1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2018, Intel Corporation. */
4 /* The driver transmit and receive code */
6 #include <linux/prefetch.h>
10 #define ICE_RX_HDR_SIZE 256
13 * ice_unmap_and_free_tx_buf - Release a Tx buffer
14 * @ring: the ring that owns the buffer
15 * @tx_buf: the buffer to free
18 ice_unmap_and_free_tx_buf(struct ice_ring *ring, struct ice_tx_buf *tx_buf)
21 dev_kfree_skb_any(tx_buf->skb);
22 if (dma_unmap_len(tx_buf, len))
23 dma_unmap_single(ring->dev,
24 dma_unmap_addr(tx_buf, dma),
25 dma_unmap_len(tx_buf, len),
27 } else if (dma_unmap_len(tx_buf, len)) {
28 dma_unmap_page(ring->dev,
29 dma_unmap_addr(tx_buf, dma),
30 dma_unmap_len(tx_buf, len),
34 tx_buf->next_to_watch = NULL;
36 dma_unmap_len_set(tx_buf, len, 0);
37 /* tx_buf must be completely set up in the transmit path */
40 static struct netdev_queue *txring_txq(const struct ice_ring *ring)
42 return netdev_get_tx_queue(ring->netdev, ring->q_index);
46 * ice_clean_tx_ring - Free any empty Tx buffers
47 * @tx_ring: ring to be cleaned
49 void ice_clean_tx_ring(struct ice_ring *tx_ring)
54 /* ring already cleared, nothing to do */
58 /* Free all the Tx ring sk_bufss */
59 for (i = 0; i < tx_ring->count; i++)
60 ice_unmap_and_free_tx_buf(tx_ring, &tx_ring->tx_buf[i]);
62 size = sizeof(struct ice_tx_buf) * tx_ring->count;
63 memset(tx_ring->tx_buf, 0, size);
65 /* Zero out the descriptor ring */
66 memset(tx_ring->desc, 0, tx_ring->size);
68 tx_ring->next_to_use = 0;
69 tx_ring->next_to_clean = 0;
74 /* cleanup Tx queue statistics */
75 netdev_tx_reset_queue(txring_txq(tx_ring));
79 * ice_free_tx_ring - Free Tx resources per queue
80 * @tx_ring: Tx descriptor ring for a specific queue
82 * Free all transmit software resources
84 void ice_free_tx_ring(struct ice_ring *tx_ring)
86 ice_clean_tx_ring(tx_ring);
87 devm_kfree(tx_ring->dev, tx_ring->tx_buf);
88 tx_ring->tx_buf = NULL;
91 dmam_free_coherent(tx_ring->dev, tx_ring->size,
92 tx_ring->desc, tx_ring->dma);
98 * ice_clean_tx_irq - Reclaim resources after transmit completes
99 * @vsi: the VSI we care about
100 * @tx_ring: Tx ring to clean
101 * @napi_budget: Used to determine if we are in netpoll
103 * Returns true if there's any budget left (e.g. the clean is finished)
105 static bool ice_clean_tx_irq(struct ice_vsi *vsi, struct ice_ring *tx_ring,
108 unsigned int total_bytes = 0, total_pkts = 0;
109 unsigned int budget = vsi->work_lmt;
110 s16 i = tx_ring->next_to_clean;
111 struct ice_tx_desc *tx_desc;
112 struct ice_tx_buf *tx_buf;
114 tx_buf = &tx_ring->tx_buf[i];
115 tx_desc = ICE_TX_DESC(tx_ring, i);
119 struct ice_tx_desc *eop_desc = tx_buf->next_to_watch;
121 /* if next_to_watch is not set then there is no work pending */
125 smp_rmb(); /* prevent any other reads prior to eop_desc */
127 /* if the descriptor isn't done, no work yet to do */
128 if (!(eop_desc->cmd_type_offset_bsz &
129 cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
132 /* clear next_to_watch to prevent false hangs */
133 tx_buf->next_to_watch = NULL;
135 /* update the statistics for this packet */
136 total_bytes += tx_buf->bytecount;
137 total_pkts += tx_buf->gso_segs;
140 napi_consume_skb(tx_buf->skb, napi_budget);
142 /* unmap skb header data */
143 dma_unmap_single(tx_ring->dev,
144 dma_unmap_addr(tx_buf, dma),
145 dma_unmap_len(tx_buf, len),
148 /* clear tx_buf data */
150 dma_unmap_len_set(tx_buf, len, 0);
152 /* unmap remaining buffers */
153 while (tx_desc != eop_desc) {
159 tx_buf = tx_ring->tx_buf;
160 tx_desc = ICE_TX_DESC(tx_ring, 0);
163 /* unmap any remaining paged data */
164 if (dma_unmap_len(tx_buf, len)) {
165 dma_unmap_page(tx_ring->dev,
166 dma_unmap_addr(tx_buf, dma),
167 dma_unmap_len(tx_buf, len),
169 dma_unmap_len_set(tx_buf, len, 0);
173 /* move us one more past the eop_desc for start of next pkt */
179 tx_buf = tx_ring->tx_buf;
180 tx_desc = ICE_TX_DESC(tx_ring, 0);
185 /* update budget accounting */
187 } while (likely(budget));
190 tx_ring->next_to_clean = i;
191 u64_stats_update_begin(&tx_ring->syncp);
192 tx_ring->stats.bytes += total_bytes;
193 tx_ring->stats.pkts += total_pkts;
194 u64_stats_update_end(&tx_ring->syncp);
195 tx_ring->q_vector->tx.total_bytes += total_bytes;
196 tx_ring->q_vector->tx.total_pkts += total_pkts;
198 netdev_tx_completed_queue(txring_txq(tx_ring), total_pkts,
201 #define TX_WAKE_THRESHOLD ((s16)(DESC_NEEDED * 2))
202 if (unlikely(total_pkts && netif_carrier_ok(tx_ring->netdev) &&
203 (ICE_DESC_UNUSED(tx_ring) >= TX_WAKE_THRESHOLD))) {
204 /* Make sure that anybody stopping the queue after this
205 * sees the new next_to_clean.
208 if (__netif_subqueue_stopped(tx_ring->netdev,
210 !test_bit(__ICE_DOWN, vsi->state)) {
211 netif_wake_subqueue(tx_ring->netdev,
213 ++tx_ring->tx_stats.restart_q;
221 * ice_setup_tx_ring - Allocate the Tx descriptors
222 * @tx_ring: the tx ring to set up
224 * Return 0 on success, negative on error
226 int ice_setup_tx_ring(struct ice_ring *tx_ring)
228 struct device *dev = tx_ring->dev;
234 /* warn if we are about to overwrite the pointer */
235 WARN_ON(tx_ring->tx_buf);
236 bi_size = sizeof(struct ice_tx_buf) * tx_ring->count;
237 tx_ring->tx_buf = devm_kzalloc(dev, bi_size, GFP_KERNEL);
238 if (!tx_ring->tx_buf)
241 /* round up to nearest 4K */
242 tx_ring->size = tx_ring->count * sizeof(struct ice_tx_desc);
243 tx_ring->size = ALIGN(tx_ring->size, 4096);
244 tx_ring->desc = dmam_alloc_coherent(dev, tx_ring->size, &tx_ring->dma,
246 if (!tx_ring->desc) {
247 dev_err(dev, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
252 tx_ring->next_to_use = 0;
253 tx_ring->next_to_clean = 0;
254 tx_ring->tx_stats.prev_pkt = -1;
258 devm_kfree(dev, tx_ring->tx_buf);
259 tx_ring->tx_buf = NULL;
264 * ice_clean_rx_ring - Free Rx buffers
265 * @rx_ring: ring to be cleaned
267 void ice_clean_rx_ring(struct ice_ring *rx_ring)
269 struct device *dev = rx_ring->dev;
273 /* ring already cleared, nothing to do */
274 if (!rx_ring->rx_buf)
277 /* Free all the Rx ring sk_buffs */
278 for (i = 0; i < rx_ring->count; i++) {
279 struct ice_rx_buf *rx_buf = &rx_ring->rx_buf[i];
282 dev_kfree_skb(rx_buf->skb);
288 dma_unmap_page(dev, rx_buf->dma, PAGE_SIZE, DMA_FROM_DEVICE);
289 __free_pages(rx_buf->page, 0);
292 rx_buf->page_offset = 0;
295 size = sizeof(struct ice_rx_buf) * rx_ring->count;
296 memset(rx_ring->rx_buf, 0, size);
298 /* Zero out the descriptor ring */
299 memset(rx_ring->desc, 0, rx_ring->size);
301 rx_ring->next_to_alloc = 0;
302 rx_ring->next_to_clean = 0;
303 rx_ring->next_to_use = 0;
307 * ice_free_rx_ring - Free Rx resources
308 * @rx_ring: ring to clean the resources from
310 * Free all receive software resources
312 void ice_free_rx_ring(struct ice_ring *rx_ring)
314 ice_clean_rx_ring(rx_ring);
315 devm_kfree(rx_ring->dev, rx_ring->rx_buf);
316 rx_ring->rx_buf = NULL;
319 dmam_free_coherent(rx_ring->dev, rx_ring->size,
320 rx_ring->desc, rx_ring->dma);
321 rx_ring->desc = NULL;
326 * ice_setup_rx_ring - Allocate the Rx descriptors
327 * @rx_ring: the rx ring to set up
329 * Return 0 on success, negative on error
331 int ice_setup_rx_ring(struct ice_ring *rx_ring)
333 struct device *dev = rx_ring->dev;
339 /* warn if we are about to overwrite the pointer */
340 WARN_ON(rx_ring->rx_buf);
341 bi_size = sizeof(struct ice_rx_buf) * rx_ring->count;
342 rx_ring->rx_buf = devm_kzalloc(dev, bi_size, GFP_KERNEL);
343 if (!rx_ring->rx_buf)
346 /* round up to nearest 4K */
347 rx_ring->size = rx_ring->count * sizeof(union ice_32byte_rx_desc);
348 rx_ring->size = ALIGN(rx_ring->size, 4096);
349 rx_ring->desc = dmam_alloc_coherent(dev, rx_ring->size, &rx_ring->dma,
351 if (!rx_ring->desc) {
352 dev_err(dev, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
357 rx_ring->next_to_use = 0;
358 rx_ring->next_to_clean = 0;
362 devm_kfree(dev, rx_ring->rx_buf);
363 rx_ring->rx_buf = NULL;
368 * ice_release_rx_desc - Store the new tail and head values
369 * @rx_ring: ring to bump
370 * @val: new head index
372 static void ice_release_rx_desc(struct ice_ring *rx_ring, u32 val)
374 rx_ring->next_to_use = val;
376 /* update next to alloc since we have filled the ring */
377 rx_ring->next_to_alloc = val;
379 /* Force memory writes to complete before letting h/w
380 * know there are new descriptors to fetch. (Only
381 * applicable for weak-ordered memory model archs,
385 writel(val, rx_ring->tail);
389 * ice_alloc_mapped_page - recycle or make a new page
390 * @rx_ring: ring to use
391 * @bi: rx_buf struct to modify
393 * Returns true if the page was successfully allocated or
396 static bool ice_alloc_mapped_page(struct ice_ring *rx_ring,
397 struct ice_rx_buf *bi)
399 struct page *page = bi->page;
402 /* since we are recycling buffers we should seldom need to alloc */
404 rx_ring->rx_stats.page_reuse_count++;
408 /* alloc new page for storage */
409 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
410 if (unlikely(!page)) {
411 rx_ring->rx_stats.alloc_page_failed++;
415 /* map page for use */
416 dma = dma_map_page(rx_ring->dev, page, 0, PAGE_SIZE, DMA_FROM_DEVICE);
418 /* if mapping failed free memory back to system since
419 * there isn't much point in holding memory we can't use
421 if (dma_mapping_error(rx_ring->dev, dma)) {
422 __free_pages(page, 0);
423 rx_ring->rx_stats.alloc_page_failed++;
435 * ice_alloc_rx_bufs - Replace used receive buffers
436 * @rx_ring: ring to place buffers on
437 * @cleaned_count: number of buffers to replace
439 * Returns false if all allocations were successful, true if any fail
441 bool ice_alloc_rx_bufs(struct ice_ring *rx_ring, u16 cleaned_count)
443 union ice_32b_rx_flex_desc *rx_desc;
444 u16 ntu = rx_ring->next_to_use;
445 struct ice_rx_buf *bi;
447 /* do nothing if no valid netdev defined */
448 if (!rx_ring->netdev || !cleaned_count)
451 /* get the RX descriptor and buffer based on next_to_use */
452 rx_desc = ICE_RX_DESC(rx_ring, ntu);
453 bi = &rx_ring->rx_buf[ntu];
456 if (!ice_alloc_mapped_page(rx_ring, bi))
459 /* Refresh the desc even if buffer_addrs didn't change
460 * because each write-back erases this info.
462 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
467 if (unlikely(ntu == rx_ring->count)) {
468 rx_desc = ICE_RX_DESC(rx_ring, 0);
469 bi = rx_ring->rx_buf;
473 /* clear the status bits for the next_to_use descriptor */
474 rx_desc->wb.status_error0 = 0;
477 } while (cleaned_count);
479 if (rx_ring->next_to_use != ntu)
480 ice_release_rx_desc(rx_ring, ntu);
485 if (rx_ring->next_to_use != ntu)
486 ice_release_rx_desc(rx_ring, ntu);
488 /* make sure to come back via polling to try again after
495 * ice_page_is_reserved - check if reuse is possible
496 * @page: page struct to check
498 static bool ice_page_is_reserved(struct page *page)
500 return (page_to_nid(page) != numa_mem_id()) || page_is_pfmemalloc(page);
504 * ice_add_rx_frag - Add contents of Rx buffer to sk_buff
505 * @rx_buf: buffer containing page to add
506 * @rx_desc: descriptor containing length of buffer written by hardware
507 * @skb: sk_buf to place the data into
509 * This function will add the data contained in rx_buf->page to the skb.
510 * This is done either through a direct copy if the data in the buffer is
511 * less than the skb header size, otherwise it will just attach the page as
514 * The function will then update the page offset if necessary and return
515 * true if the buffer can be reused by the adapter.
517 static bool ice_add_rx_frag(struct ice_rx_buf *rx_buf,
518 union ice_32b_rx_flex_desc *rx_desc,
521 #if (PAGE_SIZE < 8192)
522 unsigned int truesize = ICE_RXBUF_2048;
524 unsigned int last_offset = PAGE_SIZE - ICE_RXBUF_2048;
525 unsigned int truesize;
526 #endif /* PAGE_SIZE < 8192) */
531 size = le16_to_cpu(rx_desc->wb.pkt_len) &
532 ICE_RX_FLX_DESC_PKT_LEN_M;
536 #if (PAGE_SIZE >= 8192)
537 truesize = ALIGN(size, L1_CACHE_BYTES);
538 #endif /* PAGE_SIZE >= 8192) */
540 /* will the data fit in the skb we allocated? if so, just
541 * copy it as it is pretty small anyway
543 if (size <= ICE_RX_HDR_SIZE && !skb_is_nonlinear(skb)) {
544 unsigned char *va = page_address(page) + rx_buf->page_offset;
546 memcpy(__skb_put(skb, size), va, ALIGN(size, sizeof(long)));
548 /* page is not reserved, we can reuse buffer as-is */
549 if (likely(!ice_page_is_reserved(page)))
552 /* this page cannot be reused so discard it */
553 __free_pages(page, 0);
557 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page,
558 rx_buf->page_offset, size, truesize);
560 /* avoid re-using remote pages */
561 if (unlikely(ice_page_is_reserved(page)))
564 #if (PAGE_SIZE < 8192)
565 /* if we are only owner of page we can reuse it */
566 if (unlikely(page_count(page) != 1))
569 /* flip page offset to other buffer */
570 rx_buf->page_offset ^= truesize;
572 /* move offset up to the next cache line */
573 rx_buf->page_offset += truesize;
575 if (rx_buf->page_offset > last_offset)
577 #endif /* PAGE_SIZE < 8192) */
579 /* Even if we own the page, we are not allowed to use atomic_set()
580 * This would break get_page_unless_zero() users.
582 get_page(rx_buf->page);
588 * ice_reuse_rx_page - page flip buffer and store it back on the ring
589 * @rx_ring: rx descriptor ring to store buffers on
590 * @old_buf: donor buffer to have page reused
592 * Synchronizes page for reuse by the adapter
594 static void ice_reuse_rx_page(struct ice_ring *rx_ring,
595 struct ice_rx_buf *old_buf)
597 u16 nta = rx_ring->next_to_alloc;
598 struct ice_rx_buf *new_buf;
600 new_buf = &rx_ring->rx_buf[nta];
602 /* update, and store next to alloc */
604 rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
606 /* transfer page from old buffer to new buffer */
611 * ice_fetch_rx_buf - Allocate skb and populate it
612 * @rx_ring: rx descriptor ring to transact packets on
613 * @rx_desc: descriptor containing info written by hardware
615 * This function allocates an skb on the fly, and populates it with the page
616 * data from the current receive descriptor, taking care to set up the skb
617 * correctly, as well as handling calling the page recycle function if
620 static struct sk_buff *ice_fetch_rx_buf(struct ice_ring *rx_ring,
621 union ice_32b_rx_flex_desc *rx_desc)
623 struct ice_rx_buf *rx_buf;
627 rx_buf = &rx_ring->rx_buf[rx_ring->next_to_clean];
634 u8 *page_addr = page_address(page) + rx_buf->page_offset;
636 /* prefetch first cache line of first page */
638 #if L1_CACHE_BYTES < 128
639 prefetch((void *)(page_addr + L1_CACHE_BYTES));
640 #endif /* L1_CACHE_BYTES */
642 /* allocate a skb to store the frags */
643 skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
645 GFP_ATOMIC | __GFP_NOWARN);
646 if (unlikely(!skb)) {
647 rx_ring->rx_stats.alloc_buf_failed++;
651 /* we will be copying header into skb->data in
652 * pskb_may_pull so it is in our interest to prefetch
653 * it now to avoid a possible cache miss
655 prefetchw(skb->data);
657 skb_record_rx_queue(skb, rx_ring->q_index);
659 /* we are reusing so sync this buffer for CPU use */
660 dma_sync_single_range_for_cpu(rx_ring->dev, rx_buf->dma,
668 /* pull page into skb */
669 if (ice_add_rx_frag(rx_buf, rx_desc, skb)) {
670 /* hand second half of page back to the ring */
671 ice_reuse_rx_page(rx_ring, rx_buf);
672 rx_ring->rx_stats.page_reuse_count++;
674 /* we are not reusing the buffer so unmap it */
675 dma_unmap_page(rx_ring->dev, rx_buf->dma, PAGE_SIZE,
679 /* clear contents of buffer_info */
686 * ice_pull_tail - ice specific version of skb_pull_tail
687 * @skb: pointer to current skb being adjusted
689 * This function is an ice specific version of __pskb_pull_tail. The
690 * main difference between this version and the original function is that
691 * this function can make several assumptions about the state of things
692 * that allow for significant optimizations versus the standard function.
693 * As a result we can do things like drop a frag and maintain an accurate
694 * truesize for the skb.
696 static void ice_pull_tail(struct sk_buff *skb)
698 struct skb_frag_struct *frag = &skb_shinfo(skb)->frags[0];
699 unsigned int pull_len;
702 /* it is valid to use page_address instead of kmap since we are
703 * working with pages allocated out of the lomem pool per
704 * alloc_page(GFP_ATOMIC)
706 va = skb_frag_address(frag);
708 /* we need the header to contain the greater of either ETH_HLEN or
709 * 60 bytes if the skb->len is less than 60 for skb_pad.
711 pull_len = eth_get_headlen(va, ICE_RX_HDR_SIZE);
713 /* align pull length to size of long to optimize memcpy performance */
714 skb_copy_to_linear_data(skb, va, ALIGN(pull_len, sizeof(long)));
716 /* update all of the pointers */
717 skb_frag_size_sub(frag, pull_len);
718 frag->page_offset += pull_len;
719 skb->data_len -= pull_len;
720 skb->tail += pull_len;
724 * ice_cleanup_headers - Correct empty headers
725 * @skb: pointer to current skb being fixed
727 * Also address the case where we are pulling data in on pages only
728 * and as such no data is present in the skb header.
730 * In addition if skb is not at least 60 bytes we need to pad it so that
731 * it is large enough to qualify as a valid Ethernet frame.
733 * Returns true if an error was encountered and skb was freed.
735 static bool ice_cleanup_headers(struct sk_buff *skb)
737 /* place header in linear portion of buffer */
738 if (skb_is_nonlinear(skb))
741 /* if eth_skb_pad returns an error the skb was freed */
742 if (eth_skb_pad(skb))
749 * ice_test_staterr - tests bits in Rx descriptor status and error fields
750 * @rx_desc: pointer to receive descriptor (in le64 format)
751 * @stat_err_bits: value to mask
753 * This function does some fast chicanery in order to return the
754 * value of the mask which is really only used for boolean tests.
755 * The status_error_len doesn't need to be shifted because it begins
758 static bool ice_test_staterr(union ice_32b_rx_flex_desc *rx_desc,
759 const u16 stat_err_bits)
761 return !!(rx_desc->wb.status_error0 &
762 cpu_to_le16(stat_err_bits));
766 * ice_is_non_eop - process handling of non-EOP buffers
767 * @rx_ring: Rx ring being processed
768 * @rx_desc: Rx descriptor for current buffer
769 * @skb: Current socket buffer containing buffer in progress
771 * This function updates next to clean. If the buffer is an EOP buffer
772 * this function exits returning false, otherwise it will place the
773 * sk_buff in the next buffer to be chained and return true indicating
774 * that this is in fact a non-EOP buffer.
776 static bool ice_is_non_eop(struct ice_ring *rx_ring,
777 union ice_32b_rx_flex_desc *rx_desc,
780 u32 ntc = rx_ring->next_to_clean + 1;
782 /* fetch, update, and store next to clean */
783 ntc = (ntc < rx_ring->count) ? ntc : 0;
784 rx_ring->next_to_clean = ntc;
786 prefetch(ICE_RX_DESC(rx_ring, ntc));
788 /* if we are the last buffer then there is nothing else to do */
789 #define ICE_RXD_EOF BIT(ICE_RX_FLEX_DESC_STATUS0_EOF_S)
790 if (likely(ice_test_staterr(rx_desc, ICE_RXD_EOF)))
793 /* place skb in next buffer to be received */
794 rx_ring->rx_buf[ntc].skb = skb;
795 rx_ring->rx_stats.non_eop_descs++;
801 * ice_ptype_to_htype - get a hash type
802 * @ptype: the ptype value from the descriptor
804 * Returns a hash type to be used by skb_set_hash
806 static enum pkt_hash_types ice_ptype_to_htype(u8 __always_unused ptype)
808 return PKT_HASH_TYPE_NONE;
812 * ice_rx_hash - set the hash value in the skb
813 * @rx_ring: descriptor ring
814 * @rx_desc: specific descriptor
815 * @skb: pointer to current skb
816 * @rx_ptype: the ptype value from the descriptor
819 ice_rx_hash(struct ice_ring *rx_ring, union ice_32b_rx_flex_desc *rx_desc,
820 struct sk_buff *skb, u8 rx_ptype)
822 struct ice_32b_rx_flex_desc_nic *nic_mdid;
825 if (!(rx_ring->netdev->features & NETIF_F_RXHASH))
828 if (rx_desc->wb.rxdid != ICE_RXDID_FLEX_NIC)
831 nic_mdid = (struct ice_32b_rx_flex_desc_nic *)rx_desc;
832 hash = le32_to_cpu(nic_mdid->rss_hash);
833 skb_set_hash(skb, hash, ice_ptype_to_htype(rx_ptype));
837 * ice_rx_csum - Indicate in skb if checksum is good
838 * @vsi: the VSI we care about
839 * @skb: skb currently being received and modified
840 * @rx_desc: the receive descriptor
841 * @ptype: the packet type decoded by hardware
843 * skb->protocol must be set before this function is called
845 static void ice_rx_csum(struct ice_vsi *vsi, struct sk_buff *skb,
846 union ice_32b_rx_flex_desc *rx_desc, u8 ptype)
848 struct ice_rx_ptype_decoded decoded;
849 u32 rx_error, rx_status;
852 rx_status = le16_to_cpu(rx_desc->wb.status_error0);
853 rx_error = rx_status;
855 decoded = ice_decode_rx_desc_ptype(ptype);
857 /* Start with CHECKSUM_NONE and by default csum_level = 0 */
858 skb->ip_summed = CHECKSUM_NONE;
859 skb_checksum_none_assert(skb);
861 /* check if Rx checksum is enabled */
862 if (!(vsi->netdev->features & NETIF_F_RXCSUM))
865 /* check if HW has decoded the packet and checksum */
866 if (!(rx_status & BIT(ICE_RX_FLEX_DESC_STATUS0_L3L4P_S)))
869 if (!(decoded.known && decoded.outer_ip))
872 ipv4 = (decoded.outer_ip == ICE_RX_PTYPE_OUTER_IP) &&
873 (decoded.outer_ip_ver == ICE_RX_PTYPE_OUTER_IPV4);
874 ipv6 = (decoded.outer_ip == ICE_RX_PTYPE_OUTER_IP) &&
875 (decoded.outer_ip_ver == ICE_RX_PTYPE_OUTER_IPV6);
877 if (ipv4 && (rx_error & (BIT(ICE_RX_FLEX_DESC_STATUS0_XSUM_IPE_S) |
878 BIT(ICE_RX_FLEX_DESC_STATUS0_XSUM_EIPE_S))))
880 else if (ipv6 && (rx_status &
881 (BIT(ICE_RX_FLEX_DESC_STATUS0_IPV6EXADD_S))))
884 /* check for L4 errors and handle packets that were not able to be
885 * checksummed due to arrival speed
887 if (rx_error & BIT(ICE_RX_FLEX_DESC_STATUS0_XSUM_L4E_S))
890 /* Only report checksum unnecessary for TCP, UDP, or SCTP */
891 switch (decoded.inner_prot) {
892 case ICE_RX_PTYPE_INNER_PROT_TCP:
893 case ICE_RX_PTYPE_INNER_PROT_UDP:
894 case ICE_RX_PTYPE_INNER_PROT_SCTP:
895 skb->ip_summed = CHECKSUM_UNNECESSARY;
902 vsi->back->hw_csum_rx_error++;
906 * ice_process_skb_fields - Populate skb header fields from Rx descriptor
907 * @rx_ring: rx descriptor ring packet is being transacted on
908 * @rx_desc: pointer to the EOP Rx descriptor
909 * @skb: pointer to current skb being populated
910 * @ptype: the packet type decoded by hardware
912 * This function checks the ring, descriptor, and packet information in
913 * order to populate the hash, checksum, VLAN, protocol, and
914 * other fields within the skb.
916 static void ice_process_skb_fields(struct ice_ring *rx_ring,
917 union ice_32b_rx_flex_desc *rx_desc,
918 struct sk_buff *skb, u8 ptype)
920 ice_rx_hash(rx_ring, rx_desc, skb, ptype);
922 /* modifies the skb - consumes the enet header */
923 skb->protocol = eth_type_trans(skb, rx_ring->netdev);
925 ice_rx_csum(rx_ring->vsi, skb, rx_desc, ptype);
929 * ice_receive_skb - Send a completed packet up the stack
930 * @rx_ring: rx ring in play
931 * @skb: packet to send up
932 * @vlan_tag: vlan tag for packet
934 * This function sends the completed packet (via. skb) up the stack using
935 * gro receive functions (with/without vlan tag)
937 static void ice_receive_skb(struct ice_ring *rx_ring, struct sk_buff *skb,
940 if ((rx_ring->netdev->features & NETIF_F_HW_VLAN_CTAG_RX) &&
941 (vlan_tag & VLAN_VID_MASK)) {
942 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
944 napi_gro_receive(&rx_ring->q_vector->napi, skb);
948 * ice_clean_rx_irq - Clean completed descriptors from Rx ring - bounce buf
949 * @rx_ring: rx descriptor ring to transact packets on
950 * @budget: Total limit on number of packets to process
952 * This function provides a "bounce buffer" approach to Rx interrupt
953 * processing. The advantage to this is that on systems that have
954 * expensive overhead for IOMMU access this provides a means of avoiding
955 * it by maintaining the mapping of the page to the system.
957 * Returns amount of work completed
959 static int ice_clean_rx_irq(struct ice_ring *rx_ring, int budget)
961 unsigned int total_rx_bytes = 0, total_rx_pkts = 0;
962 u16 cleaned_count = ICE_DESC_UNUSED(rx_ring);
963 bool failure = false;
965 /* start the loop to process RX packets bounded by 'budget' */
966 while (likely(total_rx_pkts < (unsigned int)budget)) {
967 union ice_32b_rx_flex_desc *rx_desc;
973 /* return some buffers to hardware, one at a time is too slow */
974 if (cleaned_count >= ICE_RX_BUF_WRITE) {
976 ice_alloc_rx_bufs(rx_ring, cleaned_count);
980 /* get the RX desc from RX ring based on 'next_to_clean' */
981 rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
983 /* status_error_len will always be zero for unused descriptors
984 * because it's cleared in cleanup, and overlaps with hdr_addr
985 * which is always zero because packet split isn't used, if the
986 * hardware wrote DD then it will be non-zero
988 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
989 if (!ice_test_staterr(rx_desc, stat_err_bits))
992 /* This memory barrier is needed to keep us from reading
993 * any other fields out of the rx_desc until we know the
998 /* allocate (if needed) and populate skb */
999 skb = ice_fetch_rx_buf(rx_ring, rx_desc);
1005 /* skip if it is NOP desc */
1006 if (ice_is_non_eop(rx_ring, rx_desc, skb))
1009 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_RXE_S);
1010 if (unlikely(ice_test_staterr(rx_desc, stat_err_bits))) {
1011 dev_kfree_skb_any(skb);
1015 rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
1016 ICE_RX_FLEX_DESC_PTYPE_M;
1018 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_L2TAG1P_S);
1019 if (ice_test_staterr(rx_desc, stat_err_bits))
1020 vlan_tag = le16_to_cpu(rx_desc->wb.l2tag1);
1022 /* correct empty headers and pad skb if needed (to make valid
1025 if (ice_cleanup_headers(skb)) {
1030 /* probably a little skewed due to removing CRC */
1031 total_rx_bytes += skb->len;
1033 /* populate checksum, VLAN, and protocol */
1034 ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
1036 /* send completed skb up the stack */
1037 ice_receive_skb(rx_ring, skb, vlan_tag);
1039 /* update budget accounting */
1043 /* update queue and vector specific stats */
1044 u64_stats_update_begin(&rx_ring->syncp);
1045 rx_ring->stats.pkts += total_rx_pkts;
1046 rx_ring->stats.bytes += total_rx_bytes;
1047 u64_stats_update_end(&rx_ring->syncp);
1048 rx_ring->q_vector->rx.total_pkts += total_rx_pkts;
1049 rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
1051 /* guarantee a trip back through this routine if there was a failure */
1052 return failure ? budget : (int)total_rx_pkts;
1056 * ice_napi_poll - NAPI polling Rx/Tx cleanup routine
1057 * @napi: napi struct with our devices info in it
1058 * @budget: amount of work driver is allowed to do this pass, in packets
1060 * This function will clean all queues associated with a q_vector.
1062 * Returns the amount of work done
1064 int ice_napi_poll(struct napi_struct *napi, int budget)
1066 struct ice_q_vector *q_vector =
1067 container_of(napi, struct ice_q_vector, napi);
1068 struct ice_vsi *vsi = q_vector->vsi;
1069 struct ice_pf *pf = vsi->back;
1070 bool clean_complete = true;
1071 int budget_per_ring = 0;
1072 struct ice_ring *ring;
1075 /* Since the actual Tx work is minimal, we can give the Tx a larger
1076 * budget and be more aggressive about cleaning up the Tx descriptors.
1078 ice_for_each_ring(ring, q_vector->tx)
1079 if (!ice_clean_tx_irq(vsi, ring, budget))
1080 clean_complete = false;
1082 /* Handle case where we are called by netpoll with a budget of 0 */
1086 /* We attempt to distribute budget to each Rx queue fairly, but don't
1087 * allow the budget to go below 1 because that would exit polling early.
1089 if (q_vector->num_ring_rx)
1090 budget_per_ring = max(budget / q_vector->num_ring_rx, 1);
1092 ice_for_each_ring(ring, q_vector->rx) {
1095 cleaned = ice_clean_rx_irq(ring, budget_per_ring);
1096 work_done += cleaned;
1097 /* if we clean as many as budgeted, we must not be done */
1098 if (cleaned >= budget_per_ring)
1099 clean_complete = false;
1102 /* If work not completed, return budget and polling will return */
1103 if (!clean_complete)
1106 /* Work is done so exit the polling mode and re-enable the interrupt */
1107 napi_complete_done(napi, work_done);
1108 if (test_bit(ICE_FLAG_MSIX_ENA, pf->flags))
1109 ice_irq_dynamic_ena(&vsi->back->hw, vsi, q_vector);
1113 /* helper function for building cmd/type/offset */
1115 build_ctob(u64 td_cmd, u64 td_offset, unsigned int size, u64 td_tag)
1117 return cpu_to_le64(ICE_TX_DESC_DTYPE_DATA |
1118 (td_cmd << ICE_TXD_QW1_CMD_S) |
1119 (td_offset << ICE_TXD_QW1_OFFSET_S) |
1120 ((u64)size << ICE_TXD_QW1_TX_BUF_SZ_S) |
1121 (td_tag << ICE_TXD_QW1_L2TAG1_S));
1125 * __ice_maybe_stop_tx - 2nd level check for tx stop conditions
1126 * @tx_ring: the ring to be checked
1127 * @size: the size buffer we want to assure is available
1129 * Returns -EBUSY if a stop is needed, else 0
1131 static int __ice_maybe_stop_tx(struct ice_ring *tx_ring, unsigned int size)
1133 netif_stop_subqueue(tx_ring->netdev, tx_ring->q_index);
1134 /* Memory barrier before checking head and tail */
1137 /* Check again in a case another CPU has just made room available. */
1138 if (likely(ICE_DESC_UNUSED(tx_ring) < size))
1141 /* A reprieve! - use start_subqueue because it doesn't call schedule */
1142 netif_start_subqueue(tx_ring->netdev, tx_ring->q_index);
1143 ++tx_ring->tx_stats.restart_q;
1148 * ice_maybe_stop_tx - 1st level check for tx stop conditions
1149 * @tx_ring: the ring to be checked
1150 * @size: the size buffer we want to assure is available
1152 * Returns 0 if stop is not needed
1154 static int ice_maybe_stop_tx(struct ice_ring *tx_ring, unsigned int size)
1156 if (likely(ICE_DESC_UNUSED(tx_ring) >= size))
1158 return __ice_maybe_stop_tx(tx_ring, size);
1162 * ice_tx_map - Build the Tx descriptor
1163 * @tx_ring: ring to send buffer on
1164 * @first: first buffer info buffer to use
1165 * @off: pointer to struct that holds offload parameters
1167 * This function loops over the skb data pointed to by *first
1168 * and gets a physical address for each memory location and programs
1169 * it and the length into the transmit descriptor.
1172 ice_tx_map(struct ice_ring *tx_ring, struct ice_tx_buf *first,
1173 struct ice_tx_offload_params *off)
1175 u64 td_offset, td_tag, td_cmd;
1176 u16 i = tx_ring->next_to_use;
1177 struct skb_frag_struct *frag;
1178 unsigned int data_len, size;
1179 struct ice_tx_desc *tx_desc;
1180 struct ice_tx_buf *tx_buf;
1181 struct sk_buff *skb;
1184 td_tag = off->td_l2tag1;
1185 td_cmd = off->td_cmd;
1186 td_offset = off->td_offset;
1189 data_len = skb->data_len;
1190 size = skb_headlen(skb);
1192 tx_desc = ICE_TX_DESC(tx_ring, i);
1194 if (first->tx_flags & ICE_TX_FLAGS_HW_VLAN) {
1195 td_cmd |= (u64)ICE_TX_DESC_CMD_IL2TAG1;
1196 td_tag = (first->tx_flags & ICE_TX_FLAGS_VLAN_M) >>
1197 ICE_TX_FLAGS_VLAN_S;
1200 dma = dma_map_single(tx_ring->dev, skb->data, size, DMA_TO_DEVICE);
1204 for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
1205 unsigned int max_data = ICE_MAX_DATA_PER_TXD_ALIGNED;
1207 if (dma_mapping_error(tx_ring->dev, dma))
1210 /* record length, and DMA address */
1211 dma_unmap_len_set(tx_buf, len, size);
1212 dma_unmap_addr_set(tx_buf, dma, dma);
1214 /* align size to end of page */
1215 max_data += -dma & (ICE_MAX_READ_REQ_SIZE - 1);
1216 tx_desc->buf_addr = cpu_to_le64(dma);
1218 /* account for data chunks larger than the hardware
1221 while (unlikely(size > ICE_MAX_DATA_PER_TXD)) {
1222 tx_desc->cmd_type_offset_bsz =
1223 build_ctob(td_cmd, td_offset, max_data, td_tag);
1228 if (i == tx_ring->count) {
1229 tx_desc = ICE_TX_DESC(tx_ring, 0);
1236 max_data = ICE_MAX_DATA_PER_TXD_ALIGNED;
1237 tx_desc->buf_addr = cpu_to_le64(dma);
1240 if (likely(!data_len))
1243 tx_desc->cmd_type_offset_bsz = build_ctob(td_cmd, td_offset,
1249 if (i == tx_ring->count) {
1250 tx_desc = ICE_TX_DESC(tx_ring, 0);
1254 size = skb_frag_size(frag);
1257 dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
1260 tx_buf = &tx_ring->tx_buf[i];
1263 /* record bytecount for BQL */
1264 netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount);
1266 /* record SW timestamp if HW timestamp is not available */
1267 skb_tx_timestamp(first->skb);
1270 if (i == tx_ring->count)
1273 /* write last descriptor with RS and EOP bits */
1274 td_cmd |= (u64)(ICE_TX_DESC_CMD_EOP | ICE_TX_DESC_CMD_RS);
1275 tx_desc->cmd_type_offset_bsz =
1276 build_ctob(td_cmd, td_offset, size, td_tag);
1278 /* Force memory writes to complete before letting h/w know there
1279 * are new descriptors to fetch.
1281 * We also use this memory barrier to make certain all of the
1282 * status bits have been updated before next_to_watch is written.
1286 /* set next_to_watch value indicating a packet is present */
1287 first->next_to_watch = tx_desc;
1289 tx_ring->next_to_use = i;
1291 ice_maybe_stop_tx(tx_ring, DESC_NEEDED);
1293 /* notify HW of packet */
1294 if (netif_xmit_stopped(txring_txq(tx_ring)) || !skb->xmit_more) {
1295 writel(i, tx_ring->tail);
1297 /* we need this if more than one processor can write to our tail
1298 * at a time, it synchronizes IO on IA64/Altix systems
1306 /* clear dma mappings for failed tx_buf map */
1308 tx_buf = &tx_ring->tx_buf[i];
1309 ice_unmap_and_free_tx_buf(tx_ring, tx_buf);
1310 if (tx_buf == first)
1317 tx_ring->next_to_use = i;
1321 * ice_tx_csum - Enable Tx checksum offloads
1322 * @first: pointer to the first descriptor
1323 * @off: pointer to struct that holds offload parameters
1325 * Returns 0 or error (negative) if checksum offload can't happen, 1 otherwise.
1328 int ice_tx_csum(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
1330 u32 l4_len = 0, l3_len = 0, l2_len = 0;
1331 struct sk_buff *skb = first->skb;
1341 __be16 frag_off, protocol;
1342 unsigned char *exthdr;
1343 u32 offset, cmd = 0;
1346 if (skb->ip_summed != CHECKSUM_PARTIAL)
1349 ip.hdr = skb_network_header(skb);
1350 l4.hdr = skb_transport_header(skb);
1352 /* compute outer L2 header size */
1353 l2_len = ip.hdr - skb->data;
1354 offset = (l2_len / 2) << ICE_TX_DESC_LEN_MACLEN_S;
1356 if (skb->encapsulation)
1359 /* Enable IP checksum offloads */
1360 protocol = vlan_get_protocol(skb);
1361 if (protocol == htons(ETH_P_IP)) {
1362 l4_proto = ip.v4->protocol;
1363 /* the stack computes the IP header already, the only time we
1364 * need the hardware to recompute it is in the case of TSO.
1366 if (first->tx_flags & ICE_TX_FLAGS_TSO)
1367 cmd |= ICE_TX_DESC_CMD_IIPT_IPV4_CSUM;
1369 cmd |= ICE_TX_DESC_CMD_IIPT_IPV4;
1371 } else if (protocol == htons(ETH_P_IPV6)) {
1372 cmd |= ICE_TX_DESC_CMD_IIPT_IPV6;
1373 exthdr = ip.hdr + sizeof(*ip.v6);
1374 l4_proto = ip.v6->nexthdr;
1375 if (l4.hdr != exthdr)
1376 ipv6_skip_exthdr(skb, exthdr - skb->data, &l4_proto,
1382 /* compute inner L3 header size */
1383 l3_len = l4.hdr - ip.hdr;
1384 offset |= (l3_len / 4) << ICE_TX_DESC_LEN_IPLEN_S;
1386 /* Enable L4 checksum offloads */
1389 /* enable checksum offloads */
1390 cmd |= ICE_TX_DESC_CMD_L4T_EOFT_TCP;
1391 l4_len = l4.tcp->doff;
1392 offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1395 /* enable UDP checksum offload */
1396 cmd |= ICE_TX_DESC_CMD_L4T_EOFT_UDP;
1397 l4_len = (sizeof(struct udphdr) >> 2);
1398 offset |= l4_len << ICE_TX_DESC_LEN_L4_LEN_S;
1402 if (first->tx_flags & ICE_TX_FLAGS_TSO)
1404 skb_checksum_help(skb);
1409 off->td_offset |= offset;
1414 * ice_tx_prepare_vlan_flags - prepare generic TX VLAN tagging flags for HW
1415 * @tx_ring: ring to send buffer on
1416 * @first: pointer to struct ice_tx_buf
1418 * Checks the skb and set up correspondingly several generic transmit flags
1419 * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
1421 * Returns error code indicate the frame should be dropped upon error and the
1422 * otherwise returns 0 to indicate the flags has been set properly.
1425 ice_tx_prepare_vlan_flags(struct ice_ring *tx_ring, struct ice_tx_buf *first)
1427 struct sk_buff *skb = first->skb;
1428 __be16 protocol = skb->protocol;
1430 if (protocol == htons(ETH_P_8021Q) &&
1431 !(tx_ring->netdev->features & NETIF_F_HW_VLAN_CTAG_TX)) {
1432 /* when HW VLAN acceleration is turned off by the user the
1433 * stack sets the protocol to 8021q so that the driver
1434 * can take any steps required to support the SW only
1435 * VLAN handling. In our case the driver doesn't need
1436 * to take any further steps so just set the protocol
1437 * to the encapsulated ethertype.
1439 skb->protocol = vlan_get_protocol(skb);
1443 /* if we have a HW VLAN tag being added, default to the HW one */
1444 if (skb_vlan_tag_present(skb)) {
1445 first->tx_flags |= skb_vlan_tag_get(skb) << ICE_TX_FLAGS_VLAN_S;
1446 first->tx_flags |= ICE_TX_FLAGS_HW_VLAN;
1447 } else if (protocol == htons(ETH_P_8021Q)) {
1448 struct vlan_hdr *vhdr, _vhdr;
1450 /* for SW VLAN, check the next protocol and store the tag */
1451 vhdr = (struct vlan_hdr *)skb_header_pointer(skb, ETH_HLEN,
1457 first->tx_flags |= ntohs(vhdr->h_vlan_TCI) <<
1458 ICE_TX_FLAGS_VLAN_S;
1459 first->tx_flags |= ICE_TX_FLAGS_SW_VLAN;
1467 * ice_tso - computes mss and TSO length to prepare for TSO
1468 * @first: pointer to struct ice_tx_buf
1469 * @off: pointer to struct that holds offload parameters
1471 * Returns 0 or error (negative) if TSO can't happen, 1 otherwise.
1474 int ice_tso(struct ice_tx_buf *first, struct ice_tx_offload_params *off)
1476 struct sk_buff *skb = first->skb;
1486 u64 cd_mss, cd_tso_len;
1487 u32 paylen, l4_start;
1490 if (skb->ip_summed != CHECKSUM_PARTIAL)
1493 if (!skb_is_gso(skb))
1496 err = skb_cow_head(skb, 0);
1500 ip.hdr = skb_network_header(skb);
1501 l4.hdr = skb_transport_header(skb);
1503 /* initialize outer IP header fields */
1504 if (ip.v4->version == 4) {
1508 ip.v6->payload_len = 0;
1511 /* determine offset of transport header */
1512 l4_start = l4.hdr - skb->data;
1514 /* remove payload length from checksum */
1515 paylen = skb->len - l4_start;
1516 csum_replace_by_diff(&l4.tcp->check, (__force __wsum)htonl(paylen));
1518 /* compute length of segmentation header */
1519 off->header_len = (l4.tcp->doff * 4) + l4_start;
1521 /* update gso_segs and bytecount */
1522 first->gso_segs = skb_shinfo(skb)->gso_segs;
1523 first->bytecount += (first->gso_segs - 1) * off->header_len;
1525 cd_tso_len = skb->len - off->header_len;
1526 cd_mss = skb_shinfo(skb)->gso_size;
1528 /* record cdesc_qw1 with TSO parameters */
1529 off->cd_qw1 |= ICE_TX_DESC_DTYPE_CTX |
1530 (ICE_TX_CTX_DESC_TSO << ICE_TXD_CTX_QW1_CMD_S) |
1531 (cd_tso_len << ICE_TXD_CTX_QW1_TSO_LEN_S) |
1532 (cd_mss << ICE_TXD_CTX_QW1_MSS_S);
1533 first->tx_flags |= ICE_TX_FLAGS_TSO;
1538 * ice_txd_use_count - estimate the number of descriptors needed for Tx
1539 * @size: transmit request size in bytes
1541 * Due to hardware alignment restrictions (4K alignment), we need to
1542 * assume that we can have no more than 12K of data per descriptor, even
1543 * though each descriptor can take up to 16K - 1 bytes of aligned memory.
1544 * Thus, we need to divide by 12K. But division is slow! Instead,
1545 * we decompose the operation into shifts and one relatively cheap
1546 * multiply operation.
1548 * To divide by 12K, we first divide by 4K, then divide by 3:
1549 * To divide by 4K, shift right by 12 bits
1550 * To divide by 3, multiply by 85, then divide by 256
1551 * (Divide by 256 is done by shifting right by 8 bits)
1552 * Finally, we add one to round up. Because 256 isn't an exact multiple of
1553 * 3, we'll underestimate near each multiple of 12K. This is actually more
1554 * accurate as we have 4K - 1 of wiggle room that we can fit into the last
1555 * segment. For our purposes this is accurate out to 1M which is orders of
1556 * magnitude greater than our largest possible GSO size.
1558 * This would then be implemented as:
1559 * return (((size >> 12) * 85) >> 8) + ICE_DESCS_FOR_SKB_DATA_PTR;
1561 * Since multiplication and division are commutative, we can reorder
1563 * return ((size * 85) >> 20) + ICE_DESCS_FOR_SKB_DATA_PTR;
1565 static unsigned int ice_txd_use_count(unsigned int size)
1567 return ((size * 85) >> 20) + ICE_DESCS_FOR_SKB_DATA_PTR;
1571 * ice_xmit_desc_count - calculate number of tx descriptors needed
1574 * Returns number of data descriptors needed for this skb.
1576 static unsigned int ice_xmit_desc_count(struct sk_buff *skb)
1578 const struct skb_frag_struct *frag = &skb_shinfo(skb)->frags[0];
1579 unsigned int nr_frags = skb_shinfo(skb)->nr_frags;
1580 unsigned int count = 0, size = skb_headlen(skb);
1583 count += ice_txd_use_count(size);
1588 size = skb_frag_size(frag++);
1595 * __ice_chk_linearize - Check if there are more than 8 buffers per packet
1598 * Note: This HW can't DMA more than 8 buffers to build a packet on the wire
1599 * and so we need to figure out the cases where we need to linearize the skb.
1601 * For TSO we need to count the TSO header and segment payload separately.
1602 * As such we need to check cases where we have 7 fragments or more as we
1603 * can potentially require 9 DMA transactions, 1 for the TSO header, 1 for
1604 * the segment payload in the first descriptor, and another 7 for the
1607 static bool __ice_chk_linearize(struct sk_buff *skb)
1609 const struct skb_frag_struct *frag, *stale;
1612 /* no need to check if number of frags is less than 7 */
1613 nr_frags = skb_shinfo(skb)->nr_frags;
1614 if (nr_frags < (ICE_MAX_BUF_TXD - 1))
1617 /* We need to walk through the list and validate that each group
1618 * of 6 fragments totals at least gso_size.
1620 nr_frags -= ICE_MAX_BUF_TXD - 2;
1621 frag = &skb_shinfo(skb)->frags[0];
1623 /* Initialize size to the negative value of gso_size minus 1. We
1624 * use this as the worst case scenerio in which the frag ahead
1625 * of us only provides one byte which is why we are limited to 6
1626 * descriptors for a single transmit as the header and previous
1627 * fragment are already consuming 2 descriptors.
1629 sum = 1 - skb_shinfo(skb)->gso_size;
1631 /* Add size of frags 0 through 4 to create our initial sum */
1632 sum += skb_frag_size(frag++);
1633 sum += skb_frag_size(frag++);
1634 sum += skb_frag_size(frag++);
1635 sum += skb_frag_size(frag++);
1636 sum += skb_frag_size(frag++);
1638 /* Walk through fragments adding latest fragment, testing it, and
1639 * then removing stale fragments from the sum.
1641 stale = &skb_shinfo(skb)->frags[0];
1643 sum += skb_frag_size(frag++);
1645 /* if sum is negative we failed to make sufficient progress */
1652 sum -= skb_frag_size(stale++);
1659 * ice_chk_linearize - Check if there are more than 8 fragments per packet
1661 * @count: number of buffers used
1663 * Note: Our HW can't scatter-gather more than 8 fragments to build
1664 * a packet on the wire and so we need to figure out the cases where we
1665 * need to linearize the skb.
1667 static bool ice_chk_linearize(struct sk_buff *skb, unsigned int count)
1669 /* Both TSO and single send will work if count is less than 8 */
1670 if (likely(count < ICE_MAX_BUF_TXD))
1673 if (skb_is_gso(skb))
1674 return __ice_chk_linearize(skb);
1676 /* we can support up to 8 data buffers for a single send */
1677 return count != ICE_MAX_BUF_TXD;
1681 * ice_xmit_frame_ring - Sends buffer on Tx ring
1683 * @tx_ring: ring to send buffer on
1685 * Returns NETDEV_TX_OK if sent, else an error code
1688 ice_xmit_frame_ring(struct sk_buff *skb, struct ice_ring *tx_ring)
1690 struct ice_tx_offload_params offload = { 0 };
1691 struct ice_tx_buf *first;
1695 count = ice_xmit_desc_count(skb);
1696 if (ice_chk_linearize(skb, count)) {
1697 if (__skb_linearize(skb))
1699 count = ice_txd_use_count(skb->len);
1700 tx_ring->tx_stats.tx_linearize++;
1703 /* need: 1 descriptor per page * PAGE_SIZE/ICE_MAX_DATA_PER_TXD,
1704 * + 1 desc for skb_head_len/ICE_MAX_DATA_PER_TXD,
1705 * + 4 desc gap to avoid the cache line where head is,
1706 * + 1 desc for context descriptor,
1707 * otherwise try next time
1709 if (ice_maybe_stop_tx(tx_ring, count + ICE_DESCS_PER_CACHE_LINE +
1710 ICE_DESCS_FOR_CTX_DESC)) {
1711 tx_ring->tx_stats.tx_busy++;
1712 return NETDEV_TX_BUSY;
1715 offload.tx_ring = tx_ring;
1717 /* record the location of the first descriptor for this packet */
1718 first = &tx_ring->tx_buf[tx_ring->next_to_use];
1720 first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN);
1721 first->gso_segs = 1;
1722 first->tx_flags = 0;
1724 /* prepare the VLAN tagging flags for Tx */
1725 if (ice_tx_prepare_vlan_flags(tx_ring, first))
1728 /* set up TSO offload */
1729 tso = ice_tso(first, &offload);
1733 /* always set up Tx checksum offload */
1734 csum = ice_tx_csum(first, &offload);
1738 if (tso || offload.cd_tunnel_params) {
1739 struct ice_tx_ctx_desc *cdesc;
1740 int i = tx_ring->next_to_use;
1742 /* grab the next descriptor */
1743 cdesc = ICE_TX_CTX_DESC(tx_ring, i);
1745 tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
1747 /* setup context descriptor */
1748 cdesc->tunneling_params = cpu_to_le32(offload.cd_tunnel_params);
1749 cdesc->l2tag2 = cpu_to_le16(offload.cd_l2tag2);
1750 cdesc->rsvd = cpu_to_le16(0);
1751 cdesc->qw1 = cpu_to_le64(offload.cd_qw1);
1754 ice_tx_map(tx_ring, first, &offload);
1755 return NETDEV_TX_OK;
1758 dev_kfree_skb_any(skb);
1759 return NETDEV_TX_OK;
1763 * ice_start_xmit - Selects the correct VSI and Tx queue to send buffer
1765 * @netdev: network interface device structure
1767 * Returns NETDEV_TX_OK if sent, else an error code
1769 netdev_tx_t ice_start_xmit(struct sk_buff *skb, struct net_device *netdev)
1771 struct ice_netdev_priv *np = netdev_priv(netdev);
1772 struct ice_vsi *vsi = np->vsi;
1773 struct ice_ring *tx_ring;
1775 tx_ring = vsi->tx_rings[skb->queue_mapping];
1777 /* hardware can't handle really short frames, hardware padding works
1780 if (skb_put_padto(skb, ICE_MIN_TX_LEN))
1781 return NETDEV_TX_OK;
1783 return ice_xmit_frame_ring(skb, tx_ring);