xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_ib.c (revision 275876e2)
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  *          Christian König
28  */
29 #include <drm/drmP.h>
30 #include "radeon.h"
31 
32 /*
33  * IB
34  * IBs (Indirect Buffers) and areas of GPU accessible memory where
35  * commands are stored.  You can put a pointer to the IB in the
36  * command ring and the hw will fetch the commands from the IB
37  * and execute them.  Generally userspace acceleration drivers
38  * produce command buffers which are send to the kernel and
39  * put in IBs for execution by the requested ring.
40  */
41 static int radeon_debugfs_sa_init(struct radeon_device *rdev);
42 
43 /**
44  * radeon_ib_get - request an IB (Indirect Buffer)
45  *
46  * @rdev: radeon_device pointer
47  * @ring: ring index the IB is associated with
48  * @ib: IB object returned
49  * @size: requested IB size
50  *
51  * Request an IB (all asics).  IBs are allocated using the
52  * suballocator.
53  * Returns 0 on success, error on failure.
54  */
55 int radeon_ib_get(struct radeon_device *rdev, int ring,
56 		  struct radeon_ib *ib, struct radeon_vm *vm,
57 		  unsigned size)
58 {
59 	int r;
60 
61 	r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo, &ib->sa_bo, size, 256);
62 	if (r) {
63 		dev_err(rdev->dev, "failed to get a new IB (%d)\n", r);
64 		return r;
65 	}
66 
67 	r = radeon_semaphore_create(rdev, &ib->semaphore);
68 	if (r) {
69 		return r;
70 	}
71 
72 	ib->ring = ring;
73 	ib->fence = NULL;
74 	ib->ptr = radeon_sa_bo_cpu_addr(ib->sa_bo);
75 	ib->vm = vm;
76 	if (vm) {
77 		/* ib pool is bound at RADEON_VA_IB_OFFSET in virtual address
78 		 * space and soffset is the offset inside the pool bo
79 		 */
80 		ib->gpu_addr = ib->sa_bo->soffset + RADEON_VA_IB_OFFSET;
81 	} else {
82 		ib->gpu_addr = radeon_sa_bo_gpu_addr(ib->sa_bo);
83 	}
84 	ib->is_const_ib = false;
85 
86 	return 0;
87 }
88 
89 /**
90  * radeon_ib_free - free an IB (Indirect Buffer)
91  *
92  * @rdev: radeon_device pointer
93  * @ib: IB object to free
94  *
95  * Free an IB (all asics).
96  */
97 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib)
98 {
99 	radeon_semaphore_free(rdev, &ib->semaphore, ib->fence);
100 	radeon_sa_bo_free(rdev, &ib->sa_bo, ib->fence);
101 	radeon_fence_unref(&ib->fence);
102 }
103 
104 /**
105  * radeon_ib_schedule - schedule an IB (Indirect Buffer) on the ring
106  *
107  * @rdev: radeon_device pointer
108  * @ib: IB object to schedule
109  * @const_ib: Const IB to schedule (SI only)
110  *
111  * Schedule an IB on the associated ring (all asics).
112  * Returns 0 on success, error on failure.
113  *
114  * On SI, there are two parallel engines fed from the primary ring,
115  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
116  * resource descriptors have moved to memory, the CE allows you to
117  * prime the caches while the DE is updating register state so that
118  * the resource descriptors will be already in cache when the draw is
119  * processed.  To accomplish this, the userspace driver submits two
120  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
121  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
122  * to SI there was just a DE IB.
123  */
124 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
125 		       struct radeon_ib *const_ib)
126 {
127 	struct radeon_ring *ring = &rdev->ring[ib->ring];
128 	int r = 0;
129 
130 	if (!ib->length_dw || !ring->ready) {
131 		/* TODO: Nothings in the ib we should report. */
132 		dev_err(rdev->dev, "couldn't schedule ib\n");
133 		return -EINVAL;
134 	}
135 
136 	/* 64 dwords should be enough for fence too */
137 	r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_SYNCS * 8);
138 	if (r) {
139 		dev_err(rdev->dev, "scheduling IB failed (%d).\n", r);
140 		return r;
141 	}
142 
143 	/* grab a vm id if necessary */
144 	if (ib->vm) {
145 		struct radeon_fence *vm_id_fence;
146 		vm_id_fence = radeon_vm_grab_id(rdev, ib->vm, ib->ring);
147         	radeon_semaphore_sync_to(ib->semaphore, vm_id_fence);
148 	}
149 
150 	/* sync with other rings */
151 	r = radeon_semaphore_sync_rings(rdev, ib->semaphore, ib->ring);
152 	if (r) {
153 		dev_err(rdev->dev, "failed to sync rings (%d)\n", r);
154 		radeon_ring_unlock_undo(rdev, ring);
155 		return r;
156 	}
157 
158 	if (ib->vm)
159 		radeon_vm_flush(rdev, ib->vm, ib->ring);
160 
161 	if (const_ib) {
162 		radeon_ring_ib_execute(rdev, const_ib->ring, const_ib);
163 		radeon_semaphore_free(rdev, &const_ib->semaphore, NULL);
164 	}
165 	radeon_ring_ib_execute(rdev, ib->ring, ib);
166 	r = radeon_fence_emit(rdev, &ib->fence, ib->ring);
167 	if (r) {
168 		dev_err(rdev->dev, "failed to emit fence for new IB (%d)\n", r);
169 		radeon_ring_unlock_undo(rdev, ring);
170 		return r;
171 	}
172 	if (const_ib) {
173 		const_ib->fence = radeon_fence_ref(ib->fence);
174 	}
175 
176 	if (ib->vm)
177 		radeon_vm_fence(rdev, ib->vm, ib->fence);
178 
179 	radeon_ring_unlock_commit(rdev, ring);
180 	return 0;
181 }
182 
183 /**
184  * radeon_ib_pool_init - Init the IB (Indirect Buffer) pool
185  *
186  * @rdev: radeon_device pointer
187  *
188  * Initialize the suballocator to manage a pool of memory
189  * for use as IBs (all asics).
190  * Returns 0 on success, error on failure.
191  */
192 int radeon_ib_pool_init(struct radeon_device *rdev)
193 {
194 	int r;
195 
196 	if (rdev->ib_pool_ready) {
197 		return 0;
198 	}
199 
200 	if (rdev->family >= CHIP_BONAIRE) {
201 		r = radeon_sa_bo_manager_init(rdev, &rdev->ring_tmp_bo,
202 					      RADEON_IB_POOL_SIZE*64*1024,
203 					      RADEON_GPU_PAGE_SIZE,
204 					      RADEON_GEM_DOMAIN_GTT,
205 					      RADEON_GEM_GTT_WC);
206 	} else {
207 		/* Before CIK, it's better to stick to cacheable GTT due
208 		 * to the command stream checking
209 		 */
210 		r = radeon_sa_bo_manager_init(rdev, &rdev->ring_tmp_bo,
211 					      RADEON_IB_POOL_SIZE*64*1024,
212 					      RADEON_GPU_PAGE_SIZE,
213 					      RADEON_GEM_DOMAIN_GTT, 0);
214 	}
215 	if (r) {
216 		return r;
217 	}
218 
219 	r = radeon_sa_bo_manager_start(rdev, &rdev->ring_tmp_bo);
220 	if (r) {
221 		return r;
222 	}
223 
224 	rdev->ib_pool_ready = true;
225 	if (radeon_debugfs_sa_init(rdev)) {
226 		dev_err(rdev->dev, "failed to register debugfs file for SA\n");
227 	}
228 	return 0;
229 }
230 
231 /**
232  * radeon_ib_pool_fini - Free the IB (Indirect Buffer) pool
233  *
234  * @rdev: radeon_device pointer
235  *
236  * Tear down the suballocator managing the pool of memory
237  * for use as IBs (all asics).
238  */
239 void radeon_ib_pool_fini(struct radeon_device *rdev)
240 {
241 	if (rdev->ib_pool_ready) {
242 		radeon_sa_bo_manager_suspend(rdev, &rdev->ring_tmp_bo);
243 		radeon_sa_bo_manager_fini(rdev, &rdev->ring_tmp_bo);
244 		rdev->ib_pool_ready = false;
245 	}
246 }
247 
248 /**
249  * radeon_ib_ring_tests - test IBs on the rings
250  *
251  * @rdev: radeon_device pointer
252  *
253  * Test an IB (Indirect Buffer) on each ring.
254  * If the test fails, disable the ring.
255  * Returns 0 on success, error if the primary GFX ring
256  * IB test fails.
257  */
258 int radeon_ib_ring_tests(struct radeon_device *rdev)
259 {
260 	unsigned i;
261 	int r;
262 
263 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
264 		struct radeon_ring *ring = &rdev->ring[i];
265 
266 		if (!ring->ready)
267 			continue;
268 
269 		r = radeon_ib_test(rdev, i, ring);
270 		if (r) {
271 			ring->ready = false;
272 			rdev->needs_reset = false;
273 
274 			if (i == RADEON_RING_TYPE_GFX_INDEX) {
275 				/* oh, oh, that's really bad */
276 				DRM_ERROR("radeon: failed testing IB on GFX ring (%d).\n", r);
277 		                rdev->accel_working = false;
278 				return r;
279 
280 			} else {
281 				/* still not good, but we can live with it */
282 				DRM_ERROR("radeon: failed testing IB on ring %d (%d).\n", i, r);
283 			}
284 		}
285 	}
286 	return 0;
287 }
288 
289 /*
290  * Debugfs info
291  */
292 #if defined(CONFIG_DEBUG_FS)
293 
294 static int radeon_debugfs_sa_info(struct seq_file *m, void *data)
295 {
296 	struct drm_info_node *node = (struct drm_info_node *) m->private;
297 	struct drm_device *dev = node->minor->dev;
298 	struct radeon_device *rdev = dev->dev_private;
299 
300 	radeon_sa_bo_dump_debug_info(&rdev->ring_tmp_bo, m);
301 
302 	return 0;
303 
304 }
305 
306 static struct drm_info_list radeon_debugfs_sa_list[] = {
307         {"radeon_sa_info", &radeon_debugfs_sa_info, 0, NULL},
308 };
309 
310 #endif
311 
312 static int radeon_debugfs_sa_init(struct radeon_device *rdev)
313 {
314 #if defined(CONFIG_DEBUG_FS)
315 	return radeon_debugfs_add_files(rdev, radeon_debugfs_sa_list, 1);
316 #else
317 	return 0;
318 #endif
319 }
320