xref: /linux/drivers/gpu/drm/radeon/radeon_ib.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 <linux/debugfs.h>
31 
32 #include <drm/drm_file.h>
33 
34 #include "radeon.h"
35 
36 /*
37  * IB
38  * IBs (Indirect Buffers) and areas of GPU accessible memory where
39  * commands are stored.  You can put a pointer to the IB in the
40  * command ring and the hw will fetch the commands from the IB
41  * and execute them.  Generally userspace acceleration drivers
42  * produce command buffers which are send to the kernel and
43  * put in IBs for execution by the requested ring.
44  */
45 static void radeon_debugfs_sa_init(struct radeon_device *rdev);
46 
47 /**
48  * radeon_ib_get - request an IB (Indirect Buffer)
49  *
50  * @rdev: radeon_device pointer
51  * @ring: ring index the IB is associated with
52  * @vm: requested vm
53  * @ib: IB object returned
54  * @size: requested IB size
55  *
56  * Request an IB (all asics).  IBs are allocated using the
57  * suballocator.
58  * Returns 0 on success, error on failure.
59  */
60 int radeon_ib_get(struct radeon_device *rdev, int ring,
61 		  struct radeon_ib *ib, struct radeon_vm *vm,
62 		  unsigned size)
63 {
64 	int r;
65 
66 	r = radeon_sa_bo_new(&rdev->ring_tmp_bo, &ib->sa_bo, size, 256);
67 	if (r) {
68 		dev_err(rdev->dev, "failed to get a new IB (%d)\n", r);
69 		return r;
70 	}
71 
72 	radeon_sync_create(&ib->sync);
73 
74 	ib->ring = ring;
75 	ib->fence = NULL;
76 	ib->ptr = radeon_sa_bo_cpu_addr(ib->sa_bo);
77 	ib->vm = vm;
78 	if (vm) {
79 		/* ib pool is bound at RADEON_VA_IB_OFFSET in virtual address
80 		 * space and soffset is the offset inside the pool bo
81 		 */
82 		ib->gpu_addr = drm_suballoc_soffset(ib->sa_bo) + RADEON_VA_IB_OFFSET;
83 	} else {
84 		ib->gpu_addr = radeon_sa_bo_gpu_addr(ib->sa_bo);
85 	}
86 	ib->is_const_ib = false;
87 
88 	return 0;
89 }
90 
91 /**
92  * radeon_ib_free - free an IB (Indirect Buffer)
93  *
94  * @rdev: radeon_device pointer
95  * @ib: IB object to free
96  *
97  * Free an IB (all asics).
98  */
99 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib)
100 {
101 	radeon_sync_free(rdev, &ib->sync, ib->fence);
102 	radeon_sa_bo_free(&ib->sa_bo, ib->fence);
103 	radeon_fence_unref(&ib->fence);
104 }
105 
106 /**
107  * radeon_ib_schedule - schedule an IB (Indirect Buffer) on the ring
108  *
109  * @rdev: radeon_device pointer
110  * @ib: IB object to schedule
111  * @const_ib: Const IB to schedule (SI only)
112  * @hdp_flush: Whether or not to perform an HDP cache flush
113  *
114  * Schedule an IB on the associated ring (all asics).
115  * Returns 0 on success, error on failure.
116  *
117  * On SI, there are two parallel engines fed from the primary ring,
118  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
119  * resource descriptors have moved to memory, the CE allows you to
120  * prime the caches while the DE is updating register state so that
121  * the resource descriptors will be already in cache when the draw is
122  * processed.  To accomplish this, the userspace driver submits two
123  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
124  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
125  * to SI there was just a DE IB.
126  */
127 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
128 		       struct radeon_ib *const_ib, bool hdp_flush)
129 {
130 	struct radeon_ring *ring = &rdev->ring[ib->ring];
131 	int r = 0;
132 
133 	if (!ib->length_dw || !ring->ready) {
134 		/* TODO: Nothings in the ib we should report. */
135 		dev_err(rdev->dev, "couldn't schedule ib\n");
136 		return -EINVAL;
137 	}
138 
139 	/* 64 dwords should be enough for fence too */
140 	r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_SYNCS * 8);
141 	if (r) {
142 		dev_err(rdev->dev, "scheduling IB failed (%d).\n", r);
143 		return r;
144 	}
145 
146 	/* grab a vm id if necessary */
147 	if (ib->vm) {
148 		struct radeon_fence *vm_id_fence;
149 		vm_id_fence = radeon_vm_grab_id(rdev, ib->vm, ib->ring);
150 		radeon_sync_fence(&ib->sync, vm_id_fence);
151 	}
152 
153 	/* sync with other rings */
154 	r = radeon_sync_rings(rdev, &ib->sync, ib->ring);
155 	if (r) {
156 		dev_err(rdev->dev, "failed to sync rings (%d)\n", r);
157 		radeon_ring_unlock_undo(rdev, ring);
158 		return r;
159 	}
160 
161 	if (ib->vm)
162 		radeon_vm_flush(rdev, ib->vm, ib->ring,
163 				ib->sync.last_vm_update);
164 
165 	if (const_ib) {
166 		radeon_ring_ib_execute(rdev, const_ib->ring, const_ib);
167 		radeon_sync_free(rdev, &const_ib->sync, NULL);
168 	}
169 	radeon_ring_ib_execute(rdev, ib->ring, ib);
170 	r = radeon_fence_emit(rdev, &ib->fence, ib->ring);
171 	if (r) {
172 		dev_err(rdev->dev, "failed to emit fence for new IB (%d)\n", r);
173 		radeon_ring_unlock_undo(rdev, ring);
174 		return r;
175 	}
176 	if (const_ib) {
177 		const_ib->fence = radeon_fence_ref(ib->fence);
178 	}
179 
180 	if (ib->vm)
181 		radeon_vm_fence(rdev, ib->vm, ib->fence);
182 
183 	radeon_ring_unlock_commit(rdev, ring, hdp_flush);
184 	return 0;
185 }
186 
187 /**
188  * radeon_ib_pool_init - Init the IB (Indirect Buffer) pool
189  *
190  * @rdev: radeon_device pointer
191  *
192  * Initialize the suballocator to manage a pool of memory
193  * for use as IBs (all asics).
194  * Returns 0 on success, error on failure.
195  */
196 int radeon_ib_pool_init(struct radeon_device *rdev)
197 {
198 	int r;
199 
200 	if (rdev->ib_pool_ready) {
201 		return 0;
202 	}
203 
204 	if (rdev->family >= CHIP_BONAIRE) {
205 		r = radeon_sa_bo_manager_init(rdev, &rdev->ring_tmp_bo,
206 					      RADEON_IB_POOL_SIZE*64*1024, 256,
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, 256,
215 					      RADEON_GEM_DOMAIN_GTT, 0);
216 	}
217 	if (r) {
218 		return r;
219 	}
220 
221 	r = radeon_sa_bo_manager_start(rdev, &rdev->ring_tmp_bo);
222 	if (r) {
223 		return r;
224 	}
225 
226 	rdev->ib_pool_ready = true;
227 	radeon_debugfs_sa_init(rdev);
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 			radeon_fence_driver_force_completion(rdev, i);
272 			ring->ready = false;
273 			rdev->needs_reset = false;
274 
275 			if (i == RADEON_RING_TYPE_GFX_INDEX) {
276 				/* oh, oh, that's really bad */
277 				DRM_ERROR("radeon: failed testing IB on GFX ring (%d).\n", r);
278 				rdev->accel_working = false;
279 				return r;
280 
281 			} else {
282 				/* still not good, but we can live with it */
283 				DRM_ERROR("radeon: failed testing IB on ring %d (%d).\n", i, r);
284 			}
285 		}
286 	}
287 	return 0;
288 }
289 
290 /*
291  * Debugfs info
292  */
293 #if defined(CONFIG_DEBUG_FS)
294 
295 static int radeon_debugfs_sa_info_show(struct seq_file *m, void *unused)
296 {
297 	struct radeon_device *rdev = m->private;
298 
299 	radeon_sa_bo_dump_debug_info(&rdev->ring_tmp_bo, m);
300 
301 	return 0;
302 
303 }
304 
305 DEFINE_SHOW_ATTRIBUTE(radeon_debugfs_sa_info);
306 
307 #endif
308 
309 static void radeon_debugfs_sa_init(struct radeon_device *rdev)
310 {
311 #if defined(CONFIG_DEBUG_FS)
312 	struct dentry *root = rdev_to_drm(rdev)->primary->debugfs_root;
313 
314 	debugfs_create_file("radeon_sa_info", 0444, root, rdev,
315 			    &radeon_debugfs_sa_info_fops);
316 #endif
317 }
318