xref: /openbmc/linux/drivers/gpu/drm/radeon/radeon_ib.c (revision d9fd5a71)
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 
30 #include <drm/drm_debugfs.h>
31 #include <drm/drm_file.h>
32 
33 #include "radeon.h"
34 
35 /*
36  * IB
37  * IBs (Indirect Buffers) and areas of GPU accessible memory where
38  * commands are stored.  You can put a pointer to the IB in the
39  * command ring and the hw will fetch the commands from the IB
40  * and execute them.  Generally userspace acceleration drivers
41  * produce command buffers which are send to the kernel and
42  * put in IBs for execution by the requested ring.
43  */
44 static int radeon_debugfs_sa_init(struct radeon_device *rdev);
45 
46 /**
47  * radeon_ib_get - request an IB (Indirect Buffer)
48  *
49  * @rdev: radeon_device pointer
50  * @ring: ring index the IB is associated with
51  * @vm: requested vm
52  * @ib: IB object returned
53  * @size: requested IB size
54  *
55  * Request an IB (all asics).  IBs are allocated using the
56  * suballocator.
57  * Returns 0 on success, error on failure.
58  */
59 int radeon_ib_get(struct radeon_device *rdev, int ring,
60 		  struct radeon_ib *ib, struct radeon_vm *vm,
61 		  unsigned size)
62 {
63 	int r;
64 
65 	r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo, &ib->sa_bo, size, 256);
66 	if (r) {
67 		dev_err(rdev->dev, "failed to get a new IB (%d)\n", r);
68 		return r;
69 	}
70 
71 	radeon_sync_create(&ib->sync);
72 
73 	ib->ring = ring;
74 	ib->fence = NULL;
75 	ib->ptr = radeon_sa_bo_cpu_addr(ib->sa_bo);
76 	ib->vm = vm;
77 	if (vm) {
78 		/* ib pool is bound at RADEON_VA_IB_OFFSET in virtual address
79 		 * space and soffset is the offset inside the pool bo
80 		 */
81 		ib->gpu_addr = ib->sa_bo->soffset + RADEON_VA_IB_OFFSET;
82 	} else {
83 		ib->gpu_addr = radeon_sa_bo_gpu_addr(ib->sa_bo);
84 	}
85 	ib->is_const_ib = false;
86 
87 	return 0;
88 }
89 
90 /**
91  * radeon_ib_free - free an IB (Indirect Buffer)
92  *
93  * @rdev: radeon_device pointer
94  * @ib: IB object to free
95  *
96  * Free an IB (all asics).
97  */
98 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib)
99 {
100 	radeon_sync_free(rdev, &ib->sync, ib->fence);
101 	radeon_sa_bo_free(rdev, &ib->sa_bo, ib->fence);
102 	radeon_fence_unref(&ib->fence);
103 }
104 
105 /**
106  * radeon_ib_schedule - schedule an IB (Indirect Buffer) on the ring
107  *
108  * @rdev: radeon_device pointer
109  * @ib: IB object to schedule
110  * @const_ib: Const IB to schedule (SI only)
111  * @hdp_flush: Whether or not to perform an HDP cache flush
112  *
113  * Schedule an IB on the associated ring (all asics).
114  * Returns 0 on success, error on failure.
115  *
116  * On SI, there are two parallel engines fed from the primary ring,
117  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
118  * resource descriptors have moved to memory, the CE allows you to
119  * prime the caches while the DE is updating register state so that
120  * the resource descriptors will be already in cache when the draw is
121  * processed.  To accomplish this, the userspace driver submits two
122  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
123  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
124  * to SI there was just a DE IB.
125  */
126 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
127 		       struct radeon_ib *const_ib, bool hdp_flush)
128 {
129 	struct radeon_ring *ring = &rdev->ring[ib->ring];
130 	int r = 0;
131 
132 	if (!ib->length_dw || !ring->ready) {
133 		/* TODO: Nothings in the ib we should report. */
134 		dev_err(rdev->dev, "couldn't schedule ib\n");
135 		return -EINVAL;
136 	}
137 
138 	/* 64 dwords should be enough for fence too */
139 	r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_SYNCS * 8);
140 	if (r) {
141 		dev_err(rdev->dev, "scheduling IB failed (%d).\n", r);
142 		return r;
143 	}
144 
145 	/* grab a vm id if necessary */
146 	if (ib->vm) {
147 		struct radeon_fence *vm_id_fence;
148 		vm_id_fence = radeon_vm_grab_id(rdev, ib->vm, ib->ring);
149 		radeon_sync_fence(&ib->sync, vm_id_fence);
150 	}
151 
152 	/* sync with other rings */
153 	r = radeon_sync_rings(rdev, &ib->sync, ib->ring);
154 	if (r) {
155 		dev_err(rdev->dev, "failed to sync rings (%d)\n", r);
156 		radeon_ring_unlock_undo(rdev, ring);
157 		return r;
158 	}
159 
160 	if (ib->vm)
161 		radeon_vm_flush(rdev, ib->vm, ib->ring,
162 				ib->sync.last_vm_update);
163 
164 	if (const_ib) {
165 		radeon_ring_ib_execute(rdev, const_ib->ring, const_ib);
166 		radeon_sync_free(rdev, &const_ib->sync, NULL);
167 	}
168 	radeon_ring_ib_execute(rdev, ib->ring, ib);
169 	r = radeon_fence_emit(rdev, &ib->fence, ib->ring);
170 	if (r) {
171 		dev_err(rdev->dev, "failed to emit fence for new IB (%d)\n", r);
172 		radeon_ring_unlock_undo(rdev, ring);
173 		return r;
174 	}
175 	if (const_ib) {
176 		const_ib->fence = radeon_fence_ref(ib->fence);
177 	}
178 
179 	if (ib->vm)
180 		radeon_vm_fence(rdev, ib->vm, ib->fence);
181 
182 	radeon_ring_unlock_commit(rdev, ring, hdp_flush);
183 	return 0;
184 }
185 
186 /**
187  * radeon_ib_pool_init - Init the IB (Indirect Buffer) pool
188  *
189  * @rdev: radeon_device pointer
190  *
191  * Initialize the suballocator to manage a pool of memory
192  * for use as IBs (all asics).
193  * Returns 0 on success, error on failure.
194  */
195 int radeon_ib_pool_init(struct radeon_device *rdev)
196 {
197 	int r;
198 
199 	if (rdev->ib_pool_ready) {
200 		return 0;
201 	}
202 
203 	if (rdev->family >= CHIP_BONAIRE) {
204 		r = radeon_sa_bo_manager_init(rdev, &rdev->ring_tmp_bo,
205 					      RADEON_IB_POOL_SIZE*64*1024,
206 					      RADEON_GPU_PAGE_SIZE,
207 					      RADEON_GEM_DOMAIN_GTT,
208 					      RADEON_GEM_GTT_WC);
209 	} else {
210 		/* Before CIK, it's better to stick to cacheable GTT due
211 		 * to the command stream checking
212 		 */
213 		r = radeon_sa_bo_manager_init(rdev, &rdev->ring_tmp_bo,
214 					      RADEON_IB_POOL_SIZE*64*1024,
215 					      RADEON_GPU_PAGE_SIZE,
216 					      RADEON_GEM_DOMAIN_GTT, 0);
217 	}
218 	if (r) {
219 		return r;
220 	}
221 
222 	r = radeon_sa_bo_manager_start(rdev, &rdev->ring_tmp_bo);
223 	if (r) {
224 		return r;
225 	}
226 
227 	rdev->ib_pool_ready = true;
228 	if (radeon_debugfs_sa_init(rdev)) {
229 		dev_err(rdev->dev, "failed to register debugfs file for SA\n");
230 	}
231 	return 0;
232 }
233 
234 /**
235  * radeon_ib_pool_fini - Free the IB (Indirect Buffer) pool
236  *
237  * @rdev: radeon_device pointer
238  *
239  * Tear down the suballocator managing the pool of memory
240  * for use as IBs (all asics).
241  */
242 void radeon_ib_pool_fini(struct radeon_device *rdev)
243 {
244 	if (rdev->ib_pool_ready) {
245 		radeon_sa_bo_manager_suspend(rdev, &rdev->ring_tmp_bo);
246 		radeon_sa_bo_manager_fini(rdev, &rdev->ring_tmp_bo);
247 		rdev->ib_pool_ready = false;
248 	}
249 }
250 
251 /**
252  * radeon_ib_ring_tests - test IBs on the rings
253  *
254  * @rdev: radeon_device pointer
255  *
256  * Test an IB (Indirect Buffer) on each ring.
257  * If the test fails, disable the ring.
258  * Returns 0 on success, error if the primary GFX ring
259  * IB test fails.
260  */
261 int radeon_ib_ring_tests(struct radeon_device *rdev)
262 {
263 	unsigned i;
264 	int r;
265 
266 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
267 		struct radeon_ring *ring = &rdev->ring[i];
268 
269 		if (!ring->ready)
270 			continue;
271 
272 		r = radeon_ib_test(rdev, i, ring);
273 		if (r) {
274 			radeon_fence_driver_force_completion(rdev, i);
275 			ring->ready = false;
276 			rdev->needs_reset = false;
277 
278 			if (i == RADEON_RING_TYPE_GFX_INDEX) {
279 				/* oh, oh, that's really bad */
280 				DRM_ERROR("radeon: failed testing IB on GFX ring (%d).\n", r);
281 				rdev->accel_working = false;
282 				return r;
283 
284 			} else {
285 				/* still not good, but we can live with it */
286 				DRM_ERROR("radeon: failed testing IB on ring %d (%d).\n", i, r);
287 			}
288 		}
289 	}
290 	return 0;
291 }
292 
293 /*
294  * Debugfs info
295  */
296 #if defined(CONFIG_DEBUG_FS)
297 
298 static int radeon_debugfs_sa_info(struct seq_file *m, void *data)
299 {
300 	struct drm_info_node *node = (struct drm_info_node *) m->private;
301 	struct drm_device *dev = node->minor->dev;
302 	struct radeon_device *rdev = dev->dev_private;
303 
304 	radeon_sa_bo_dump_debug_info(&rdev->ring_tmp_bo, m);
305 
306 	return 0;
307 
308 }
309 
310 static struct drm_info_list radeon_debugfs_sa_list[] = {
311 	{"radeon_sa_info", &radeon_debugfs_sa_info, 0, NULL},
312 };
313 
314 #endif
315 
316 static int radeon_debugfs_sa_init(struct radeon_device *rdev)
317 {
318 #if defined(CONFIG_DEBUG_FS)
319 	return radeon_debugfs_add_files(rdev, radeon_debugfs_sa_list, 1);
320 #else
321 	return 0;
322 #endif
323 }
324