xref: /linux/drivers/net/ethernet/sfc/tx_common.c (revision f8e17c17b81070f38062dce79ca7f4541851dadd)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3  * Driver for Solarflare network controllers and boards
4  * Copyright 2018 Solarflare Communications Inc.
5  *
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.
9  */
10 
11 #include "net_driver.h"
12 #include "efx.h"
13 #include "nic.h"
14 #include "tx_common.h"
15 
16 static unsigned int efx_tx_cb_page_count(struct efx_tx_queue *tx_queue)
17 {
18 	return DIV_ROUND_UP(tx_queue->ptr_mask + 1,
19 			    PAGE_SIZE >> EFX_TX_CB_ORDER);
20 }
21 
22 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
23 {
24 	struct efx_nic *efx = tx_queue->efx;
25 	unsigned int entries;
26 	int rc;
27 
28 	/* Create the smallest power-of-two aligned ring */
29 	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
30 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
31 	tx_queue->ptr_mask = entries - 1;
32 
33 	netif_dbg(efx, probe, efx->net_dev,
34 		  "creating TX queue %d size %#x mask %#x\n",
35 		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
36 
37 	/* Allocate software ring */
38 	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
39 				   GFP_KERNEL);
40 	if (!tx_queue->buffer)
41 		return -ENOMEM;
42 
43 	tx_queue->cb_page = kcalloc(efx_tx_cb_page_count(tx_queue),
44 				    sizeof(tx_queue->cb_page[0]), GFP_KERNEL);
45 	if (!tx_queue->cb_page) {
46 		rc = -ENOMEM;
47 		goto fail1;
48 	}
49 
50 	/* Allocate hardware ring */
51 	rc = efx_nic_probe_tx(tx_queue);
52 	if (rc)
53 		goto fail2;
54 
55 	return 0;
56 
57 fail2:
58 	kfree(tx_queue->cb_page);
59 	tx_queue->cb_page = NULL;
60 fail1:
61 	kfree(tx_queue->buffer);
62 	tx_queue->buffer = NULL;
63 	return rc;
64 }
65 
66 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
67 {
68 	struct efx_nic *efx = tx_queue->efx;
69 
70 	netif_dbg(efx, drv, efx->net_dev,
71 		  "initialising TX queue %d\n", tx_queue->queue);
72 
73 	tx_queue->insert_count = 0;
74 	tx_queue->write_count = 0;
75 	tx_queue->packet_write_count = 0;
76 	tx_queue->old_write_count = 0;
77 	tx_queue->read_count = 0;
78 	tx_queue->old_read_count = 0;
79 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
80 	tx_queue->xmit_more_available = false;
81 	tx_queue->timestamping = (efx_ptp_use_mac_tx_timestamps(efx) &&
82 				  tx_queue->channel == efx_ptp_channel(efx));
83 	tx_queue->completed_desc_ptr = tx_queue->ptr_mask;
84 	tx_queue->completed_timestamp_major = 0;
85 	tx_queue->completed_timestamp_minor = 0;
86 
87 	tx_queue->xdp_tx = efx_channel_is_xdp_tx(tx_queue->channel);
88 
89 	/* Set up default function pointers. These may get replaced by
90 	 * efx_nic_init_tx() based off NIC/queue capabilities.
91 	 */
92 	tx_queue->handle_tso = efx_enqueue_skb_tso;
93 
94 	/* Set up TX descriptor ring */
95 	efx_nic_init_tx(tx_queue);
96 
97 	tx_queue->initialised = true;
98 }
99 
100 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
101 {
102 	struct efx_tx_buffer *buffer;
103 
104 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
105 		  "shutting down TX queue %d\n", tx_queue->queue);
106 
107 	if (!tx_queue->buffer)
108 		return;
109 
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 
114 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
115 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
116 
117 		++tx_queue->read_count;
118 	}
119 	tx_queue->xmit_more_available = false;
120 	netdev_tx_reset_queue(tx_queue->core_txq);
121 }
122 
123 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
124 {
125 	int i;
126 
127 	if (!tx_queue->buffer)
128 		return;
129 
130 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
131 		  "destroying TX queue %d\n", tx_queue->queue);
132 	efx_nic_remove_tx(tx_queue);
133 
134 	if (tx_queue->cb_page) {
135 		for (i = 0; i < efx_tx_cb_page_count(tx_queue); i++)
136 			efx_nic_free_buffer(tx_queue->efx,
137 					    &tx_queue->cb_page[i]);
138 		kfree(tx_queue->cb_page);
139 		tx_queue->cb_page = NULL;
140 	}
141 
142 	kfree(tx_queue->buffer);
143 	tx_queue->buffer = NULL;
144 }
145 
146 void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
147 			struct efx_tx_buffer *buffer,
148 			unsigned int *pkts_compl,
149 			unsigned int *bytes_compl)
150 {
151 	if (buffer->unmap_len) {
152 		struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
153 		dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
154 
155 		if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
156 			dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
157 					 DMA_TO_DEVICE);
158 		else
159 			dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
160 				       DMA_TO_DEVICE);
161 		buffer->unmap_len = 0;
162 	}
163 
164 	if (buffer->flags & EFX_TX_BUF_SKB) {
165 		struct sk_buff *skb = (struct sk_buff *)buffer->skb;
166 
167 		EFX_WARN_ON_PARANOID(!pkts_compl || !bytes_compl);
168 		(*pkts_compl)++;
169 		(*bytes_compl) += skb->len;
170 		if (tx_queue->timestamping &&
171 		    (tx_queue->completed_timestamp_major ||
172 		     tx_queue->completed_timestamp_minor)) {
173 			struct skb_shared_hwtstamps hwtstamp;
174 
175 			hwtstamp.hwtstamp =
176 				efx_ptp_nic_to_kernel_time(tx_queue);
177 			skb_tstamp_tx(skb, &hwtstamp);
178 
179 			tx_queue->completed_timestamp_major = 0;
180 			tx_queue->completed_timestamp_minor = 0;
181 		}
182 		dev_consume_skb_any((struct sk_buff *)buffer->skb);
183 		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
184 			   "TX queue %d transmission id %x complete\n",
185 			   tx_queue->queue, tx_queue->read_count);
186 	} else if (buffer->flags & EFX_TX_BUF_XDP) {
187 		xdp_return_frame_rx_napi(buffer->xdpf);
188 	}
189 
190 	buffer->len = 0;
191 	buffer->flags = 0;
192 }
193 
194 /* Remove packets from the TX queue
195  *
196  * This removes packets from the TX queue, up to and including the
197  * specified index.
198  */
199 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
200 				unsigned int index,
201 				unsigned int *pkts_compl,
202 				unsigned int *bytes_compl)
203 {
204 	struct efx_nic *efx = tx_queue->efx;
205 	unsigned int stop_index, read_ptr;
206 
207 	stop_index = (index + 1) & tx_queue->ptr_mask;
208 	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
209 
210 	while (read_ptr != stop_index) {
211 		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
212 
213 		if (!(buffer->flags & EFX_TX_BUF_OPTION) &&
214 		    unlikely(buffer->len == 0)) {
215 			netif_err(efx, tx_err, efx->net_dev,
216 				  "TX queue %d spurious TX completion id %x\n",
217 				  tx_queue->queue, read_ptr);
218 			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
219 			return;
220 		}
221 
222 		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
223 
224 		++tx_queue->read_count;
225 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
226 	}
227 }
228 
229 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
230 {
231 	unsigned int fill_level, pkts_compl = 0, bytes_compl = 0;
232 	struct efx_nic *efx = tx_queue->efx;
233 	struct efx_tx_queue *txq2;
234 
235 	EFX_WARN_ON_ONCE_PARANOID(index > tx_queue->ptr_mask);
236 
237 	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
238 	tx_queue->pkts_compl += pkts_compl;
239 	tx_queue->bytes_compl += bytes_compl;
240 
241 	if (pkts_compl > 1)
242 		++tx_queue->merge_events;
243 
244 	/* See if we need to restart the netif queue.  This memory
245 	 * barrier ensures that we write read_count (inside
246 	 * efx_dequeue_buffers()) before reading the queue status.
247 	 */
248 	smp_mb();
249 	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
250 	    likely(efx->port_enabled) &&
251 	    likely(netif_device_present(efx->net_dev))) {
252 		txq2 = efx_tx_queue_partner(tx_queue);
253 		fill_level = max(tx_queue->insert_count - tx_queue->read_count,
254 				 txq2->insert_count - txq2->read_count);
255 		if (fill_level <= efx->txq_wake_thresh)
256 			netif_tx_wake_queue(tx_queue->core_txq);
257 	}
258 
259 	/* Check whether the hardware queue is now empty */
260 	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
261 		tx_queue->old_write_count = READ_ONCE(tx_queue->write_count);
262 		if (tx_queue->read_count == tx_queue->old_write_count) {
263 			smp_mb();
264 			tx_queue->empty_read_count =
265 				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
266 		}
267 	}
268 }
269 
270 /* Remove buffers put into a tx_queue for the current packet.
271  * None of the buffers must have an skb attached.
272  */
273 void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
274 			unsigned int insert_count)
275 {
276 	struct efx_tx_buffer *buffer;
277 	unsigned int bytes_compl = 0;
278 	unsigned int pkts_compl = 0;
279 
280 	/* Work backwards until we hit the original insert pointer value */
281 	while (tx_queue->insert_count != insert_count) {
282 		--tx_queue->insert_count;
283 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
284 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
285 	}
286 }
287 
288 struct efx_tx_buffer *efx_tx_map_chunk(struct efx_tx_queue *tx_queue,
289 				       dma_addr_t dma_addr, size_t len)
290 {
291 	const struct efx_nic_type *nic_type = tx_queue->efx->type;
292 	struct efx_tx_buffer *buffer;
293 	unsigned int dma_len;
294 
295 	/* Map the fragment taking account of NIC-dependent DMA limits. */
296 	do {
297 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
298 		dma_len = nic_type->tx_limit_len(tx_queue, dma_addr, len);
299 
300 		buffer->len = dma_len;
301 		buffer->dma_addr = dma_addr;
302 		buffer->flags = EFX_TX_BUF_CONT;
303 		len -= dma_len;
304 		dma_addr += dma_len;
305 		++tx_queue->insert_count;
306 	} while (len);
307 
308 	return buffer;
309 }
310 
311 /* Map all data from an SKB for DMA and create descriptors on the queue. */
312 int efx_tx_map_data(struct efx_tx_queue *tx_queue, struct sk_buff *skb,
313 		    unsigned int segment_count)
314 {
315 	struct efx_nic *efx = tx_queue->efx;
316 	struct device *dma_dev = &efx->pci_dev->dev;
317 	unsigned int frag_index, nr_frags;
318 	dma_addr_t dma_addr, unmap_addr;
319 	unsigned short dma_flags;
320 	size_t len, unmap_len;
321 
322 	nr_frags = skb_shinfo(skb)->nr_frags;
323 	frag_index = 0;
324 
325 	/* Map header data. */
326 	len = skb_headlen(skb);
327 	dma_addr = dma_map_single(dma_dev, skb->data, len, DMA_TO_DEVICE);
328 	dma_flags = EFX_TX_BUF_MAP_SINGLE;
329 	unmap_len = len;
330 	unmap_addr = dma_addr;
331 
332 	if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
333 		return -EIO;
334 
335 	if (segment_count) {
336 		/* For TSO we need to put the header in to a separate
337 		 * descriptor. Map this separately if necessary.
338 		 */
339 		size_t header_len = skb_transport_header(skb) - skb->data +
340 				(tcp_hdr(skb)->doff << 2u);
341 
342 		if (header_len != len) {
343 			tx_queue->tso_long_headers++;
344 			efx_tx_map_chunk(tx_queue, dma_addr, header_len);
345 			len -= header_len;
346 			dma_addr += header_len;
347 		}
348 	}
349 
350 	/* Add descriptors for each fragment. */
351 	do {
352 		struct efx_tx_buffer *buffer;
353 		skb_frag_t *fragment;
354 
355 		buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
356 
357 		/* The final descriptor for a fragment is responsible for
358 		 * unmapping the whole fragment.
359 		 */
360 		buffer->flags = EFX_TX_BUF_CONT | dma_flags;
361 		buffer->unmap_len = unmap_len;
362 		buffer->dma_offset = buffer->dma_addr - unmap_addr;
363 
364 		if (frag_index >= nr_frags) {
365 			/* Store SKB details with the final buffer for
366 			 * the completion.
367 			 */
368 			buffer->skb = skb;
369 			buffer->flags = EFX_TX_BUF_SKB | dma_flags;
370 			return 0;
371 		}
372 
373 		/* Move on to the next fragment. */
374 		fragment = &skb_shinfo(skb)->frags[frag_index++];
375 		len = skb_frag_size(fragment);
376 		dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
377 					    DMA_TO_DEVICE);
378 		dma_flags = 0;
379 		unmap_len = len;
380 		unmap_addr = dma_addr;
381 
382 		if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
383 			return -EIO;
384 	} while (1);
385 }
386 
387 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
388 {
389 	/* Header and payload descriptor for each output segment, plus
390 	 * one for every input fragment boundary within a segment
391 	 */
392 	unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
393 
394 	/* Possibly one more per segment for option descriptors */
395 	if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
396 		max_descs += EFX_TSO_MAX_SEGS;
397 
398 	/* Possibly more for PCIe page boundaries within input fragments */
399 	if (PAGE_SIZE > EFX_PAGE_SIZE)
400 		max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
401 				   DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE));
402 
403 	return max_descs;
404 }
405