1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright (C) 2017 Intel Deutschland GmbH
4  * Copyright (C) 2018-2020 Intel Corporation
5  */
6 #include <net/tso.h>
7 #include <linux/tcp.h>
8 
9 #include "iwl-debug.h"
10 #include "iwl-csr.h"
11 #include "iwl-io.h"
12 #include "internal.h"
13 #include "fw/api/tx.h"
14 #include "queue/tx.h"
15 
16 /*************** HOST COMMAND QUEUE FUNCTIONS   *****/
17 
18 /*
19  * iwl_pcie_gen2_enqueue_hcmd - enqueue a uCode command
20  * @priv: device private data point
21  * @cmd: a pointer to the ucode command structure
22  *
23  * The function returns < 0 values to indicate the operation
24  * failed. On success, it returns the index (>= 0) of command in the
25  * command queue.
26  */
27 int iwl_pcie_gen2_enqueue_hcmd(struct iwl_trans *trans,
28 			       struct iwl_host_cmd *cmd)
29 {
30 	struct iwl_trans_pcie *trans_pcie = IWL_TRANS_GET_PCIE_TRANS(trans);
31 	struct iwl_txq *txq = trans->txqs.txq[trans->txqs.cmd.q_id];
32 	struct iwl_device_cmd *out_cmd;
33 	struct iwl_cmd_meta *out_meta;
34 	void *dup_buf = NULL;
35 	dma_addr_t phys_addr;
36 	int i, cmd_pos, idx;
37 	u16 copy_size, cmd_size, tb0_size;
38 	bool had_nocopy = false;
39 	u8 group_id = iwl_cmd_groupid(cmd->id);
40 	const u8 *cmddata[IWL_MAX_CMD_TBS_PER_TFD];
41 	u16 cmdlen[IWL_MAX_CMD_TBS_PER_TFD];
42 	struct iwl_tfh_tfd *tfd;
43 	unsigned long flags;
44 
45 	copy_size = sizeof(struct iwl_cmd_header_wide);
46 	cmd_size = sizeof(struct iwl_cmd_header_wide);
47 
48 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
49 		cmddata[i] = cmd->data[i];
50 		cmdlen[i] = cmd->len[i];
51 
52 		if (!cmd->len[i])
53 			continue;
54 
55 		/* need at least IWL_FIRST_TB_SIZE copied */
56 		if (copy_size < IWL_FIRST_TB_SIZE) {
57 			int copy = IWL_FIRST_TB_SIZE - copy_size;
58 
59 			if (copy > cmdlen[i])
60 				copy = cmdlen[i];
61 			cmdlen[i] -= copy;
62 			cmddata[i] += copy;
63 			copy_size += copy;
64 		}
65 
66 		if (cmd->dataflags[i] & IWL_HCMD_DFL_NOCOPY) {
67 			had_nocopy = true;
68 			if (WARN_ON(cmd->dataflags[i] & IWL_HCMD_DFL_DUP)) {
69 				idx = -EINVAL;
70 				goto free_dup_buf;
71 			}
72 		} else if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP) {
73 			/*
74 			 * This is also a chunk that isn't copied
75 			 * to the static buffer so set had_nocopy.
76 			 */
77 			had_nocopy = true;
78 
79 			/* only allowed once */
80 			if (WARN_ON(dup_buf)) {
81 				idx = -EINVAL;
82 				goto free_dup_buf;
83 			}
84 
85 			dup_buf = kmemdup(cmddata[i], cmdlen[i],
86 					  GFP_ATOMIC);
87 			if (!dup_buf)
88 				return -ENOMEM;
89 		} else {
90 			/* NOCOPY must not be followed by normal! */
91 			if (WARN_ON(had_nocopy)) {
92 				idx = -EINVAL;
93 				goto free_dup_buf;
94 			}
95 			copy_size += cmdlen[i];
96 		}
97 		cmd_size += cmd->len[i];
98 	}
99 
100 	/*
101 	 * If any of the command structures end up being larger than the
102 	 * TFD_MAX_PAYLOAD_SIZE and they aren't dynamically allocated into
103 	 * separate TFDs, then we will need to increase the size of the buffers
104 	 */
105 	if (WARN(copy_size > TFD_MAX_PAYLOAD_SIZE,
106 		 "Command %s (%#x) is too large (%d bytes)\n",
107 		 iwl_get_cmd_string(trans, cmd->id), cmd->id, copy_size)) {
108 		idx = -EINVAL;
109 		goto free_dup_buf;
110 	}
111 
112 	spin_lock_irqsave(&txq->lock, flags);
113 
114 	idx = iwl_txq_get_cmd_index(txq, txq->write_ptr);
115 	tfd = iwl_txq_get_tfd(trans, txq, txq->write_ptr);
116 	memset(tfd, 0, sizeof(*tfd));
117 
118 	if (iwl_txq_space(trans, txq) < ((cmd->flags & CMD_ASYNC) ? 2 : 1)) {
119 		spin_unlock_irqrestore(&txq->lock, flags);
120 
121 		IWL_ERR(trans, "No space in command queue\n");
122 		iwl_op_mode_cmd_queue_full(trans->op_mode);
123 		idx = -ENOSPC;
124 		goto free_dup_buf;
125 	}
126 
127 	out_cmd = txq->entries[idx].cmd;
128 	out_meta = &txq->entries[idx].meta;
129 
130 	/* re-initialize to NULL */
131 	memset(out_meta, 0, sizeof(*out_meta));
132 	if (cmd->flags & CMD_WANT_SKB)
133 		out_meta->source = cmd;
134 
135 	/* set up the header */
136 	out_cmd->hdr_wide.cmd = iwl_cmd_opcode(cmd->id);
137 	out_cmd->hdr_wide.group_id = group_id;
138 	out_cmd->hdr_wide.version = iwl_cmd_version(cmd->id);
139 	out_cmd->hdr_wide.length =
140 		cpu_to_le16(cmd_size - sizeof(struct iwl_cmd_header_wide));
141 	out_cmd->hdr_wide.reserved = 0;
142 	out_cmd->hdr_wide.sequence =
143 		cpu_to_le16(QUEUE_TO_SEQ(trans->txqs.cmd.q_id) |
144 					 INDEX_TO_SEQ(txq->write_ptr));
145 
146 	cmd_pos = sizeof(struct iwl_cmd_header_wide);
147 	copy_size = sizeof(struct iwl_cmd_header_wide);
148 
149 	/* and copy the data that needs to be copied */
150 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
151 		int copy;
152 
153 		if (!cmd->len[i])
154 			continue;
155 
156 		/* copy everything if not nocopy/dup */
157 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
158 					   IWL_HCMD_DFL_DUP))) {
159 			copy = cmd->len[i];
160 
161 			memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
162 			cmd_pos += copy;
163 			copy_size += copy;
164 			continue;
165 		}
166 
167 		/*
168 		 * Otherwise we need at least IWL_FIRST_TB_SIZE copied
169 		 * in total (for bi-directional DMA), but copy up to what
170 		 * we can fit into the payload for debug dump purposes.
171 		 */
172 		copy = min_t(int, TFD_MAX_PAYLOAD_SIZE - cmd_pos, cmd->len[i]);
173 
174 		memcpy((u8 *)out_cmd + cmd_pos, cmd->data[i], copy);
175 		cmd_pos += copy;
176 
177 		/* However, treat copy_size the proper way, we need it below */
178 		if (copy_size < IWL_FIRST_TB_SIZE) {
179 			copy = IWL_FIRST_TB_SIZE - copy_size;
180 
181 			if (copy > cmd->len[i])
182 				copy = cmd->len[i];
183 			copy_size += copy;
184 		}
185 	}
186 
187 	IWL_DEBUG_HC(trans,
188 		     "Sending command %s (%.2x.%.2x), seq: 0x%04X, %d bytes at %d[%d]:%d\n",
189 		     iwl_get_cmd_string(trans, cmd->id), group_id,
190 		     out_cmd->hdr.cmd, le16_to_cpu(out_cmd->hdr.sequence),
191 		     cmd_size, txq->write_ptr, idx, trans->txqs.cmd.q_id);
192 
193 	/* start the TFD with the minimum copy bytes */
194 	tb0_size = min_t(int, copy_size, IWL_FIRST_TB_SIZE);
195 	memcpy(&txq->first_tb_bufs[idx], out_cmd, tb0_size);
196 	iwl_txq_gen2_set_tb(trans, tfd, iwl_txq_get_first_tb_dma(txq, idx),
197 			    tb0_size);
198 
199 	/* map first command fragment, if any remains */
200 	if (copy_size > tb0_size) {
201 		phys_addr = dma_map_single(trans->dev,
202 					   (u8 *)out_cmd + tb0_size,
203 					   copy_size - tb0_size,
204 					   DMA_TO_DEVICE);
205 		if (dma_mapping_error(trans->dev, phys_addr)) {
206 			idx = -ENOMEM;
207 			iwl_txq_gen2_tfd_unmap(trans, out_meta, tfd);
208 			goto out;
209 		}
210 		iwl_txq_gen2_set_tb(trans, tfd, phys_addr,
211 				    copy_size - tb0_size);
212 	}
213 
214 	/* map the remaining (adjusted) nocopy/dup fragments */
215 	for (i = 0; i < IWL_MAX_CMD_TBS_PER_TFD; i++) {
216 		void *data = (void *)(uintptr_t)cmddata[i];
217 
218 		if (!cmdlen[i])
219 			continue;
220 		if (!(cmd->dataflags[i] & (IWL_HCMD_DFL_NOCOPY |
221 					   IWL_HCMD_DFL_DUP)))
222 			continue;
223 		if (cmd->dataflags[i] & IWL_HCMD_DFL_DUP)
224 			data = dup_buf;
225 		phys_addr = dma_map_single(trans->dev, data,
226 					   cmdlen[i], DMA_TO_DEVICE);
227 		if (dma_mapping_error(trans->dev, phys_addr)) {
228 			idx = -ENOMEM;
229 			iwl_txq_gen2_tfd_unmap(trans, out_meta, tfd);
230 			goto out;
231 		}
232 		iwl_txq_gen2_set_tb(trans, tfd, phys_addr, cmdlen[i]);
233 	}
234 
235 	BUILD_BUG_ON(IWL_TFH_NUM_TBS > sizeof(out_meta->tbs) * BITS_PER_BYTE);
236 	out_meta->flags = cmd->flags;
237 	if (WARN_ON_ONCE(txq->entries[idx].free_buf))
238 		kfree_sensitive(txq->entries[idx].free_buf);
239 	txq->entries[idx].free_buf = dup_buf;
240 
241 	trace_iwlwifi_dev_hcmd(trans->dev, cmd, cmd_size, &out_cmd->hdr_wide);
242 
243 	/* start timer if queue currently empty */
244 	if (txq->read_ptr == txq->write_ptr && txq->wd_timeout)
245 		mod_timer(&txq->stuck_timer, jiffies + txq->wd_timeout);
246 
247 	spin_lock(&trans_pcie->reg_lock);
248 	/* Increment and update queue's write index */
249 	txq->write_ptr = iwl_txq_inc_wrap(trans, txq->write_ptr);
250 	iwl_txq_inc_wr_ptr(trans, txq);
251 	spin_unlock(&trans_pcie->reg_lock);
252 
253 out:
254 	spin_unlock_irqrestore(&txq->lock, flags);
255 free_dup_buf:
256 	if (idx < 0)
257 		kfree(dup_buf);
258 	return idx;
259 }
260