Linux Audio

Check our new training course

Yocto / OpenEmbedded training

Mar 24-27, 2025, special US time zones
Register
Loading...
v3.1
  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 "drmP.h"
 28#include "drm.h"
 29#include "nouveau_drv.h"
 30#include "nouveau_dma.h"
 31#include "nouveau_ramht.h"
 32
 33void
 34nouveau_dma_pre_init(struct nouveau_channel *chan)
 35{
 36	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
 37	struct nouveau_bo *pushbuf = chan->pushbuf_bo;
 38
 39	if (dev_priv->card_type >= NV_50) {
 40		const int ib_size = pushbuf->bo.mem.size / 2;
 41
 42		chan->dma.ib_base = (pushbuf->bo.mem.size - ib_size) >> 2;
 43		chan->dma.ib_max = (ib_size / 8) - 1;
 44		chan->dma.ib_put = 0;
 45		chan->dma.ib_free = chan->dma.ib_max - chan->dma.ib_put;
 46
 47		chan->dma.max = (pushbuf->bo.mem.size - ib_size) >> 2;
 48	} else {
 49		chan->dma.max  = (pushbuf->bo.mem.size >> 2) - 2;
 50	}
 51
 52	chan->dma.put  = 0;
 53	chan->dma.cur  = chan->dma.put;
 54	chan->dma.free = chan->dma.max - chan->dma.cur;
 55}
 56
 57int
 58nouveau_dma_init(struct nouveau_channel *chan)
 59{
 60	struct drm_device *dev = chan->dev;
 61	struct drm_nouveau_private *dev_priv = dev->dev_private;
 62	int ret, i;
 63
 64	if (dev_priv->card_type >= NV_C0) {
 65		ret = nouveau_gpuobj_gr_new(chan, 0x9039, 0x9039);
 66		if (ret)
 67			return ret;
 68
 69		ret = RING_SPACE(chan, 2);
 70		if (ret)
 71			return ret;
 72
 73		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0000, 1);
 74		OUT_RING  (chan, 0x00009039);
 75		FIRE_RING (chan);
 76		return 0;
 77	}
 78
 79	/* Create NV_MEMORY_TO_MEMORY_FORMAT for buffer moves */
 80	ret = nouveau_gpuobj_gr_new(chan, NvM2MF, dev_priv->card_type < NV_50 ?
 81				    0x0039 : 0x5039);
 82	if (ret)
 83		return ret;
 84
 85	/* NV_MEMORY_TO_MEMORY_FORMAT requires a notifier object */
 86	ret = nouveau_notifier_alloc(chan, NvNotify0, 32, 0xfe0, 0x1000,
 87				     &chan->m2mf_ntfy);
 88	if (ret)
 89		return ret;
 90
 91	/* Insert NOPS for NOUVEAU_DMA_SKIPS */
 92	ret = RING_SPACE(chan, NOUVEAU_DMA_SKIPS);
 93	if (ret)
 94		return ret;
 95
 96	for (i = 0; i < NOUVEAU_DMA_SKIPS; i++)
 97		OUT_RING(chan, 0);
 98
 99	/* Initialise NV_MEMORY_TO_MEMORY_FORMAT */
100	ret = RING_SPACE(chan, 6);
101	if (ret)
102		return ret;
103	BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NAME, 1);
104	OUT_RING  (chan, NvM2MF);
105	BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_NOTIFY, 3);
106	OUT_RING  (chan, NvNotify0);
107	OUT_RING  (chan, chan->vram_handle);
108	OUT_RING  (chan, chan->gart_handle);
109
110	/* Sit back and pray the channel works.. */
111	FIRE_RING(chan);
112
113	return 0;
114}
115
116void
117OUT_RINGp(struct nouveau_channel *chan, const void *data, unsigned nr_dwords)
118{
119	bool is_iomem;
120	u32 *mem = ttm_kmap_obj_virtual(&chan->pushbuf_bo->kmap, &is_iomem);
121	mem = &mem[chan->dma.cur];
122	if (is_iomem)
123		memcpy_toio((void __force __iomem *)mem, data, nr_dwords * 4);
124	else
125		memcpy(mem, data, nr_dwords * 4);
126	chan->dma.cur += nr_dwords;
127}
128
129/* Fetch and adjust GPU GET pointer
130 *
131 * Returns:
132 *  value >= 0, the adjusted GET pointer
133 *  -EINVAL if GET pointer currently outside main push buffer
134 *  -EBUSY if timeout exceeded
135 */
136static inline int
137READ_GET(struct nouveau_channel *chan, uint32_t *prev_get, uint32_t *timeout)
138{
139	uint32_t val;
140
141	val = nvchan_rd32(chan, chan->user_get);
 
 
142
143	/* reset counter as long as GET is still advancing, this is
144	 * to avoid misdetecting a GPU lockup if the GPU happens to
145	 * just be processing an operation that takes a long time
146	 */
147	if (val != *prev_get) {
148		*prev_get = val;
149		*timeout = 0;
150	}
151
152	if ((++*timeout & 0xff) == 0) {
153		DRM_UDELAY(1);
154		if (*timeout > 100000)
155			return -EBUSY;
156	}
157
158	if (val < chan->pushbuf_base ||
159	    val > chan->pushbuf_base + (chan->dma.max << 2))
160		return -EINVAL;
161
162	return (val - chan->pushbuf_base) >> 2;
163}
164
165void
166nv50_dma_push(struct nouveau_channel *chan, struct nouveau_bo *bo,
167	      int delta, int length)
168{
169	struct nouveau_bo *pb = chan->pushbuf_bo;
170	struct nouveau_vma *vma;
171	int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base;
172	u64 offset;
173
174	vma = nouveau_bo_vma_find(bo, chan->vm);
175	BUG_ON(!vma);
176	offset = vma->offset + delta;
177
178	BUG_ON(chan->dma.ib_free < 1);
 
179	nouveau_bo_wr32(pb, ip++, lower_32_bits(offset));
180	nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8);
181
182	chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max;
183
184	DRM_MEMORYBARRIER();
185	/* Flush writes. */
186	nouveau_bo_rd32(pb, 0);
187
188	nvchan_wr32(chan, 0x8c, chan->dma.ib_put);
 
 
189	chan->dma.ib_free--;
190}
191
192static int
193nv50_dma_push_wait(struct nouveau_channel *chan, int count)
194{
195	uint32_t cnt = 0, prev_get = 0;
196
197	while (chan->dma.ib_free < count) {
198		uint32_t get = nvchan_rd32(chan, 0x88);
199		if (get != prev_get) {
200			prev_get = get;
201			cnt = 0;
202		}
203
204		if ((++cnt & 0xff) == 0) {
205			DRM_UDELAY(1);
206			if (cnt > 100000)
207				return -EBUSY;
208		}
209
210		chan->dma.ib_free = get - chan->dma.ib_put;
211		if (chan->dma.ib_free <= 0)
212			chan->dma.ib_free += chan->dma.ib_max;
213	}
214
215	return 0;
216}
217
218static int
219nv50_dma_wait(struct nouveau_channel *chan, int slots, int count)
220{
221	uint32_t cnt = 0, prev_get = 0;
222	int ret;
223
224	ret = nv50_dma_push_wait(chan, slots + 1);
225	if (unlikely(ret))
226		return ret;
227
228	while (chan->dma.free < count) {
229		int get = READ_GET(chan, &prev_get, &cnt);
230		if (unlikely(get < 0)) {
231			if (get == -EINVAL)
232				continue;
233
234			return get;
235		}
236
237		if (get <= chan->dma.cur) {
238			chan->dma.free = chan->dma.max - chan->dma.cur;
239			if (chan->dma.free >= count)
240				break;
241
242			FIRE_RING(chan);
243			do {
244				get = READ_GET(chan, &prev_get, &cnt);
245				if (unlikely(get < 0)) {
246					if (get == -EINVAL)
247						continue;
248					return get;
249				}
250			} while (get == 0);
251			chan->dma.cur = 0;
252			chan->dma.put = 0;
253		}
254
255		chan->dma.free = get - chan->dma.cur - 1;
256	}
257
258	return 0;
259}
260
261int
262nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size)
263{
264	uint32_t prev_get = 0, cnt = 0;
265	int get;
266
267	if (chan->dma.ib_max)
268		return nv50_dma_wait(chan, slots, size);
269
270	while (chan->dma.free < size) {
271		get = READ_GET(chan, &prev_get, &cnt);
272		if (unlikely(get == -EBUSY))
273			return -EBUSY;
274
275		/* loop until we have a usable GET pointer.  the value
276		 * we read from the GPU may be outside the main ring if
277		 * PFIFO is processing a buffer called from the main ring,
278		 * discard these values until something sensible is seen.
279		 *
280		 * the other case we discard GET is while the GPU is fetching
281		 * from the SKIPS area, so the code below doesn't have to deal
282		 * with some fun corner cases.
283		 */
284		if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS)
285			continue;
286
287		if (get <= chan->dma.cur) {
288			/* engine is fetching behind us, or is completely
289			 * idle (GET == PUT) so we have free space up until
290			 * the end of the push buffer
291			 *
292			 * we can only hit that path once per call due to
293			 * looping back to the beginning of the push buffer,
294			 * we'll hit the fetching-ahead-of-us path from that
295			 * point on.
296			 *
297			 * the *one* exception to that rule is if we read
298			 * GET==PUT, in which case the below conditional will
299			 * always succeed and break us out of the wait loop.
300			 */
301			chan->dma.free = chan->dma.max - chan->dma.cur;
302			if (chan->dma.free >= size)
303				break;
304
305			/* not enough space left at the end of the push buffer,
306			 * instruct the GPU to jump back to the start right
307			 * after processing the currently pending commands.
308			 */
309			OUT_RING(chan, chan->pushbuf_base | 0x20000000);
310
311			/* wait for GET to depart from the skips area.
312			 * prevents writing GET==PUT and causing a race
313			 * condition that causes us to think the GPU is
314			 * idle when it's not.
315			 */
316			do {
317				get = READ_GET(chan, &prev_get, &cnt);
318				if (unlikely(get == -EBUSY))
319					return -EBUSY;
320				if (unlikely(get == -EINVAL))
321					continue;
322			} while (get <= NOUVEAU_DMA_SKIPS);
323			WRITE_PUT(NOUVEAU_DMA_SKIPS);
324
325			/* we're now submitting commands at the start of
326			 * the push buffer.
327			 */
328			chan->dma.cur  =
329			chan->dma.put  = NOUVEAU_DMA_SKIPS;
330		}
331
332		/* engine fetching ahead of us, we have space up until the
333		 * current GET pointer.  the "- 1" is to ensure there's
334		 * space left to emit a jump back to the beginning of the
335		 * push buffer if we require it.  we can never get GET == PUT
336		 * here, so this is safe.
337		 */
338		chan->dma.free = get - chan->dma.cur - 1;
339	}
340
341	return 0;
342}
343
v5.9
  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/* Fetch and adjust GPU GET pointer
 34 *
 35 * Returns:
 36 *  value >= 0, the adjusted GET pointer
 37 *  -EINVAL if GET pointer currently outside main push buffer
 38 *  -EBUSY if timeout exceeded
 39 */
 40static inline int
 41READ_GET(struct nouveau_channel *chan, uint64_t *prev_get, int *timeout)
 42{
 43	uint64_t val;
 44
 45	val = nvif_rd32(&chan->user, chan->user_get);
 46        if (chan->user_get_hi)
 47                val |= (uint64_t)nvif_rd32(&chan->user, chan->user_get_hi) << 32;
 48
 49	/* reset counter as long as GET is still advancing, this is
 50	 * to avoid misdetecting a GPU lockup if the GPU happens to
 51	 * just be processing an operation that takes a long time
 52	 */
 53	if (val != *prev_get) {
 54		*prev_get = val;
 55		*timeout = 0;
 56	}
 57
 58	if ((++*timeout & 0xff) == 0) {
 59		udelay(1);
 60		if (*timeout > 100000)
 61			return -EBUSY;
 62	}
 63
 64	if (val < chan->push.addr ||
 65	    val > chan->push.addr + (chan->dma.max << 2))
 66		return -EINVAL;
 67
 68	return (val - chan->push.addr) >> 2;
 69}
 70
 71void
 72nv50_dma_push(struct nouveau_channel *chan, u64 offset, int length)
 
 73{
 74	struct nvif_user *user = &chan->drm->client.device.user;
 75	struct nouveau_bo *pb = chan->push.buffer;
 76	int ip = (chan->dma.ib_put * 2) + chan->dma.ib_base;
 
 
 
 
 
 77
 78	BUG_ON(chan->dma.ib_free < 1);
 79
 80	nouveau_bo_wr32(pb, ip++, lower_32_bits(offset));
 81	nouveau_bo_wr32(pb, ip++, upper_32_bits(offset) | length << 8);
 82
 83	chan->dma.ib_put = (chan->dma.ib_put + 1) & chan->dma.ib_max;
 84
 85	mb();
 86	/* Flush writes. */
 87	nouveau_bo_rd32(pb, 0);
 88
 89	nvif_wr32(&chan->user, 0x8c, chan->dma.ib_put);
 90	if (user->func && user->func->doorbell)
 91		user->func->doorbell(user, chan->token);
 92	chan->dma.ib_free--;
 93}
 94
 95static int
 96nv50_dma_push_wait(struct nouveau_channel *chan, int count)
 97{
 98	uint32_t cnt = 0, prev_get = 0;
 99
100	while (chan->dma.ib_free < count) {
101		uint32_t get = nvif_rd32(&chan->user, 0x88);
102		if (get != prev_get) {
103			prev_get = get;
104			cnt = 0;
105		}
106
107		if ((++cnt & 0xff) == 0) {
108			udelay(1);
109			if (cnt > 100000)
110				return -EBUSY;
111		}
112
113		chan->dma.ib_free = get - chan->dma.ib_put;
114		if (chan->dma.ib_free <= 0)
115			chan->dma.ib_free += chan->dma.ib_max;
116	}
117
118	return 0;
119}
120
121static int
122nv50_dma_wait(struct nouveau_channel *chan, int slots, int count)
123{
124	uint64_t prev_get = 0;
125	int ret, cnt = 0;
126
127	ret = nv50_dma_push_wait(chan, slots + 1);
128	if (unlikely(ret))
129		return ret;
130
131	while (chan->dma.free < count) {
132		int get = READ_GET(chan, &prev_get, &cnt);
133		if (unlikely(get < 0)) {
134			if (get == -EINVAL)
135				continue;
136
137			return get;
138		}
139
140		if (get <= chan->dma.cur) {
141			chan->dma.free = chan->dma.max - chan->dma.cur;
142			if (chan->dma.free >= count)
143				break;
144
145			FIRE_RING(chan);
146			do {
147				get = READ_GET(chan, &prev_get, &cnt);
148				if (unlikely(get < 0)) {
149					if (get == -EINVAL)
150						continue;
151					return get;
152				}
153			} while (get == 0);
154			chan->dma.cur = 0;
155			chan->dma.put = 0;
156		}
157
158		chan->dma.free = get - chan->dma.cur - 1;
159	}
160
161	return 0;
162}
163
164int
165nouveau_dma_wait(struct nouveau_channel *chan, int slots, int size)
166{
167	uint64_t prev_get = 0;
168	int cnt = 0, get;
169
170	if (chan->dma.ib_max)
171		return nv50_dma_wait(chan, slots, size);
172
173	while (chan->dma.free < size) {
174		get = READ_GET(chan, &prev_get, &cnt);
175		if (unlikely(get == -EBUSY))
176			return -EBUSY;
177
178		/* loop until we have a usable GET pointer.  the value
179		 * we read from the GPU may be outside the main ring if
180		 * PFIFO is processing a buffer called from the main ring,
181		 * discard these values until something sensible is seen.
182		 *
183		 * the other case we discard GET is while the GPU is fetching
184		 * from the SKIPS area, so the code below doesn't have to deal
185		 * with some fun corner cases.
186		 */
187		if (unlikely(get == -EINVAL) || get < NOUVEAU_DMA_SKIPS)
188			continue;
189
190		if (get <= chan->dma.cur) {
191			/* engine is fetching behind us, or is completely
192			 * idle (GET == PUT) so we have free space up until
193			 * the end of the push buffer
194			 *
195			 * we can only hit that path once per call due to
196			 * looping back to the beginning of the push buffer,
197			 * we'll hit the fetching-ahead-of-us path from that
198			 * point on.
199			 *
200			 * the *one* exception to that rule is if we read
201			 * GET==PUT, in which case the below conditional will
202			 * always succeed and break us out of the wait loop.
203			 */
204			chan->dma.free = chan->dma.max - chan->dma.cur;
205			if (chan->dma.free >= size)
206				break;
207
208			/* not enough space left at the end of the push buffer,
209			 * instruct the GPU to jump back to the start right
210			 * after processing the currently pending commands.
211			 */
212			OUT_RING(chan, chan->push.addr | 0x20000000);
213
214			/* wait for GET to depart from the skips area.
215			 * prevents writing GET==PUT and causing a race
216			 * condition that causes us to think the GPU is
217			 * idle when it's not.
218			 */
219			do {
220				get = READ_GET(chan, &prev_get, &cnt);
221				if (unlikely(get == -EBUSY))
222					return -EBUSY;
223				if (unlikely(get == -EINVAL))
224					continue;
225			} while (get <= NOUVEAU_DMA_SKIPS);
226			WRITE_PUT(NOUVEAU_DMA_SKIPS);
227
228			/* we're now submitting commands at the start of
229			 * the push buffer.
230			 */
231			chan->dma.cur  =
232			chan->dma.put  = NOUVEAU_DMA_SKIPS;
233		}
234
235		/* engine fetching ahead of us, we have space up until the
236		 * current GET pointer.  the "- 1" is to ensure there's
237		 * space left to emit a jump back to the beginning of the
238		 * push buffer if we require it.  we can never get GET == PUT
239		 * here, so this is safe.
240		 */
241		chan->dma.free = get - chan->dma.cur - 1;
242	}
243
244	return 0;
245}
246