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;
64 if (dev_priv->card_type >= NV_C0) {
65 ret = nouveau_gpuobj_gr_new(chan, 0x9039, 0x9039);
69 ret = RING_SPACE(chan, 2);
73 BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0000, 1);
74 OUT_RING (chan, 0x00009039);
79 /* Create NV_MEMORY_TO_MEMORY_FORMAT for buffer moves */
80 ret = nouveau_gpuobj_gr_new(chan, NvM2MF, dev_priv->card_type < NV_50 ?
85 /* NV_MEMORY_TO_MEMORY_FORMAT requires a notifier object */
86 ret = nouveau_notifier_alloc(chan, NvNotify0, 32, &chan->m2mf_ntfy);
90 /* Insert NOPS for NOUVEAU_DMA_SKIPS */
91 ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
95 for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
98 /* Initialise NV_MEMORY_TO_MEMORY_FORMAT */
99 ret = RING_SPACE(chan, 4);
102 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NAME, 1);
103 OUT_RING(chan, NvM2MF);
104 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_NOTIFY, 1);
105 OUT_RING(chan, NvNotify0);
107 /* Sit back and pray the channel works.. */
114 OUT_RINGp(struct nouveau_channel *chan, const void *data, unsigned nr_dwords)
117 u32 *mem = ttm_kmap_obj_virtual(&chan->pushbuf_bo->kmap, &is_iomem);
118 mem = &mem[chan->dma.cur];
120 memcpy_toio((void __force __iomem *)mem, data, nr_dwords * 4);
122 memcpy(mem, data, nr_dwords * 4);
123 chan->dma.cur += nr_dwords;
126 /* Fetch and adjust GPU GET pointer
129 * value >= 0, the adjusted GET pointer
130 * -EINVAL if GET pointer currently outside main push buffer
131 * -EBUSY if timeout exceeded
134 READ_GET(struct nouveau_channel *chan, uint32_t *prev_get, uint32_t *timeout)
138 val = nvchan_rd32(chan, chan->user_get);
140 /* reset counter as long as GET is still advancing, this is
141 * to avoid misdetecting a GPU lockup if the GPU happens to
142 * just be processing an operation that takes a long time
144 if (val != *prev_get) {
149 if ((++*timeout & 0xff) == 0) {
151 if (*timeout > 100000)
155 if (val < chan->pushbuf_base ||
156 val > chan->pushbuf_base + (chan->dma.max << 2))
159 return (val - chan->pushbuf_base) >> 2;
163 nv50_dma_push(struct nouveau_channel *chan, struct nouveau_bo *bo,
164 int delta, int length)
166 struct nouveau_bo *pb = chan->pushbuf_bo;
167 uint64_t offset = bo->bo.offset + delta;
168 int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base;
170 BUG_ON(chan->dma.ib_free < 1);
171 nouveau_bo_wr32(pb, ip++, lower_32_bits(offset));
172 nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8);
174 chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max;
178 nouveau_bo_rd32(pb, 0);
180 nvchan_wr32(chan, 0x8c, chan->dma.ib_put);
185 nv50_dma_push_wait(struct nouveau_channel *chan, int count)
187 uint32_t cnt = 0, prev_get = 0;
189 while (chan->dma.ib_free < count) {
190 uint32_t get = nvchan_rd32(chan, 0x88);
191 if (get != prev_get) {
196 if ((++cnt & 0xff) == 0) {
202 chan->dma.ib_free = get - chan->dma.ib_put;
203 if (chan->dma.ib_free <= 0)
204 chan->dma.ib_free += chan->dma.ib_max;
211 nv50_dma_wait(struct nouveau_channel *chan, int slots, int count)
213 uint32_t cnt = 0, prev_get = 0;
216 ret = nv50_dma_push_wait(chan, slots + 1);
220 while (chan->dma.free < count) {
221 int get = READ_GET(chan, &prev_get, &cnt);
222 if (unlikely(get < 0)) {
229 if (get <= chan->dma.cur) {
230 chan->dma.free = chan->dma.max - chan->dma.cur;
231 if (chan->dma.free >= count)
236 get = READ_GET(chan, &prev_get, &cnt);
237 if (unlikely(get < 0)) {
247 chan->dma.free = get - chan->dma.cur - 1;
254 nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size)
256 uint32_t prev_get = 0, cnt = 0;
259 if (chan->dma.ib_max)
260 return nv50_dma_wait(chan, slots, size);
262 while (chan->dma.free < size) {
263 get = READ_GET(chan, &prev_get, &cnt);
264 if (unlikely(get == -EBUSY))
267 /* loop until we have a usable GET pointer. the value
268 * we read from the GPU may be outside the main ring if
269 * PFIFO is processing a buffer called from the main ring,
270 * discard these values until something sensible is seen.
272 * the other case we discard GET is while the GPU is fetching
273 * from the SKIPS area, so the code below doesn't have to deal
274 * with some fun corner cases.
276 if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS)
279 if (get <= chan->dma.cur) {
280 /* engine is fetching behind us, or is completely
281 * idle (GET == PUT) so we have free space up until
282 * the end of the push buffer
284 * we can only hit that path once per call due to
285 * looping back to the beginning of the push buffer,
286 * we'll hit the fetching-ahead-of-us path from that
289 * the *one* exception to that rule is if we read
290 * GET==PUT, in which case the below conditional will
291 * always succeed and break us out of the wait loop.
293 chan->dma.free = chan->dma.max - chan->dma.cur;
294 if (chan->dma.free >= size)
297 /* not enough space left at the end of the push buffer,
298 * instruct the GPU to jump back to the start right
299 * after processing the currently pending commands.
301 OUT_RING(chan, chan->pushbuf_base | 0x20000000);
303 /* wait for GET to depart from the skips area.
304 * prevents writing GET==PUT and causing a race
305 * condition that causes us to think the GPU is
306 * idle when it's not.
309 get = READ_GET(chan, &prev_get, &cnt);
310 if (unlikely(get == -EBUSY))
312 if (unlikely(get == -EINVAL))
314 } while (get <= NOUVEAU_DMA_SKIPS);
315 WRITE_PUT(NOUVEAU_DMA_SKIPS);
317 /* we're now submitting commands at the start of
321 chan->dma.put = NOUVEAU_DMA_SKIPS;
324 /* engine fetching ahead of us, we have space up until the
325 * current GET pointer. the "- 1" is to ensure there's
326 * space left to emit a jump back to the beginning of the
327 * push buffer if we require it. we can never get GET == PUT
328 * here, so this is safe.
330 chan->dma.free = get - chan->dma.cur - 1;