1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2005-2006 Fen Systems Ltd.
5 * Copyright 2005-2013 Solarflare Communications Inc.
12 #include <linux/ipv6.h>
13 #include <linux/slab.h>
15 #include <linux/if_ether.h>
16 #include <linux/highmem.h>
17 #include <linux/cache.h>
18 #include "net_driver.h"
23 #include "tx_common.h"
24 #include "workarounds.h"
25 #include "ef10_regs.h"
29 #define EFX_PIOBUF_SIZE_DEF ALIGN(256, L1_CACHE_BYTES)
30 unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF;
32 #endif /* EFX_USE_PIO */
34 static inline u8 *efx_tx_get_copy_buffer(struct efx_tx_queue *tx_queue,
35 struct efx_tx_buffer *buffer)
37 unsigned int index = efx_tx_queue_get_insert_index(tx_queue);
38 struct efx_buffer *page_buf =
39 &tx_queue->cb_page[index >> (PAGE_SHIFT - EFX_TX_CB_ORDER)];
41 ((index << EFX_TX_CB_ORDER) + NET_IP_ALIGN) & (PAGE_SIZE - 1);
43 if (unlikely(!page_buf->addr) &&
44 efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
47 buffer->dma_addr = page_buf->dma_addr + offset;
48 buffer->unmap_len = 0;
49 return (u8 *)page_buf->addr + offset;
52 u8 *efx_tx_get_copy_buffer_limited(struct efx_tx_queue *tx_queue,
53 struct efx_tx_buffer *buffer, size_t len)
55 if (len > EFX_TX_CB_SIZE)
57 return efx_tx_get_copy_buffer(tx_queue, buffer);
60 static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
62 /* We need to consider all queues that the net core sees as one */
63 struct efx_nic *efx = txq1->efx;
64 struct efx_tx_queue *txq2;
65 unsigned int fill_level;
67 fill_level = efx_channel_tx_old_fill_level(txq1->channel);
68 if (likely(fill_level < efx->txq_stop_thresh))
71 /* We used the stale old_read_count above, which gives us a
72 * pessimistic estimate of the fill level (which may even
73 * validly be >= efx->txq_entries). Now try again using
74 * read_count (more likely to be a cache miss).
76 * If we read read_count and then conditionally stop the
77 * queue, it is possible for the completion path to race with
78 * us and complete all outstanding descriptors in the middle,
79 * after which there will be no more completions to wake it.
80 * Therefore we stop the queue first, then read read_count
81 * (with a memory barrier to ensure the ordering), then
82 * restart the queue if the fill level turns out to be low
85 netif_tx_stop_queue(txq1->core_txq);
87 efx_for_each_channel_tx_queue(txq2, txq1->channel)
88 txq2->old_read_count = READ_ONCE(txq2->read_count);
90 fill_level = efx_channel_tx_old_fill_level(txq1->channel);
91 EFX_WARN_ON_ONCE_PARANOID(fill_level >= efx->txq_entries);
92 if (likely(fill_level < efx->txq_stop_thresh)) {
94 if (likely(!efx->loopback_selftest))
95 netif_tx_start_queue(txq1->core_txq);
99 static int efx_enqueue_skb_copy(struct efx_tx_queue *tx_queue,
102 unsigned int copy_len = skb->len;
103 struct efx_tx_buffer *buffer;
107 EFX_WARN_ON_ONCE_PARANOID(copy_len > EFX_TX_CB_SIZE);
109 buffer = efx_tx_queue_get_insert_buffer(tx_queue);
111 copy_buffer = efx_tx_get_copy_buffer(tx_queue, buffer);
112 if (unlikely(!copy_buffer))
115 rc = skb_copy_bits(skb, 0, copy_buffer, copy_len);
116 EFX_WARN_ON_PARANOID(rc);
117 buffer->len = copy_len;
120 buffer->flags = EFX_TX_BUF_SKB;
122 ++tx_queue->insert_count;
128 struct efx_short_copy_buffer {
130 u8 buf[L1_CACHE_BYTES];
133 /* Copy to PIO, respecting that writes to PIO buffers must be dword aligned.
134 * Advances piobuf pointer. Leaves additional data in the copy buffer.
136 static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf,
138 struct efx_short_copy_buffer *copy_buf)
140 int block_len = len & ~(sizeof(copy_buf->buf) - 1);
142 __iowrite64_copy(*piobuf, data, block_len >> 3);
143 *piobuf += block_len;
148 BUG_ON(copy_buf->used);
149 BUG_ON(len > sizeof(copy_buf->buf));
150 memcpy(copy_buf->buf, data, len);
151 copy_buf->used = len;
155 /* Copy to PIO, respecting dword alignment, popping data from copy buffer first.
156 * Advances piobuf pointer. Leaves additional data in the copy buffer.
158 static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf,
160 struct efx_short_copy_buffer *copy_buf)
162 if (copy_buf->used) {
163 /* if the copy buffer is partially full, fill it up and write */
165 min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len);
167 memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf);
168 copy_buf->used += copy_to_buf;
170 /* if we didn't fill it up then we're done for now */
171 if (copy_buf->used < sizeof(copy_buf->buf))
174 __iowrite64_copy(*piobuf, copy_buf->buf,
175 sizeof(copy_buf->buf) >> 3);
176 *piobuf += sizeof(copy_buf->buf);
182 efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf);
185 static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf,
186 struct efx_short_copy_buffer *copy_buf)
188 /* if there's anything in it, write the whole buffer, including junk */
190 __iowrite64_copy(piobuf, copy_buf->buf,
191 sizeof(copy_buf->buf) >> 3);
194 /* Traverse skb structure and copy fragments in to PIO buffer.
195 * Advances piobuf pointer.
197 static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb,
199 struct efx_short_copy_buffer *copy_buf)
203 efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb),
206 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
207 skb_frag_t *f = &skb_shinfo(skb)->frags[i];
210 vaddr = kmap_atomic(skb_frag_page(f));
212 efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + skb_frag_off(f),
213 skb_frag_size(f), copy_buf);
214 kunmap_atomic(vaddr);
217 EFX_WARN_ON_ONCE_PARANOID(skb_shinfo(skb)->frag_list);
220 static int efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue,
223 struct efx_tx_buffer *buffer =
224 efx_tx_queue_get_insert_buffer(tx_queue);
225 u8 __iomem *piobuf = tx_queue->piobuf;
227 /* Copy to PIO buffer. Ensure the writes are padded to the end
228 * of a cache line, as this is required for write-combining to be
229 * effective on at least x86.
232 if (skb_shinfo(skb)->nr_frags) {
233 /* The size of the copy buffer will ensure all writes
234 * are the size of a cache line.
236 struct efx_short_copy_buffer copy_buf;
240 efx_skb_copy_bits_to_pio(tx_queue->efx, skb,
242 efx_flush_copy_buffer(tx_queue->efx, piobuf, ©_buf);
244 /* Pad the write to the size of a cache line.
245 * We can do this because we know the skb_shared_info struct is
246 * after the source, and the destination buffer is big enough.
248 BUILD_BUG_ON(L1_CACHE_BYTES >
249 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
250 __iowrite64_copy(tx_queue->piobuf, skb->data,
251 ALIGN(skb->len, L1_CACHE_BYTES) >> 3);
255 buffer->flags = EFX_TX_BUF_SKB | EFX_TX_BUF_OPTION;
257 EFX_POPULATE_QWORD_5(buffer->option,
258 ESF_DZ_TX_DESC_IS_OPT, 1,
259 ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO,
260 ESF_DZ_TX_PIO_CONT, 0,
261 ESF_DZ_TX_PIO_BYTE_CNT, skb->len,
262 ESF_DZ_TX_PIO_BUF_ADDR,
263 tx_queue->piobuf_offset);
264 ++tx_queue->insert_count;
268 /* Decide whether we can use TX PIO, ie. write packet data directly into
269 * a buffer on the device. This can reduce latency at the expense of
270 * throughput, so we only do this if both hardware and software TX rings
271 * are empty, including all queues for the channel. This also ensures that
272 * only one packet at a time can be using the PIO buffer. If the xmit_more
273 * flag is set then we don't use this - there'll be another packet along
274 * shortly and we want to hold off the doorbell.
276 static bool efx_tx_may_pio(struct efx_tx_queue *tx_queue)
278 struct efx_channel *channel = tx_queue->channel;
280 if (!tx_queue->piobuf)
283 EFX_WARN_ON_ONCE_PARANOID(!channel->efx->type->option_descriptors);
285 efx_for_each_channel_tx_queue(tx_queue, channel)
286 if (!efx_nic_tx_is_empty(tx_queue, tx_queue->packet_write_count))
291 #endif /* EFX_USE_PIO */
293 /* Send any pending traffic for a channel. xmit_more is shared across all
294 * queues for a channel, so we must check all of them.
296 static void efx_tx_send_pending(struct efx_channel *channel)
298 struct efx_tx_queue *q;
300 efx_for_each_channel_tx_queue(q, channel) {
302 efx_nic_push_buffers(q);
307 * Add a socket buffer to a TX queue
309 * This maps all fragments of a socket buffer for DMA and adds them to
310 * the TX queue. The queue's insert pointer will be incremented by
311 * the number of fragments in the socket buffer.
313 * If any DMA mapping fails, any mapped fragments will be unmapped,
314 * the queue's insert pointer will be restored to its original value.
316 * This function is split out from efx_hard_start_xmit to allow the
317 * loopback test to direct packets via specific TX queues.
319 * Returns NETDEV_TX_OK.
320 * You must hold netif_tx_lock() to call this function.
322 netdev_tx_t __efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
324 unsigned int old_insert_count = tx_queue->insert_count;
325 bool xmit_more = netdev_xmit_more();
326 bool data_mapped = false;
327 unsigned int segments;
328 unsigned int skb_len;
332 segments = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 0;
334 segments = 0; /* Don't use TSO for a single segment. */
336 /* Handle TSO first - it's *possible* (although unlikely) that we might
337 * be passed a packet to segment that's smaller than the copybreak/PIO
341 switch (tx_queue->tso_version) {
343 rc = efx_enqueue_skb_tso(tx_queue, skb, &data_mapped);
346 rc = efx_ef10_tx_tso_desc(tx_queue, skb, &data_mapped);
348 case 0: /* No TSO on this queue, SW fallback needed */
354 rc = efx_tx_tso_fallback(tx_queue, skb);
355 tx_queue->tso_fallbacks++;
362 } else if (skb_len <= efx_piobuf_size && !xmit_more &&
363 efx_tx_may_pio(tx_queue)) {
364 /* Use PIO for short packets with an empty queue. */
365 if (efx_enqueue_skb_pio(tx_queue, skb))
367 tx_queue->pio_packets++;
370 } else if (skb->data_len && skb_len <= EFX_TX_CB_SIZE) {
371 /* Pad short packets or coalesce short fragmented packets. */
372 if (efx_enqueue_skb_copy(tx_queue, skb))
374 tx_queue->cb_packets++;
378 /* Map for DMA and create descriptors if we haven't done so already. */
379 if (!data_mapped && (efx_tx_map_data(tx_queue, skb, segments)))
382 efx_tx_maybe_stop_queue(tx_queue);
384 tx_queue->xmit_pending = true;
386 /* Pass off to hardware */
387 if (__netdev_tx_sent_queue(tx_queue->core_txq, skb_len, xmit_more))
388 efx_tx_send_pending(tx_queue->channel);
391 tx_queue->tso_bursts++;
392 tx_queue->tso_packets += segments;
393 tx_queue->tx_packets += segments;
395 tx_queue->tx_packets++;
402 efx_enqueue_unwind(tx_queue, old_insert_count);
403 dev_kfree_skb_any(skb);
405 /* If we're not expecting another transmit and we had something to push
406 * on this queue or a partner queue then we need to push here to get the
407 * previous packets out.
410 efx_tx_send_pending(tx_queue->channel);
415 /* Transmit a packet from an XDP buffer
417 * Returns number of packets sent on success, error code otherwise.
418 * Runs in NAPI context, either in our poll (for XDP TX) or a different NIC
419 * (for XDP redirect).
421 int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
424 struct efx_tx_buffer *tx_buffer;
425 struct efx_tx_queue *tx_queue;
426 struct xdp_frame *xdpf;
433 if (unlikely(n && !xdpfs))
438 cpu = raw_smp_processor_id();
439 if (unlikely(cpu >= efx->xdp_tx_queue_count))
442 tx_queue = efx->xdp_tx_queues[cpu];
443 if (unlikely(!tx_queue))
446 if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
447 HARD_TX_LOCK(efx->net_dev, tx_queue->core_txq, cpu);
449 /* If we're borrowing net stack queues we have to handle stop-restart
450 * or we might block the queue and it will be considered as frozen
452 if (efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_BORROWED) {
453 if (netif_tx_queue_stopped(tx_queue->core_txq))
455 efx_tx_maybe_stop_queue(tx_queue);
458 /* Check for available space. We should never need multiple
459 * descriptors per frame.
461 space = efx->txq_entries +
462 tx_queue->read_count - tx_queue->insert_count;
464 for (i = 0; i < n; i++) {
470 /* We'll want a descriptor for this tx. */
471 prefetchw(__efx_tx_queue_get_insert_buffer(tx_queue));
476 dma_addr = dma_map_single(&efx->pci_dev->dev,
479 if (dma_mapping_error(&efx->pci_dev->dev, dma_addr))
482 /* Create descriptor and set up for unmapping DMA. */
483 tx_buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
484 tx_buffer->xdpf = xdpf;
485 tx_buffer->flags = EFX_TX_BUF_XDP |
486 EFX_TX_BUF_MAP_SINGLE;
487 tx_buffer->dma_offset = 0;
488 tx_buffer->unmap_len = len;
489 tx_queue->tx_packets++;
492 /* Pass mapped frames to hardware. */
494 efx_nic_push_buffers(tx_queue);
497 if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
498 HARD_TX_UNLOCK(efx->net_dev, tx_queue->core_txq);
500 return i == 0 ? -EIO : i;
503 /* Initiate a packet transmission. We use one channel per CPU
504 * (sharing when we have more CPUs than channels).
506 * Context: non-blocking.
507 * Should always return NETDEV_TX_OK and consume the skb.
509 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
510 struct net_device *net_dev)
512 struct efx_nic *efx = netdev_priv(net_dev);
513 struct efx_tx_queue *tx_queue;
514 unsigned index, type;
516 EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
518 index = skb_get_queue_mapping(skb);
519 type = efx_tx_csum_type_skb(skb);
520 if (index >= efx->n_tx_channels) {
521 index -= efx->n_tx_channels;
522 type |= EFX_TXQ_TYPE_HIGHPRI;
525 /* PTP "event" packet */
526 if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
527 unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
528 /* There may be existing transmits on the channel that are
529 * waiting for this packet to trigger the doorbell write.
530 * We need to send the packets at this point.
532 efx_tx_send_pending(efx_get_tx_channel(efx, index));
533 return efx_ptp_tx(efx, skb);
536 tx_queue = efx_get_tx_queue(efx, index, type);
537 if (WARN_ON_ONCE(!tx_queue)) {
538 /* We don't have a TXQ of the right type.
539 * This should never happen, as we don't advertise offload
540 * features unless we can support them.
542 dev_kfree_skb_any(skb);
543 /* If we're not expecting another transmit and we had something to push
544 * on this queue or a partner queue then we need to push here to get the
545 * previous packets out.
547 if (!netdev_xmit_more())
548 efx_tx_send_pending(tx_queue->channel);
552 return __efx_enqueue_skb(tx_queue, skb);
555 void efx_xmit_done_single(struct efx_tx_queue *tx_queue)
557 unsigned int pkts_compl = 0, bytes_compl = 0;
558 unsigned int read_ptr;
559 bool finished = false;
561 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
564 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
566 if (!efx_tx_buffer_in_use(buffer)) {
567 struct efx_nic *efx = tx_queue->efx;
569 netif_err(efx, hw, efx->net_dev,
570 "TX queue %d spurious single TX completion\n",
572 efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
576 /* Need to check the flag before dequeueing. */
577 if (buffer->flags & EFX_TX_BUF_SKB)
579 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
581 ++tx_queue->read_count;
582 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
585 tx_queue->pkts_compl += pkts_compl;
586 tx_queue->bytes_compl += bytes_compl;
588 EFX_WARN_ON_PARANOID(pkts_compl != 1);
590 efx_xmit_done_check_empty(tx_queue);
593 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
595 struct efx_nic *efx = tx_queue->efx;
597 /* Must be inverse of queue lookup in efx_hard_start_xmit() */
599 netdev_get_tx_queue(efx->net_dev,
600 tx_queue->channel->channel +
601 ((tx_queue->type & EFX_TXQ_TYPE_HIGHPRI) ?
602 efx->n_tx_channels : 0));
605 int efx_setup_tc(struct net_device *net_dev, enum tc_setup_type type,
608 struct efx_nic *efx = netdev_priv(net_dev);
609 struct tc_mqprio_qopt *mqprio = type_data;
612 if (type != TC_SETUP_QDISC_MQPRIO)
615 /* Only Siena supported highpri queues */
616 if (efx_nic_rev(efx) > EFX_REV_SIENA_A0)
619 num_tc = mqprio->num_tc;
621 if (num_tc > EFX_MAX_TX_TC)
624 mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
626 if (num_tc == net_dev->num_tc)
629 for (tc = 0; tc < num_tc; tc++) {
630 net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
631 net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
634 net_dev->num_tc = num_tc;
636 return netif_set_real_num_tx_queues(net_dev,
637 max_t(int, num_tc, 1) *