1 /*
2  * Copyright (C) 2007 Ben Skeggs.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining
6  * a copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sublicense, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * The above copyright notice and this permission notice (including the
14  * next paragraph) shall be included in all copies or substantial
15  * portions of the Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20  * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
21  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
22  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
23  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24  *
25  */
26 
27 #include "drmP.h"
28 #include "drm.h"
29 #include "nouveau_drv.h"
30 #include "nouveau_dma.h"
31 #include "nouveau_ramht.h"
32 
33 void
34 nouveau_dma_init(struct nouveau_channel *chan)
35 {
36 	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
37 	struct nouveau_bo *pushbuf = chan->pushbuf_bo;
38 
39 	if (dev_priv->card_type >= NV_50) {
40 		const int ib_size = pushbuf->bo.mem.size / 2;
41 
42 		chan->dma.ib_base = (pushbuf->bo.mem.size - ib_size) >> 2;
43 		chan->dma.ib_max = (ib_size / 8) - 1;
44 		chan->dma.ib_put = 0;
45 		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
46 
47 		chan->dma.max = (pushbuf->bo.mem.size - ib_size) >> 2;
48 	} else {
49 		chan->dma.max  = (pushbuf->bo.mem.size >> 2) - 2;
50 	}
51 
52 	chan->dma.put  = 0;
53 	chan->dma.cur  = chan->dma.put;
54 	chan->dma.free = chan->dma.max - chan->dma.cur;
55 }
56 
57 void
58 OUT_RINGp(struct nouveau_channel *chan, const void *data, unsigned nr_dwords)
59 {
60 	bool is_iomem;
61 	u32 *mem = ttm_kmap_obj_virtual(&chan->pushbuf_bo->kmap, &is_iomem);
62 	mem = &mem[chan->dma.cur];
63 	if (is_iomem)
64 		memcpy_toio((void __force __iomem *)mem, data, nr_dwords * 4);
65 	else
66 		memcpy(mem, data, nr_dwords * 4);
67 	chan->dma.cur += nr_dwords;
68 }
69 
70 /* Fetch and adjust GPU GET pointer
71  *
72  * Returns:
73  *  value >= 0, the adjusted GET pointer
74  *  -EINVAL if GET pointer currently outside main push buffer
75  *  -EBUSY if timeout exceeded
76  */
77 static inline int
78 READ_GET(struct nouveau_channel *chan, uint64_t *prev_get, int *timeout)
79 {
80 	uint64_t val;
81 
82 	val = nvchan_rd32(chan, chan->user_get);
83         if (chan->user_get_hi)
84                 val |= (uint64_t)nvchan_rd32(chan, chan->user_get_hi) << 32;
85 
86 	/* reset counter as long as GET is still advancing, this is
87 	 * to avoid misdetecting a GPU lockup if the GPU happens to
88 	 * just be processing an operation that takes a long time
89 	 */
90 	if (val != *prev_get) {
91 		*prev_get = val;
92 		*timeout = 0;
93 	}
94 
95 	if ((++*timeout & 0xff) == 0) {
96 		DRM_UDELAY(1);
97 		if (*timeout > 100000)
98 			return -EBUSY;
99 	}
100 
101 	if (val < chan->pushbuf_base ||
102 	    val > chan->pushbuf_base + (chan->dma.max << 2))
103 		return -EINVAL;
104 
105 	return (val - chan->pushbuf_base) >> 2;
106 }
107 
108 void
109 nv50_dma_push(struct nouveau_channel *chan, struct nouveau_bo *bo,
110 	      int delta, int length)
111 {
112 	struct nouveau_bo *pb = chan->pushbuf_bo;
113 	struct nouveau_vma *vma;
114 	int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base;
115 	u64 offset;
116 
117 	vma = nouveau_bo_vma_find(bo, chan->vm);
118 	BUG_ON(!vma);
119 	offset = vma->offset + delta;
120 
121 	BUG_ON(chan->dma.ib_free < 1);
122 	nouveau_bo_wr32(pb, ip++, lower_32_bits(offset));
123 	nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8);
124 
125 	chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max;
126 
127 	DRM_MEMORYBARRIER();
128 	/* Flush writes. */
129 	nouveau_bo_rd32(pb, 0);
130 
131 	nvchan_wr32(chan, 0x8c, chan->dma.ib_put);
132 	chan->dma.ib_free--;
133 }
134 
135 static int
136 nv50_dma_push_wait(struct nouveau_channel *chan, int count)
137 {
138 	uint32_t cnt = 0, prev_get = 0;
139 
140 	while (chan->dma.ib_free < count) {
141 		uint32_t get = nvchan_rd32(chan, 0x88);
142 		if (get != prev_get) {
143 			prev_get = get;
144 			cnt = 0;
145 		}
146 
147 		if ((++cnt & 0xff) == 0) {
148 			DRM_UDELAY(1);
149 			if (cnt > 100000)
150 				return -EBUSY;
151 		}
152 
153 		chan->dma.ib_free = get - chan->dma.ib_put;
154 		if (chan->dma.ib_free <= 0)
155 			chan->dma.ib_free += chan->dma.ib_max;
156 	}
157 
158 	return 0;
159 }
160 
161 static int
162 nv50_dma_wait(struct nouveau_channel *chan, int slots, int count)
163 {
164 	uint64_t prev_get = 0;
165 	int ret, cnt = 0;
166 
167 	ret = nv50_dma_push_wait(chan, slots + 1);
168 	if (unlikely(ret))
169 		return ret;
170 
171 	while (chan->dma.free < count) {
172 		int get = READ_GET(chan, &prev_get, &cnt);
173 		if (unlikely(get < 0)) {
174 			if (get == -EINVAL)
175 				continue;
176 
177 			return get;
178 		}
179 
180 		if (get <= chan->dma.cur) {
181 			chan->dma.free = chan->dma.max - chan->dma.cur;
182 			if (chan->dma.free >= count)
183 				break;
184 
185 			FIRE_RING(chan);
186 			do {
187 				get = READ_GET(chan, &prev_get, &cnt);
188 				if (unlikely(get < 0)) {
189 					if (get == -EINVAL)
190 						continue;
191 					return get;
192 				}
193 			} while (get == 0);
194 			chan->dma.cur = 0;
195 			chan->dma.put = 0;
196 		}
197 
198 		chan->dma.free = get - chan->dma.cur - 1;
199 	}
200 
201 	return 0;
202 }
203 
204 int
205 nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size)
206 {
207 	uint64_t prev_get = 0;
208 	int cnt = 0, get;
209 
210 	if (chan->dma.ib_max)
211 		return nv50_dma_wait(chan, slots, size);
212 
213 	while (chan->dma.free < size) {
214 		get = READ_GET(chan, &prev_get, &cnt);
215 		if (unlikely(get == -EBUSY))
216 			return -EBUSY;
217 
218 		/* loop until we have a usable GET pointer.  the value
219 		 * we read from the GPU may be outside the main ring if
220 		 * PFIFO is processing a buffer called from the main ring,
221 		 * discard these values until something sensible is seen.
222 		 *
223 		 * the other case we discard GET is while the GPU is fetching
224 		 * from the SKIPS area, so the code below doesn't have to deal
225 		 * with some fun corner cases.
226 		 */
227 		if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS)
228 			continue;
229 
230 		if (get <= chan->dma.cur) {
231 			/* engine is fetching behind us, or is completely
232 			 * idle (GET == PUT) so we have free space up until
233 			 * the end of the push buffer
234 			 *
235 			 * we can only hit that path once per call due to
236 			 * looping back to the beginning of the push buffer,
237 			 * we'll hit the fetching-ahead-of-us path from that
238 			 * point on.
239 			 *
240 			 * the *one* exception to that rule is if we read
241 			 * GET==PUT, in which case the below conditional will
242 			 * always succeed and break us out of the wait loop.
243 			 */
244 			chan->dma.free = chan->dma.max - chan->dma.cur;
245 			if (chan->dma.free >= size)
246 				break;
247 
248 			/* not enough space left at the end of the push buffer,
249 			 * instruct the GPU to jump back to the start right
250 			 * after processing the currently pending commands.
251 			 */
252 			OUT_RING(chan, chan->pushbuf_base | 0x20000000);
253 
254 			/* wait for GET to depart from the skips area.
255 			 * prevents writing GET==PUT and causing a race
256 			 * condition that causes us to think the GPU is
257 			 * idle when it's not.
258 			 */
259 			do {
260 				get = READ_GET(chan, &prev_get, &cnt);
261 				if (unlikely(get == -EBUSY))
262 					return -EBUSY;
263 				if (unlikely(get == -EINVAL))
264 					continue;
265 			} while (get <= NOUVEAU_DMA_SKIPS);
266 			WRITE_PUT(NOUVEAU_DMA_SKIPS);
267 
268 			/* we're now submitting commands at the start of
269 			 * the push buffer.
270 			 */
271 			chan->dma.cur  =
272 			chan->dma.put  = NOUVEAU_DMA_SKIPS;
273 		}
274 
275 		/* engine fetching ahead of us, we have space up until the
276 		 * current GET pointer.  the "- 1" is to ensure there's
277 		 * space left to emit a jump back to the beginning of the
278 		 * push buffer if we require it.  we can never get GET == PUT
279 		 * here, so this is safe.
280 		 */
281 		chan->dma.free = get - chan->dma.cur - 1;
282 	}
283 
284 	return 0;
285 }
286 
287