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