2 * Copyright (C) 2007 Ben Skeggs.
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:
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.
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.
29 #include "nouveau_drv.h"
30 #include "nouveau_dma.h"
31 #include "nouveau_ramht.h"
34 nouveau_dma_pre_init(struct nouveau_channel
*chan
)
36 struct drm_nouveau_private
*dev_priv
= chan
->dev
->dev_private
;
37 struct nouveau_bo
*pushbuf
= chan
->pushbuf_bo
;
39 if (dev_priv
->card_type
== NV_50
) {
40 const int ib_size
= pushbuf
->bo
.mem
.size
/ 2;
42 chan
->dma
.ib_base
= (pushbuf
->bo
.mem
.size
- ib_size
) >> 2;
43 chan
->dma
.ib_max
= (ib_size
/ 8) - 1;
45 chan
->dma
.ib_free
= chan
->dma
.ib_max
- chan
->dma
.ib_put
;
47 chan
->dma
.max
= (pushbuf
->bo
.mem
.size
- ib_size
) >> 2;
49 chan
->dma
.max
= (pushbuf
->bo
.mem
.size
>> 2) - 2;
53 chan
->dma
.cur
= chan
->dma
.put
;
54 chan
->dma
.free
= chan
->dma
.max
- chan
->dma
.cur
;
58 nouveau_dma_init(struct nouveau_channel
*chan
)
60 struct drm_device
*dev
= chan
->dev
;
61 struct drm_nouveau_private
*dev_priv
= dev
->dev_private
;
62 struct nouveau_gpuobj
*obj
= NULL
;
65 /* Create NV_MEMORY_TO_MEMORY_FORMAT for buffer moves */
66 ret
= nouveau_gpuobj_gr_new(chan
, dev_priv
->card_type
< NV_50
?
67 0x0039 : 0x5039, &obj
);
71 ret
= nouveau_ramht_insert(chan
, NvM2MF
, obj
);
72 nouveau_gpuobj_ref(NULL
, &obj
);
76 /* NV_MEMORY_TO_MEMORY_FORMAT requires a notifier object */
77 ret
= nouveau_notifier_alloc(chan
, NvNotify0
, 32, &chan
->m2mf_ntfy
);
82 ret
= nouveau_bo_map(chan
->pushbuf_bo
);
86 /* Insert NOPS for NOUVEAU_DMA_SKIPS */
87 ret
= RING_SPACE(chan
, NOUVEAU_DMA_SKIPS
);
91 for (i
= 0; i
< NOUVEAU_DMA_SKIPS
; i
++)
94 /* Initialise NV_MEMORY_TO_MEMORY_FORMAT */
95 ret
= RING_SPACE(chan
, 4);
98 BEGIN_RING(chan
, NvSubM2MF
, NV_MEMORY_TO_MEMORY_FORMAT_NAME
, 1);
99 OUT_RING(chan
, NvM2MF
);
100 BEGIN_RING(chan
, NvSubM2MF
, NV_MEMORY_TO_MEMORY_FORMAT_DMA_NOTIFY
, 1);
101 OUT_RING(chan
, NvNotify0
);
103 /* Sit back and pray the channel works.. */
110 OUT_RINGp(struct nouveau_channel
*chan
, const void *data
, unsigned nr_dwords
)
113 u32
*mem
= ttm_kmap_obj_virtual(&chan
->pushbuf_bo
->kmap
, &is_iomem
);
114 mem
= &mem
[chan
->dma
.cur
];
116 memcpy_toio((void __force __iomem
*)mem
, data
, nr_dwords
* 4);
118 memcpy(mem
, data
, nr_dwords
* 4);
119 chan
->dma
.cur
+= nr_dwords
;
122 /* Fetch and adjust GPU GET pointer
125 * value >= 0, the adjusted GET pointer
126 * -EINVAL if GET pointer currently outside main push buffer
127 * -EBUSY if timeout exceeded
130 READ_GET(struct nouveau_channel
*chan
, uint32_t *prev_get
, uint32_t *timeout
)
134 val
= nvchan_rd32(chan
, chan
->user_get
);
136 /* reset counter as long as GET is still advancing, this is
137 * to avoid misdetecting a GPU lockup if the GPU happens to
138 * just be processing an operation that takes a long time
140 if (val
!= *prev_get
) {
145 if ((++*timeout
& 0xff) == 0) {
147 if (*timeout
> 100000)
151 if (val
< chan
->pushbuf_base
||
152 val
> chan
->pushbuf_base
+ (chan
->dma
.max
<< 2))
155 return (val
- chan
->pushbuf_base
) >> 2;
159 nv50_dma_push(struct nouveau_channel
*chan
, struct nouveau_bo
*bo
,
160 int delta
, int length
)
162 struct nouveau_bo
*pb
= chan
->pushbuf_bo
;
163 uint64_t offset
= bo
->bo
.offset
+ delta
;
164 int ip
= (chan
->dma
.ib_put
* 2) + chan
->dma
.ib_base
;
166 BUG_ON(chan
->dma
.ib_free
< 1);
167 nouveau_bo_wr32(pb
, ip
++, lower_32_bits(offset
));
168 nouveau_bo_wr32(pb
, ip
++, upper_32_bits(offset
) | length
<< 8);
170 chan
->dma
.ib_put
= (chan
->dma
.ib_put
+ 1) & chan
->dma
.ib_max
;
174 nouveau_bo_rd32(pb
, 0);
176 nvchan_wr32(chan
, 0x8c, chan
->dma
.ib_put
);
181 nv50_dma_push_wait(struct nouveau_channel
*chan
, int count
)
183 uint32_t cnt
= 0, prev_get
= 0;
185 while (chan
->dma
.ib_free
< count
) {
186 uint32_t get
= nvchan_rd32(chan
, 0x88);
187 if (get
!= prev_get
) {
192 if ((++cnt
& 0xff) == 0) {
198 chan
->dma
.ib_free
= get
- chan
->dma
.ib_put
;
199 if (chan
->dma
.ib_free
<= 0)
200 chan
->dma
.ib_free
+= chan
->dma
.ib_max
;
207 nv50_dma_wait(struct nouveau_channel
*chan
, int slots
, int count
)
209 uint32_t cnt
= 0, prev_get
= 0;
212 ret
= nv50_dma_push_wait(chan
, slots
+ 1);
216 while (chan
->dma
.free
< count
) {
217 int get
= READ_GET(chan
, &prev_get
, &cnt
);
218 if (unlikely(get
< 0)) {
225 if (get
<= chan
->dma
.cur
) {
226 chan
->dma
.free
= chan
->dma
.max
- chan
->dma
.cur
;
227 if (chan
->dma
.free
>= count
)
232 get
= READ_GET(chan
, &prev_get
, &cnt
);
233 if (unlikely(get
< 0)) {
243 chan
->dma
.free
= get
- chan
->dma
.cur
- 1;
250 nouveau_dma_wait(struct nouveau_channel
*chan
, int slots
, int size
)
252 uint32_t prev_get
= 0, cnt
= 0;
255 if (chan
->dma
.ib_max
)
256 return nv50_dma_wait(chan
, slots
, size
);
258 while (chan
->dma
.free
< size
) {
259 get
= READ_GET(chan
, &prev_get
, &cnt
);
260 if (unlikely(get
== -EBUSY
))
263 /* loop until we have a usable GET pointer. the value
264 * we read from the GPU may be outside the main ring if
265 * PFIFO is processing a buffer called from the main ring,
266 * discard these values until something sensible is seen.
268 * the other case we discard GET is while the GPU is fetching
269 * from the SKIPS area, so the code below doesn't have to deal
270 * with some fun corner cases.
272 if (unlikely(get
== -EINVAL
) || get
< NOUVEAU_DMA_SKIPS
)
275 if (get
<= chan
->dma
.cur
) {
276 /* engine is fetching behind us, or is completely
277 * idle (GET == PUT) so we have free space up until
278 * the end of the push buffer
280 * we can only hit that path once per call due to
281 * looping back to the beginning of the push buffer,
282 * we'll hit the fetching-ahead-of-us path from that
285 * the *one* exception to that rule is if we read
286 * GET==PUT, in which case the below conditional will
287 * always succeed and break us out of the wait loop.
289 chan
->dma
.free
= chan
->dma
.max
- chan
->dma
.cur
;
290 if (chan
->dma
.free
>= size
)
293 /* not enough space left at the end of the push buffer,
294 * instruct the GPU to jump back to the start right
295 * after processing the currently pending commands.
297 OUT_RING(chan
, chan
->pushbuf_base
| 0x20000000);
299 /* wait for GET to depart from the skips area.
300 * prevents writing GET==PUT and causing a race
301 * condition that causes us to think the GPU is
302 * idle when it's not.
305 get
= READ_GET(chan
, &prev_get
, &cnt
);
306 if (unlikely(get
== -EBUSY
))
308 if (unlikely(get
== -EINVAL
))
310 } while (get
<= NOUVEAU_DMA_SKIPS
);
311 WRITE_PUT(NOUVEAU_DMA_SKIPS
);
313 /* we're now submitting commands at the start of
317 chan
->dma
.put
= NOUVEAU_DMA_SKIPS
;
320 /* engine fetching ahead of us, we have space up until the
321 * current GET pointer. the "- 1" is to ensure there's
322 * space left to emit a jump back to the beginning of the
323 * push buffer if we require it. we can never get GET == PUT
324 * here, so this is safe.
326 chan
->dma
.free
= get
- chan
->dma
.cur
- 1;