1 // SPDX-License-Identifier: GPL-2.0-only OR BSD-3-Clause
2 
3 /* Packet transmit logic for Mellanox Gigabit Ethernet driver
4  *
5  * Copyright (C) 2020-2021 NVIDIA CORPORATION & AFFILIATES
6  */
7 
8 #include <linux/skbuff.h>
9 
10 #include "mlxbf_gige.h"
11 #include "mlxbf_gige_regs.h"
12 
13 /* Transmit Initialization
14  * 1) Allocates TX WQE array using coherent DMA mapping
15  * 2) Allocates TX completion counter using coherent DMA mapping
16  */
mlxbf_gige_tx_init(struct mlxbf_gige * priv)17 int mlxbf_gige_tx_init(struct mlxbf_gige *priv)
18 {
19 	size_t size;
20 
21 	size = MLXBF_GIGE_TX_WQE_SZ * priv->tx_q_entries;
22 	priv->tx_wqe_base = dma_alloc_coherent(priv->dev, size,
23 					       &priv->tx_wqe_base_dma,
24 					       GFP_KERNEL);
25 	if (!priv->tx_wqe_base)
26 		return -ENOMEM;
27 
28 	priv->tx_wqe_next = priv->tx_wqe_base;
29 
30 	/* Write TX WQE base address into MMIO reg */
31 	writeq(priv->tx_wqe_base_dma, priv->base + MLXBF_GIGE_TX_WQ_BASE);
32 
33 	/* Allocate address for TX completion count */
34 	priv->tx_cc = dma_alloc_coherent(priv->dev, MLXBF_GIGE_TX_CC_SZ,
35 					 &priv->tx_cc_dma, GFP_KERNEL);
36 	if (!priv->tx_cc) {
37 		dma_free_coherent(priv->dev, size,
38 				  priv->tx_wqe_base, priv->tx_wqe_base_dma);
39 		return -ENOMEM;
40 	}
41 
42 	/* Write TX CC base address into MMIO reg */
43 	writeq(priv->tx_cc_dma, priv->base + MLXBF_GIGE_TX_CI_UPDATE_ADDRESS);
44 
45 	writeq(ilog2(priv->tx_q_entries),
46 	       priv->base + MLXBF_GIGE_TX_WQ_SIZE_LOG2);
47 
48 	priv->prev_tx_ci = 0;
49 	priv->tx_pi = 0;
50 
51 	return 0;
52 }
53 
54 /* Transmit Deinitialization
55  * This routine will free allocations done by mlxbf_gige_tx_init(),
56  * namely the TX WQE array and the TX completion counter
57  */
mlxbf_gige_tx_deinit(struct mlxbf_gige * priv)58 void mlxbf_gige_tx_deinit(struct mlxbf_gige *priv)
59 {
60 	u64 *tx_wqe_addr;
61 	size_t size;
62 	int i;
63 
64 	tx_wqe_addr = priv->tx_wqe_base;
65 
66 	for (i = 0; i < priv->tx_q_entries; i++) {
67 		if (priv->tx_skb[i]) {
68 			dma_unmap_single(priv->dev, *tx_wqe_addr,
69 					 priv->tx_skb[i]->len, DMA_TO_DEVICE);
70 			dev_kfree_skb(priv->tx_skb[i]);
71 			priv->tx_skb[i] = NULL;
72 		}
73 		tx_wqe_addr += 2;
74 	}
75 
76 	size = MLXBF_GIGE_TX_WQE_SZ * priv->tx_q_entries;
77 	dma_free_coherent(priv->dev, size,
78 			  priv->tx_wqe_base, priv->tx_wqe_base_dma);
79 
80 	dma_free_coherent(priv->dev, MLXBF_GIGE_TX_CC_SZ,
81 			  priv->tx_cc, priv->tx_cc_dma);
82 
83 	priv->tx_wqe_base = NULL;
84 	priv->tx_wqe_base_dma = 0;
85 	priv->tx_cc = NULL;
86 	priv->tx_cc_dma = 0;
87 	priv->tx_wqe_next = NULL;
88 	writeq(0, priv->base + MLXBF_GIGE_TX_WQ_BASE);
89 	writeq(0, priv->base + MLXBF_GIGE_TX_CI_UPDATE_ADDRESS);
90 }
91 
92 /* Function that returns status of TX ring:
93  *          0: TX ring is full, i.e. there are no
94  *             available un-used entries in TX ring.
95  *   non-null: TX ring is not full, i.e. there are
96  *             some available entries in TX ring.
97  *             The non-null value is a measure of
98  *             how many TX entries are available, but
99  *             it is not the exact number of available
100  *             entries (see below).
101  *
102  * The algorithm makes the assumption that if
103  * (prev_tx_ci == tx_pi) then the TX ring is empty.
104  * An empty ring actually has (tx_q_entries-1)
105  * entries, which allows the algorithm to differentiate
106  * the case of an empty ring vs. a full ring.
107  */
mlxbf_gige_tx_buffs_avail(struct mlxbf_gige * priv)108 static u16 mlxbf_gige_tx_buffs_avail(struct mlxbf_gige *priv)
109 {
110 	unsigned long flags;
111 	u16 avail;
112 
113 	spin_lock_irqsave(&priv->lock, flags);
114 
115 	if (priv->prev_tx_ci == priv->tx_pi)
116 		avail = priv->tx_q_entries - 1;
117 	else
118 		avail = ((priv->tx_q_entries + priv->prev_tx_ci - priv->tx_pi)
119 			  % priv->tx_q_entries) - 1;
120 
121 	spin_unlock_irqrestore(&priv->lock, flags);
122 
123 	return avail;
124 }
125 
mlxbf_gige_handle_tx_complete(struct mlxbf_gige * priv)126 bool mlxbf_gige_handle_tx_complete(struct mlxbf_gige *priv)
127 {
128 	struct net_device_stats *stats;
129 	u16 tx_wqe_index;
130 	u64 *tx_wqe_addr;
131 	u64 tx_status;
132 	u16 tx_ci;
133 
134 	tx_status = readq(priv->base + MLXBF_GIGE_TX_STATUS);
135 	if (tx_status & MLXBF_GIGE_TX_STATUS_DATA_FIFO_FULL)
136 		priv->stats.tx_fifo_full++;
137 	tx_ci = readq(priv->base + MLXBF_GIGE_TX_CONSUMER_INDEX);
138 	stats = &priv->netdev->stats;
139 
140 	/* Transmit completion logic needs to loop until the completion
141 	 * index (in SW) equals TX consumer index (from HW).  These
142 	 * parameters are unsigned 16-bit values and the wrap case needs
143 	 * to be supported, that is TX consumer index wrapped from 0xFFFF
144 	 * to 0 while TX completion index is still < 0xFFFF.
145 	 */
146 	for (; priv->prev_tx_ci != tx_ci; priv->prev_tx_ci++) {
147 		tx_wqe_index = priv->prev_tx_ci % priv->tx_q_entries;
148 		/* Each TX WQE is 16 bytes. The 8 MSB store the 2KB TX
149 		 * buffer address and the 8 LSB contain information
150 		 * about the TX WQE.
151 		 */
152 		tx_wqe_addr = priv->tx_wqe_base +
153 			       (tx_wqe_index * MLXBF_GIGE_TX_WQE_SZ_QWORDS);
154 
155 		stats->tx_packets++;
156 		stats->tx_bytes += MLXBF_GIGE_TX_WQE_PKT_LEN(tx_wqe_addr);
157 
158 		dma_unmap_single(priv->dev, *tx_wqe_addr,
159 				 priv->tx_skb[tx_wqe_index]->len, DMA_TO_DEVICE);
160 		dev_consume_skb_any(priv->tx_skb[tx_wqe_index]);
161 		priv->tx_skb[tx_wqe_index] = NULL;
162 
163 		/* Ensure completion of updates across all cores */
164 		mb();
165 	}
166 
167 	/* Since the TX ring was likely just drained, check if TX queue
168 	 * had previously been stopped and now that there are TX buffers
169 	 * available the TX queue can be awakened.
170 	 */
171 	if (netif_queue_stopped(priv->netdev) &&
172 	    mlxbf_gige_tx_buffs_avail(priv))
173 		netif_wake_queue(priv->netdev);
174 
175 	return true;
176 }
177 
178 /* Function to advance the tx_wqe_next pointer to next TX WQE */
mlxbf_gige_update_tx_wqe_next(struct mlxbf_gige * priv)179 void mlxbf_gige_update_tx_wqe_next(struct mlxbf_gige *priv)
180 {
181 	/* Advance tx_wqe_next pointer */
182 	priv->tx_wqe_next += MLXBF_GIGE_TX_WQE_SZ_QWORDS;
183 
184 	/* Check if 'next' pointer is beyond end of TX ring */
185 	/* If so, set 'next' back to 'base' pointer of ring */
186 	if (priv->tx_wqe_next == (priv->tx_wqe_base +
187 				  (priv->tx_q_entries * MLXBF_GIGE_TX_WQE_SZ_QWORDS)))
188 		priv->tx_wqe_next = priv->tx_wqe_base;
189 }
190 
mlxbf_gige_start_xmit(struct sk_buff * skb,struct net_device * netdev)191 netdev_tx_t mlxbf_gige_start_xmit(struct sk_buff *skb,
192 				  struct net_device *netdev)
193 {
194 	struct mlxbf_gige *priv = netdev_priv(netdev);
195 	long buff_addr, start_dma_page, end_dma_page;
196 	struct sk_buff *tx_skb;
197 	dma_addr_t tx_buf_dma;
198 	unsigned long flags;
199 	u64 *tx_wqe_addr;
200 	u64 word2;
201 
202 	/* If needed, linearize TX SKB as hardware DMA expects this */
203 	if (skb->len > MLXBF_GIGE_DEFAULT_BUF_SZ || skb_linearize(skb)) {
204 		dev_kfree_skb(skb);
205 		netdev->stats.tx_dropped++;
206 		return NETDEV_TX_OK;
207 	}
208 
209 	buff_addr = (long)skb->data;
210 	start_dma_page = buff_addr >> MLXBF_GIGE_DMA_PAGE_SHIFT;
211 	end_dma_page   = (buff_addr + skb->len - 1) >> MLXBF_GIGE_DMA_PAGE_SHIFT;
212 
213 	/* Verify that payload pointer and data length of SKB to be
214 	 * transmitted does not violate the hardware DMA limitation.
215 	 */
216 	if (start_dma_page != end_dma_page) {
217 		/* DMA operation would fail as-is, alloc new aligned SKB */
218 		tx_skb = mlxbf_gige_alloc_skb(priv, skb->len,
219 					      &tx_buf_dma, DMA_TO_DEVICE);
220 		if (!tx_skb) {
221 			/* Free original skb, could not alloc new aligned SKB */
222 			dev_kfree_skb(skb);
223 			netdev->stats.tx_dropped++;
224 			return NETDEV_TX_OK;
225 		}
226 
227 		skb_put_data(tx_skb, skb->data, skb->len);
228 
229 		/* Free the original SKB */
230 		dev_kfree_skb(skb);
231 	} else {
232 		tx_skb = skb;
233 		tx_buf_dma = dma_map_single(priv->dev, skb->data,
234 					    skb->len, DMA_TO_DEVICE);
235 		if (dma_mapping_error(priv->dev, tx_buf_dma)) {
236 			dev_kfree_skb(skb);
237 			netdev->stats.tx_dropped++;
238 			return NETDEV_TX_OK;
239 		}
240 	}
241 
242 	/* Get address of TX WQE */
243 	tx_wqe_addr = priv->tx_wqe_next;
244 
245 	mlxbf_gige_update_tx_wqe_next(priv);
246 
247 	/* Put PA of buffer address into first 64-bit word of TX WQE */
248 	*tx_wqe_addr = tx_buf_dma;
249 
250 	/* Set TX WQE pkt_len appropriately
251 	 * NOTE: GigE silicon will automatically pad up to
252 	 *       minimum packet length if needed.
253 	 */
254 	word2 = tx_skb->len & MLXBF_GIGE_TX_WQE_PKT_LEN_MASK;
255 
256 	/* Write entire 2nd word of TX WQE */
257 	*(tx_wqe_addr + 1) = word2;
258 
259 	spin_lock_irqsave(&priv->lock, flags);
260 	priv->tx_skb[priv->tx_pi % priv->tx_q_entries] = tx_skb;
261 	priv->tx_pi++;
262 	spin_unlock_irqrestore(&priv->lock, flags);
263 
264 	if (!netdev_xmit_more()) {
265 		/* Create memory barrier before write to TX PI */
266 		wmb();
267 		writeq(priv->tx_pi, priv->base + MLXBF_GIGE_TX_PRODUCER_INDEX);
268 	}
269 
270 	/* Check if the last TX entry was just used */
271 	if (!mlxbf_gige_tx_buffs_avail(priv)) {
272 		/* TX ring is full, inform stack */
273 		netif_stop_queue(netdev);
274 
275 		/* Since there is no separate "TX complete" interrupt, need
276 		 * to explicitly schedule NAPI poll.  This will trigger logic
277 		 * which processes TX completions, and will hopefully drain
278 		 * the TX ring allowing the TX queue to be awakened.
279 		 */
280 		napi_schedule(&priv->napi);
281 	}
282 
283 	return NETDEV_TX_OK;
284 }
285