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"
33 nouveau_dma_pre_init(struct nouveau_channel *chan)
35 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
36 struct nouveau_bo *pushbuf = chan->pushbuf_bo;
38 if (dev_priv->card_type == NV_50) {
39 const int ib_size = pushbuf->bo.mem.size / 2;
41 chan->dma.ib_base = (pushbuf->bo.mem.size - ib_size) >> 2;
42 chan->dma.ib_max = (ib_size / 8) - 1;
44 chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
46 chan->dma.max = (pushbuf->bo.mem.size - ib_size) >> 2;
48 chan->dma.max = (pushbuf->bo.mem.size >> 2) - 2;
52 chan->dma.cur = chan->dma.put;
53 chan->dma.free = chan->dma.max - chan->dma.cur;
57 nouveau_dma_init(struct nouveau_channel *chan)
59 struct drm_device *dev = chan->dev;
60 struct drm_nouveau_private *dev_priv = dev->dev_private;
61 struct nouveau_gpuobj *m2mf = NULL;
62 struct nouveau_gpuobj *nvsw = 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, &m2mf);
71 ret = nouveau_gpuobj_ref_add(dev, chan, NvM2MF, m2mf, NULL);
75 /* Create an NV_SW object for various sync purposes */
76 ret = nouveau_gpuobj_sw_new(chan, NV_SW, &nvsw);
80 ret = nouveau_gpuobj_ref_add(dev, chan, NvSw, nvsw, NULL);
84 /* NV_MEMORY_TO_MEMORY_FORMAT requires a notifier object */
85 ret = nouveau_notifier_alloc(chan, NvNotify0, 32, &chan->m2mf_ntfy);
90 ret = nouveau_bo_map(chan->pushbuf_bo);
94 /* Map M2MF notifier object - fbcon. */
95 if (drm_core_check_feature(dev, DRIVER_MODESET)) {
96 ret = nouveau_bo_map(chan->notifier_bo);
101 /* Insert NOPS for NOUVEAU_DMA_SKIPS */
102 ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
106 for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
109 /* Initialise NV_MEMORY_TO_MEMORY_FORMAT */
110 ret = RING_SPACE(chan, 4);
113 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NAME, 1);
114 OUT_RING(chan, NvM2MF);
115 BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_NOTIFY, 1);
116 OUT_RING(chan, NvNotify0);
118 /* Initialise NV_SW */
119 ret = RING_SPACE(chan, 2);
122 BEGIN_RING(chan, NvSubSw, 0, 1);
123 OUT_RING(chan, NvSw);
125 /* Sit back and pray the channel works.. */
132 OUT_RINGp(struct nouveau_channel *chan, const void *data, unsigned nr_dwords)
135 u32 *mem = ttm_kmap_obj_virtual(&chan->pushbuf_bo->kmap, &is_iomem);
136 mem = &mem[chan->dma.cur];
138 memcpy_toio((void __force __iomem *)mem, data, nr_dwords * 4);
140 memcpy(mem, data, nr_dwords * 4);
141 chan->dma.cur += nr_dwords;
144 /* Fetch and adjust GPU GET pointer
147 * value >= 0, the adjusted GET pointer
148 * -EINVAL if GET pointer currently outside main push buffer
149 * -EBUSY if timeout exceeded
152 READ_GET(struct nouveau_channel *chan, uint32_t *prev_get, uint32_t *timeout)
156 val = nvchan_rd32(chan, chan->user_get);
158 /* reset counter as long as GET is still advancing, this is
159 * to avoid misdetecting a GPU lockup if the GPU happens to
160 * just be processing an operation that takes a long time
162 if (val != *prev_get) {
167 if ((++*timeout & 0xff) == 0) {
169 if (*timeout > 100000)
173 if (val < chan->pushbuf_base ||
174 val > chan->pushbuf_base + (chan->dma.max << 2))
177 return (val - chan->pushbuf_base) >> 2;
181 nv50_dma_push(struct nouveau_channel *chan, struct nouveau_bo *bo,
182 int delta, int length)
184 struct nouveau_bo *pb = chan->pushbuf_bo;
185 uint64_t offset = bo->bo.offset + delta;
186 int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base;
188 BUG_ON(chan->dma.ib_free < 1);
189 nouveau_bo_wr32(pb, ip++, lower_32_bits(offset));
190 nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8);
192 chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max;
193 nvchan_wr32(chan, 0x8c, chan->dma.ib_put);
198 nv50_dma_push_wait(struct nouveau_channel *chan, int count)
200 uint32_t cnt = 0, prev_get = 0;
202 while (chan->dma.ib_free < count) {
203 uint32_t get = nvchan_rd32(chan, 0x88);
204 if (get != prev_get) {
209 if ((++cnt & 0xff) == 0) {
215 chan->dma.ib_free = get - chan->dma.ib_put;
216 if (chan->dma.ib_free <= 0)
217 chan->dma.ib_free += chan->dma.ib_max + 1;
224 nv50_dma_wait(struct nouveau_channel *chan, int slots, int count)
226 uint32_t cnt = 0, prev_get = 0;
229 ret = nv50_dma_push_wait(chan, slots + 1);
233 while (chan->dma.free < count) {
234 int get = READ_GET(chan, &prev_get, &cnt);
235 if (unlikely(get < 0)) {
242 if (get <= chan->dma.cur) {
243 chan->dma.free = chan->dma.max - chan->dma.cur;
244 if (chan->dma.free >= count)
249 get = READ_GET(chan, &prev_get, &cnt);
250 if (unlikely(get < 0)) {
260 chan->dma.free = get - chan->dma.cur - 1;
267 nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size)
269 uint32_t prev_get = 0, cnt = 0;
272 if (chan->dma.ib_max)
273 return nv50_dma_wait(chan, slots, size);
275 while (chan->dma.free < size) {
276 get = READ_GET(chan, &prev_get, &cnt);
277 if (unlikely(get == -EBUSY))
280 /* loop until we have a usable GET pointer. the value
281 * we read from the GPU may be outside the main ring if
282 * PFIFO is processing a buffer called from the main ring,
283 * discard these values until something sensible is seen.
285 * the other case we discard GET is while the GPU is fetching
286 * from the SKIPS area, so the code below doesn't have to deal
287 * with some fun corner cases.
289 if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS)
292 if (get <= chan->dma.cur) {
293 /* engine is fetching behind us, or is completely
294 * idle (GET == PUT) so we have free space up until
295 * the end of the push buffer
297 * we can only hit that path once per call due to
298 * looping back to the beginning of the push buffer,
299 * we'll hit the fetching-ahead-of-us path from that
302 * the *one* exception to that rule is if we read
303 * GET==PUT, in which case the below conditional will
304 * always succeed and break us out of the wait loop.
306 chan->dma.free = chan->dma.max - chan->dma.cur;
307 if (chan->dma.free >= size)
310 /* not enough space left at the end of the push buffer,
311 * instruct the GPU to jump back to the start right
312 * after processing the currently pending commands.
314 OUT_RING(chan, chan->pushbuf_base | 0x20000000);
316 /* wait for GET to depart from the skips area.
317 * prevents writing GET==PUT and causing a race
318 * condition that causes us to think the GPU is
319 * idle when it's not.
322 get = READ_GET(chan, &prev_get, &cnt);
323 if (unlikely(get == -EBUSY))
325 if (unlikely(get == -EINVAL))
327 } while (get <= NOUVEAU_DMA_SKIPS);
328 WRITE_PUT(NOUVEAU_DMA_SKIPS);
330 /* we're now submitting commands at the start of
334 chan->dma.put = NOUVEAU_DMA_SKIPS;
337 /* engine fetching ahead of us, we have space up until the
338 * current GET pointer. the "- 1" is to ensure there's
339 * space left to emit a jump back to the beginning of the
340 * push buffer if we require it. we can never get GET == PUT
341 * here, so this is safe.
343 chan->dma.free = get - chan->dma.cur - 1;