xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision a8fe58cec351c25e09c393bf46117c0c47b5a17c)
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_ring *ring, struct amdgpu_vm *vm,
59 		  unsigned size, struct amdgpu_ib *ib)
60 {
61 	struct amdgpu_device *adev = ring->adev;
62 	int r;
63 
64 	if (size) {
65 		r = amdgpu_sa_bo_new(&adev->ring_tmp_bo,
66 				      &ib->sa_bo, size, 256);
67 		if (r) {
68 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
69 			return r;
70 		}
71 
72 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
73 
74 		if (!vm)
75 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
76 	}
77 
78 	amdgpu_sync_create(&ib->sync);
79 
80 	ib->ring = ring;
81 	ib->vm = vm;
82 
83 	return 0;
84 }
85 
86 /**
87  * amdgpu_ib_free - free an IB (Indirect Buffer)
88  *
89  * @adev: amdgpu_device pointer
90  * @ib: IB object to free
91  *
92  * Free an IB (all asics).
93  */
94 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib)
95 {
96 	amdgpu_sync_free(adev, &ib->sync, &ib->fence->base);
97 	amdgpu_sa_bo_free(adev, &ib->sa_bo, &ib->fence->base);
98 	if (ib->fence)
99 		fence_put(&ib->fence->base);
100 }
101 
102 /**
103  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
104  *
105  * @adev: amdgpu_device pointer
106  * @num_ibs: number of IBs to schedule
107  * @ibs: IB objects to schedule
108  * @owner: owner for creating the fences
109  *
110  * Schedule an IB on the associated ring (all asics).
111  * Returns 0 on success, error on failure.
112  *
113  * On SI, there are two parallel engines fed from the primary ring,
114  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
115  * resource descriptors have moved to memory, the CE allows you to
116  * prime the caches while the DE is updating register state so that
117  * the resource descriptors will be already in cache when the draw is
118  * processed.  To accomplish this, the userspace driver submits two
119  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
120  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
121  * to SI there was just a DE IB.
122  */
123 int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
124 		       struct amdgpu_ib *ibs, void *owner)
125 {
126 	struct amdgpu_ib *ib = &ibs[0];
127 	struct amdgpu_ring *ring;
128 	struct amdgpu_ctx *ctx, *old_ctx;
129 	struct amdgpu_vm *vm;
130 	unsigned i;
131 	int r = 0;
132 
133 	if (num_ibs == 0)
134 		return -EINVAL;
135 
136 	ring = ibs->ring;
137 	ctx = ibs->ctx;
138 	vm = ibs->vm;
139 
140 	if (!ring->ready) {
141 		dev_err(adev->dev, "couldn't schedule ib\n");
142 		return -EINVAL;
143 	}
144 
145 	if (vm && !ibs->grabbed_vmid) {
146 		dev_err(adev->dev, "VM IB without ID\n");
147 		return -EINVAL;
148 	}
149 
150 	r = amdgpu_ring_alloc(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
151 	if (r) {
152 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
153 		return r;
154 	}
155 
156 	r = amdgpu_sync_wait(&ibs->sync);
157 	if (r) {
158 		amdgpu_ring_undo(ring);
159 		dev_err(adev->dev, "failed to sync wait (%d)\n", r);
160 		return r;
161 	}
162 
163 	if (vm) {
164 		/* do context switch */
165 		amdgpu_vm_flush(ring, vm, ib->sync.last_vm_update);
166 
167 		if (ring->funcs->emit_gds_switch)
168 			amdgpu_ring_emit_gds_switch(ring, ib->vm->ids[ring->idx].id,
169 						    ib->gds_base, ib->gds_size,
170 						    ib->gws_base, ib->gws_size,
171 						    ib->oa_base, ib->oa_size);
172 
173 		if (ring->funcs->emit_hdp_flush)
174 			amdgpu_ring_emit_hdp_flush(ring);
175 	}
176 
177 	old_ctx = ring->current_ctx;
178 	for (i = 0; i < num_ibs; ++i) {
179 		ib = &ibs[i];
180 
181 		if (ib->ring != ring || ib->ctx != ctx || ib->vm != vm) {
182 			ring->current_ctx = old_ctx;
183 			amdgpu_ring_undo(ring);
184 			return -EINVAL;
185 		}
186 		amdgpu_ring_emit_ib(ring, ib);
187 		ring->current_ctx = ctx;
188 	}
189 
190 	r = amdgpu_fence_emit(ring, owner, &ib->fence);
191 	if (r) {
192 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
193 		ring->current_ctx = old_ctx;
194 		amdgpu_ring_undo(ring);
195 		return r;
196 	}
197 
198 	/* wrap the last IB with fence */
199 	if (ib->user) {
200 		uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
201 		addr += ib->user->offset;
202 		amdgpu_ring_emit_fence(ring, addr, ib->sequence,
203 				       AMDGPU_FENCE_FLAG_64BIT);
204 	}
205 
206 	amdgpu_ring_commit(ring);
207 	return 0;
208 }
209 
210 /**
211  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
212  *
213  * @adev: amdgpu_device pointer
214  *
215  * Initialize the suballocator to manage a pool of memory
216  * for use as IBs (all asics).
217  * Returns 0 on success, error on failure.
218  */
219 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
220 {
221 	int r;
222 
223 	if (adev->ib_pool_ready) {
224 		return 0;
225 	}
226 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
227 				      AMDGPU_IB_POOL_SIZE*64*1024,
228 				      AMDGPU_GPU_PAGE_SIZE,
229 				      AMDGPU_GEM_DOMAIN_GTT);
230 	if (r) {
231 		return r;
232 	}
233 
234 	r = amdgpu_sa_bo_manager_start(adev, &adev->ring_tmp_bo);
235 	if (r) {
236 		return r;
237 	}
238 
239 	adev->ib_pool_ready = true;
240 	if (amdgpu_debugfs_sa_init(adev)) {
241 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
242 	}
243 	return 0;
244 }
245 
246 /**
247  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
248  *
249  * @adev: amdgpu_device pointer
250  *
251  * Tear down the suballocator managing the pool of memory
252  * for use as IBs (all asics).
253  */
254 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
255 {
256 	if (adev->ib_pool_ready) {
257 		amdgpu_sa_bo_manager_suspend(adev, &adev->ring_tmp_bo);
258 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
259 		adev->ib_pool_ready = false;
260 	}
261 }
262 
263 /**
264  * amdgpu_ib_ring_tests - test IBs on the rings
265  *
266  * @adev: amdgpu_device pointer
267  *
268  * Test an IB (Indirect Buffer) on each ring.
269  * If the test fails, disable the ring.
270  * Returns 0 on success, error if the primary GFX ring
271  * IB test fails.
272  */
273 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
274 {
275 	unsigned i;
276 	int r;
277 
278 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
279 		struct amdgpu_ring *ring = adev->rings[i];
280 
281 		if (!ring || !ring->ready)
282 			continue;
283 
284 		r = amdgpu_ring_test_ib(ring);
285 		if (r) {
286 			ring->ready = false;
287 
288 			if (ring == &adev->gfx.gfx_ring[0]) {
289 				/* oh, oh, that's really bad */
290 				DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
291 				adev->accel_working = false;
292 				return r;
293 
294 			} else {
295 				/* still not good, but we can live with it */
296 				DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
297 			}
298 		}
299 	}
300 	return 0;
301 }
302 
303 /*
304  * Debugfs info
305  */
306 #if defined(CONFIG_DEBUG_FS)
307 
308 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
309 {
310 	struct drm_info_node *node = (struct drm_info_node *) m->private;
311 	struct drm_device *dev = node->minor->dev;
312 	struct amdgpu_device *adev = dev->dev_private;
313 
314 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
315 
316 	return 0;
317 
318 }
319 
320 static struct drm_info_list amdgpu_debugfs_sa_list[] = {
321 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
322 };
323 
324 #endif
325 
326 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
327 {
328 #if defined(CONFIG_DEBUG_FS)
329 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
330 #else
331 	return 0;
332 #endif
333 }
334