1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3  * Driver for Solarflare network controllers and boards
4  * Copyright 2018 Solarflare Communications Inc.
5  * Copyright 2019-2020 Xilinx Inc.
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms of the GNU General Public License version 2 as published
9  * by the Free Software Foundation, incorporated herein by reference.
10  */
11 
12 #include <net/ip6_checksum.h>
13 
14 #include "net_driver.h"
15 #include "tx_common.h"
16 #include "nic_common.h"
17 #include "mcdi_functions.h"
18 #include "ef100_regs.h"
19 #include "io.h"
20 #include "ef100_tx.h"
21 #include "ef100_nic.h"
22 
23 int ef100_tx_probe(struct efx_tx_queue *tx_queue)
24 {
25 	/* Allocate an extra descriptor for the QMDA status completion entry */
26 	return efx_nic_alloc_buffer(tx_queue->efx, &tx_queue->txd.buf,
27 				    (tx_queue->ptr_mask + 2) *
28 				    sizeof(efx_oword_t),
29 				    GFP_KERNEL);
30 	return 0;
31 }
32 
33 void ef100_tx_init(struct efx_tx_queue *tx_queue)
34 {
35 	/* must be the inverse of lookup in efx_get_tx_channel */
36 	tx_queue->core_txq =
37 		netdev_get_tx_queue(tx_queue->efx->net_dev,
38 				    tx_queue->channel->channel -
39 				    tx_queue->efx->tx_channel_offset);
40 
41 	if (efx_mcdi_tx_init(tx_queue, false))
42 		netdev_WARN(tx_queue->efx->net_dev,
43 			    "failed to initialise TXQ %d\n", tx_queue->queue);
44 }
45 
46 static bool ef100_tx_can_tso(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
47 {
48 	struct efx_nic *efx = tx_queue->efx;
49 	struct ef100_nic_data *nic_data;
50 	struct efx_tx_buffer *buffer;
51 	struct tcphdr *tcphdr;
52 	struct iphdr *iphdr;
53 	size_t header_len;
54 	u32 mss;
55 
56 	nic_data = efx->nic_data;
57 
58 	if (!skb_is_gso_tcp(skb))
59 		return false;
60 	if (!(efx->net_dev->features & NETIF_F_TSO))
61 		return false;
62 
63 	mss = skb_shinfo(skb)->gso_size;
64 	if (unlikely(mss < 4)) {
65 		WARN_ONCE(1, "MSS of %u is too small for TSO\n", mss);
66 		return false;
67 	}
68 
69 	header_len = efx_tx_tso_header_length(skb);
70 	if (header_len > nic_data->tso_max_hdr_len)
71 		return false;
72 
73 	if (skb_shinfo(skb)->gso_segs > nic_data->tso_max_payload_num_segs) {
74 		/* net_dev->gso_max_segs should've caught this */
75 		WARN_ON_ONCE(1);
76 		return false;
77 	}
78 
79 	if (skb->data_len / mss > nic_data->tso_max_frames)
80 		return false;
81 
82 	/* net_dev->gso_max_size should've caught this */
83 	if (WARN_ON_ONCE(skb->data_len > nic_data->tso_max_payload_len))
84 		return false;
85 
86 	/* Reserve an empty buffer for the TSO V3 descriptor.
87 	 * Convey the length of the header since we already know it.
88 	 */
89 	buffer = efx_tx_queue_get_insert_buffer(tx_queue);
90 	buffer->flags = EFX_TX_BUF_TSO_V3 | EFX_TX_BUF_CONT;
91 	buffer->len = header_len;
92 	buffer->unmap_len = 0;
93 	buffer->skb = skb;
94 	++tx_queue->insert_count;
95 
96 	/* Adjust the TCP checksum to exclude the total length, since we set
97 	 * ED_INNER_IP_LEN in the descriptor.
98 	 */
99 	tcphdr = tcp_hdr(skb);
100 	if (skb_is_gso_v6(skb)) {
101 		tcphdr->check = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
102 						 &ipv6_hdr(skb)->daddr,
103 						 0, IPPROTO_TCP, 0);
104 	} else {
105 		iphdr = ip_hdr(skb);
106 		tcphdr->check = ~csum_tcpudp_magic(iphdr->saddr, iphdr->daddr,
107 						   0, IPPROTO_TCP, 0);
108 	}
109 	return true;
110 }
111 
112 static efx_oword_t *ef100_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
113 {
114 	if (likely(tx_queue->txd.buf.addr))
115 		return ((efx_oword_t *)tx_queue->txd.buf.addr) + index;
116 	else
117 		return NULL;
118 }
119 
120 void ef100_notify_tx_desc(struct efx_tx_queue *tx_queue)
121 {
122 	unsigned int write_ptr;
123 	efx_dword_t reg;
124 
125 	if (unlikely(tx_queue->notify_count == tx_queue->write_count))
126 		return;
127 
128 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
129 	/* The write pointer goes into the high word */
130 	EFX_POPULATE_DWORD_1(reg, ERF_GZ_TX_RING_PIDX, write_ptr);
131 	efx_writed_page(tx_queue->efx, &reg,
132 			ER_GZ_TX_RING_DOORBELL, tx_queue->queue);
133 	tx_queue->notify_count = tx_queue->write_count;
134 	tx_queue->xmit_more_available = false;
135 }
136 
137 static void ef100_tx_push_buffers(struct efx_tx_queue *tx_queue)
138 {
139 	ef100_notify_tx_desc(tx_queue);
140 	++tx_queue->pushes;
141 }
142 
143 static void ef100_set_tx_csum_partial(const struct sk_buff *skb,
144 				      struct efx_tx_buffer *buffer, efx_oword_t *txd)
145 {
146 	efx_oword_t csum;
147 	int csum_start;
148 
149 	if (!skb || skb->ip_summed != CHECKSUM_PARTIAL)
150 		return;
151 
152 	/* skb->csum_start has the offset from head, but we need the offset
153 	 * from data.
154 	 */
155 	csum_start = skb_checksum_start_offset(skb);
156 	EFX_POPULATE_OWORD_3(csum,
157 			     ESF_GZ_TX_SEND_CSO_PARTIAL_EN, 1,
158 			     ESF_GZ_TX_SEND_CSO_PARTIAL_START_W,
159 			     csum_start >> 1,
160 			     ESF_GZ_TX_SEND_CSO_PARTIAL_CSUM_W,
161 			     skb->csum_offset >> 1);
162 	EFX_OR_OWORD(*txd, *txd, csum);
163 }
164 
165 static void ef100_set_tx_hw_vlan(const struct sk_buff *skb, efx_oword_t *txd)
166 {
167 	u16 vlan_tci = skb_vlan_tag_get(skb);
168 	efx_oword_t vlan;
169 
170 	EFX_POPULATE_OWORD_2(vlan,
171 			     ESF_GZ_TX_SEND_VLAN_INSERT_EN, 1,
172 			     ESF_GZ_TX_SEND_VLAN_INSERT_TCI, vlan_tci);
173 	EFX_OR_OWORD(*txd, *txd, vlan);
174 }
175 
176 static void ef100_make_send_desc(struct efx_nic *efx,
177 				 const struct sk_buff *skb,
178 				 struct efx_tx_buffer *buffer, efx_oword_t *txd,
179 				 unsigned int segment_count)
180 {
181 	/* TX send descriptor */
182 	EFX_POPULATE_OWORD_3(*txd,
183 			     ESF_GZ_TX_SEND_NUM_SEGS, segment_count,
184 			     ESF_GZ_TX_SEND_LEN, buffer->len,
185 			     ESF_GZ_TX_SEND_ADDR, buffer->dma_addr);
186 
187 	if (likely(efx->net_dev->features & NETIF_F_HW_CSUM))
188 		ef100_set_tx_csum_partial(skb, buffer, txd);
189 	if (efx->net_dev->features & NETIF_F_HW_VLAN_CTAG_TX &&
190 	    skb && skb_vlan_tag_present(skb))
191 		ef100_set_tx_hw_vlan(skb, txd);
192 }
193 
194 static void ef100_make_tso_desc(struct efx_nic *efx,
195 				const struct sk_buff *skb,
196 				struct efx_tx_buffer *buffer, efx_oword_t *txd,
197 				unsigned int segment_count)
198 {
199 	u32 mangleid = (efx->net_dev->features & NETIF_F_TSO_MANGLEID) ||
200 		skb_shinfo(skb)->gso_type & SKB_GSO_TCP_FIXEDID ?
201 		ESE_GZ_TX_DESC_IP4_ID_NO_OP :
202 		ESE_GZ_TX_DESC_IP4_ID_INC_MOD16;
203 	u16 vlan_enable =  efx->net_dev->features & NETIF_F_HW_VLAN_CTAG_TX ?
204 		skb_vlan_tag_present(skb) : 0;
205 	unsigned int len, ip_offset, tcp_offset, payload_segs;
206 	u16 vlan_tci = skb_vlan_tag_get(skb);
207 	u32 mss = skb_shinfo(skb)->gso_size;
208 
209 	len = skb->len - buffer->len;
210 	/* We use 1 for the TSO descriptor and 1 for the header */
211 	payload_segs = segment_count - 2;
212 	ip_offset =  skb_network_offset(skb);
213 	tcp_offset = skb_transport_offset(skb);
214 
215 	EFX_POPULATE_OWORD_13(*txd,
216 			      ESF_GZ_TX_DESC_TYPE, ESE_GZ_TX_DESC_TYPE_TSO,
217 			      ESF_GZ_TX_TSO_MSS, mss,
218 			      ESF_GZ_TX_TSO_HDR_NUM_SEGS, 1,
219 			      ESF_GZ_TX_TSO_PAYLOAD_NUM_SEGS, payload_segs,
220 			      ESF_GZ_TX_TSO_HDR_LEN_W, buffer->len >> 1,
221 			      ESF_GZ_TX_TSO_PAYLOAD_LEN, len,
222 			      ESF_GZ_TX_TSO_CSO_INNER_L4, 1,
223 			      ESF_GZ_TX_TSO_INNER_L3_OFF_W, ip_offset >> 1,
224 			      ESF_GZ_TX_TSO_INNER_L4_OFF_W, tcp_offset >> 1,
225 			      ESF_GZ_TX_TSO_ED_INNER_IP4_ID, mangleid,
226 			      ESF_GZ_TX_TSO_ED_INNER_IP_LEN, 1,
227 			      ESF_GZ_TX_TSO_VLAN_INSERT_EN, vlan_enable,
228 			      ESF_GZ_TX_TSO_VLAN_INSERT_TCI, vlan_tci
229 		);
230 }
231 
232 static void ef100_tx_make_descriptors(struct efx_tx_queue *tx_queue,
233 				      const struct sk_buff *skb,
234 				      unsigned int segment_count)
235 {
236 	unsigned int old_write_count = tx_queue->write_count;
237 	unsigned int new_write_count = old_write_count;
238 	struct efx_tx_buffer *buffer;
239 	unsigned int next_desc_type;
240 	unsigned int write_ptr;
241 	efx_oword_t *txd;
242 	unsigned int nr_descs = tx_queue->insert_count - old_write_count;
243 
244 	if (unlikely(nr_descs == 0))
245 		return;
246 
247 	if (segment_count)
248 		next_desc_type = ESE_GZ_TX_DESC_TYPE_TSO;
249 	else
250 		next_desc_type = ESE_GZ_TX_DESC_TYPE_SEND;
251 
252 	/* if it's a raw write (such as XDP) then always SEND single frames */
253 	if (!skb)
254 		nr_descs = 1;
255 
256 	do {
257 		write_ptr = new_write_count & tx_queue->ptr_mask;
258 		buffer = &tx_queue->buffer[write_ptr];
259 		txd = ef100_tx_desc(tx_queue, write_ptr);
260 		++new_write_count;
261 
262 		/* Create TX descriptor ring entry */
263 		tx_queue->packet_write_count = new_write_count;
264 
265 		switch (next_desc_type) {
266 		case ESE_GZ_TX_DESC_TYPE_SEND:
267 			ef100_make_send_desc(tx_queue->efx, skb,
268 					     buffer, txd, nr_descs);
269 			break;
270 		case ESE_GZ_TX_DESC_TYPE_TSO:
271 			/* TX TSO descriptor */
272 			WARN_ON_ONCE(!(buffer->flags & EFX_TX_BUF_TSO_V3));
273 			ef100_make_tso_desc(tx_queue->efx, skb,
274 					    buffer, txd, nr_descs);
275 			break;
276 		default:
277 			/* TX segment descriptor */
278 			EFX_POPULATE_OWORD_3(*txd,
279 					     ESF_GZ_TX_DESC_TYPE, ESE_GZ_TX_DESC_TYPE_SEG,
280 					     ESF_GZ_TX_SEG_LEN, buffer->len,
281 					     ESF_GZ_TX_SEG_ADDR, buffer->dma_addr);
282 		}
283 		/* if it's a raw write (such as XDP) then always SEND */
284 		next_desc_type = skb ? ESE_GZ_TX_DESC_TYPE_SEG :
285 				       ESE_GZ_TX_DESC_TYPE_SEND;
286 
287 	} while (new_write_count != tx_queue->insert_count);
288 
289 	wmb(); /* Ensure descriptors are written before they are fetched */
290 
291 	tx_queue->write_count = new_write_count;
292 
293 	/* The write_count above must be updated before reading
294 	 * channel->holdoff_doorbell to avoid a race with the
295 	 * completion path, so ensure these operations are not
296 	 * re-ordered.  This also flushes the update of write_count
297 	 * back into the cache.
298 	 */
299 	smp_mb();
300 }
301 
302 void ef100_tx_write(struct efx_tx_queue *tx_queue)
303 {
304 	ef100_tx_make_descriptors(tx_queue, NULL, 0);
305 	ef100_tx_push_buffers(tx_queue);
306 }
307 
308 void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event)
309 {
310 	unsigned int tx_done =
311 		EFX_QWORD_FIELD(*p_event, ESF_GZ_EV_TXCMPL_NUM_DESC);
312 	unsigned int qlabel =
313 		EFX_QWORD_FIELD(*p_event, ESF_GZ_EV_TXCMPL_Q_LABEL);
314 	struct efx_tx_queue *tx_queue =
315 		efx_channel_get_tx_queue(channel, qlabel);
316 	unsigned int tx_index = (tx_queue->read_count + tx_done - 1) &
317 				tx_queue->ptr_mask;
318 
319 	efx_xmit_done(tx_queue, tx_index);
320 }
321 
322 /* Add a socket buffer to a TX queue
323  *
324  * You must hold netif_tx_lock() to call this function.
325  *
326  * Returns 0 on success, error code otherwise. In case of an error this
327  * function will free the SKB.
328  */
329 int ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
330 {
331 	unsigned int old_insert_count = tx_queue->insert_count;
332 	struct efx_nic *efx = tx_queue->efx;
333 	bool xmit_more = netdev_xmit_more();
334 	unsigned int fill_level;
335 	unsigned int segments;
336 	int rc;
337 
338 	if (!tx_queue->buffer || !tx_queue->ptr_mask) {
339 		netif_stop_queue(efx->net_dev);
340 		dev_kfree_skb_any(skb);
341 		return -ENODEV;
342 	}
343 
344 	segments = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 0;
345 	if (segments == 1)
346 		segments = 0;	/* Don't use TSO/GSO for a single segment. */
347 	if (segments && !ef100_tx_can_tso(tx_queue, skb)) {
348 		rc = efx_tx_tso_fallback(tx_queue, skb);
349 		tx_queue->tso_fallbacks++;
350 		if (rc)
351 			goto err;
352 		else
353 			return 0;
354 	}
355 
356 	/* Map for DMA and create descriptors */
357 	rc = efx_tx_map_data(tx_queue, skb, segments);
358 	if (rc)
359 		goto err;
360 	ef100_tx_make_descriptors(tx_queue, skb, segments);
361 
362 	fill_level = efx_channel_tx_fill_level(tx_queue->channel);
363 	if (fill_level > efx->txq_stop_thresh) {
364 		netif_tx_stop_queue(tx_queue->core_txq);
365 		/* Re-read after a memory barrier in case we've raced with
366 		 * the completion path. Otherwise there's a danger we'll never
367 		 * restart the queue if all completions have just happened.
368 		 */
369 		smp_mb();
370 		fill_level = efx_channel_tx_fill_level(tx_queue->channel);
371 		if (fill_level < efx->txq_stop_thresh)
372 			netif_tx_start_queue(tx_queue->core_txq);
373 	}
374 
375 	if (__netdev_tx_sent_queue(tx_queue->core_txq, skb->len, xmit_more))
376 		tx_queue->xmit_more_available = false; /* push doorbell */
377 	else if (tx_queue->write_count - tx_queue->notify_count > 255)
378 		/* Ensure we never push more than 256 packets at once */
379 		tx_queue->xmit_more_available = false; /* push */
380 	else
381 		tx_queue->xmit_more_available = true; /* don't push yet */
382 
383 	if (!tx_queue->xmit_more_available)
384 		ef100_tx_push_buffers(tx_queue);
385 
386 	if (segments) {
387 		tx_queue->tso_bursts++;
388 		tx_queue->tso_packets += segments;
389 		tx_queue->tx_packets  += segments;
390 	} else {
391 		tx_queue->tx_packets++;
392 	}
393 	return 0;
394 
395 err:
396 	efx_enqueue_unwind(tx_queue, old_insert_count);
397 	if (!IS_ERR_OR_NULL(skb))
398 		dev_kfree_skb_any(skb);
399 
400 	/* If we're not expecting another transmit and we had something to push
401 	 * on this queue then we need to push here to get the previous packets
402 	 * out.  We only enter this branch from before the 'Update BQL' section
403 	 * above, so xmit_more_available still refers to the old state.
404 	 */
405 	if (tx_queue->xmit_more_available && !xmit_more)
406 		ef100_tx_push_buffers(tx_queue);
407 	return rc;
408 }
409