
724 rx_desc->read.hdr_addr = 0;
726 if (i == rx_ring->count)
731 if (rx_ring->next_to_use != i)
732 i40e_release_rx_desc(rx_ring, i);
736 * i40e_receive_skb - Send a completed packet up the stack
737 * @rx_ring: rx ring in play
738 * @skb: packet to send up
739 * @vlan_tag: vlan tag for packet
741 static void i40e_receive_skb(struct i40e_ring *rx_ring,
742 struct sk_buff *skb, u16 vlan_tag)
744 struct i40e_q_vector *q_vector = rx_ring->q_vector;
746 if (vlan_tag & VLAN_VID_MASK)
747 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
749 napi_gro_receive(&q_vector->napi, skb);
753 * i40e_rx_checksum - Indicate in skb if hw indicated a good cksum
754 * @vsi: the VSI we care about
755 * @skb: skb currently being received and modified
756 * @rx_status: status value of last descriptor in packet
757 * @rx_error: error value of last descriptor in packet
758 * @rx_ptype: ptype value of last descriptor in packet
760 static inline void i40e_rx_checksum(struct i40e_vsi *vsi,
766 struct i40e_rx_ptype_decoded decoded = decode_rx_desc_ptype(rx_ptype);
767 bool ipv4 = false, ipv6 = false;
768 bool ipv4_tunnel, ipv6_tunnel;
773 ipv4_tunnel = (rx_ptype >= I40E_RX_PTYPE_GRENAT4_MAC_PAY3) &&
774 (rx_ptype <= I40E_RX_PTYPE_GRENAT4_MACVLAN_IPV6_ICMP_PAY4);
775 ipv6_tunnel = (rx_ptype >= I40E_RX_PTYPE_GRENAT6_MAC_PAY3) &&
776 (rx_ptype <= I40E_RX_PTYPE_GRENAT6_MACVLAN_IPV6_ICMP_PAY4);
778 skb->ip_summed = CHECKSUM_NONE;
780 /* Rx csum enabled and ip headers found? */
781 if (!(vsi->netdev->features & NETIF_F_RXCSUM))
784 /* did the hardware decode the packet and checksum? */
785 if (!(rx_status & BIT(I40E_RX_DESC_STATUS_L3L4P_SHIFT)))
788 /* both known and outer_ip must be set for the below code to work */
789 if (!(decoded.known && decoded.outer_ip))
792 if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
793 decoded.outer_ip_ver == I40E_RX_PTYPE_OUTER_IPV4)
795 else if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
796 decoded.outer_ip_ver == I40E_RX_PTYPE_OUTER_IPV6)
800 (rx_error & (BIT(I40E_RX_DESC_ERROR_IPE_SHIFT) |
801 BIT(I40E_RX_DESC_ERROR_EIPE_SHIFT))))
804 /* likely incorrect csum if alternate IP extension headers found */
806 rx_status & BIT(I40E_RX_DESC_STATUS_IPV6EXADD_SHIFT))
807 /* don't increment checksum err here, non-fatal err */
810 /* there was some L4 error, count error and punt packet to the stack */
811 if (rx_error & BIT(I40E_RX_DESC_ERROR_L4E_SHIFT))
814 /* handle packets that were not able to be checksummed due
815 * to arrival speed, in this case the stack can compute
818 if (rx_error & BIT(I40E_RX_DESC_ERROR_PPRS_SHIFT))
821 /* If VXLAN traffic has an outer UDPv4 checksum we need to check
822 * it in the driver, hardware does not do it for us.
823 * Since L3L4P bit was set we assume a valid IHL value (>=5)
824 * so the total length of IPv4 header is IHL*4 bytes
825 * The UDP_0 bit *may* bet set if the *inner* header is UDP
828 skb->transport_header = skb->mac_header +
829 sizeof(struct ethhdr) +
830 (ip_hdr(skb)->ihl * 4);
832 /* Add 4 bytes for VLAN tagged packets */
833 skb->transport_header += (skb->protocol == htons(ETH_P_8021Q) ||
834 skb->protocol == htons(ETH_P_8021AD))
837 if ((ip_hdr(skb)->protocol == IPPROTO_UDP) &&
838 (udp_hdr(skb)->check != 0)) {
839 rx_udp_csum = udp_csum(skb);
841 csum = csum_tcpudp_magic(iph->saddr, iph->daddr,
843 skb_transport_offset(skb)),
844 IPPROTO_UDP, rx_udp_csum);
846 if (udp_hdr(skb)->check != csum)
849 } /* else its GRE and so no outer UDP header */
852 skb->ip_summed = CHECKSUM_UNNECESSARY;
853 skb->csum_level = ipv4_tunnel || ipv6_tunnel;
858 vsi->back->hw_csum_rx_error++;
862 * i40e_rx_hash - returns the hash value from the Rx descriptor
863 * @ring: descriptor ring
864 * @rx_desc: specific descriptor
866 static inline u32 i40e_rx_hash(struct i40e_ring *ring,
867 union i40e_rx_desc *rx_desc)
869 const __le64 rss_mask =
870 cpu_to_le64((u64)I40E_RX_DESC_FLTSTAT_RSS_HASH <<
871 I40E_RX_DESC_STATUS_FLTSTAT_SHIFT);
873 if ((ring->netdev->features & NETIF_F_RXHASH) &&
874 (rx_desc->wb.qword1.status_error_len & rss_mask) == rss_mask)
875 return le32_to_cpu(rx_desc->wb.qword0.hi_dword.rss);
881 * i40e_ptype_to_hash - get a hash type
882 * @ptype: the ptype value from the descriptor
884 * Returns a hash type to be used by skb_set_hash
886 static inline enum pkt_hash_types i40e_ptype_to_hash(u8 ptype)
888 struct i40e_rx_ptype_decoded decoded = decode_rx_desc_ptype(ptype);
891 return PKT_HASH_TYPE_NONE;
893 if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
894 decoded.payload_layer == I40E_RX_PTYPE_PAYLOAD_LAYER_PAY4)
895 return PKT_HASH_TYPE_L4;
896 else if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
897 decoded.payload_layer == I40E_RX_PTYPE_PAYLOAD_LAYER_PAY3)
898 return PKT_HASH_TYPE_L3;
900 return PKT_HASH_TYPE_L2;
904 * i40e_clean_rx_irq_ps - Reclaim resources after receive; packet split
905 * @rx_ring: rx ring to clean
906 * @budget: how many cleans we're allowed
908 * Returns true if there's any budget left (e.g. the clean is finished)
910 static int i40e_clean_rx_irq_ps(struct i40e_ring *rx_ring, int budget)
912 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
913 u16 rx_packet_len, rx_header_len, rx_sph, rx_hbo;
914 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
915 const int current_node = numa_mem_id();
916 struct i40e_vsi *vsi = rx_ring->vsi;
917 u16 i = rx_ring->next_to_clean;
918 union i40e_rx_desc *rx_desc;
919 u32 rx_error, rx_status;
924 struct i40e_rx_buffer *rx_bi;
927 /* return some buffers to hardware, one at a time is too slow */
928 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
929 i40evf_alloc_rx_buffers_ps(rx_ring, cleaned_count);
933 i = rx_ring->next_to_clean;
934 rx_desc = I40E_RX_DESC(rx_ring, i);
935 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
936 rx_status = (qword & I40E_RXD_QW1_STATUS_MASK) >>
937 I40E_RXD_QW1_STATUS_SHIFT;
939 if (!(rx_status & BIT(I40E_RX_DESC_STATUS_DD_SHIFT)))
942 /* This memory barrier is needed to keep us from reading
943 * any other fields out of the rx_desc until we know the
947 rx_bi = &rx_ring->rx_bi[i];
950 skb = netdev_alloc_skb_ip_align(rx_ring->netdev,
951 rx_ring->rx_hdr_len);
953 rx_ring->rx_stats.alloc_buff_failed++;
957 /* initialize queue mapping */
958 skb_record_rx_queue(skb, rx_ring->queue_index);
959 /* we are reusing so sync this buffer for CPU use */
960 dma_sync_single_range_for_cpu(rx_ring->dev,
966 rx_packet_len = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
967 I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
968 rx_header_len = (qword & I40E_RXD_QW1_LENGTH_HBUF_MASK) >>
969 I40E_RXD_QW1_LENGTH_HBUF_SHIFT;
970 rx_sph = (qword & I40E_RXD_QW1_LENGTH_SPH_MASK) >>
971 I40E_RXD_QW1_LENGTH_SPH_SHIFT;
973 rx_error = (qword & I40E_RXD_QW1_ERROR_MASK) >>
974 I40E_RXD_QW1_ERROR_SHIFT;
975 rx_hbo = rx_error & BIT(I40E_RX_DESC_ERROR_HBO_SHIFT);
976 rx_error &= ~BIT(I40E_RX_DESC_ERROR_HBO_SHIFT);
978 rx_ptype = (qword & I40E_RXD_QW1_PTYPE_MASK) >>
979 I40E_RXD_QW1_PTYPE_SHIFT;
980 prefetch(rx_bi->page);
983 if (rx_hbo || rx_sph) {
987 len = I40E_RX_HDR_SIZE;
990 memcpy(__skb_put(skb, len), rx_bi->hdr_buf, len);
991 } else if (skb->len == 0) {
994 len = (rx_packet_len > skb_headlen(skb) ?
995 skb_headlen(skb) : rx_packet_len);
996 memcpy(__skb_put(skb, len),
997 rx_bi->page + rx_bi->page_offset,
999 rx_bi->page_offset += len;
1000 rx_packet_len -= len;
1003 /* Get the rest of the data if this was a header split */
1004 if (rx_packet_len) {
1005 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
1010 skb->len += rx_packet_len;
1011 skb->data_len += rx_packet_len;
1012 skb->truesize += rx_packet_len;
1014 if ((page_count(rx_bi->page) == 1) &&
1015 (page_to_nid(rx_bi->page) == current_node))
1016 get_page(rx_bi->page);
1020 dma_unmap_page(rx_ring->dev,
1024 rx_bi->page_dma = 0;
1026 I40E_RX_INCREMENT(rx_ring, i);
1029 !(rx_status & BIT(I40E_RX_DESC_STATUS_EOF_SHIFT)))) {
1030 struct i40e_rx_buffer *next_buffer;
1032 next_buffer = &rx_ring->rx_bi[i];
1033 next_buffer->skb = skb;
1034 rx_ring->rx_stats.non_eop_descs++;
1038 /* ERR_MASK will only have valid bits if EOP set */
1039 if (unlikely(rx_error & BIT(I40E_RX_DESC_ERROR_RXE_SHIFT))) {
1040 dev_kfree_skb_any(skb);
1044 skb_set_hash(skb, i40e_rx_hash(rx_ring, rx_desc),
1045 i40e_ptype_to_hash(rx_ptype));
1046 /* probably a little skewed due to removing CRC */
1047 total_rx_bytes += skb->len;
1050 skb->protocol = eth_type_trans(skb, rx_ring->netdev);
1052 i40e_rx_checksum(vsi, skb, rx_status, rx_error, rx_ptype);
1054 vlan_tag = rx_status & BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)
1055 ? le16_to_cpu(rx_desc->wb.qword0.lo_dword.l2tag1)
1058 if (!i40e_fcoe_handle_offload(rx_ring, rx_desc, skb)) {
1059 dev_kfree_skb_any(skb);
1063 skb_mark_napi_id(skb, &rx_ring->q_vector->napi);
1064 i40e_receive_skb(rx_ring, skb, vlan_tag);
1066 rx_desc->wb.qword1.status_error_len = 0;
1068 } while (likely(total_rx_packets < budget));
1070 u64_stats_update_begin(&rx_ring->syncp);
1071 rx_ring->stats.packets += total_rx_packets;
1072 rx_ring->stats.bytes += total_rx_bytes;
1073 u64_stats_update_end(&rx_ring->syncp);
1074 rx_ring->q_vector->rx.total_packets += total_rx_packets;
1075 rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
1077 return total_rx_packets;
1081 * i40e_clean_rx_irq_1buf - Reclaim resources after receive; single buffer
1082 * @rx_ring: rx ring to clean
1083 * @budget: how many cleans we're allowed
1085 * Returns number of packets cleaned
1087 static int i40e_clean_rx_irq_1buf(struct i40e_ring *rx_ring, int budget)
1089 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
1090 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
1091 struct i40e_vsi *vsi = rx_ring->vsi;
1092 union i40e_rx_desc *rx_desc;
1093 u32 rx_error, rx_status;
1100 struct i40e_rx_buffer *rx_bi;
1101 struct sk_buff *skb;
1103 /* return some buffers to hardware, one at a time is too slow */
1104 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
1105 i40evf_alloc_rx_buffers_1buf(rx_ring, cleaned_count);
1109 i = rx_ring->next_to_clean;
1110 rx_desc = I40E_RX_DESC(rx_ring, i);
1111 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
1112 rx_status = (qword & I40E_RXD_QW1_STATUS_MASK) >>
1113 I40E_RXD_QW1_STATUS_SHIFT;
1115 if (!(rx_status & BIT(I40E_RX_DESC_STATUS_DD_SHIFT)))
1118 /* This memory barrier is needed to keep us from reading
1119 * any other fields out of the rx_desc until we know the
1124 rx_bi = &rx_ring->rx_bi[i];
1126 prefetch(skb->data);
1128 rx_packet_len = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
1129 I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
1131 rx_error = (qword & I40E_RXD_QW1_ERROR_MASK) >>
1132 I40E_RXD_QW1_ERROR_SHIFT;
1133 rx_error &= ~BIT(I40E_RX_DESC_ERROR_HBO_SHIFT);
1135 rx_ptype = (qword & I40E_RXD_QW1_PTYPE_MASK) >>
1136 I40E_RXD_QW1_PTYPE_SHIFT;
1140 /* Get the header and possibly the whole packet
1141 * If this is an skb from previous receive dma will be 0
1143 skb_put(skb, rx_packet_len);
1144 dma_unmap_single(rx_ring->dev, rx_bi->dma, rx_ring->rx_buf_len,
1148 I40E_RX_INCREMENT(rx_ring, i);
1151 !(rx_status & BIT(I40E_RX_DESC_STATUS_EOF_SHIFT)))) {
1152 rx_ring->rx_stats.non_eop_descs++;
1156 /* ERR_MASK will only have valid bits if EOP set */
1157 if (unlikely(rx_error & BIT(I40E_RX_DESC_ERROR_RXE_SHIFT))) {
1158 dev_kfree_skb_any(skb);
1162 skb_set_hash(skb, i40e_rx_hash(rx_ring, rx_desc),
1163 i40e_ptype_to_hash(rx_ptype));
1164 /* probably a little skewed due to removing CRC */
1165 total_rx_bytes += skb->len;
1168 skb->protocol = eth_type_trans(skb, rx_ring->netdev);
1170 i40e_rx_checksum(vsi, skb, rx_status, rx_error, rx_ptype);
1172 vlan_tag = rx_status & BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)
1173 ? le16_to_cpu(rx_desc->wb.qword0.lo_dword.l2tag1)
1175 i40e_receive_skb(rx_ring, skb, vlan_tag);
1177 rx_desc->wb.qword1.status_error_len = 0;
1178 } while (likely(total_rx_packets < budget));
1180 u64_stats_update_begin(&rx_ring->syncp);
1181 rx_ring->stats.packets += total_rx_packets;
1182 rx_ring->stats.bytes += total_rx_bytes;
1183 u64_stats_update_end(&rx_ring->syncp);
1184 rx_ring->q_vector->rx.total_packets += total_rx_packets;
1185 rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
1187 return total_rx_packets;
1191 * i40e_update_enable_itr - Update itr and re-enable MSIX interrupt
1192 * @vsi: the VSI we care about
1193 * @q_vector: q_vector for which itr is being updated and interrupt enabled
1196 static inline void i40e_update_enable_itr(struct i40e_vsi *vsi,
1197 struct i40e_q_vector *q_vector)
1199 struct i40e_hw *hw = &vsi->back->hw;
1204 vector = (q_vector->v_idx + vsi->base_vector);
1205 if (ITR_IS_DYNAMIC(vsi->rx_itr_setting)) {
1206 old_itr = q_vector->rx.itr;
1207 i40e_set_new_dynamic_itr(&q_vector->rx);
1208 if (old_itr != q_vector->rx.itr) {
1209 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
1210 I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK |
1212 I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) |
1213 (q_vector->rx.itr <<
1214 I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT);
1216 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
1217 I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK |
1219 I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT);
1221 if (!test_bit(__I40E_DOWN, &vsi->state))
1222 wr32(hw, I40E_VFINT_DYN_CTLN1(vector - 1), val);
1224 i40evf_irq_enable_queues(vsi->back, 1
1225 << q_vector->v_idx);
1227 if (ITR_IS_DYNAMIC(vsi->tx_itr_setting)) {
1228 old_itr = q_vector->tx.itr;
1229 i40e_set_new_dynamic_itr(&q_vector->tx);
1230 if (old_itr != q_vector->tx.itr) {
1231 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
1232 I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK |
1234 I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) |
1235 (q_vector->tx.itr <<
1236 I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT);
1239 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
1240 I40E_VFINT_DYN_CTLN1_CLEARPBA_MASK |
1242 I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT);
1244 if (!test_bit(__I40E_DOWN, &vsi->state))
1245 wr32(hw, I40E_VFINT_DYN_CTLN1(vector - 1), val);
1247 i40evf_irq_enable_queues(vsi->back, BIT(q_vector->v_idx));
1252 * i40evf_napi_poll - NAPI polling Rx/Tx cleanup routine
1253 * @napi: napi struct with our devices info in it
1254 * @budget: amount of work driver is allowed to do this pass, in packets
1256 * This function will clean all queues associated with a q_vector.
1258 * Returns the amount of work done
1260 int i40evf_napi_poll(struct napi_struct *napi, int budget)
1262 struct i40e_q_vector *q_vector =
1263 container_of(napi, struct i40e_q_vector, napi);
1264 struct i40e_vsi *vsi = q_vector->vsi;
1265 struct i40e_ring *ring;
1266 bool clean_complete = true;
1267 bool arm_wb = false;
1268 int budget_per_ring;
1271 if (test_bit(__I40E_DOWN, &vsi->state)) {
1272 napi_complete(napi);
1276 /* Since the actual Tx work is minimal, we can give the Tx a larger
1277 * budget and be more aggressive about cleaning up the Tx descriptors.
1279 i40e_for_each_ring(ring, q_vector->tx) {
1280 clean_complete &= i40e_clean_tx_irq(ring, vsi->work_limit);
1281 arm_wb |= ring->arm_wb;
1282 ring->arm_wb = false;
1285 /* Handle case where we are called by netpoll with a budget of 0 */
1289 /* We attempt to distribute budget to each Rx queue fairly, but don't
1290 * allow the budget to go below 1 because that would exit polling early.
1292 budget_per_ring = max(budget/q_vector->num_ringpairs, 1);
1294 i40e_for_each_ring(ring, q_vector->rx) {
1297 if (ring_is_ps_enabled(ring))
1298 cleaned = i40e_clean_rx_irq_ps(ring, budget_per_ring);
1300 cleaned = i40e_clean_rx_irq_1buf(ring, budget_per_ring);
1302 work_done += cleaned;
1303 /* if we didn't clean as many as budgeted, we must be done */
1304 clean_complete &= (budget_per_ring != cleaned);
1307 /* If work not completed, return budget and polling will return */
1308 if (!clean_complete) {
1311 i40evf_force_wb(vsi, q_vector);
1315 if (vsi->back->flags & I40E_TXR_FLAGS_WB_ON_ITR)
1316 q_vector->arm_wb_state = false;
1318 /* Work is done so exit the polling mode and re-enable the interrupt */
1319 napi_complete_done(napi, work_done);
1320 i40e_update_enable_itr(vsi, q_vector);
1325 * i40evf_tx_prepare_vlan_flags - prepare generic TX VLAN tagging flags for HW
1327 * @tx_ring: ring to send buffer on
1328 * @flags: the tx flags to be set
1330 * Checks the skb and set up correspondingly several generic transmit flags
1331 * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
1333 * Returns error code indicate the frame should be dropped upon error and the
1334 * otherwise returns 0 to indicate the flags has been set properly.
1336 static inline int i40evf_tx_prepare_vlan_flags(struct sk_buff *skb,
1337 struct i40e_ring *tx_ring,
1340 __be16 protocol = skb->protocol;
1343 if (protocol == htons(ETH_P_8021Q) &&
1344 !(tx_ring->netdev->features & NETIF_F_HW_VLAN_CTAG_TX)) {
1345 /* When HW VLAN acceleration is turned off by the user the
1346 * stack sets the protocol to 8021q so that the driver
1347 * can take any steps required to support the SW only
1348 * VLAN handling. In our case the driver doesn't need
1349 * to take any further steps so just set the protocol
1350 * to the encapsulated ethertype.
1352 skb->protocol = vlan_get_protocol(skb);
1356 /* if we have a HW VLAN tag being added, default to the HW one */
1357 if (skb_vlan_tag_present(skb)) {
1358 tx_flags |= skb_vlan_tag_get(skb) << I40E_TX_FLAGS_VLAN_SHIFT;
1359 tx_flags |= I40E_TX_FLAGS_HW_VLAN;
1360 /* else if it is a SW VLAN, check the next protocol and store the tag */
1361 } else if (protocol == htons(ETH_P_8021Q)) {
1362 struct vlan_hdr *vhdr, _vhdr;
1364 vhdr = skb_header_pointer(skb, ETH_HLEN, sizeof(_vhdr), &_vhdr);
1368 protocol = vhdr->h_vlan_encapsulated_proto;
1369 tx_flags |= ntohs(vhdr->h_vlan_TCI) << I40E_TX_FLAGS_VLAN_SHIFT;
1370 tx_flags |= I40E_TX_FLAGS_SW_VLAN;
1379 * i40e_tso - set up the tso context descriptor
1380 * @tx_ring: ptr to the ring to send
1381 * @skb: ptr to the skb we're sending
1382 * @hdr_len: ptr to the size of the packet header
1383 * @cd_tunneling: ptr to context descriptor bits
1385 * Returns 0 if no TSO can happen, 1 if tso is going, or error
1387 static int i40e_tso(struct i40e_ring *tx_ring, struct sk_buff *skb,
1388 u8 *hdr_len, u64 *cd_type_cmd_tso_mss,
1391 u32 cd_cmd, cd_tso_len, cd_mss;
1392 struct ipv6hdr *ipv6h;
1393 struct tcphdr *tcph;
1398 if (!skb_is_gso(skb))
1401 err = skb_cow_head(skb, 0);
1405 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
1406 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
1408 if (iph->version == 4) {
1409 tcph = skb->encapsulation ? inner_tcp_hdr(skb) : tcp_hdr(skb);
1412 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1414 } else if (ipv6h->version == 6) {
1415 tcph = skb->encapsulation ? inner_tcp_hdr(skb) : tcp_hdr(skb);
1416 ipv6h->payload_len = 0;
1417 tcph->check = ~csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
1421 l4len = skb->encapsulation ? inner_tcp_hdrlen(skb) : tcp_hdrlen(skb);
1422 *hdr_len = (skb->encapsulation
1423 ? (skb_inner_transport_header(skb) - skb->data)
1424 : skb_transport_offset(skb)) + l4len;
1426 /* find the field values */
1427 cd_cmd = I40E_TX_CTX_DESC_TSO;
1428 cd_tso_len = skb->len - *hdr_len;
1429 cd_mss = skb_shinfo(skb)->gso_size;
1430 *cd_type_cmd_tso_mss |= ((u64)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
1432 I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
1433 ((u64)cd_mss << I40E_TXD_CTX_QW1_MSS_SHIFT);
1438 * i40e_tx_enable_csum - Enable Tx checksum offloads
1440 * @tx_flags: pointer to Tx flags currently set
1441 * @td_cmd: Tx descriptor command bits to set
1442 * @td_offset: Tx descriptor header offsets to set
1443 * @cd_tunneling: ptr to context desc bits
1445 static void i40e_tx_enable_csum(struct sk_buff *skb, u32 *tx_flags,
1446 u32 *td_cmd, u32 *td_offset,
1447 struct i40e_ring *tx_ring,
1450 struct ipv6hdr *this_ipv6_hdr;
1451 unsigned int this_tcp_hdrlen;
1452 struct iphdr *this_ip_hdr;
1453 u32 network_hdr_len;
1455 struct udphdr *oudph;
1459 if (skb->encapsulation) {
1460 switch (ip_hdr(skb)->protocol) {
1462 oudph = udp_hdr(skb);
1464 l4_tunnel = I40E_TXD_CTX_UDP_TUNNELING;
1465 *tx_flags |= I40E_TX_FLAGS_VXLAN_TUNNEL;
1470 network_hdr_len = skb_inner_network_header_len(skb);
1471 this_ip_hdr = inner_ip_hdr(skb);
1472 this_ipv6_hdr = inner_ipv6_hdr(skb);
1473 this_tcp_hdrlen = inner_tcp_hdrlen(skb);
1475 if (*tx_flags & I40E_TX_FLAGS_IPV4) {
1476 if (*tx_flags & I40E_TX_FLAGS_TSO) {
1477 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
1478 ip_hdr(skb)->check = 0;
1481 I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
1483 } else if (*tx_flags & I40E_TX_FLAGS_IPV6) {
1484 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
1485 if (*tx_flags & I40E_TX_FLAGS_TSO)
1486 ip_hdr(skb)->check = 0;
1489 /* Now set the ctx descriptor fields */
1490 *cd_tunneling |= (skb_network_header_len(skb) >> 2) <<
1491 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT |
1493 ((skb_inner_network_offset(skb) -
1494 skb_transport_offset(skb)) >> 1) <<
1495 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
1496 if (this_ip_hdr->version == 6) {
1497 *tx_flags &= ~I40E_TX_FLAGS_IPV4;
1498 *tx_flags |= I40E_TX_FLAGS_IPV6;
1502 if ((tx_ring->flags & I40E_TXR_FLAGS_OUTER_UDP_CSUM) &&
1503 (l4_tunnel == I40E_TXD_CTX_UDP_TUNNELING) &&
1504 (*cd_tunneling & I40E_TXD_CTX_QW0_EXT_IP_MASK)) {
1505 oudph->check = ~csum_tcpudp_magic(oiph->saddr,
1507 (skb->len - skb_transport_offset(skb)),
1509 *cd_tunneling |= I40E_TXD_CTX_QW0_L4T_CS_MASK;
1512 network_hdr_len = skb_network_header_len(skb);
1513 this_ip_hdr = ip_hdr(skb);
1514 this_ipv6_hdr = ipv6_hdr(skb);
1515 this_tcp_hdrlen = tcp_hdrlen(skb);
1518 /* Enable IP checksum offloads */
1519 if (*tx_flags & I40E_TX_FLAGS_IPV4) {
1520 l4_hdr = this_ip_hdr->protocol;
1521 /* the stack computes the IP header already, the only time we
1522 * need the hardware to recompute it is in the case of TSO.
1524 if (*tx_flags & I40E_TX_FLAGS_TSO) {
1525 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
1526 this_ip_hdr->check = 0;
1528 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
1530 /* Now set the td_offset for IP header length */
1531 *td_offset = (network_hdr_len >> 2) <<
1532 I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
1533 } else if (*tx_flags & I40E_TX_FLAGS_IPV6) {
1534 l4_hdr = this_ipv6_hdr->nexthdr;
1535 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
1536 /* Now set the td_offset for IP header length */
1537 *td_offset = (network_hdr_len >> 2) <<
1538 I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
1540 /* words in MACLEN + dwords in IPLEN + dwords in L4Len */
1541 *td_offset |= (skb_network_offset(skb) >> 1) <<
1542 I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
1544 /* Enable L4 checksum offloads */
1547 /* enable checksum offloads */
1548 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
1549 *td_offset |= (this_tcp_hdrlen >> 2) <<
1550 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
1553 /* enable SCTP checksum offload */
1554 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
1555 *td_offset |= (sizeof(struct sctphdr) >> 2) <<
1556 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
1559 /* enable UDP checksum offload */
1560 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
1561 *td_offset |= (sizeof(struct udphdr) >> 2) <<
1562 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
1570 * i40e_create_tx_ctx Build the Tx context descriptor
1571 * @tx_ring: ring to create the descriptor on
1572 * @cd_type_cmd_tso_mss: Quad Word 1
1573 * @cd_tunneling: Quad Word 0 - bits 0-31
1574 * @cd_l2tag2: Quad Word 0 - bits 32-63
1576 static void i40e_create_tx_ctx(struct i40e_ring *tx_ring,
1577 const u64 cd_type_cmd_tso_mss,
1578 const u32 cd_tunneling, const u32 cd_l2tag2)
1580 struct i40e_tx_context_desc *context_desc;
1581 int i = tx_ring->next_to_use;
1583 if ((cd_type_cmd_tso_mss == I40E_TX_DESC_DTYPE_CONTEXT) &&
1584 !cd_tunneling && !cd_l2tag2)
1587 /* grab the next descriptor */
1588 context_desc = I40E_TX_CTXTDESC(tx_ring, i);
1591 tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
1593 /* cpu_to_le32 and assign to struct fields */
1594 context_desc->tunneling_params = cpu_to_le32(cd_tunneling);
1595 context_desc->l2tag2 = cpu_to_le16(cd_l2tag2);
1596 context_desc->rsvd = cpu_to_le16(0);
1597 context_desc->type_cmd_tso_mss = cpu_to_le64(cd_type_cmd_tso_mss);
1601 * i40e_chk_linearize - Check if there are more than 8 fragments per packet
1603 * @tx_flags: collected send information
1605 * Note: Our HW can't scatter-gather more than 8 fragments to build
1606 * a packet on the wire and so we need to figure out the cases where we
1607 * need to linearize the skb.
1609 static bool i40e_chk_linearize(struct sk_buff *skb, u32 tx_flags)
1611 struct skb_frag_struct *frag;
1612 bool linearize = false;
1613 unsigned int size = 0;
1617 num_frags = skb_shinfo(skb)->nr_frags;
1618 gso_segs = skb_shinfo(skb)->gso_segs;
1620 if (tx_flags & (I40E_TX_FLAGS_TSO | I40E_TX_FLAGS_FSO)) {
1623 if (num_frags < (I40E_MAX_BUFFER_TXD))
1624 goto linearize_chk_done;
1625 /* try the simple math, if we have too many frags per segment */
1626 if (DIV_ROUND_UP((num_frags + gso_segs), gso_segs) >
1627 I40E_MAX_BUFFER_TXD) {
1629 goto linearize_chk_done;
1631 frag = &skb_shinfo(skb)->frags[0];
1632 /* we might still have more fragments per segment */
1634 size += skb_frag_size(frag);
1636 if ((size >= skb_shinfo(skb)->gso_size) &&
1637 (j < I40E_MAX_BUFFER_TXD)) {
1638 size = (size % skb_shinfo(skb)->gso_size);
1641 if (j == I40E_MAX_BUFFER_TXD) {
1646 } while (num_frags);
1648 if (num_frags >= I40E_MAX_BUFFER_TXD)
1657 * __i40evf_maybe_stop_tx - 2nd level check for tx stop conditions
1658 * @tx_ring: the ring to be checked
1659 * @size: the size buffer we want to assure is available
1661 * Returns -EBUSY if a stop is needed, else 0
1663 static inline int __i40evf_maybe_stop_tx(struct i40e_ring *tx_ring, int size)
1665 netif_stop_subqueue(tx_ring->netdev, tx_ring->queue_index);
1666 /* Memory barrier before checking head and tail */
1669 /* Check again in a case another CPU has just made room available. */
1670 if (likely(I40E_DESC_UNUSED(tx_ring) < size))
1673 /* A reprieve! - use start_queue because it doesn't call schedule */
1674 netif_start_subqueue(tx_ring->netdev, tx_ring->queue_index);
1675 ++tx_ring->tx_stats.restart_queue;
1680 * i40evf_maybe_stop_tx - 1st level check for tx stop conditions
1681 * @tx_ring: the ring to be checked
1682 * @size: the size buffer we want to assure is available
1684 * Returns 0 if stop is not needed
1686 static inline int i40evf_maybe_stop_tx(struct i40e_ring *tx_ring, int size)
1688 if (likely(I40E_DESC_UNUSED(tx_ring) >= size))
1690 return __i40evf_maybe_stop_tx(tx_ring, size);
1694 * i40evf_tx_map - Build the Tx descriptor
1695 * @tx_ring: ring to send buffer on
1697 * @first: first buffer info buffer to use
1698 * @tx_flags: collected send information
1699 * @hdr_len: size of the packet header
1700 * @td_cmd: the command field in the descriptor
1701 * @td_offset: offset for checksum or crc
1703 static inline void i40evf_tx_map(struct i40e_ring *tx_ring, struct sk_buff *skb,
1704 struct i40e_tx_buffer *first, u32 tx_flags,
1705 const u8 hdr_len, u32 td_cmd, u32 td_offset)
1707 unsigned int data_len = skb->data_len;
1708 unsigned int size = skb_headlen(skb);
1709 struct skb_frag_struct *frag;
1710 struct i40e_tx_buffer *tx_bi;
1711 struct i40e_tx_desc *tx_desc;
1712 u16 i = tx_ring->next_to_use;
1717 if (tx_flags & I40E_TX_FLAGS_HW_VLAN) {
1718 td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1719 td_tag = (tx_flags & I40E_TX_FLAGS_VLAN_MASK) >>
1720 I40E_TX_FLAGS_VLAN_SHIFT;
1723 if (tx_flags & (I40E_TX_FLAGS_TSO | I40E_TX_FLAGS_FSO))
1724 gso_segs = skb_shinfo(skb)->gso_segs;
1728 /* multiply data chunks by size of headers */
1729 first->bytecount = skb->len - hdr_len + (gso_segs * hdr_len);
1730 first->gso_segs = gso_segs;
1732 first->tx_flags = tx_flags;
1734 dma = dma_map_single(tx_ring->dev, skb->data, size, DMA_TO_DEVICE);
1736 tx_desc = I40E_TX_DESC(tx_ring, i);
1739 for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
1740 if (dma_mapping_error(tx_ring->dev, dma))
1743 /* record length, and DMA address */
1744 dma_unmap_len_set(tx_bi, len, size);
1745 dma_unmap_addr_set(tx_bi, dma, dma);
1747 tx_desc->buffer_addr = cpu_to_le64(dma);
1749 while (unlikely(size > I40E_MAX_DATA_PER_TXD)) {
1750 tx_desc->cmd_type_offset_bsz =
1751 build_ctob(td_cmd, td_offset,
1752 I40E_MAX_DATA_PER_TXD, td_tag);
1756 if (i == tx_ring->count) {
1757 tx_desc = I40E_TX_DESC(tx_ring, 0);
1761 dma += I40E_MAX_DATA_PER_TXD;
1762 size -= I40E_MAX_DATA_PER_TXD;
1764 tx_desc->buffer_addr = cpu_to_le64(dma);
1767 if (likely(!data_len))
1770 tx_desc->cmd_type_offset_bsz = build_ctob(td_cmd, td_offset,
1775 if (i == tx_ring->count) {
1776 tx_desc = I40E_TX_DESC(tx_ring, 0);
1780 size = skb_frag_size(frag);
1783 dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
1786 tx_bi = &tx_ring->tx_bi[i];
1789 /* Place RS bit on last descriptor of any packet that spans across the
1790 * 4th descriptor (WB_STRIDE aka 0x3) in a 64B cacheline.
1792 #define WB_STRIDE 0x3
1793 if (((i & WB_STRIDE) != WB_STRIDE) &&
1794 (first <= &tx_ring->tx_bi[i]) &&
1795 (first >= &tx_ring->tx_bi[i & ~WB_STRIDE])) {
1796 tx_desc->cmd_type_offset_bsz =
1797 build_ctob(td_cmd, td_offset, size, td_tag) |
1798 cpu_to_le64((u64)I40E_TX_DESC_CMD_EOP <<
1799 I40E_TXD_QW1_CMD_SHIFT);
1801 tx_desc->cmd_type_offset_bsz =
1802 build_ctob(td_cmd, td_offset, size, td_tag) |
1803 cpu_to_le64((u64)I40E_TXD_CMD <<
1804 I40E_TXD_QW1_CMD_SHIFT);
1807 netdev_tx_sent_queue(netdev_get_tx_queue(tx_ring->netdev,
1808 tx_ring->queue_index),
1811 /* Force memory writes to complete before letting h/w
1812 * know there are new descriptors to fetch. (Only
1813 * applicable for weak-ordered memory model archs,
1818 /* set next_to_watch value indicating a packet is present */
1819 first->next_to_watch = tx_desc;
1822 if (i == tx_ring->count)
1825 tx_ring->next_to_use = i;
1827 i40evf_maybe_stop_tx(tx_ring, DESC_NEEDED);
1828 /* notify HW of packet */
1829 if (!skb->xmit_more ||
1830 netif_xmit_stopped(netdev_get_tx_queue(tx_ring->netdev,
1831 tx_ring->queue_index)))
1832 writel(i, tx_ring->tail);
1834 prefetchw(tx_desc + 1);
1839 dev_info(tx_ring->dev, "TX DMA map failed\n");
1841 /* clear dma mappings for failed tx_bi map */
1843 tx_bi = &tx_ring->tx_bi[i];
1844 i40e_unmap_and_free_tx_resource(tx_ring, tx_bi);
1852 tx_ring->next_to_use = i;
1856 * i40evf_xmit_descriptor_count - calculate number of tx descriptors needed
1858 * @tx_ring: ring to send buffer on
1860 * Returns number of data descriptors needed for this skb. Returns 0 to indicate
1861 * there is not enough descriptors available in this ring since we need at least
1864 static inline int i40evf_xmit_descriptor_count(struct sk_buff *skb,
1865 struct i40e_ring *tx_ring)
1870 /* need: 1 descriptor per page * PAGE_SIZE/I40E_MAX_DATA_PER_TXD,
1871 * + 1 desc for skb_head_len/I40E_MAX_DATA_PER_TXD,
1872 * + 4 desc gap to avoid the cache line where head is,
1873 * + 1 desc for context descriptor,
1874 * otherwise try next time
1876 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1877 count += TXD_USE_COUNT(skb_shinfo(skb)->frags[f].size);
1879 count += TXD_USE_COUNT(skb_headlen(skb));
1880 if (i40evf_maybe_stop_tx(tx_ring, count + 4 + 1)) {
1881 tx_ring->tx_stats.tx_busy++;
1888 * i40e_xmit_frame_ring - Sends buffer on Tx ring
1890 * @tx_ring: ring to send buffer on
1892 * Returns NETDEV_TX_OK if sent, else an error code
1894 static netdev_tx_t i40e_xmit_frame_ring(struct sk_buff *skb,
1895 struct i40e_ring *tx_ring)
1897 u64 cd_type_cmd_tso_mss = I40E_TX_DESC_DTYPE_CONTEXT;
1898 u32 cd_tunneling = 0, cd_l2tag2 = 0;
1899 struct i40e_tx_buffer *first;
1907 if (0 == i40evf_xmit_descriptor_count(skb, tx_ring))
1908 return NETDEV_TX_BUSY;
1910 /* prepare the xmit flags */
1911 if (i40evf_tx_prepare_vlan_flags(skb, tx_ring, &tx_flags))
1914 /* obtain protocol of skb */
1915 protocol = vlan_get_protocol(skb);
1917 /* record the location of the first descriptor for this packet */
1918 first = &tx_ring->tx_bi[tx_ring->next_to_use];
1920 /* setup IPv4/IPv6 offloads */
1921 if (protocol == htons(ETH_P_IP))
1922 tx_flags |= I40E_TX_FLAGS_IPV4;
1923 else if (protocol == htons(ETH_P_IPV6))
1924 tx_flags |= I40E_TX_FLAGS_IPV6;
1926 tso = i40e_tso(tx_ring, skb, &hdr_len,
1927 &cd_type_cmd_tso_mss, &cd_tunneling);
1932 tx_flags |= I40E_TX_FLAGS_TSO;
1934 if (i40e_chk_linearize(skb, tx_flags)) {
1935 if (skb_linearize(skb))
1937 tx_ring->tx_stats.tx_linearize++;
1939 skb_tx_timestamp(skb);
1941 /* always enable CRC insertion offload */
1942 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1944 /* Always offload the checksum, since it's in the data descriptor */
1945 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1946 tx_flags |= I40E_TX_FLAGS_CSUM;
1948 i40e_tx_enable_csum(skb, &tx_flags, &td_cmd, &td_offset,
1949 tx_ring, &cd_tunneling);
1952 i40e_create_tx_ctx(tx_ring, cd_type_cmd_tso_mss,
1953 cd_tunneling, cd_l2tag2);
1955 i40evf_tx_map(tx_ring, skb, first, tx_flags, hdr_len,
1958 return NETDEV_TX_OK;
1961 dev_kfree_skb_any(skb);
1962 return NETDEV_TX_OK;
1966 * i40evf_xmit_frame - Selects the correct VSI and Tx queue to send buffer
1968 * @netdev: network interface device structure
1970 * Returns NETDEV_TX_OK if sent, else an error code
1972 netdev_tx_t i40evf_xmit_frame(struct sk_buff *skb, struct net_device *netdev)
1974 struct i40evf_adapter *adapter = netdev_priv(netdev);
1975 struct i40e_ring *tx_ring = adapter->tx_rings[skb->queue_mapping];
1977 /* hardware can't handle really short frames, hardware padding works
1980 if (unlikely(skb->len < I40E_MIN_TX_LEN)) {
1981 if (skb_pad(skb, I40E_MIN_TX_LEN - skb->len))
1982 return NETDEV_TX_OK;
1983 skb->len = I40E_MIN_TX_LEN;
1984 skb_set_tail_pointer(skb, I40E_MIN_TX_LEN);
1987 return i40e_xmit_frame_ring(skb, tx_ring);