xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision 781095f903f398148cd0b646d3984234a715f29e)
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 <linux/seq_file.h>
30 #include <linux/slab.h>
31 #include <drm/drmP.h>
32 #include <drm/amdgpu_drm.h>
33 #include "amdgpu.h"
34 #include "atom.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 int amdgpu_debugfs_sa_init(struct amdgpu_device *adev);
46 
47 /**
48  * amdgpu_ib_get - request an IB (Indirect Buffer)
49  *
50  * @ring: ring index the IB is associated with
51  * @size: requested IB size
52  * @ib: IB object returned
53  *
54  * Request an IB (all asics).  IBs are allocated using the
55  * suballocator.
56  * Returns 0 on success, error on failure.
57  */
58 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
59 		  unsigned size, struct amdgpu_ib *ib)
60 {
61 	int r;
62 
63 	if (size) {
64 		r = amdgpu_sa_bo_new(&adev->ring_tmp_bo,
65 				      &ib->sa_bo, size, 256);
66 		if (r) {
67 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
68 			return r;
69 		}
70 
71 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
72 
73 		if (!vm)
74 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
75 	}
76 
77 	ib->vm = vm;
78 
79 	return 0;
80 }
81 
82 /**
83  * amdgpu_ib_free - free an IB (Indirect Buffer)
84  *
85  * @adev: amdgpu_device pointer
86  * @ib: IB object to free
87  *
88  * Free an IB (all asics).
89  */
90 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib)
91 {
92 	amdgpu_sa_bo_free(adev, &ib->sa_bo, &ib->fence->base);
93 	if (ib->fence)
94 		fence_put(&ib->fence->base);
95 }
96 
97 /**
98  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
99  *
100  * @adev: amdgpu_device pointer
101  * @num_ibs: number of IBs to schedule
102  * @ibs: IB objects to schedule
103  * @owner: owner for creating the fences
104  * @f: fence created during this submission
105  *
106  * Schedule an IB on the associated ring (all asics).
107  * Returns 0 on success, error on failure.
108  *
109  * On SI, there are two parallel engines fed from the primary ring,
110  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
111  * resource descriptors have moved to memory, the CE allows you to
112  * prime the caches while the DE is updating register state so that
113  * the resource descriptors will be already in cache when the draw is
114  * processed.  To accomplish this, the userspace driver submits two
115  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
116  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
117  * to SI there was just a DE IB.
118  */
119 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
120 		       struct amdgpu_ib *ibs, void *owner,
121 		       struct fence *last_vm_update,
122 		       struct fence **f)
123 {
124 	struct amdgpu_device *adev = ring->adev;
125 	struct amdgpu_ib *ib = &ibs[0];
126 	struct amdgpu_ctx *ctx, *old_ctx;
127 	struct amdgpu_vm *vm;
128 	unsigned i;
129 	int r = 0;
130 
131 	if (num_ibs == 0)
132 		return -EINVAL;
133 
134 	ctx = ibs->ctx;
135 	vm = ibs->vm;
136 
137 	if (!ring->ready) {
138 		dev_err(adev->dev, "couldn't schedule ib\n");
139 		return -EINVAL;
140 	}
141 
142 	if (vm && !ibs->grabbed_vmid) {
143 		dev_err(adev->dev, "VM IB without ID\n");
144 		return -EINVAL;
145 	}
146 
147 	r = amdgpu_ring_alloc(ring, 256 * num_ibs);
148 	if (r) {
149 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
150 		return r;
151 	}
152 
153 	if (vm) {
154 		/* do context switch */
155 		amdgpu_vm_flush(ring, vm, last_vm_update);
156 
157 		if (ring->funcs->emit_gds_switch)
158 			amdgpu_ring_emit_gds_switch(ring, ib->vm->ids[ring->idx].id,
159 						    ib->gds_base, ib->gds_size,
160 						    ib->gws_base, ib->gws_size,
161 						    ib->oa_base, ib->oa_size);
162 
163 		if (ring->funcs->emit_hdp_flush)
164 			amdgpu_ring_emit_hdp_flush(ring);
165 	}
166 
167 	old_ctx = ring->current_ctx;
168 	for (i = 0; i < num_ibs; ++i) {
169 		ib = &ibs[i];
170 
171 		if (ib->ctx != ctx || ib->vm != vm) {
172 			ring->current_ctx = old_ctx;
173 			amdgpu_ring_undo(ring);
174 			return -EINVAL;
175 		}
176 		amdgpu_ring_emit_ib(ring, ib);
177 		ring->current_ctx = ctx;
178 	}
179 
180 	r = amdgpu_fence_emit(ring, owner, &ib->fence);
181 	if (r) {
182 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
183 		ring->current_ctx = old_ctx;
184 		amdgpu_ring_undo(ring);
185 		return r;
186 	}
187 
188 	/* wrap the last IB with fence */
189 	if (ib->user) {
190 		uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
191 		addr += ib->user->offset;
192 		amdgpu_ring_emit_fence(ring, addr, ib->sequence,
193 				       AMDGPU_FENCE_FLAG_64BIT);
194 	}
195 
196 	if (f)
197 		*f = fence_get(&ib->fence->base);
198 
199 	amdgpu_ring_commit(ring);
200 	return 0;
201 }
202 
203 /**
204  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
205  *
206  * @adev: amdgpu_device pointer
207  *
208  * Initialize the suballocator to manage a pool of memory
209  * for use as IBs (all asics).
210  * Returns 0 on success, error on failure.
211  */
212 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
213 {
214 	int r;
215 
216 	if (adev->ib_pool_ready) {
217 		return 0;
218 	}
219 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
220 				      AMDGPU_IB_POOL_SIZE*64*1024,
221 				      AMDGPU_GPU_PAGE_SIZE,
222 				      AMDGPU_GEM_DOMAIN_GTT);
223 	if (r) {
224 		return r;
225 	}
226 
227 	r = amdgpu_sa_bo_manager_start(adev, &adev->ring_tmp_bo);
228 	if (r) {
229 		return r;
230 	}
231 
232 	adev->ib_pool_ready = true;
233 	if (amdgpu_debugfs_sa_init(adev)) {
234 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
235 	}
236 	return 0;
237 }
238 
239 /**
240  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
241  *
242  * @adev: amdgpu_device pointer
243  *
244  * Tear down the suballocator managing the pool of memory
245  * for use as IBs (all asics).
246  */
247 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
248 {
249 	if (adev->ib_pool_ready) {
250 		amdgpu_sa_bo_manager_suspend(adev, &adev->ring_tmp_bo);
251 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
252 		adev->ib_pool_ready = false;
253 	}
254 }
255 
256 /**
257  * amdgpu_ib_ring_tests - test IBs on the rings
258  *
259  * @adev: amdgpu_device pointer
260  *
261  * Test an IB (Indirect Buffer) on each ring.
262  * If the test fails, disable the ring.
263  * Returns 0 on success, error if the primary GFX ring
264  * IB test fails.
265  */
266 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
267 {
268 	unsigned i;
269 	int r;
270 
271 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
272 		struct amdgpu_ring *ring = adev->rings[i];
273 
274 		if (!ring || !ring->ready)
275 			continue;
276 
277 		r = amdgpu_ring_test_ib(ring);
278 		if (r) {
279 			ring->ready = false;
280 
281 			if (ring == &adev->gfx.gfx_ring[0]) {
282 				/* oh, oh, that's really bad */
283 				DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
284 				adev->accel_working = false;
285 				return r;
286 
287 			} else {
288 				/* still not good, but we can live with it */
289 				DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
290 			}
291 		}
292 	}
293 	return 0;
294 }
295 
296 /*
297  * Debugfs info
298  */
299 #if defined(CONFIG_DEBUG_FS)
300 
301 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
302 {
303 	struct drm_info_node *node = (struct drm_info_node *) m->private;
304 	struct drm_device *dev = node->minor->dev;
305 	struct amdgpu_device *adev = dev->dev_private;
306 
307 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
308 
309 	return 0;
310 
311 }
312 
313 static struct drm_info_list amdgpu_debugfs_sa_list[] = {
314 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
315 };
316 
317 #endif
318 
319 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
320 {
321 #if defined(CONFIG_DEBUG_FS)
322 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
323 #else
324 	return 0;
325 #endif
326 }
327