1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2018 Solarflare Communications Inc.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
11 #include "net_driver.h"
13 #include "nic_common.h"
14 #include "tx_common.h"
17 static unsigned int efx_tx_cb_page_count(struct efx_tx_queue *tx_queue)
19 return DIV_ROUND_UP(tx_queue->ptr_mask + 1,
20 PAGE_SIZE >> EFX_TX_CB_ORDER);
23 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
25 struct efx_nic *efx = tx_queue->efx;
29 /* Create the smallest power-of-two aligned ring */
30 entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
31 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
32 tx_queue->ptr_mask = entries - 1;
34 netif_dbg(efx, probe, efx->net_dev,
35 "creating TX queue %d size %#x mask %#x\n",
36 tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
38 /* Allocate software ring */
39 tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
41 if (!tx_queue->buffer)
44 tx_queue->cb_page = kcalloc(efx_tx_cb_page_count(tx_queue),
45 sizeof(tx_queue->cb_page[0]), GFP_KERNEL);
46 if (!tx_queue->cb_page) {
51 /* Allocate hardware ring, determine TXQ type */
52 rc = efx_nic_probe_tx(tx_queue);
56 tx_queue->channel->tx_queue_by_type[tx_queue->type] = tx_queue;
60 kfree(tx_queue->cb_page);
61 tx_queue->cb_page = NULL;
63 kfree(tx_queue->buffer);
64 tx_queue->buffer = NULL;
68 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
70 struct efx_nic *efx = tx_queue->efx;
72 netif_dbg(efx, drv, efx->net_dev,
73 "initialising TX queue %d\n", tx_queue->queue);
75 tx_queue->insert_count = 0;
76 tx_queue->notify_count = 0;
77 tx_queue->write_count = 0;
78 tx_queue->packet_write_count = 0;
79 tx_queue->old_write_count = 0;
80 tx_queue->read_count = 0;
81 tx_queue->old_read_count = 0;
82 tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
83 tx_queue->xmit_pending = false;
84 tx_queue->timestamping = (efx_ptp_use_mac_tx_timestamps(efx) &&
85 tx_queue->channel == efx_ptp_channel(efx));
86 tx_queue->completed_timestamp_major = 0;
87 tx_queue->completed_timestamp_minor = 0;
89 tx_queue->xdp_tx = efx_channel_is_xdp_tx(tx_queue->channel);
90 tx_queue->tso_version = 0;
92 /* Set up TX descriptor ring */
93 efx_nic_init_tx(tx_queue);
95 tx_queue->initialised = true;
98 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
100 struct efx_tx_buffer *buffer;
102 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
103 "shutting down TX queue %d\n", tx_queue->queue);
105 tx_queue->initialised = false;
107 if (!tx_queue->buffer)
110 /* Free any buffers left in the ring */
111 while (tx_queue->read_count != tx_queue->write_count) {
112 unsigned int pkts_compl = 0, bytes_compl = 0;
113 unsigned int efv_pkts_compl = 0;
115 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
116 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
119 ++tx_queue->read_count;
121 tx_queue->xmit_pending = false;
122 netdev_tx_reset_queue(tx_queue->core_txq);
125 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
129 if (!tx_queue->buffer)
132 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
133 "destroying TX queue %d\n", tx_queue->queue);
134 efx_nic_remove_tx(tx_queue);
136 if (tx_queue->cb_page) {
137 for (i = 0; i < efx_tx_cb_page_count(tx_queue); i++)
138 efx_nic_free_buffer(tx_queue->efx,
139 &tx_queue->cb_page[i]);
140 kfree(tx_queue->cb_page);
141 tx_queue->cb_page = NULL;
144 kfree(tx_queue->buffer);
145 tx_queue->buffer = NULL;
146 tx_queue->channel->tx_queue_by_type[tx_queue->type] = NULL;
149 void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
150 struct efx_tx_buffer *buffer,
151 unsigned int *pkts_compl,
152 unsigned int *bytes_compl,
153 unsigned int *efv_pkts_compl)
155 if (buffer->unmap_len) {
156 struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
157 dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
159 if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
160 dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
163 dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
165 buffer->unmap_len = 0;
168 if (buffer->flags & EFX_TX_BUF_SKB) {
169 struct sk_buff *skb = (struct sk_buff *)buffer->skb;
171 if (unlikely(buffer->flags & EFX_TX_BUF_EFV)) {
172 EFX_WARN_ON_PARANOID(!efv_pkts_compl);
175 EFX_WARN_ON_PARANOID(!pkts_compl || !bytes_compl);
177 (*bytes_compl) += skb->len;
180 if (tx_queue->timestamping &&
181 (tx_queue->completed_timestamp_major ||
182 tx_queue->completed_timestamp_minor)) {
183 struct skb_shared_hwtstamps hwtstamp;
186 efx_ptp_nic_to_kernel_time(tx_queue);
187 skb_tstamp_tx(skb, &hwtstamp);
189 tx_queue->completed_timestamp_major = 0;
190 tx_queue->completed_timestamp_minor = 0;
192 dev_consume_skb_any((struct sk_buff *)buffer->skb);
193 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
194 "TX queue %d transmission id %x complete\n",
195 tx_queue->queue, tx_queue->read_count);
196 } else if (buffer->flags & EFX_TX_BUF_XDP) {
197 xdp_return_frame_rx_napi(buffer->xdpf);
204 /* Remove packets from the TX queue
206 * This removes packets from the TX queue, up to and including the
209 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
211 unsigned int *pkts_compl,
212 unsigned int *bytes_compl,
213 unsigned int *efv_pkts_compl)
215 struct efx_nic *efx = tx_queue->efx;
216 unsigned int stop_index, read_ptr;
218 stop_index = (index + 1) & tx_queue->ptr_mask;
219 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
221 while (read_ptr != stop_index) {
222 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
224 if (!efx_tx_buffer_in_use(buffer)) {
225 netif_err(efx, tx_err, efx->net_dev,
226 "TX queue %d spurious TX completion id %d\n",
227 tx_queue->queue, read_ptr);
228 efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
232 efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl,
235 ++tx_queue->read_count;
236 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
240 void efx_xmit_done_check_empty(struct efx_tx_queue *tx_queue)
242 if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
243 tx_queue->old_write_count = READ_ONCE(tx_queue->write_count);
244 if (tx_queue->read_count == tx_queue->old_write_count) {
245 /* Ensure that read_count is flushed. */
247 tx_queue->empty_read_count =
248 tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
253 int efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
255 unsigned int fill_level, pkts_compl = 0, bytes_compl = 0;
256 unsigned int efv_pkts_compl = 0;
257 struct efx_nic *efx = tx_queue->efx;
259 EFX_WARN_ON_ONCE_PARANOID(index > tx_queue->ptr_mask);
261 efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl,
263 tx_queue->pkts_compl += pkts_compl;
264 tx_queue->bytes_compl += bytes_compl;
266 if (pkts_compl + efv_pkts_compl > 1)
267 ++tx_queue->merge_events;
269 /* See if we need to restart the netif queue. This memory
270 * barrier ensures that we write read_count (inside
271 * efx_dequeue_buffers()) before reading the queue status.
274 if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
275 likely(efx->port_enabled) &&
276 likely(netif_device_present(efx->net_dev))) {
277 fill_level = efx_channel_tx_fill_level(tx_queue->channel);
278 if (fill_level <= efx->txq_wake_thresh)
279 netif_tx_wake_queue(tx_queue->core_txq);
282 efx_xmit_done_check_empty(tx_queue);
284 return pkts_compl + efv_pkts_compl;
287 /* Remove buffers put into a tx_queue for the current packet.
288 * None of the buffers must have an skb attached.
290 void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
291 unsigned int insert_count)
293 unsigned int efv_pkts_compl = 0;
294 struct efx_tx_buffer *buffer;
295 unsigned int bytes_compl = 0;
296 unsigned int pkts_compl = 0;
298 /* Work backwards until we hit the original insert pointer value */
299 while (tx_queue->insert_count != insert_count) {
300 --tx_queue->insert_count;
301 buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
302 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
307 struct efx_tx_buffer *efx_tx_map_chunk(struct efx_tx_queue *tx_queue,
308 dma_addr_t dma_addr, size_t len)
310 const struct efx_nic_type *nic_type = tx_queue->efx->type;
311 struct efx_tx_buffer *buffer;
312 unsigned int dma_len;
314 /* Map the fragment taking account of NIC-dependent DMA limits. */
316 buffer = efx_tx_queue_get_insert_buffer(tx_queue);
318 if (nic_type->tx_limit_len)
319 dma_len = nic_type->tx_limit_len(tx_queue, dma_addr, len);
323 buffer->len = dma_len;
324 buffer->dma_addr = dma_addr;
325 buffer->flags = EFX_TX_BUF_CONT;
328 ++tx_queue->insert_count;
334 int efx_tx_tso_header_length(struct sk_buff *skb)
338 if (skb->encapsulation)
339 header_len = skb_inner_transport_header(skb) -
341 (inner_tcp_hdr(skb)->doff << 2u);
343 header_len = skb_transport_header(skb) - skb->data +
344 (tcp_hdr(skb)->doff << 2u);
348 /* Map all data from an SKB for DMA and create descriptors on the queue. */
349 int efx_tx_map_data(struct efx_tx_queue *tx_queue, struct sk_buff *skb,
350 unsigned int segment_count)
352 struct efx_nic *efx = tx_queue->efx;
353 struct device *dma_dev = &efx->pci_dev->dev;
354 unsigned int frag_index, nr_frags;
355 dma_addr_t dma_addr, unmap_addr;
356 unsigned short dma_flags;
357 size_t len, unmap_len;
359 nr_frags = skb_shinfo(skb)->nr_frags;
362 /* Map header data. */
363 len = skb_headlen(skb);
364 dma_addr = dma_map_single(dma_dev, skb->data, len, DMA_TO_DEVICE);
365 dma_flags = EFX_TX_BUF_MAP_SINGLE;
367 unmap_addr = dma_addr;
369 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
373 /* For TSO we need to put the header in to a separate
374 * descriptor. Map this separately if necessary.
376 size_t header_len = efx_tx_tso_header_length(skb);
378 if (header_len != len) {
379 tx_queue->tso_long_headers++;
380 efx_tx_map_chunk(tx_queue, dma_addr, header_len);
382 dma_addr += header_len;
386 /* Add descriptors for each fragment. */
388 struct efx_tx_buffer *buffer;
389 skb_frag_t *fragment;
391 buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
393 /* The final descriptor for a fragment is responsible for
394 * unmapping the whole fragment.
396 buffer->flags = EFX_TX_BUF_CONT | dma_flags;
397 buffer->unmap_len = unmap_len;
398 buffer->dma_offset = buffer->dma_addr - unmap_addr;
400 if (frag_index >= nr_frags) {
401 /* Store SKB details with the final buffer for
405 buffer->flags = EFX_TX_BUF_SKB | dma_flags;
409 /* Move on to the next fragment. */
410 fragment = &skb_shinfo(skb)->frags[frag_index++];
411 len = skb_frag_size(fragment);
412 dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
416 unmap_addr = dma_addr;
418 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
423 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
425 /* Header and payload descriptor for each output segment, plus
426 * one for every input fragment boundary within a segment
428 unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
430 /* Possibly one more per segment for option descriptors */
431 if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
432 max_descs += EFX_TSO_MAX_SEGS;
434 /* Possibly more for PCIe page boundaries within input fragments */
435 if (PAGE_SIZE > EFX_PAGE_SIZE)
436 max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
437 DIV_ROUND_UP(GSO_LEGACY_MAX_SIZE,
444 * Fallback to software TSO.
446 * This is used if we are unable to send a GSO packet through hardware TSO.
447 * This should only ever happen due to per-queue restrictions - unsupported
448 * packets should first be filtered by the feature flags.
450 * Returns 0 on success, error code otherwise.
452 int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
454 struct sk_buff *segments, *next;
456 segments = skb_gso_segment(skb, 0);
457 if (IS_ERR(segments))
458 return PTR_ERR(segments);
460 dev_consume_skb_any(skb);
462 skb_list_walk_safe(segments, skb, next) {
463 skb_mark_not_on_list(skb);
464 efx_enqueue_skb(tx_queue, skb);