1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * AMD Passthrough DMA device driver
4  * -- Based on the CCP driver
5  *
6  * Copyright (C) 2016,2021 Advanced Micro Devices, Inc.
7  *
8  * Author: Sanjay R Mehta <sanju.mehta@amd.com>
9  * Author: Gary R Hook <gary.hook@amd.com>
10  */
11 
12 #include "ptdma.h"
13 #include "../dmaengine.h"
14 #include "../virt-dma.h"
15 
16 static inline struct pt_dma_chan *to_pt_chan(struct dma_chan *dma_chan)
17 {
18 	return container_of(dma_chan, struct pt_dma_chan, vc.chan);
19 }
20 
21 static inline struct pt_dma_desc *to_pt_desc(struct virt_dma_desc *vd)
22 {
23 	return container_of(vd, struct pt_dma_desc, vd);
24 }
25 
26 static void pt_free_chan_resources(struct dma_chan *dma_chan)
27 {
28 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
29 
30 	vchan_free_chan_resources(&chan->vc);
31 }
32 
33 static void pt_synchronize(struct dma_chan *dma_chan)
34 {
35 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
36 
37 	vchan_synchronize(&chan->vc);
38 }
39 
40 static void pt_do_cleanup(struct virt_dma_desc *vd)
41 {
42 	struct pt_dma_desc *desc = to_pt_desc(vd);
43 	struct pt_device *pt = desc->pt;
44 
45 	kmem_cache_free(pt->dma_desc_cache, desc);
46 }
47 
48 static int pt_dma_start_desc(struct pt_dma_desc *desc)
49 {
50 	struct pt_passthru_engine *pt_engine;
51 	struct pt_device *pt;
52 	struct pt_cmd *pt_cmd;
53 	struct pt_cmd_queue *cmd_q;
54 
55 	desc->issued_to_hw = 1;
56 
57 	pt_cmd = &desc->pt_cmd;
58 	pt = pt_cmd->pt;
59 	cmd_q = &pt->cmd_q;
60 	pt_engine = &pt_cmd->passthru;
61 
62 	pt->tdata.cmd = pt_cmd;
63 
64 	/* Execute the command */
65 	pt_cmd->ret = pt_core_perform_passthru(cmd_q, pt_engine);
66 
67 	return 0;
68 }
69 
70 static struct pt_dma_desc *pt_next_dma_desc(struct pt_dma_chan *chan)
71 {
72 	/* Get the next DMA descriptor on the active list */
73 	struct virt_dma_desc *vd = vchan_next_desc(&chan->vc);
74 
75 	return vd ? to_pt_desc(vd) : NULL;
76 }
77 
78 static struct pt_dma_desc *pt_handle_active_desc(struct pt_dma_chan *chan,
79 						 struct pt_dma_desc *desc)
80 {
81 	struct dma_async_tx_descriptor *tx_desc;
82 	struct virt_dma_desc *vd;
83 	unsigned long flags;
84 
85 	/* Loop over descriptors until one is found with commands */
86 	do {
87 		if (desc) {
88 			if (!desc->issued_to_hw) {
89 				/* No errors, keep going */
90 				if (desc->status != DMA_ERROR)
91 					return desc;
92 			}
93 
94 			tx_desc = &desc->vd.tx;
95 			vd = &desc->vd;
96 		} else {
97 			tx_desc = NULL;
98 		}
99 
100 		spin_lock_irqsave(&chan->vc.lock, flags);
101 
102 		if (desc) {
103 			if (desc->status != DMA_ERROR)
104 				desc->status = DMA_COMPLETE;
105 
106 			dma_cookie_complete(tx_desc);
107 			dma_descriptor_unmap(tx_desc);
108 			list_del(&desc->vd.node);
109 		}
110 
111 		desc = pt_next_dma_desc(chan);
112 
113 		spin_unlock_irqrestore(&chan->vc.lock, flags);
114 
115 		if (tx_desc) {
116 			dmaengine_desc_get_callback_invoke(tx_desc, NULL);
117 			dma_run_dependencies(tx_desc);
118 			vchan_vdesc_fini(vd);
119 		}
120 	} while (desc);
121 
122 	return NULL;
123 }
124 
125 static void pt_cmd_callback(void *data, int err)
126 {
127 	struct pt_dma_desc *desc = data;
128 	struct dma_chan *dma_chan;
129 	struct pt_dma_chan *chan;
130 	int ret;
131 
132 	if (err == -EINPROGRESS)
133 		return;
134 
135 	dma_chan = desc->vd.tx.chan;
136 	chan = to_pt_chan(dma_chan);
137 
138 	if (err)
139 		desc->status = DMA_ERROR;
140 
141 	while (true) {
142 		/* Check for DMA descriptor completion */
143 		desc = pt_handle_active_desc(chan, desc);
144 
145 		/* Don't submit cmd if no descriptor or DMA is paused */
146 		if (!desc)
147 			break;
148 
149 		ret = pt_dma_start_desc(desc);
150 		if (!ret)
151 			break;
152 
153 		desc->status = DMA_ERROR;
154 	}
155 }
156 
157 static struct pt_dma_desc *pt_alloc_dma_desc(struct pt_dma_chan *chan,
158 					     unsigned long flags)
159 {
160 	struct pt_dma_desc *desc;
161 
162 	desc = kmem_cache_zalloc(chan->pt->dma_desc_cache, GFP_NOWAIT);
163 	if (!desc)
164 		return NULL;
165 
166 	vchan_tx_prep(&chan->vc, &desc->vd, flags);
167 
168 	desc->pt = chan->pt;
169 	desc->issued_to_hw = 0;
170 	desc->status = DMA_IN_PROGRESS;
171 
172 	return desc;
173 }
174 
175 static struct pt_dma_desc *pt_create_desc(struct dma_chan *dma_chan,
176 					  dma_addr_t dst,
177 					  dma_addr_t src,
178 					  unsigned int len,
179 					  unsigned long flags)
180 {
181 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
182 	struct pt_passthru_engine *pt_engine;
183 	struct pt_dma_desc *desc;
184 	struct pt_cmd *pt_cmd;
185 
186 	desc = pt_alloc_dma_desc(chan, flags);
187 	if (!desc)
188 		return NULL;
189 
190 	pt_cmd = &desc->pt_cmd;
191 	pt_cmd->pt = chan->pt;
192 	pt_engine = &pt_cmd->passthru;
193 	pt_cmd->engine = PT_ENGINE_PASSTHRU;
194 	pt_engine->src_dma = src;
195 	pt_engine->dst_dma = dst;
196 	pt_engine->src_len = len;
197 	pt_cmd->pt_cmd_callback = pt_cmd_callback;
198 	pt_cmd->data = desc;
199 
200 	desc->len = len;
201 
202 	return desc;
203 }
204 
205 static struct dma_async_tx_descriptor *
206 pt_prep_dma_memcpy(struct dma_chan *dma_chan, dma_addr_t dst,
207 		   dma_addr_t src, size_t len, unsigned long flags)
208 {
209 	struct pt_dma_desc *desc;
210 
211 	desc = pt_create_desc(dma_chan, dst, src, len, flags);
212 	if (!desc)
213 		return NULL;
214 
215 	return &desc->vd.tx;
216 }
217 
218 static struct dma_async_tx_descriptor *
219 pt_prep_dma_interrupt(struct dma_chan *dma_chan, unsigned long flags)
220 {
221 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
222 	struct pt_dma_desc *desc;
223 
224 	desc = pt_alloc_dma_desc(chan, flags);
225 	if (!desc)
226 		return NULL;
227 
228 	return &desc->vd.tx;
229 }
230 
231 static void pt_issue_pending(struct dma_chan *dma_chan)
232 {
233 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
234 	struct pt_dma_desc *desc;
235 	unsigned long flags;
236 
237 	spin_lock_irqsave(&chan->vc.lock, flags);
238 
239 	vchan_issue_pending(&chan->vc);
240 
241 	desc = pt_next_dma_desc(chan);
242 
243 	spin_unlock_irqrestore(&chan->vc.lock, flags);
244 
245 	/* If there was nothing active, start processing */
246 	if (desc)
247 		pt_cmd_callback(desc, 0);
248 }
249 
250 static int pt_pause(struct dma_chan *dma_chan)
251 {
252 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
253 	unsigned long flags;
254 
255 	spin_lock_irqsave(&chan->vc.lock, flags);
256 	pt_stop_queue(&chan->pt->cmd_q);
257 	spin_unlock_irqrestore(&chan->vc.lock, flags);
258 
259 	return 0;
260 }
261 
262 static int pt_resume(struct dma_chan *dma_chan)
263 {
264 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
265 	struct pt_dma_desc *desc = NULL;
266 	unsigned long flags;
267 
268 	spin_lock_irqsave(&chan->vc.lock, flags);
269 	pt_start_queue(&chan->pt->cmd_q);
270 	desc = pt_next_dma_desc(chan);
271 	spin_unlock_irqrestore(&chan->vc.lock, flags);
272 
273 	/* If there was something active, re-start */
274 	if (desc)
275 		pt_cmd_callback(desc, 0);
276 
277 	return 0;
278 }
279 
280 static int pt_terminate_all(struct dma_chan *dma_chan)
281 {
282 	struct pt_dma_chan *chan = to_pt_chan(dma_chan);
283 	unsigned long flags;
284 	LIST_HEAD(head);
285 
286 	spin_lock_irqsave(&chan->vc.lock, flags);
287 	vchan_get_all_descriptors(&chan->vc, &head);
288 	spin_unlock_irqrestore(&chan->vc.lock, flags);
289 
290 	vchan_dma_desc_free_list(&chan->vc, &head);
291 	vchan_free_chan_resources(&chan->vc);
292 
293 	return 0;
294 }
295 
296 int pt_dmaengine_register(struct pt_device *pt)
297 {
298 	struct pt_dma_chan *chan;
299 	struct dma_device *dma_dev = &pt->dma_dev;
300 	char *cmd_cache_name;
301 	char *desc_cache_name;
302 	int ret;
303 
304 	pt->pt_dma_chan = devm_kzalloc(pt->dev, sizeof(*pt->pt_dma_chan),
305 				       GFP_KERNEL);
306 	if (!pt->pt_dma_chan)
307 		return -ENOMEM;
308 
309 	cmd_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
310 					"%s-dmaengine-cmd-cache",
311 					dev_name(pt->dev));
312 	if (!cmd_cache_name)
313 		return -ENOMEM;
314 
315 	desc_cache_name = devm_kasprintf(pt->dev, GFP_KERNEL,
316 					 "%s-dmaengine-desc-cache",
317 					 dev_name(pt->dev));
318 	if (!desc_cache_name) {
319 		ret = -ENOMEM;
320 		goto err_cache;
321 	}
322 
323 	pt->dma_desc_cache = kmem_cache_create(desc_cache_name,
324 					       sizeof(struct pt_dma_desc), 0,
325 					       SLAB_HWCACHE_ALIGN, NULL);
326 	if (!pt->dma_desc_cache) {
327 		ret = -ENOMEM;
328 		goto err_cache;
329 	}
330 
331 	dma_dev->dev = pt->dev;
332 	dma_dev->src_addr_widths = DMA_SLAVE_BUSWIDTH_64_BYTES;
333 	dma_dev->dst_addr_widths = DMA_SLAVE_BUSWIDTH_64_BYTES;
334 	dma_dev->directions = DMA_MEM_TO_MEM;
335 	dma_dev->residue_granularity = DMA_RESIDUE_GRANULARITY_DESCRIPTOR;
336 	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
337 	dma_cap_set(DMA_INTERRUPT, dma_dev->cap_mask);
338 
339 	/*
340 	 * PTDMA is intended to be used with the AMD NTB devices, hence
341 	 * marking it as DMA_PRIVATE.
342 	 */
343 	dma_cap_set(DMA_PRIVATE, dma_dev->cap_mask);
344 
345 	INIT_LIST_HEAD(&dma_dev->channels);
346 
347 	chan = pt->pt_dma_chan;
348 	chan->pt = pt;
349 
350 	/* Set base and prep routines */
351 	dma_dev->device_free_chan_resources = pt_free_chan_resources;
352 	dma_dev->device_prep_dma_memcpy = pt_prep_dma_memcpy;
353 	dma_dev->device_prep_dma_interrupt = pt_prep_dma_interrupt;
354 	dma_dev->device_issue_pending = pt_issue_pending;
355 	dma_dev->device_tx_status = dma_cookie_status;
356 	dma_dev->device_pause = pt_pause;
357 	dma_dev->device_resume = pt_resume;
358 	dma_dev->device_terminate_all = pt_terminate_all;
359 	dma_dev->device_synchronize = pt_synchronize;
360 
361 	chan->vc.desc_free = pt_do_cleanup;
362 	vchan_init(&chan->vc, dma_dev);
363 
364 	dma_set_mask_and_coherent(pt->dev, DMA_BIT_MASK(64));
365 
366 	ret = dma_async_device_register(dma_dev);
367 	if (ret)
368 		goto err_reg;
369 
370 	return 0;
371 
372 err_reg:
373 	kmem_cache_destroy(pt->dma_desc_cache);
374 
375 err_cache:
376 	kmem_cache_destroy(pt->dma_cmd_cache);
377 
378 	return ret;
379 }
380 
381 void pt_dmaengine_unregister(struct pt_device *pt)
382 {
383 	struct dma_device *dma_dev = &pt->dma_dev;
384 
385 	dma_async_device_unregister(dma_dev);
386 
387 	kmem_cache_destroy(pt->dma_desc_cache);
388 	kmem_cache_destroy(pt->dma_cmd_cache);
389 }
390