1 /*
2  * Copyright 2008 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  */
28 #include <linux/seq_file.h>
29 #include "drmP.h"
30 #include "radeon_drm.h"
31 #include "radeon_reg.h"
32 #include "radeon.h"
33 #include "atom.h"
34 
35 int radeon_debugfs_ib_init(struct radeon_device *rdev);
36 
37 /*
38  * IB.
39  */
40 int radeon_ib_get(struct radeon_device *rdev, struct radeon_ib **ib)
41 {
42 	struct radeon_fence *fence;
43 	struct radeon_ib *nib;
44 	int r = 0, i, c;
45 
46 	*ib = NULL;
47 	r = radeon_fence_create(rdev, &fence);
48 	if (r) {
49 		dev_err(rdev->dev, "failed to create fence for new IB\n");
50 		return r;
51 	}
52 	mutex_lock(&rdev->ib_pool.mutex);
53 	for (i = rdev->ib_pool.head_id, c = 0, nib = NULL; c < RADEON_IB_POOL_SIZE; c++, i++) {
54 		i &= (RADEON_IB_POOL_SIZE - 1);
55 		if (rdev->ib_pool.ibs[i].free) {
56 			nib = &rdev->ib_pool.ibs[i];
57 			break;
58 		}
59 	}
60 	if (nib == NULL) {
61 		/* This should never happen, it means we allocated all
62 		 * IB and haven't scheduled one yet, return EBUSY to
63 		 * userspace hoping that on ioctl recall we get better
64 		 * luck
65 		 */
66 		dev_err(rdev->dev, "no free indirect buffer !\n");
67 		mutex_unlock(&rdev->ib_pool.mutex);
68 		radeon_fence_unref(&fence);
69 		return -EBUSY;
70 	}
71 	rdev->ib_pool.head_id = (nib->idx + 1) & (RADEON_IB_POOL_SIZE - 1);
72 	nib->free = false;
73 	if (nib->fence) {
74 		mutex_unlock(&rdev->ib_pool.mutex);
75 		r = radeon_fence_wait(nib->fence, false);
76 		if (r) {
77 			dev_err(rdev->dev, "error waiting fence of IB(%u:0x%016lX:%u)\n",
78 				nib->idx, (unsigned long)nib->gpu_addr, nib->length_dw);
79 			mutex_lock(&rdev->ib_pool.mutex);
80 			nib->free = true;
81 			mutex_unlock(&rdev->ib_pool.mutex);
82 			radeon_fence_unref(&fence);
83 			return r;
84 		}
85 		mutex_lock(&rdev->ib_pool.mutex);
86 	}
87 	radeon_fence_unref(&nib->fence);
88 	nib->fence = fence;
89 	nib->length_dw = 0;
90 	mutex_unlock(&rdev->ib_pool.mutex);
91 	*ib = nib;
92 	return 0;
93 }
94 
95 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib **ib)
96 {
97 	struct radeon_ib *tmp = *ib;
98 
99 	*ib = NULL;
100 	if (tmp == NULL) {
101 		return;
102 	}
103 	if (!tmp->fence->emited)
104 		radeon_fence_unref(&tmp->fence);
105 	mutex_lock(&rdev->ib_pool.mutex);
106 	tmp->free = true;
107 	mutex_unlock(&rdev->ib_pool.mutex);
108 }
109 
110 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib)
111 {
112 	int r = 0;
113 
114 	if (!ib->length_dw || !rdev->cp.ready) {
115 		/* TODO: Nothings in the ib we should report. */
116 		DRM_ERROR("radeon: couldn't schedule IB(%u).\n", ib->idx);
117 		return -EINVAL;
118 	}
119 
120 	/* 64 dwords should be enough for fence too */
121 	r = radeon_ring_lock(rdev, 64);
122 	if (r) {
123 		DRM_ERROR("radeon: scheduling IB failled (%d).\n", r);
124 		return r;
125 	}
126 	radeon_ring_ib_execute(rdev, ib);
127 	radeon_fence_emit(rdev, ib->fence);
128 	mutex_lock(&rdev->ib_pool.mutex);
129 	/* once scheduled IB is considered free and protected by the fence */
130 	ib->free = true;
131 	mutex_unlock(&rdev->ib_pool.mutex);
132 	radeon_ring_unlock_commit(rdev);
133 	return 0;
134 }
135 
136 int radeon_ib_pool_init(struct radeon_device *rdev)
137 {
138 	void *ptr;
139 	uint64_t gpu_addr;
140 	int i;
141 	int r = 0;
142 
143 	if (rdev->ib_pool.robj)
144 		return 0;
145 	/* Allocate 1M object buffer */
146 	r = radeon_bo_create(rdev, NULL,  RADEON_IB_POOL_SIZE*64*1024,
147 				true, RADEON_GEM_DOMAIN_GTT,
148 				&rdev->ib_pool.robj);
149 	if (r) {
150 		DRM_ERROR("radeon: failed to ib pool (%d).\n", r);
151 		return r;
152 	}
153 	r = radeon_bo_reserve(rdev->ib_pool.robj, false);
154 	if (unlikely(r != 0))
155 		return r;
156 	r = radeon_bo_pin(rdev->ib_pool.robj, RADEON_GEM_DOMAIN_GTT, &gpu_addr);
157 	if (r) {
158 		radeon_bo_unreserve(rdev->ib_pool.robj);
159 		DRM_ERROR("radeon: failed to pin ib pool (%d).\n", r);
160 		return r;
161 	}
162 	r = radeon_bo_kmap(rdev->ib_pool.robj, &ptr);
163 	radeon_bo_unreserve(rdev->ib_pool.robj);
164 	if (r) {
165 		DRM_ERROR("radeon: failed to map ib poll (%d).\n", r);
166 		return r;
167 	}
168 	for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
169 		unsigned offset;
170 
171 		offset = i * 64 * 1024;
172 		rdev->ib_pool.ibs[i].gpu_addr = gpu_addr + offset;
173 		rdev->ib_pool.ibs[i].ptr = ptr + offset;
174 		rdev->ib_pool.ibs[i].idx = i;
175 		rdev->ib_pool.ibs[i].length_dw = 0;
176 		rdev->ib_pool.ibs[i].free = true;
177 	}
178 	rdev->ib_pool.head_id = 0;
179 	rdev->ib_pool.ready = true;
180 	DRM_INFO("radeon: ib pool ready.\n");
181 	if (radeon_debugfs_ib_init(rdev)) {
182 		DRM_ERROR("Failed to register debugfs file for IB !\n");
183 	}
184 	return r;
185 }
186 
187 void radeon_ib_pool_fini(struct radeon_device *rdev)
188 {
189 	int r;
190 
191 	if (!rdev->ib_pool.ready) {
192 		return;
193 	}
194 	mutex_lock(&rdev->ib_pool.mutex);
195 	if (rdev->ib_pool.robj) {
196 		r = radeon_bo_reserve(rdev->ib_pool.robj, false);
197 		if (likely(r == 0)) {
198 			radeon_bo_kunmap(rdev->ib_pool.robj);
199 			radeon_bo_unpin(rdev->ib_pool.robj);
200 			radeon_bo_unreserve(rdev->ib_pool.robj);
201 		}
202 		radeon_bo_unref(&rdev->ib_pool.robj);
203 		rdev->ib_pool.robj = NULL;
204 	}
205 	mutex_unlock(&rdev->ib_pool.mutex);
206 }
207 
208 
209 /*
210  * Ring.
211  */
212 void radeon_ring_free_size(struct radeon_device *rdev)
213 {
214 	if (rdev->family >= CHIP_R600)
215 		rdev->cp.rptr = RREG32(R600_CP_RB_RPTR);
216 	else
217 		rdev->cp.rptr = RREG32(RADEON_CP_RB_RPTR);
218 	/* This works because ring_size is a power of 2 */
219 	rdev->cp.ring_free_dw = (rdev->cp.rptr + (rdev->cp.ring_size / 4));
220 	rdev->cp.ring_free_dw -= rdev->cp.wptr;
221 	rdev->cp.ring_free_dw &= rdev->cp.ptr_mask;
222 	if (!rdev->cp.ring_free_dw) {
223 		rdev->cp.ring_free_dw = rdev->cp.ring_size / 4;
224 	}
225 }
226 
227 int radeon_ring_lock(struct radeon_device *rdev, unsigned ndw)
228 {
229 	int r;
230 
231 	/* Align requested size with padding so unlock_commit can
232 	 * pad safely */
233 	ndw = (ndw + rdev->cp.align_mask) & ~rdev->cp.align_mask;
234 	mutex_lock(&rdev->cp.mutex);
235 	while (ndw > (rdev->cp.ring_free_dw - 1)) {
236 		radeon_ring_free_size(rdev);
237 		if (ndw < rdev->cp.ring_free_dw) {
238 			break;
239 		}
240 		r = radeon_fence_wait_next(rdev);
241 		if (r) {
242 			mutex_unlock(&rdev->cp.mutex);
243 			return r;
244 		}
245 	}
246 	rdev->cp.count_dw = ndw;
247 	rdev->cp.wptr_old = rdev->cp.wptr;
248 	return 0;
249 }
250 
251 void radeon_ring_unlock_commit(struct radeon_device *rdev)
252 {
253 	unsigned count_dw_pad;
254 	unsigned i;
255 
256 	/* We pad to match fetch size */
257 	count_dw_pad = (rdev->cp.align_mask + 1) -
258 		       (rdev->cp.wptr & rdev->cp.align_mask);
259 	for (i = 0; i < count_dw_pad; i++) {
260 		radeon_ring_write(rdev, 2 << 30);
261 	}
262 	DRM_MEMORYBARRIER();
263 	radeon_cp_commit(rdev);
264 	mutex_unlock(&rdev->cp.mutex);
265 }
266 
267 void radeon_ring_unlock_undo(struct radeon_device *rdev)
268 {
269 	rdev->cp.wptr = rdev->cp.wptr_old;
270 	mutex_unlock(&rdev->cp.mutex);
271 }
272 
273 int radeon_ring_init(struct radeon_device *rdev, unsigned ring_size)
274 {
275 	int r;
276 
277 	rdev->cp.ring_size = ring_size;
278 	/* Allocate ring buffer */
279 	if (rdev->cp.ring_obj == NULL) {
280 		r = radeon_bo_create(rdev, NULL, rdev->cp.ring_size, true,
281 					RADEON_GEM_DOMAIN_GTT,
282 					&rdev->cp.ring_obj);
283 		if (r) {
284 			dev_err(rdev->dev, "(%d) ring create failed\n", r);
285 			return r;
286 		}
287 		r = radeon_bo_reserve(rdev->cp.ring_obj, false);
288 		if (unlikely(r != 0))
289 			return r;
290 		r = radeon_bo_pin(rdev->cp.ring_obj, RADEON_GEM_DOMAIN_GTT,
291 					&rdev->cp.gpu_addr);
292 		if (r) {
293 			radeon_bo_unreserve(rdev->cp.ring_obj);
294 			dev_err(rdev->dev, "(%d) ring pin failed\n", r);
295 			return r;
296 		}
297 		r = radeon_bo_kmap(rdev->cp.ring_obj,
298 				       (void **)&rdev->cp.ring);
299 		radeon_bo_unreserve(rdev->cp.ring_obj);
300 		if (r) {
301 			dev_err(rdev->dev, "(%d) ring map failed\n", r);
302 			return r;
303 		}
304 	}
305 	rdev->cp.ptr_mask = (rdev->cp.ring_size / 4) - 1;
306 	rdev->cp.ring_free_dw = rdev->cp.ring_size / 4;
307 	return 0;
308 }
309 
310 void radeon_ring_fini(struct radeon_device *rdev)
311 {
312 	int r;
313 
314 	mutex_lock(&rdev->cp.mutex);
315 	if (rdev->cp.ring_obj) {
316 		r = radeon_bo_reserve(rdev->cp.ring_obj, false);
317 		if (likely(r == 0)) {
318 			radeon_bo_kunmap(rdev->cp.ring_obj);
319 			radeon_bo_unpin(rdev->cp.ring_obj);
320 			radeon_bo_unreserve(rdev->cp.ring_obj);
321 		}
322 		radeon_bo_unref(&rdev->cp.ring_obj);
323 		rdev->cp.ring = NULL;
324 		rdev->cp.ring_obj = NULL;
325 	}
326 	mutex_unlock(&rdev->cp.mutex);
327 }
328 
329 
330 /*
331  * Debugfs info
332  */
333 #if defined(CONFIG_DEBUG_FS)
334 static int radeon_debugfs_ib_info(struct seq_file *m, void *data)
335 {
336 	struct drm_info_node *node = (struct drm_info_node *) m->private;
337 	struct radeon_ib *ib = node->info_ent->data;
338 	unsigned i;
339 
340 	if (ib == NULL) {
341 		return 0;
342 	}
343 	seq_printf(m, "IB %04u\n", ib->idx);
344 	seq_printf(m, "IB fence %p\n", ib->fence);
345 	seq_printf(m, "IB size %05u dwords\n", ib->length_dw);
346 	for (i = 0; i < ib->length_dw; i++) {
347 		seq_printf(m, "[%05u]=0x%08X\n", i, ib->ptr[i]);
348 	}
349 	return 0;
350 }
351 
352 static struct drm_info_list radeon_debugfs_ib_list[RADEON_IB_POOL_SIZE];
353 static char radeon_debugfs_ib_names[RADEON_IB_POOL_SIZE][32];
354 #endif
355 
356 int radeon_debugfs_ib_init(struct radeon_device *rdev)
357 {
358 #if defined(CONFIG_DEBUG_FS)
359 	unsigned i;
360 
361 	for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
362 		sprintf(radeon_debugfs_ib_names[i], "radeon_ib_%04u", i);
363 		radeon_debugfs_ib_list[i].name = radeon_debugfs_ib_names[i];
364 		radeon_debugfs_ib_list[i].show = &radeon_debugfs_ib_info;
365 		radeon_debugfs_ib_list[i].driver_features = 0;
366 		radeon_debugfs_ib_list[i].data = &rdev->ib_pool.ibs[i];
367 	}
368 	return radeon_debugfs_add_files(rdev, radeon_debugfs_ib_list,
369 					RADEON_IB_POOL_SIZE);
370 #else
371 	return 0;
372 #endif
373 }
374