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 "nouveau_drm.h"
28 #include "nouveau_dma.h"
29 
30 void
31 OUT_RINGp(struct nouveau_channel *chan, const void *data, unsigned nr_dwords)
32 {
33 	bool is_iomem;
34 	u32 *mem = ttm_kmap_obj_virtual(&chan->push.buffer->kmap, &is_iomem);
35 	mem = &mem[chan->dma.cur];
36 	if (is_iomem)
37 		memcpy_toio((void __force __iomem *)mem, data, nr_dwords * 4);
38 	else
39 		memcpy(mem, data, nr_dwords * 4);
40 	chan->dma.cur += nr_dwords;
41 }
42 
43 /* Fetch and adjust GPU GET pointer
44  *
45  * Returns:
46  *  value >= 0, the adjusted GET pointer
47  *  -EINVAL if GET pointer currently outside main push buffer
48  *  -EBUSY if timeout exceeded
49  */
50 static inline int
51 READ_GET(struct nouveau_channel *chan, uint64_t *prev_get, int *timeout)
52 {
53 	uint64_t val;
54 
55 	val = nvif_rd32(chan, chan->user_get);
56         if (chan->user_get_hi)
57                 val |= (uint64_t)nvif_rd32(chan, chan->user_get_hi) << 32;
58 
59 	/* reset counter as long as GET is still advancing, this is
60 	 * to avoid misdetecting a GPU lockup if the GPU happens to
61 	 * just be processing an operation that takes a long time
62 	 */
63 	if (val != *prev_get) {
64 		*prev_get = val;
65 		*timeout = 0;
66 	}
67 
68 	if ((++*timeout & 0xff) == 0) {
69 		udelay(1);
70 		if (*timeout > 100000)
71 			return -EBUSY;
72 	}
73 
74 	if (val < chan->push.vma.offset ||
75 	    val > chan->push.vma.offset + (chan->dma.max << 2))
76 		return -EINVAL;
77 
78 	return (val - chan->push.vma.offset) >> 2;
79 }
80 
81 void
82 nv50_dma_push(struct nouveau_channel *chan, struct nouveau_bo *bo,
83 	      int delta, int length)
84 {
85 	struct nouveau_cli *cli = (void *)nvif_client(&chan->device->base);
86 	struct nouveau_bo *pb = chan->push.buffer;
87 	struct nouveau_vma *vma;
88 	int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base;
89 	u64 offset;
90 
91 	vma = nouveau_bo_vma_find(bo, cli->vm);
92 	BUG_ON(!vma);
93 	offset = vma->offset + delta;
94 
95 	BUG_ON(chan->dma.ib_free < 1);
96 
97 	nouveau_bo_wr32(pb, ip++, lower_32_bits(offset));
98 	nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8);
99 
100 	chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max;
101 
102 	mb();
103 	/* Flush writes. */
104 	nouveau_bo_rd32(pb, 0);
105 
106 	nvif_wr32(chan, 0x8c, chan->dma.ib_put);
107 	chan->dma.ib_free--;
108 }
109 
110 static int
111 nv50_dma_push_wait(struct nouveau_channel *chan, int count)
112 {
113 	uint32_t cnt = 0, prev_get = 0;
114 
115 	while (chan->dma.ib_free < count) {
116 		uint32_t get = nvif_rd32(chan, 0x88);
117 		if (get != prev_get) {
118 			prev_get = get;
119 			cnt = 0;
120 		}
121 
122 		if ((++cnt & 0xff) == 0) {
123 			DRM_UDELAY(1);
124 			if (cnt > 100000)
125 				return -EBUSY;
126 		}
127 
128 		chan->dma.ib_free = get - chan->dma.ib_put;
129 		if (chan->dma.ib_free <= 0)
130 			chan->dma.ib_free += chan->dma.ib_max;
131 	}
132 
133 	return 0;
134 }
135 
136 static int
137 nv50_dma_wait(struct nouveau_channel *chan, int slots, int count)
138 {
139 	uint64_t prev_get = 0;
140 	int ret, cnt = 0;
141 
142 	ret = nv50_dma_push_wait(chan, slots + 1);
143 	if (unlikely(ret))
144 		return ret;
145 
146 	while (chan->dma.free < count) {
147 		int get = READ_GET(chan, &prev_get, &cnt);
148 		if (unlikely(get < 0)) {
149 			if (get == -EINVAL)
150 				continue;
151 
152 			return get;
153 		}
154 
155 		if (get <= chan->dma.cur) {
156 			chan->dma.free = chan->dma.max - chan->dma.cur;
157 			if (chan->dma.free >= count)
158 				break;
159 
160 			FIRE_RING(chan);
161 			do {
162 				get = READ_GET(chan, &prev_get, &cnt);
163 				if (unlikely(get < 0)) {
164 					if (get == -EINVAL)
165 						continue;
166 					return get;
167 				}
168 			} while (get == 0);
169 			chan->dma.cur = 0;
170 			chan->dma.put = 0;
171 		}
172 
173 		chan->dma.free = get - chan->dma.cur - 1;
174 	}
175 
176 	return 0;
177 }
178 
179 int
180 nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size)
181 {
182 	uint64_t prev_get = 0;
183 	int cnt = 0, get;
184 
185 	if (chan->dma.ib_max)
186 		return nv50_dma_wait(chan, slots, size);
187 
188 	while (chan->dma.free < size) {
189 		get = READ_GET(chan, &prev_get, &cnt);
190 		if (unlikely(get == -EBUSY))
191 			return -EBUSY;
192 
193 		/* loop until we have a usable GET pointer.  the value
194 		 * we read from the GPU may be outside the main ring if
195 		 * PFIFO is processing a buffer called from the main ring,
196 		 * discard these values until something sensible is seen.
197 		 *
198 		 * the other case we discard GET is while the GPU is fetching
199 		 * from the SKIPS area, so the code below doesn't have to deal
200 		 * with some fun corner cases.
201 		 */
202 		if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS)
203 			continue;
204 
205 		if (get <= chan->dma.cur) {
206 			/* engine is fetching behind us, or is completely
207 			 * idle (GET == PUT) so we have free space up until
208 			 * the end of the push buffer
209 			 *
210 			 * we can only hit that path once per call due to
211 			 * looping back to the beginning of the push buffer,
212 			 * we'll hit the fetching-ahead-of-us path from that
213 			 * point on.
214 			 *
215 			 * the *one* exception to that rule is if we read
216 			 * GET==PUT, in which case the below conditional will
217 			 * always succeed and break us out of the wait loop.
218 			 */
219 			chan->dma.free = chan->dma.max - chan->dma.cur;
220 			if (chan->dma.free >= size)
221 				break;
222 
223 			/* not enough space left at the end of the push buffer,
224 			 * instruct the GPU to jump back to the start right
225 			 * after processing the currently pending commands.
226 			 */
227 			OUT_RING(chan, chan->push.vma.offset | 0x20000000);
228 
229 			/* wait for GET to depart from the skips area.
230 			 * prevents writing GET==PUT and causing a race
231 			 * condition that causes us to think the GPU is
232 			 * idle when it's not.
233 			 */
234 			do {
235 				get = READ_GET(chan, &prev_get, &cnt);
236 				if (unlikely(get == -EBUSY))
237 					return -EBUSY;
238 				if (unlikely(get == -EINVAL))
239 					continue;
240 			} while (get <= NOUVEAU_DMA_SKIPS);
241 			WRITE_PUT(NOUVEAU_DMA_SKIPS);
242 
243 			/* we're now submitting commands at the start of
244 			 * the push buffer.
245 			 */
246 			chan->dma.cur  =
247 			chan->dma.put  = NOUVEAU_DMA_SKIPS;
248 		}
249 
250 		/* engine fetching ahead of us, we have space up until the
251 		 * current GET pointer.  the "- 1" is to ensure there's
252 		 * space left to emit a jump back to the beginning of the
253 		 * push buffer if we require it.  we can never get GET == PUT
254 		 * here, so this is safe.
255 		 */
256 		chan->dma.free = get - chan->dma.cur - 1;
257 	}
258 
259 	return 0;
260 }
261 
262