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
133 * Since there is no access to the ring head register
134 * in XL710, we need to use our local copies
136 u32 i40evf_get_tx_pending(struct i40e_ring *ring)
140 head = i40e_get_head(ring);
141 tail = readl(ring->tail);
144 return (head < tail) ?
145 tail - head : (tail + ring->count - head);
150 #define WB_STRIDE 0x3
153 * i40e_clean_tx_irq - Reclaim resources after transmit completes
154 * @tx_ring: tx ring to clean
155 * @budget: how many cleans we're allowed
157 * Returns true if there's any budget left (e.g. the clean is finished)
159 static bool i40e_clean_tx_irq(struct i40e_ring *tx_ring, int budget)
161 u16 i = tx_ring->next_to_clean;
162 struct i40e_tx_buffer *tx_buf;
163 struct i40e_tx_desc *tx_head;
164 struct i40e_tx_desc *tx_desc;
165 unsigned int total_packets = 0;
166 unsigned int total_bytes = 0;
168 tx_buf = &tx_ring->tx_bi[i];
169 tx_desc = I40E_TX_DESC(tx_ring, i);
172 tx_head = I40E_TX_DESC(tx_ring, i40e_get_head(tx_ring));
175 struct i40e_tx_desc *eop_desc = tx_buf->next_to_watch;
177 /* if next_to_watch is not set then there is no work pending */
181 /* prevent any other reads prior to eop_desc */
182 read_barrier_depends();
184 /* we have caught up to head, no work left to do */
185 if (tx_head == tx_desc)
188 /* clear next_to_watch to prevent false hangs */
189 tx_buf->next_to_watch = NULL;
191 /* update the statistics for this packet */
192 total_bytes += tx_buf->bytecount;
193 total_packets += tx_buf->gso_segs;
196 dev_kfree_skb_any(tx_buf->skb);
198 /* unmap skb header data */
199 dma_unmap_single(tx_ring->dev,
200 dma_unmap_addr(tx_buf, dma),
201 dma_unmap_len(tx_buf, len),
204 /* clear tx_buffer data */
206 dma_unmap_len_set(tx_buf, len, 0);
208 /* unmap remaining buffers */
209 while (tx_desc != eop_desc) {
216 tx_buf = tx_ring->tx_bi;
217 tx_desc = I40E_TX_DESC(tx_ring, 0);
220 /* unmap any remaining paged data */
221 if (dma_unmap_len(tx_buf, len)) {
222 dma_unmap_page(tx_ring->dev,
223 dma_unmap_addr(tx_buf, dma),
224 dma_unmap_len(tx_buf, len),
226 dma_unmap_len_set(tx_buf, len, 0);
230 /* move us one more past the eop_desc for start of next pkt */
236 tx_buf = tx_ring->tx_bi;
237 tx_desc = I40E_TX_DESC(tx_ring, 0);
242 /* update budget accounting */
244 } while (likely(budget));
247 tx_ring->next_to_clean = i;
248 u64_stats_update_begin(&tx_ring->syncp);
249 tx_ring->stats.bytes += total_bytes;
250 tx_ring->stats.packets += total_packets;
251 u64_stats_update_end(&tx_ring->syncp);
252 tx_ring->q_vector->tx.total_bytes += total_bytes;
253 tx_ring->q_vector->tx.total_packets += total_packets;
255 if (tx_ring->flags & I40E_TXR_FLAGS_WB_ON_ITR) {
257 /* check to see if there are < 4 descriptors
258 * waiting to be written back, then kick the hardware to force
259 * them to be written back in case we stay in NAPI.
260 * In this mode on X722 we do not enable Interrupt.
262 j = i40evf_get_tx_pending(tx_ring);
265 ((j / (WB_STRIDE + 1)) == 0) && (j > 0) &&
266 !test_bit(__I40E_DOWN, &tx_ring->vsi->state) &&
267 (I40E_DESC_UNUSED(tx_ring) != tx_ring->count))
268 tx_ring->arm_wb = true;
271 netdev_tx_completed_queue(netdev_get_tx_queue(tx_ring->netdev,
272 tx_ring->queue_index),
273 total_packets, total_bytes);
275 #define TX_WAKE_THRESHOLD (DESC_NEEDED * 2)
276 if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) &&
277 (I40E_DESC_UNUSED(tx_ring) >= TX_WAKE_THRESHOLD))) {
278 /* Make sure that anybody stopping the queue after this
279 * sees the new next_to_clean.
282 if (__netif_subqueue_stopped(tx_ring->netdev,
283 tx_ring->queue_index) &&
284 !test_bit(__I40E_DOWN, &tx_ring->vsi->state)) {
285 netif_wake_subqueue(tx_ring->netdev,
286 tx_ring->queue_index);
287 ++tx_ring->tx_stats.restart_queue;
295 * i40evf_enable_wb_on_itr - Arm hardware to do a wb, interrupts are not enabled
296 * @vsi: the VSI we care about
297 * @q_vector: the vector on which to enable writeback
300 static void i40e_enable_wb_on_itr(struct i40e_vsi *vsi,
301 struct i40e_q_vector *q_vector)
303 u16 flags = q_vector->tx.ring[0].flags;
306 if (!(flags & I40E_TXR_FLAGS_WB_ON_ITR))
309 if (q_vector->arm_wb_state)
312 val = I40E_VFINT_DYN_CTLN1_WB_ON_ITR_MASK |
313 I40E_VFINT_DYN_CTLN1_ITR_INDX_MASK; /* set noitr */
316 I40E_VFINT_DYN_CTLN1(q_vector->v_idx +
317 vsi->base_vector - 1), val);
318 q_vector->arm_wb_state = true;
322 * i40evf_force_wb - Issue SW Interrupt so HW does a wb
323 * @vsi: the VSI we care about
324 * @q_vector: the vector on which to force writeback
327 void i40evf_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector)
329 u32 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
330 I40E_VFINT_DYN_CTLN1_ITR_INDX_MASK | /* set noitr */
331 I40E_VFINT_DYN_CTLN1_SWINT_TRIG_MASK |
332 I40E_VFINT_DYN_CTLN1_SW_ITR_INDX_ENA_MASK
333 /* allow 00 to be written to the index */;
336 I40E_VFINT_DYN_CTLN1(q_vector->v_idx + vsi->base_vector - 1),
341 * i40e_set_new_dynamic_itr - Find new ITR level
342 * @rc: structure containing ring performance data
344 * Returns true if ITR changed, false if not
346 * Stores a new ITR value based on packets and byte counts during
347 * the last interrupt. The advantage of per interrupt computation
348 * is faster updates and more accurate ITR for the current traffic
349 * pattern. Constants in this function were computed based on
350 * theoretical maximum wire speed and thresholds were set based on
351 * testing data as well as attempting to minimize response time
352 * while increasing bulk throughput.
354 static bool i40e_set_new_dynamic_itr(struct i40e_ring_container *rc)
356 enum i40e_latency_range new_latency_range = rc->latency_range;
357 struct i40e_q_vector *qv = rc->ring->q_vector;
358 u32 new_itr = rc->itr;
362 if (rc->total_packets == 0 || !rc->itr)
365 /* simple throttlerate management
366 * 0-10MB/s lowest (50000 ints/s)
367 * 10-20MB/s low (20000 ints/s)
368 * 20-1249MB/s bulk (18000 ints/s)
369 * > 40000 Rx packets per second (8000 ints/s)
371 * The math works out because the divisor is in 10^(-6) which
372 * turns the bytes/us input value into MB/s values, but
373 * make sure to use usecs, as the register values written
374 * are in 2 usec increments in the ITR registers, and make sure
375 * to use the smoothed values that the countdown timer gives us.
377 usecs = (rc->itr << 1) * ITR_COUNTDOWN_START;
378 bytes_per_int = rc->total_bytes / usecs;
380 switch (new_latency_range) {
381 case I40E_LOWEST_LATENCY:
382 if (bytes_per_int > 10)
383 new_latency_range = I40E_LOW_LATENCY;
385 case I40E_LOW_LATENCY:
386 if (bytes_per_int > 20)
387 new_latency_range = I40E_BULK_LATENCY;
388 else if (bytes_per_int <= 10)
389 new_latency_range = I40E_LOWEST_LATENCY;
391 case I40E_BULK_LATENCY:
392 case I40E_ULTRA_LATENCY:
394 if (bytes_per_int <= 20)
395 new_latency_range = I40E_LOW_LATENCY;
399 /* this is to adjust RX more aggressively when streaming small
400 * packets. The value of 40000 was picked as it is just beyond
401 * what the hardware can receive per second if in low latency
404 #define RX_ULTRA_PACKET_RATE 40000
406 if ((((rc->total_packets * 1000000) / usecs) > RX_ULTRA_PACKET_RATE) &&
408 new_latency_range = I40E_ULTRA_LATENCY;
410 rc->latency_range = new_latency_range;
412 switch (new_latency_range) {
413 case I40E_LOWEST_LATENCY:
414 new_itr = I40E_ITR_50K;
416 case I40E_LOW_LATENCY:
417 new_itr = I40E_ITR_20K;
419 case I40E_BULK_LATENCY:
420 new_itr = I40E_ITR_18K;
422 case I40E_ULTRA_LATENCY:
423 new_itr = I40E_ITR_8K;
430 rc->total_packets = 0;
432 if (new_itr != rc->itr) {
441 * i40evf_setup_tx_descriptors - Allocate the Tx descriptors
442 * @tx_ring: the tx ring to set up
444 * Return 0 on success, negative on error
446 int i40evf_setup_tx_descriptors(struct i40e_ring *tx_ring)
448 struct device *dev = tx_ring->dev;
454 /* warn if we are about to overwrite the pointer */
455 WARN_ON(tx_ring->tx_bi);
456 bi_size = sizeof(struct i40e_tx_buffer) * tx_ring->count;
457 tx_ring->tx_bi = kzalloc(bi_size, GFP_KERNEL);
461 /* round up to nearest 4K */
462 tx_ring->size = tx_ring->count * sizeof(struct i40e_tx_desc);
463 /* add u32 for head writeback, align after this takes care of
464 * guaranteeing this is at least one cache line in size
466 tx_ring->size += sizeof(u32);
467 tx_ring->size = ALIGN(tx_ring->size, 4096);
468 tx_ring->desc = dma_alloc_coherent(dev, tx_ring->size,
469 &tx_ring->dma, GFP_KERNEL);
470 if (!tx_ring->desc) {
471 dev_info(dev, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
476 tx_ring->next_to_use = 0;
477 tx_ring->next_to_clean = 0;
481 kfree(tx_ring->tx_bi);
482 tx_ring->tx_bi = NULL;
487 * i40evf_clean_rx_ring - Free Rx buffers
488 * @rx_ring: ring to be cleaned
490 void i40evf_clean_rx_ring(struct i40e_ring *rx_ring)
492 struct device *dev = rx_ring->dev;
493 struct i40e_rx_buffer *rx_bi;
494 unsigned long bi_size;
497 /* ring already cleared, nothing to do */
501 if (ring_is_ps_enabled(rx_ring)) {
502 int bufsz = ALIGN(rx_ring->rx_hdr_len, 256) * rx_ring->count;
504 rx_bi = &rx_ring->rx_bi[0];
505 if (rx_bi->hdr_buf) {
506 dma_free_coherent(dev,
510 for (i = 0; i < rx_ring->count; i++) {
511 rx_bi = &rx_ring->rx_bi[i];
513 rx_bi->hdr_buf = NULL;
517 /* Free all the Rx ring sk_buffs */
518 for (i = 0; i < rx_ring->count; i++) {
519 rx_bi = &rx_ring->rx_bi[i];
521 dma_unmap_single(dev,
528 dev_kfree_skb(rx_bi->skb);
532 if (rx_bi->page_dma) {
539 __free_page(rx_bi->page);
541 rx_bi->page_offset = 0;
545 bi_size = sizeof(struct i40e_rx_buffer) * rx_ring->count;
546 memset(rx_ring->rx_bi, 0, bi_size);
548 /* Zero out the descriptor ring */
549 memset(rx_ring->desc, 0, rx_ring->size);
551 rx_ring->next_to_clean = 0;
552 rx_ring->next_to_use = 0;
556 * i40evf_free_rx_resources - Free Rx resources
557 * @rx_ring: ring to clean the resources from
559 * Free all receive software resources
561 void i40evf_free_rx_resources(struct i40e_ring *rx_ring)
563 i40evf_clean_rx_ring(rx_ring);
564 kfree(rx_ring->rx_bi);
565 rx_ring->rx_bi = NULL;
568 dma_free_coherent(rx_ring->dev, rx_ring->size,
569 rx_ring->desc, rx_ring->dma);
570 rx_ring->desc = NULL;
575 * i40evf_alloc_rx_headers - allocate rx header buffers
576 * @rx_ring: ring to alloc buffers
578 * Allocate rx header buffers for the entire ring. As these are static,
579 * this is only called when setting up a new ring.
581 void i40evf_alloc_rx_headers(struct i40e_ring *rx_ring)
583 struct device *dev = rx_ring->dev;
584 struct i40e_rx_buffer *rx_bi;
590 if (rx_ring->rx_bi[0].hdr_buf)
592 /* Make sure the buffers don't cross cache line boundaries. */
593 buf_size = ALIGN(rx_ring->rx_hdr_len, 256);
594 buffer = dma_alloc_coherent(dev, buf_size * rx_ring->count,
598 for (i = 0; i < rx_ring->count; i++) {
599 rx_bi = &rx_ring->rx_bi[i];
600 rx_bi->dma = dma + (i * buf_size);
601 rx_bi->hdr_buf = buffer + (i * buf_size);
606 * i40evf_setup_rx_descriptors - Allocate Rx descriptors
607 * @rx_ring: Rx descriptor ring (for a specific queue) to setup
609 * Returns 0 on success, negative on failure
611 int i40evf_setup_rx_descriptors(struct i40e_ring *rx_ring)
613 struct device *dev = rx_ring->dev;
616 /* warn if we are about to overwrite the pointer */
617 WARN_ON(rx_ring->rx_bi);
618 bi_size = sizeof(struct i40e_rx_buffer) * rx_ring->count;
619 rx_ring->rx_bi = kzalloc(bi_size, GFP_KERNEL);
623 u64_stats_init(&rx_ring->syncp);
625 /* Round up to nearest 4K */
626 rx_ring->size = ring_is_16byte_desc_enabled(rx_ring)
627 ? rx_ring->count * sizeof(union i40e_16byte_rx_desc)
628 : rx_ring->count * sizeof(union i40e_32byte_rx_desc);
629 rx_ring->size = ALIGN(rx_ring->size, 4096);
630 rx_ring->desc = dma_alloc_coherent(dev, rx_ring->size,
631 &rx_ring->dma, GFP_KERNEL);
633 if (!rx_ring->desc) {
634 dev_info(dev, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
639 rx_ring->next_to_clean = 0;
640 rx_ring->next_to_use = 0;
644 kfree(rx_ring->rx_bi);
645 rx_ring->rx_bi = NULL;
650 * i40e_release_rx_desc - Store the new tail and head values
651 * @rx_ring: ring to bump
652 * @val: new head index
654 static inline void i40e_release_rx_desc(struct i40e_ring *rx_ring, u32 val)
656 rx_ring->next_to_use = val;
657 /* Force memory writes to complete before letting h/w
658 * know there are new descriptors to fetch. (Only
659 * applicable for weak-ordered memory model archs,
663 writel(val, rx_ring->tail);
667 * i40evf_alloc_rx_buffers_ps - Replace used receive buffers; packet split
668 * @rx_ring: ring to place buffers on
669 * @cleaned_count: number of buffers to replace
671 * Returns true if any errors on allocation
673 bool i40evf_alloc_rx_buffers_ps(struct i40e_ring *rx_ring, u16 cleaned_count)
675 u16 i = rx_ring->next_to_use;
676 union i40e_rx_desc *rx_desc;
677 struct i40e_rx_buffer *bi;
678 const int current_node = numa_node_id();
680 /* do nothing if no valid netdev defined */
681 if (!rx_ring->netdev || !cleaned_count)
684 while (cleaned_count--) {
685 rx_desc = I40E_RX_DESC(rx_ring, i);
686 bi = &rx_ring->rx_bi[i];
688 if (bi->skb) /* desc is in use */
691 /* If we've been moved to a different NUMA node, release the
692 * page so we can get a new one on the current node.
694 if (bi->page && page_to_nid(bi->page) != current_node) {
695 dma_unmap_page(rx_ring->dev,
699 __free_page(bi->page);
702 rx_ring->rx_stats.realloc_count++;
703 } else if (bi->page) {
704 rx_ring->rx_stats.page_reuse_count++;
708 bi->page = alloc_page(GFP_ATOMIC);
710 rx_ring->rx_stats.alloc_page_failed++;
713 bi->page_dma = dma_map_page(rx_ring->dev,
718 if (dma_mapping_error(rx_ring->dev, bi->page_dma)) {
719 rx_ring->rx_stats.alloc_page_failed++;
720 __free_page(bi->page);
729 /* Refresh the desc even if buffer_addrs didn't change
730 * because each write-back erases this info.
732 rx_desc->read.pkt_addr =
733 cpu_to_le64(bi->page_dma + bi->page_offset);
734 rx_desc->read.hdr_addr = cpu_to_le64(bi->dma);
736 if (i == rx_ring->count)
740 if (rx_ring->next_to_use != i)
741 i40e_release_rx_desc(rx_ring, i);
746 if (rx_ring->next_to_use != i)
747 i40e_release_rx_desc(rx_ring, i);
749 /* make sure to come back via polling to try again after
756 * i40evf_alloc_rx_buffers_1buf - Replace used receive buffers; single buffer
757 * @rx_ring: ring to place buffers on
758 * @cleaned_count: number of buffers to replace
760 * Returns true if any errors on allocation
762 bool i40evf_alloc_rx_buffers_1buf(struct i40e_ring *rx_ring, u16 cleaned_count)
764 u16 i = rx_ring->next_to_use;
765 union i40e_rx_desc *rx_desc;
766 struct i40e_rx_buffer *bi;
769 /* do nothing if no valid netdev defined */
770 if (!rx_ring->netdev || !cleaned_count)
773 while (cleaned_count--) {
774 rx_desc = I40E_RX_DESC(rx_ring, i);
775 bi = &rx_ring->rx_bi[i];
779 skb = __netdev_alloc_skb_ip_align(rx_ring->netdev,
784 rx_ring->rx_stats.alloc_buff_failed++;
787 /* initialize queue mapping */
788 skb_record_rx_queue(skb, rx_ring->queue_index);
793 bi->dma = dma_map_single(rx_ring->dev,
797 if (dma_mapping_error(rx_ring->dev, bi->dma)) {
798 rx_ring->rx_stats.alloc_buff_failed++;
800 dev_kfree_skb(bi->skb);
806 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
807 rx_desc->read.hdr_addr = 0;
809 if (i == rx_ring->count)
813 if (rx_ring->next_to_use != i)
814 i40e_release_rx_desc(rx_ring, i);
819 if (rx_ring->next_to_use != i)
820 i40e_release_rx_desc(rx_ring, i);
822 /* make sure to come back via polling to try again after
829 * i40e_receive_skb - Send a completed packet up the stack
830 * @rx_ring: rx ring in play
831 * @skb: packet to send up
832 * @vlan_tag: vlan tag for packet
834 static void i40e_receive_skb(struct i40e_ring *rx_ring,
835 struct sk_buff *skb, u16 vlan_tag)
837 struct i40e_q_vector *q_vector = rx_ring->q_vector;
839 if (vlan_tag & VLAN_VID_MASK)
840 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
842 napi_gro_receive(&q_vector->napi, skb);
846 * i40e_rx_checksum - Indicate in skb if hw indicated a good cksum
847 * @vsi: the VSI we care about
848 * @skb: skb currently being received and modified
849 * @rx_status: status value of last descriptor in packet
850 * @rx_error: error value of last descriptor in packet
851 * @rx_ptype: ptype value of last descriptor in packet
853 static inline void i40e_rx_checksum(struct i40e_vsi *vsi,
859 struct i40e_rx_ptype_decoded decoded = decode_rx_desc_ptype(rx_ptype);
860 bool ipv4 = false, ipv6 = false;
861 bool ipv4_tunnel, ipv6_tunnel;
866 ipv4_tunnel = (rx_ptype >= I40E_RX_PTYPE_GRENAT4_MAC_PAY3) &&
867 (rx_ptype <= I40E_RX_PTYPE_GRENAT4_MACVLAN_IPV6_ICMP_PAY4);
868 ipv6_tunnel = (rx_ptype >= I40E_RX_PTYPE_GRENAT6_MAC_PAY3) &&
869 (rx_ptype <= I40E_RX_PTYPE_GRENAT6_MACVLAN_IPV6_ICMP_PAY4);
871 skb->ip_summed = CHECKSUM_NONE;
873 /* Rx csum enabled and ip headers found? */
874 if (!(vsi->netdev->features & NETIF_F_RXCSUM))
877 /* did the hardware decode the packet and checksum? */
878 if (!(rx_status & BIT(I40E_RX_DESC_STATUS_L3L4P_SHIFT)))
881 /* both known and outer_ip must be set for the below code to work */
882 if (!(decoded.known && decoded.outer_ip))
885 if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
886 decoded.outer_ip_ver == I40E_RX_PTYPE_OUTER_IPV4)
888 else if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
889 decoded.outer_ip_ver == I40E_RX_PTYPE_OUTER_IPV6)
893 (rx_error & (BIT(I40E_RX_DESC_ERROR_IPE_SHIFT) |
894 BIT(I40E_RX_DESC_ERROR_EIPE_SHIFT))))
897 /* likely incorrect csum if alternate IP extension headers found */
899 rx_status & BIT(I40E_RX_DESC_STATUS_IPV6EXADD_SHIFT))
900 /* don't increment checksum err here, non-fatal err */
903 /* there was some L4 error, count error and punt packet to the stack */
904 if (rx_error & BIT(I40E_RX_DESC_ERROR_L4E_SHIFT))
907 /* handle packets that were not able to be checksummed due
908 * to arrival speed, in this case the stack can compute
911 if (rx_error & BIT(I40E_RX_DESC_ERROR_PPRS_SHIFT))
914 /* If VXLAN traffic has an outer UDPv4 checksum we need to check
915 * it in the driver, hardware does not do it for us.
916 * Since L3L4P bit was set we assume a valid IHL value (>=5)
917 * so the total length of IPv4 header is IHL*4 bytes
918 * The UDP_0 bit *may* bet set if the *inner* header is UDP
921 skb->transport_header = skb->mac_header +
922 sizeof(struct ethhdr) +
923 (ip_hdr(skb)->ihl * 4);
925 /* Add 4 bytes for VLAN tagged packets */
926 skb->transport_header += (skb->protocol == htons(ETH_P_8021Q) ||
927 skb->protocol == htons(ETH_P_8021AD))
930 if ((ip_hdr(skb)->protocol == IPPROTO_UDP) &&
931 (udp_hdr(skb)->check != 0)) {
932 rx_udp_csum = udp_csum(skb);
934 csum = csum_tcpudp_magic(iph->saddr, iph->daddr,
936 skb_transport_offset(skb)),
937 IPPROTO_UDP, rx_udp_csum);
939 if (udp_hdr(skb)->check != csum)
942 } /* else its GRE and so no outer UDP header */
945 skb->ip_summed = CHECKSUM_UNNECESSARY;
946 skb->csum_level = ipv4_tunnel || ipv6_tunnel;
951 vsi->back->hw_csum_rx_error++;
955 * i40e_ptype_to_htype - get a hash type
956 * @ptype: the ptype value from the descriptor
958 * Returns a hash type to be used by skb_set_hash
960 static inline enum pkt_hash_types i40e_ptype_to_htype(u8 ptype)
962 struct i40e_rx_ptype_decoded decoded = decode_rx_desc_ptype(ptype);
965 return PKT_HASH_TYPE_NONE;
967 if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
968 decoded.payload_layer == I40E_RX_PTYPE_PAYLOAD_LAYER_PAY4)
969 return PKT_HASH_TYPE_L4;
970 else if (decoded.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
971 decoded.payload_layer == I40E_RX_PTYPE_PAYLOAD_LAYER_PAY3)
972 return PKT_HASH_TYPE_L3;
974 return PKT_HASH_TYPE_L2;
978 * i40e_rx_hash - set the hash value in the skb
979 * @ring: descriptor ring
980 * @rx_desc: specific descriptor
982 static inline void i40e_rx_hash(struct i40e_ring *ring,
983 union i40e_rx_desc *rx_desc,
988 const __le64 rss_mask =
989 cpu_to_le64((u64)I40E_RX_DESC_FLTSTAT_RSS_HASH <<
990 I40E_RX_DESC_STATUS_FLTSTAT_SHIFT);
992 if (ring->netdev->features & NETIF_F_RXHASH)
995 if ((rx_desc->wb.qword1.status_error_len & rss_mask) == rss_mask) {
996 hash = le32_to_cpu(rx_desc->wb.qword0.hi_dword.rss);
997 skb_set_hash(skb, hash, i40e_ptype_to_htype(rx_ptype));
1002 * i40e_clean_rx_irq_ps - Reclaim resources after receive; packet split
1003 * @rx_ring: rx ring to clean
1004 * @budget: how many cleans we're allowed
1006 * Returns true if there's any budget left (e.g. the clean is finished)
1008 static int i40e_clean_rx_irq_ps(struct i40e_ring *rx_ring, const int budget)
1010 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
1011 u16 rx_packet_len, rx_header_len, rx_sph, rx_hbo;
1012 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
1013 struct i40e_vsi *vsi = rx_ring->vsi;
1014 u16 i = rx_ring->next_to_clean;
1015 union i40e_rx_desc *rx_desc;
1016 u32 rx_error, rx_status;
1017 bool failure = false;
1023 struct i40e_rx_buffer *rx_bi;
1024 struct sk_buff *skb;
1026 /* return some buffers to hardware, one at a time is too slow */
1027 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
1028 failure = failure ||
1029 i40evf_alloc_rx_buffers_ps(rx_ring,
1034 i = rx_ring->next_to_clean;
1035 rx_desc = I40E_RX_DESC(rx_ring, i);
1036 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
1037 rx_status = (qword & I40E_RXD_QW1_STATUS_MASK) >>
1038 I40E_RXD_QW1_STATUS_SHIFT;
1040 if (!(rx_status & BIT(I40E_RX_DESC_STATUS_DD_SHIFT)))
1043 /* This memory barrier is needed to keep us from reading
1044 * any other fields out of the rx_desc until we know the
1048 /* sync header buffer for reading */
1049 dma_sync_single_range_for_cpu(rx_ring->dev,
1050 rx_ring->rx_bi[0].dma,
1051 i * rx_ring->rx_hdr_len,
1052 rx_ring->rx_hdr_len,
1054 rx_bi = &rx_ring->rx_bi[i];
1057 skb = __netdev_alloc_skb_ip_align(rx_ring->netdev,
1058 rx_ring->rx_hdr_len,
1062 rx_ring->rx_stats.alloc_buff_failed++;
1067 /* initialize queue mapping */
1068 skb_record_rx_queue(skb, rx_ring->queue_index);
1069 /* we are reusing so sync this buffer for CPU use */
1070 dma_sync_single_range_for_cpu(rx_ring->dev,
1071 rx_ring->rx_bi[0].dma,
1072 i * rx_ring->rx_hdr_len,
1073 rx_ring->rx_hdr_len,
1076 rx_packet_len = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
1077 I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
1078 rx_header_len = (qword & I40E_RXD_QW1_LENGTH_HBUF_MASK) >>
1079 I40E_RXD_QW1_LENGTH_HBUF_SHIFT;
1080 rx_sph = (qword & I40E_RXD_QW1_LENGTH_SPH_MASK) >>
1081 I40E_RXD_QW1_LENGTH_SPH_SHIFT;
1083 rx_error = (qword & I40E_RXD_QW1_ERROR_MASK) >>
1084 I40E_RXD_QW1_ERROR_SHIFT;
1085 rx_hbo = rx_error & BIT(I40E_RX_DESC_ERROR_HBO_SHIFT);
1086 rx_error &= ~BIT(I40E_RX_DESC_ERROR_HBO_SHIFT);
1088 rx_ptype = (qword & I40E_RXD_QW1_PTYPE_MASK) >>
1089 I40E_RXD_QW1_PTYPE_SHIFT;
1090 /* sync half-page for reading */
1091 dma_sync_single_range_for_cpu(rx_ring->dev,
1096 prefetch(page_address(rx_bi->page) + rx_bi->page_offset);
1100 if (rx_hbo || rx_sph) {
1104 len = I40E_RX_HDR_SIZE;
1106 len = rx_header_len;
1107 memcpy(__skb_put(skb, len), rx_bi->hdr_buf, len);
1108 } else if (skb->len == 0) {
1110 unsigned char *va = page_address(rx_bi->page) +
1113 len = min(rx_packet_len, rx_ring->rx_hdr_len);
1114 memcpy(__skb_put(skb, len), va, len);
1116 rx_packet_len -= len;
1118 /* Get the rest of the data if this was a header split */
1119 if (rx_packet_len) {
1120 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
1122 rx_bi->page_offset + copysize,
1123 rx_packet_len, I40E_RXBUFFER_2048);
1125 get_page(rx_bi->page);
1126 /* switch to the other half-page here; the allocation
1127 * code programs the right addr into HW. If we haven't
1128 * used this half-page, the address won't be changed,
1129 * and HW can just use it next time through.
1131 rx_bi->page_offset ^= PAGE_SIZE / 2;
1132 /* If the page count is more than 2, then both halves
1133 * of the page are used and we need to free it. Do it
1134 * here instead of in the alloc code. Otherwise one
1135 * of the half-pages might be released between now and
1136 * then, and we wouldn't know which one to use.
1138 if (page_count(rx_bi->page) > 2) {
1139 dma_unmap_page(rx_ring->dev,
1143 __free_page(rx_bi->page);
1145 rx_bi->page_dma = 0;
1146 rx_ring->rx_stats.realloc_count++;
1150 I40E_RX_INCREMENT(rx_ring, i);
1153 !(rx_status & BIT(I40E_RX_DESC_STATUS_EOF_SHIFT)))) {
1154 struct i40e_rx_buffer *next_buffer;
1156 next_buffer = &rx_ring->rx_bi[i];
1157 next_buffer->skb = skb;
1158 rx_ring->rx_stats.non_eop_descs++;
1162 /* ERR_MASK will only have valid bits if EOP set */
1163 if (unlikely(rx_error & BIT(I40E_RX_DESC_ERROR_RXE_SHIFT))) {
1164 dev_kfree_skb_any(skb);
1168 i40e_rx_hash(rx_ring, rx_desc, skb, rx_ptype);
1170 /* probably a little skewed due to removing CRC */
1171 total_rx_bytes += skb->len;
1174 skb->protocol = eth_type_trans(skb, rx_ring->netdev);
1176 i40e_rx_checksum(vsi, skb, rx_status, rx_error, rx_ptype);
1178 vlan_tag = rx_status & BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)
1179 ? le16_to_cpu(rx_desc->wb.qword0.lo_dword.l2tag1)
1182 if (!i40e_fcoe_handle_offload(rx_ring, rx_desc, skb)) {
1183 dev_kfree_skb_any(skb);
1187 i40e_receive_skb(rx_ring, skb, vlan_tag);
1189 rx_desc->wb.qword1.status_error_len = 0;
1191 } while (likely(total_rx_packets < budget));
1193 u64_stats_update_begin(&rx_ring->syncp);
1194 rx_ring->stats.packets += total_rx_packets;
1195 rx_ring->stats.bytes += total_rx_bytes;
1196 u64_stats_update_end(&rx_ring->syncp);
1197 rx_ring->q_vector->rx.total_packets += total_rx_packets;
1198 rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
1200 return failure ? budget : total_rx_packets;
1204 * i40e_clean_rx_irq_1buf - Reclaim resources after receive; single buffer
1205 * @rx_ring: rx ring to clean
1206 * @budget: how many cleans we're allowed
1208 * Returns number of packets cleaned
1210 static int i40e_clean_rx_irq_1buf(struct i40e_ring *rx_ring, int budget)
1212 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
1213 u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
1214 struct i40e_vsi *vsi = rx_ring->vsi;
1215 union i40e_rx_desc *rx_desc;
1216 u32 rx_error, rx_status;
1218 bool failure = false;
1224 struct i40e_rx_buffer *rx_bi;
1225 struct sk_buff *skb;
1227 /* return some buffers to hardware, one at a time is too slow */
1228 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
1229 failure = failure ||
1230 i40evf_alloc_rx_buffers_1buf(rx_ring,
1235 i = rx_ring->next_to_clean;
1236 rx_desc = I40E_RX_DESC(rx_ring, i);
1237 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
1238 rx_status = (qword & I40E_RXD_QW1_STATUS_MASK) >>
1239 I40E_RXD_QW1_STATUS_SHIFT;
1241 if (!(rx_status & BIT(I40E_RX_DESC_STATUS_DD_SHIFT)))
1244 /* This memory barrier is needed to keep us from reading
1245 * any other fields out of the rx_desc until we know the
1250 rx_bi = &rx_ring->rx_bi[i];
1252 prefetch(skb->data);
1254 rx_packet_len = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
1255 I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
1257 rx_error = (qword & I40E_RXD_QW1_ERROR_MASK) >>
1258 I40E_RXD_QW1_ERROR_SHIFT;
1259 rx_error &= ~BIT(I40E_RX_DESC_ERROR_HBO_SHIFT);
1261 rx_ptype = (qword & I40E_RXD_QW1_PTYPE_MASK) >>
1262 I40E_RXD_QW1_PTYPE_SHIFT;
1266 /* Get the header and possibly the whole packet
1267 * If this is an skb from previous receive dma will be 0
1269 skb_put(skb, rx_packet_len);
1270 dma_unmap_single(rx_ring->dev, rx_bi->dma, rx_ring->rx_buf_len,
1274 I40E_RX_INCREMENT(rx_ring, i);
1277 !(rx_status & BIT(I40E_RX_DESC_STATUS_EOF_SHIFT)))) {
1278 rx_ring->rx_stats.non_eop_descs++;
1282 /* ERR_MASK will only have valid bits if EOP set */
1283 if (unlikely(rx_error & BIT(I40E_RX_DESC_ERROR_RXE_SHIFT))) {
1284 dev_kfree_skb_any(skb);
1288 i40e_rx_hash(rx_ring, rx_desc, skb, rx_ptype);
1289 /* probably a little skewed due to removing CRC */
1290 total_rx_bytes += skb->len;
1293 skb->protocol = eth_type_trans(skb, rx_ring->netdev);
1295 i40e_rx_checksum(vsi, skb, rx_status, rx_error, rx_ptype);
1297 vlan_tag = rx_status & BIT(I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)
1298 ? le16_to_cpu(rx_desc->wb.qword0.lo_dword.l2tag1)
1300 i40e_receive_skb(rx_ring, skb, vlan_tag);
1302 rx_desc->wb.qword1.status_error_len = 0;
1303 } while (likely(total_rx_packets < budget));
1305 u64_stats_update_begin(&rx_ring->syncp);
1306 rx_ring->stats.packets += total_rx_packets;
1307 rx_ring->stats.bytes += total_rx_bytes;
1308 u64_stats_update_end(&rx_ring->syncp);
1309 rx_ring->q_vector->rx.total_packets += total_rx_packets;
1310 rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
1312 return failure ? budget : total_rx_packets;
1315 static u32 i40e_buildreg_itr(const int type, const u16 itr)
1319 val = I40E_VFINT_DYN_CTLN1_INTENA_MASK |
1320 /* Don't clear PBA because that can cause lost interrupts that
1321 * came in while we were cleaning/polling
1323 (type << I40E_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) |
1324 (itr << I40E_VFINT_DYN_CTLN1_INTERVAL_SHIFT);
1329 /* a small macro to shorten up some long lines */
1330 #define INTREG I40E_VFINT_DYN_CTLN1
1333 * i40e_update_enable_itr - Update itr and re-enable MSIX interrupt
1334 * @vsi: the VSI we care about
1335 * @q_vector: q_vector for which itr is being updated and interrupt enabled
1338 static inline void i40e_update_enable_itr(struct i40e_vsi *vsi,
1339 struct i40e_q_vector *q_vector)
1341 struct i40e_hw *hw = &vsi->back->hw;
1342 bool rx = false, tx = false;
1346 vector = (q_vector->v_idx + vsi->base_vector);
1348 /* avoid dynamic calculation if in countdown mode OR if
1349 * all dynamic is disabled
1351 rxval = txval = i40e_buildreg_itr(I40E_ITR_NONE, 0);
1353 if (q_vector->itr_countdown > 0 ||
1354 (!ITR_IS_DYNAMIC(vsi->rx_itr_setting) &&
1355 !ITR_IS_DYNAMIC(vsi->tx_itr_setting))) {
1359 if (ITR_IS_DYNAMIC(vsi->rx_itr_setting)) {
1360 rx = i40e_set_new_dynamic_itr(&q_vector->rx);
1361 rxval = i40e_buildreg_itr(I40E_RX_ITR, q_vector->rx.itr);
1364 if (ITR_IS_DYNAMIC(vsi->tx_itr_setting)) {
1365 tx = i40e_set_new_dynamic_itr(&q_vector->tx);
1366 txval = i40e_buildreg_itr(I40E_TX_ITR, q_vector->tx.itr);
1370 /* get the higher of the two ITR adjustments and
1371 * use the same value for both ITR registers
1372 * when in adaptive mode (Rx and/or Tx)
1374 u16 itr = max(q_vector->tx.itr, q_vector->rx.itr);
1376 q_vector->tx.itr = q_vector->rx.itr = itr;
1377 txval = i40e_buildreg_itr(I40E_TX_ITR, itr);
1379 rxval = i40e_buildreg_itr(I40E_RX_ITR, itr);
1383 /* only need to enable the interrupt once, but need
1384 * to possibly update both ITR values
1387 /* set the INTENA_MSK_MASK so that this first write
1388 * won't actually enable the interrupt, instead just
1389 * updating the ITR (it's bit 31 PF and VF)
1392 /* don't check _DOWN because interrupt isn't being enabled */
1393 wr32(hw, INTREG(vector - 1), rxval);
1397 if (!test_bit(__I40E_DOWN, &vsi->state))
1398 wr32(hw, INTREG(vector - 1), txval);
1400 if (q_vector->itr_countdown)
1401 q_vector->itr_countdown--;
1403 q_vector->itr_countdown = ITR_COUNTDOWN_START;
1407 * i40evf_napi_poll - NAPI polling Rx/Tx cleanup routine
1408 * @napi: napi struct with our devices info in it
1409 * @budget: amount of work driver is allowed to do this pass, in packets
1411 * This function will clean all queues associated with a q_vector.
1413 * Returns the amount of work done
1415 int i40evf_napi_poll(struct napi_struct *napi, int budget)
1417 struct i40e_q_vector *q_vector =
1418 container_of(napi, struct i40e_q_vector, napi);
1419 struct i40e_vsi *vsi = q_vector->vsi;
1420 struct i40e_ring *ring;
1421 bool clean_complete = true;
1422 bool arm_wb = false;
1423 int budget_per_ring;
1426 if (test_bit(__I40E_DOWN, &vsi->state)) {
1427 napi_complete(napi);
1431 /* Since the actual Tx work is minimal, we can give the Tx a larger
1432 * budget and be more aggressive about cleaning up the Tx descriptors.
1434 i40e_for_each_ring(ring, q_vector->tx) {
1435 clean_complete &= i40e_clean_tx_irq(ring, vsi->work_limit);
1436 arm_wb = arm_wb || ring->arm_wb;
1437 ring->arm_wb = false;
1440 /* Handle case where we are called by netpoll with a budget of 0 */
1444 /* We attempt to distribute budget to each Rx queue fairly, but don't
1445 * allow the budget to go below 1 because that would exit polling early.
1447 budget_per_ring = max(budget/q_vector->num_ringpairs, 1);
1449 i40e_for_each_ring(ring, q_vector->rx) {
1452 if (ring_is_ps_enabled(ring))
1453 cleaned = i40e_clean_rx_irq_ps(ring, budget_per_ring);
1455 cleaned = i40e_clean_rx_irq_1buf(ring, budget_per_ring);
1457 work_done += cleaned;
1458 /* if we didn't clean as many as budgeted, we must be done */
1459 clean_complete &= (budget_per_ring != cleaned);
1462 /* If work not completed, return budget and polling will return */
1463 if (!clean_complete) {
1466 q_vector->tx.ring[0].tx_stats.tx_force_wb++;
1467 i40e_enable_wb_on_itr(vsi, q_vector);
1472 if (vsi->back->flags & I40E_TXR_FLAGS_WB_ON_ITR)
1473 q_vector->arm_wb_state = false;
1475 /* Work is done so exit the polling mode and re-enable the interrupt */
1476 napi_complete_done(napi, work_done);
1477 i40e_update_enable_itr(vsi, q_vector);
1482 * i40evf_tx_prepare_vlan_flags - prepare generic TX VLAN tagging flags for HW
1484 * @tx_ring: ring to send buffer on
1485 * @flags: the tx flags to be set
1487 * Checks the skb and set up correspondingly several generic transmit flags
1488 * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
1490 * Returns error code indicate the frame should be dropped upon error and the
1491 * otherwise returns 0 to indicate the flags has been set properly.
1493 static inline int i40evf_tx_prepare_vlan_flags(struct sk_buff *skb,
1494 struct i40e_ring *tx_ring,
1497 __be16 protocol = skb->protocol;
1500 if (protocol == htons(ETH_P_8021Q) &&
1501 !(tx_ring->netdev->features & NETIF_F_HW_VLAN_CTAG_TX)) {
1502 /* When HW VLAN acceleration is turned off by the user the
1503 * stack sets the protocol to 8021q so that the driver
1504 * can take any steps required to support the SW only
1505 * VLAN handling. In our case the driver doesn't need
1506 * to take any further steps so just set the protocol
1507 * to the encapsulated ethertype.
1509 skb->protocol = vlan_get_protocol(skb);
1513 /* if we have a HW VLAN tag being added, default to the HW one */
1514 if (skb_vlan_tag_present(skb)) {
1515 tx_flags |= skb_vlan_tag_get(skb) << I40E_TX_FLAGS_VLAN_SHIFT;
1516 tx_flags |= I40E_TX_FLAGS_HW_VLAN;
1517 /* else if it is a SW VLAN, check the next protocol and store the tag */
1518 } else if (protocol == htons(ETH_P_8021Q)) {
1519 struct vlan_hdr *vhdr, _vhdr;
1521 vhdr = skb_header_pointer(skb, ETH_HLEN, sizeof(_vhdr), &_vhdr);
1525 protocol = vhdr->h_vlan_encapsulated_proto;
1526 tx_flags |= ntohs(vhdr->h_vlan_TCI) << I40E_TX_FLAGS_VLAN_SHIFT;
1527 tx_flags |= I40E_TX_FLAGS_SW_VLAN;
1536 * i40e_tso - set up the tso context descriptor
1537 * @tx_ring: ptr to the ring to send
1538 * @skb: ptr to the skb we're sending
1539 * @hdr_len: ptr to the size of the packet header
1540 * @cd_type_cmd_tso_mss: Quad Word 1
1542 * Returns 0 if no TSO can happen, 1 if tso is going, or error
1544 static int i40e_tso(struct i40e_ring *tx_ring, struct sk_buff *skb,
1545 u8 *hdr_len, u64 *cd_type_cmd_tso_mss)
1547 u32 cd_cmd, cd_tso_len, cd_mss;
1548 struct ipv6hdr *ipv6h;
1549 struct tcphdr *tcph;
1554 if (skb->ip_summed != CHECKSUM_PARTIAL)
1557 if (!skb_is_gso(skb))
1560 err = skb_cow_head(skb, 0);
1564 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
1565 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
1567 if (iph->version == 4) {
1568 tcph = skb->encapsulation ? inner_tcp_hdr(skb) : tcp_hdr(skb);
1571 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1573 } else if (ipv6h->version == 6) {
1574 tcph = skb->encapsulation ? inner_tcp_hdr(skb) : tcp_hdr(skb);
1575 ipv6h->payload_len = 0;
1576 tcph->check = ~csum_ipv6_magic(&ipv6h->saddr, &ipv6h->daddr,
1580 l4len = skb->encapsulation ? inner_tcp_hdrlen(skb) : tcp_hdrlen(skb);
1581 *hdr_len = (skb->encapsulation
1582 ? (skb_inner_transport_header(skb) - skb->data)
1583 : skb_transport_offset(skb)) + l4len;
1585 /* find the field values */
1586 cd_cmd = I40E_TX_CTX_DESC_TSO;
1587 cd_tso_len = skb->len - *hdr_len;
1588 cd_mss = skb_shinfo(skb)->gso_size;
1589 *cd_type_cmd_tso_mss |= ((u64)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
1591 I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
1592 ((u64)cd_mss << I40E_TXD_CTX_QW1_MSS_SHIFT);
1597 * i40e_tx_enable_csum - Enable Tx checksum offloads
1599 * @tx_flags: pointer to Tx flags currently set
1600 * @td_cmd: Tx descriptor command bits to set
1601 * @td_offset: Tx descriptor header offsets to set
1602 * @cd_tunneling: ptr to context desc bits
1604 static void i40e_tx_enable_csum(struct sk_buff *skb, u32 *tx_flags,
1605 u32 *td_cmd, u32 *td_offset,
1606 struct i40e_ring *tx_ring,
1609 struct ipv6hdr *this_ipv6_hdr;
1610 unsigned int this_tcp_hdrlen;
1611 struct iphdr *this_ip_hdr;
1612 u32 network_hdr_len;
1614 struct udphdr *oudph;
1618 if (skb->encapsulation) {
1619 switch (ip_hdr(skb)->protocol) {
1621 oudph = udp_hdr(skb);
1623 l4_tunnel = I40E_TXD_CTX_UDP_TUNNELING;
1624 *tx_flags |= I40E_TX_FLAGS_VXLAN_TUNNEL;
1629 network_hdr_len = skb_inner_network_header_len(skb);
1630 this_ip_hdr = inner_ip_hdr(skb);
1631 this_ipv6_hdr = inner_ipv6_hdr(skb);
1632 this_tcp_hdrlen = inner_tcp_hdrlen(skb);
1634 if (*tx_flags & I40E_TX_FLAGS_IPV4) {
1635 if (*tx_flags & I40E_TX_FLAGS_TSO) {
1636 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
1637 ip_hdr(skb)->check = 0;
1640 I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
1642 } else if (*tx_flags & I40E_TX_FLAGS_IPV6) {
1643 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
1644 if (*tx_flags & I40E_TX_FLAGS_TSO)
1645 ip_hdr(skb)->check = 0;
1648 /* Now set the ctx descriptor fields */
1649 *cd_tunneling |= (skb_network_header_len(skb) >> 2) <<
1650 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT |
1652 ((skb_inner_network_offset(skb) -
1653 skb_transport_offset(skb)) >> 1) <<
1654 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
1655 if (this_ip_hdr->version == 6) {
1656 *tx_flags &= ~I40E_TX_FLAGS_IPV4;
1657 *tx_flags |= I40E_TX_FLAGS_IPV6;
1660 if ((tx_ring->flags & I40E_TXR_FLAGS_OUTER_UDP_CSUM) &&
1661 (l4_tunnel == I40E_TXD_CTX_UDP_TUNNELING) &&
1662 (*cd_tunneling & I40E_TXD_CTX_QW0_EXT_IP_MASK)) {
1663 oudph->check = ~csum_tcpudp_magic(oiph->saddr,
1665 (skb->len - skb_transport_offset(skb)),
1667 *cd_tunneling |= I40E_TXD_CTX_QW0_L4T_CS_MASK;
1670 network_hdr_len = skb_network_header_len(skb);
1671 this_ip_hdr = ip_hdr(skb);
1672 this_ipv6_hdr = ipv6_hdr(skb);
1673 this_tcp_hdrlen = tcp_hdrlen(skb);
1676 /* Enable IP checksum offloads */
1677 if (*tx_flags & I40E_TX_FLAGS_IPV4) {
1678 l4_hdr = this_ip_hdr->protocol;
1679 /* the stack computes the IP header already, the only time we
1680 * need the hardware to recompute it is in the case of TSO.
1682 if (*tx_flags & I40E_TX_FLAGS_TSO) {
1683 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
1684 this_ip_hdr->check = 0;
1686 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
1688 /* Now set the td_offset for IP header length */
1689 *td_offset = (network_hdr_len >> 2) <<
1690 I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
1691 } else if (*tx_flags & I40E_TX_FLAGS_IPV6) {
1692 l4_hdr = this_ipv6_hdr->nexthdr;
1693 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
1694 /* Now set the td_offset for IP header length */
1695 *td_offset = (network_hdr_len >> 2) <<
1696 I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
1698 /* words in MACLEN + dwords in IPLEN + dwords in L4Len */
1699 *td_offset |= (skb_network_offset(skb) >> 1) <<
1700 I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
1702 /* Enable L4 checksum offloads */
1705 /* enable checksum offloads */
1706 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
1707 *td_offset |= (this_tcp_hdrlen >> 2) <<
1708 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
1711 /* enable SCTP checksum offload */
1712 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
1713 *td_offset |= (sizeof(struct sctphdr) >> 2) <<
1714 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
1717 /* enable UDP checksum offload */
1718 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
1719 *td_offset |= (sizeof(struct udphdr) >> 2) <<
1720 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
1728 * i40e_create_tx_ctx Build the Tx context descriptor
1729 * @tx_ring: ring to create the descriptor on
1730 * @cd_type_cmd_tso_mss: Quad Word 1
1731 * @cd_tunneling: Quad Word 0 - bits 0-31
1732 * @cd_l2tag2: Quad Word 0 - bits 32-63
1734 static void i40e_create_tx_ctx(struct i40e_ring *tx_ring,
1735 const u64 cd_type_cmd_tso_mss,
1736 const u32 cd_tunneling, const u32 cd_l2tag2)
1738 struct i40e_tx_context_desc *context_desc;
1739 int i = tx_ring->next_to_use;
1741 if ((cd_type_cmd_tso_mss == I40E_TX_DESC_DTYPE_CONTEXT) &&
1742 !cd_tunneling && !cd_l2tag2)
1745 /* grab the next descriptor */
1746 context_desc = I40E_TX_CTXTDESC(tx_ring, i);
1749 tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
1751 /* cpu_to_le32 and assign to struct fields */
1752 context_desc->tunneling_params = cpu_to_le32(cd_tunneling);
1753 context_desc->l2tag2 = cpu_to_le16(cd_l2tag2);
1754 context_desc->rsvd = cpu_to_le16(0);
1755 context_desc->type_cmd_tso_mss = cpu_to_le64(cd_type_cmd_tso_mss);
1759 * i40e_chk_linearize - Check if there are more than 8 fragments per packet
1761 * @tx_flags: collected send information
1763 * Note: Our HW can't scatter-gather more than 8 fragments to build
1764 * a packet on the wire and so we need to figure out the cases where we
1765 * need to linearize the skb.
1767 static bool i40e_chk_linearize(struct sk_buff *skb, u32 tx_flags)
1769 struct skb_frag_struct *frag;
1770 bool linearize = false;
1771 unsigned int size = 0;
1775 num_frags = skb_shinfo(skb)->nr_frags;
1776 gso_segs = skb_shinfo(skb)->gso_segs;
1778 if (tx_flags & (I40E_TX_FLAGS_TSO | I40E_TX_FLAGS_FSO)) {
1781 if (num_frags < (I40E_MAX_BUFFER_TXD))
1782 goto linearize_chk_done;
1783 /* try the simple math, if we have too many frags per segment */
1784 if (DIV_ROUND_UP((num_frags + gso_segs), gso_segs) >
1785 I40E_MAX_BUFFER_TXD) {
1787 goto linearize_chk_done;
1789 frag = &skb_shinfo(skb)->frags[0];
1790 /* we might still have more fragments per segment */
1792 size += skb_frag_size(frag);
1794 if ((size >= skb_shinfo(skb)->gso_size) &&
1795 (j < I40E_MAX_BUFFER_TXD)) {
1796 size = (size % skb_shinfo(skb)->gso_size);
1799 if (j == I40E_MAX_BUFFER_TXD) {
1804 } while (num_frags);
1806 if (num_frags >= I40E_MAX_BUFFER_TXD)
1815 * __i40evf_maybe_stop_tx - 2nd level check for tx stop conditions
1816 * @tx_ring: the ring to be checked
1817 * @size: the size buffer we want to assure is available
1819 * Returns -EBUSY if a stop is needed, else 0
1821 static inline int __i40evf_maybe_stop_tx(struct i40e_ring *tx_ring, int size)
1823 netif_stop_subqueue(tx_ring->netdev, tx_ring->queue_index);
1824 /* Memory barrier before checking head and tail */
1827 /* Check again in a case another CPU has just made room available. */
1828 if (likely(I40E_DESC_UNUSED(tx_ring) < size))
1831 /* A reprieve! - use start_queue because it doesn't call schedule */
1832 netif_start_subqueue(tx_ring->netdev, tx_ring->queue_index);
1833 ++tx_ring->tx_stats.restart_queue;
1838 * i40evf_maybe_stop_tx - 1st level check for tx stop conditions
1839 * @tx_ring: the ring to be checked
1840 * @size: the size buffer we want to assure is available
1842 * Returns 0 if stop is not needed
1844 static inline int i40evf_maybe_stop_tx(struct i40e_ring *tx_ring, int size)
1846 if (likely(I40E_DESC_UNUSED(tx_ring) >= size))
1848 return __i40evf_maybe_stop_tx(tx_ring, size);
1852 * i40evf_tx_map - Build the Tx descriptor
1853 * @tx_ring: ring to send buffer on
1855 * @first: first buffer info buffer to use
1856 * @tx_flags: collected send information
1857 * @hdr_len: size of the packet header
1858 * @td_cmd: the command field in the descriptor
1859 * @td_offset: offset for checksum or crc
1861 static inline void i40evf_tx_map(struct i40e_ring *tx_ring, struct sk_buff *skb,
1862 struct i40e_tx_buffer *first, u32 tx_flags,
1863 const u8 hdr_len, u32 td_cmd, u32 td_offset)
1865 unsigned int data_len = skb->data_len;
1866 unsigned int size = skb_headlen(skb);
1867 struct skb_frag_struct *frag;
1868 struct i40e_tx_buffer *tx_bi;
1869 struct i40e_tx_desc *tx_desc;
1870 u16 i = tx_ring->next_to_use;
1875 bool tail_bump = true;
1878 if (tx_flags & I40E_TX_FLAGS_HW_VLAN) {
1879 td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1880 td_tag = (tx_flags & I40E_TX_FLAGS_VLAN_MASK) >>
1881 I40E_TX_FLAGS_VLAN_SHIFT;
1884 if (tx_flags & (I40E_TX_FLAGS_TSO | I40E_TX_FLAGS_FSO))
1885 gso_segs = skb_shinfo(skb)->gso_segs;
1889 /* multiply data chunks by size of headers */
1890 first->bytecount = skb->len - hdr_len + (gso_segs * hdr_len);
1891 first->gso_segs = gso_segs;
1893 first->tx_flags = tx_flags;
1895 dma = dma_map_single(tx_ring->dev, skb->data, size, DMA_TO_DEVICE);
1897 tx_desc = I40E_TX_DESC(tx_ring, i);
1900 for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
1901 if (dma_mapping_error(tx_ring->dev, dma))
1904 /* record length, and DMA address */
1905 dma_unmap_len_set(tx_bi, len, size);
1906 dma_unmap_addr_set(tx_bi, dma, dma);
1908 tx_desc->buffer_addr = cpu_to_le64(dma);
1910 while (unlikely(size > I40E_MAX_DATA_PER_TXD)) {
1911 tx_desc->cmd_type_offset_bsz =
1912 build_ctob(td_cmd, td_offset,
1913 I40E_MAX_DATA_PER_TXD, td_tag);
1919 if (i == tx_ring->count) {
1920 tx_desc = I40E_TX_DESC(tx_ring, 0);
1924 dma += I40E_MAX_DATA_PER_TXD;
1925 size -= I40E_MAX_DATA_PER_TXD;
1927 tx_desc->buffer_addr = cpu_to_le64(dma);
1930 if (likely(!data_len))
1933 tx_desc->cmd_type_offset_bsz = build_ctob(td_cmd, td_offset,
1940 if (i == tx_ring->count) {
1941 tx_desc = I40E_TX_DESC(tx_ring, 0);
1945 size = skb_frag_size(frag);
1948 dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
1951 tx_bi = &tx_ring->tx_bi[i];
1954 /* set next_to_watch value indicating a packet is present */
1955 first->next_to_watch = tx_desc;
1958 if (i == tx_ring->count)
1961 tx_ring->next_to_use = i;
1963 netdev_tx_sent_queue(netdev_get_tx_queue(tx_ring->netdev,
1964 tx_ring->queue_index),
1966 i40evf_maybe_stop_tx(tx_ring, DESC_NEEDED);
1968 /* Algorithm to optimize tail and RS bit setting:
1969 * if xmit_more is supported
1970 * if xmit_more is true
1971 * do not update tail and do not mark RS bit.
1972 * if xmit_more is false and last xmit_more was false
1973 * if every packet spanned less than 4 desc
1974 * then set RS bit on 4th packet and update tail
1977 * update tail and set RS bit on every packet.
1978 * if xmit_more is false and last_xmit_more was true
1979 * update tail and set RS bit.
1981 * Optimization: wmb to be issued only in case of tail update.
1982 * Also optimize the Descriptor WB path for RS bit with the same
1985 * Note: If there are less than 4 packets
1986 * pending and interrupts were disabled the service task will
1987 * trigger a force WB.
1989 if (skb->xmit_more &&
1990 !netif_xmit_stopped(netdev_get_tx_queue(tx_ring->netdev,
1991 tx_ring->queue_index))) {
1992 tx_ring->flags |= I40E_TXR_FLAGS_LAST_XMIT_MORE_SET;
1994 } else if (!skb->xmit_more &&
1995 !netif_xmit_stopped(netdev_get_tx_queue(tx_ring->netdev,
1996 tx_ring->queue_index)) &&
1997 (!(tx_ring->flags & I40E_TXR_FLAGS_LAST_XMIT_MORE_SET)) &&
1998 (tx_ring->packet_stride < WB_STRIDE) &&
1999 (desc_count < WB_STRIDE)) {
2000 tx_ring->packet_stride++;
2002 tx_ring->packet_stride = 0;
2003 tx_ring->flags &= ~I40E_TXR_FLAGS_LAST_XMIT_MORE_SET;
2007 tx_ring->packet_stride = 0;
2009 tx_desc->cmd_type_offset_bsz =
2010 build_ctob(td_cmd, td_offset, size, td_tag) |
2011 cpu_to_le64((u64)(do_rs ? I40E_TXD_CMD :
2012 I40E_TX_DESC_CMD_EOP) <<
2013 I40E_TXD_QW1_CMD_SHIFT);
2015 /* notify HW of packet */
2017 prefetchw(tx_desc + 1);
2020 /* Force memory writes to complete before letting h/w
2021 * know there are new descriptors to fetch. (Only
2022 * applicable for weak-ordered memory model archs,
2026 writel(i, tx_ring->tail);
2032 dev_info(tx_ring->dev, "TX DMA map failed\n");
2034 /* clear dma mappings for failed tx_bi map */
2036 tx_bi = &tx_ring->tx_bi[i];
2037 i40e_unmap_and_free_tx_resource(tx_ring, tx_bi);
2045 tx_ring->next_to_use = i;
2049 * i40evf_xmit_descriptor_count - calculate number of tx descriptors needed
2051 * @tx_ring: ring to send buffer on
2053 * Returns number of data descriptors needed for this skb. Returns 0 to indicate
2054 * there is not enough descriptors available in this ring since we need at least
2057 static inline int i40evf_xmit_descriptor_count(struct sk_buff *skb,
2058 struct i40e_ring *tx_ring)
2063 /* need: 1 descriptor per page * PAGE_SIZE/I40E_MAX_DATA_PER_TXD,
2064 * + 1 desc for skb_head_len/I40E_MAX_DATA_PER_TXD,
2065 * + 4 desc gap to avoid the cache line where head is,
2066 * + 1 desc for context descriptor,
2067 * otherwise try next time
2069 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
2070 count += TXD_USE_COUNT(skb_shinfo(skb)->frags[f].size);
2072 count += TXD_USE_COUNT(skb_headlen(skb));
2073 if (i40evf_maybe_stop_tx(tx_ring, count + 4 + 1)) {
2074 tx_ring->tx_stats.tx_busy++;
2081 * i40e_xmit_frame_ring - Sends buffer on Tx ring
2083 * @tx_ring: ring to send buffer on
2085 * Returns NETDEV_TX_OK if sent, else an error code
2087 static netdev_tx_t i40e_xmit_frame_ring(struct sk_buff *skb,
2088 struct i40e_ring *tx_ring)
2090 u64 cd_type_cmd_tso_mss = I40E_TX_DESC_DTYPE_CONTEXT;
2091 u32 cd_tunneling = 0, cd_l2tag2 = 0;
2092 struct i40e_tx_buffer *first;
2100 /* prefetch the data, we'll need it later */
2101 prefetch(skb->data);
2103 if (0 == i40evf_xmit_descriptor_count(skb, tx_ring))
2104 return NETDEV_TX_BUSY;
2106 /* prepare the xmit flags */
2107 if (i40evf_tx_prepare_vlan_flags(skb, tx_ring, &tx_flags))
2110 /* obtain protocol of skb */
2111 protocol = vlan_get_protocol(skb);
2113 /* record the location of the first descriptor for this packet */
2114 first = &tx_ring->tx_bi[tx_ring->next_to_use];
2116 /* setup IPv4/IPv6 offloads */
2117 if (protocol == htons(ETH_P_IP))
2118 tx_flags |= I40E_TX_FLAGS_IPV4;
2119 else if (protocol == htons(ETH_P_IPV6))
2120 tx_flags |= I40E_TX_FLAGS_IPV6;
2122 tso = i40e_tso(tx_ring, skb, &hdr_len, &cd_type_cmd_tso_mss);
2127 tx_flags |= I40E_TX_FLAGS_TSO;
2129 if (i40e_chk_linearize(skb, tx_flags)) {
2130 if (skb_linearize(skb))
2132 tx_ring->tx_stats.tx_linearize++;
2134 skb_tx_timestamp(skb);
2136 /* always enable CRC insertion offload */
2137 td_cmd |= I40E_TX_DESC_CMD_ICRC;
2139 /* Always offload the checksum, since it's in the data descriptor */
2140 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2141 tx_flags |= I40E_TX_FLAGS_CSUM;
2143 i40e_tx_enable_csum(skb, &tx_flags, &td_cmd, &td_offset,
2144 tx_ring, &cd_tunneling);
2147 i40e_create_tx_ctx(tx_ring, cd_type_cmd_tso_mss,
2148 cd_tunneling, cd_l2tag2);
2150 i40evf_tx_map(tx_ring, skb, first, tx_flags, hdr_len,
2153 return NETDEV_TX_OK;
2156 dev_kfree_skb_any(skb);
2157 return NETDEV_TX_OK;
2161 * i40evf_xmit_frame - Selects the correct VSI and Tx queue to send buffer
2163 * @netdev: network interface device structure
2165 * Returns NETDEV_TX_OK if sent, else an error code
2167 netdev_tx_t i40evf_xmit_frame(struct sk_buff *skb, struct net_device *netdev)
2169 struct i40evf_adapter *adapter = netdev_priv(netdev);
2170 struct i40e_ring *tx_ring = &adapter->tx_rings[skb->queue_mapping];
2172 /* hardware can't handle really short frames, hardware padding works
2175 if (unlikely(skb->len < I40E_MIN_TX_LEN)) {
2176 if (skb_pad(skb, I40E_MIN_TX_LEN - skb->len))
2177 return NETDEV_TX_OK;
2178 skb->len = I40E_MIN_TX_LEN;
2179 skb_set_tail_pointer(skb, I40E_MIN_TX_LEN);
2182 return i40e_xmit_frame_ring(skb, tx_ring);