xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
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, &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 	amdgpu_fence_unref(&ib->fence);
99 }
100 
101 /**
102  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
103  *
104  * @adev: amdgpu_device pointer
105  * @num_ibs: number of IBs to schedule
106  * @ibs: IB objects to schedule
107  * @owner: owner for creating the fences
108  *
109  * Schedule an IB on the associated ring (all asics).
110  * Returns 0 on success, error on failure.
111  *
112  * On SI, there are two parallel engines fed from the primary ring,
113  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
114  * resource descriptors have moved to memory, the CE allows you to
115  * prime the caches while the DE is updating register state so that
116  * the resource descriptors will be already in cache when the draw is
117  * processed.  To accomplish this, the userspace driver submits two
118  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
119  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
120  * to SI there was just a DE IB.
121  */
122 int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
123 		       struct amdgpu_ib *ibs, void *owner)
124 {
125 	struct amdgpu_ib *ib = &ibs[0];
126 	struct amdgpu_ring *ring;
127 	struct amdgpu_ctx *ctx, *old_ctx;
128 	struct amdgpu_vm *vm;
129 	unsigned i;
130 	int r = 0;
131 
132 	if (num_ibs == 0)
133 		return -EINVAL;
134 
135 	ring = ibs->ring;
136 	ctx = ibs->ctx;
137 	vm = ibs->vm;
138 
139 	if (!ring->ready) {
140 		dev_err(adev->dev, "couldn't schedule ib\n");
141 		return -EINVAL;
142 	}
143 	r = amdgpu_sync_wait(&ibs->sync);
144 	if (r) {
145 		dev_err(adev->dev, "IB sync failed (%d).\n", r);
146 		return r;
147 	}
148 	r = amdgpu_ring_lock(ring, (256 + AMDGPU_NUM_SYNCS * 8) * num_ibs);
149 	if (r) {
150 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
151 		return r;
152 	}
153 
154 	if (vm) {
155 		/* grab a vm id if necessary */
156 		r = amdgpu_vm_grab_id(ibs->vm, ibs->ring, &ibs->sync);
157 		if (r) {
158 			amdgpu_ring_unlock_undo(ring);
159 			return r;
160 		}
161 	}
162 
163 	r = amdgpu_sync_rings(&ibs->sync, ring);
164 	if (r) {
165 		amdgpu_ring_unlock_undo(ring);
166 		dev_err(adev->dev, "failed to sync rings (%d)\n", r);
167 		return r;
168 	}
169 
170 	if (vm) {
171 		/* do context switch */
172 		amdgpu_vm_flush(ring, vm, ib->sync.last_vm_update);
173 
174 		if (ring->funcs->emit_gds_switch)
175 			amdgpu_ring_emit_gds_switch(ring, ib->vm->ids[ring->idx].id,
176 						    ib->gds_base, ib->gds_size,
177 						    ib->gws_base, ib->gws_size,
178 						    ib->oa_base, ib->oa_size);
179 
180 		if (ring->funcs->emit_hdp_flush)
181 			amdgpu_ring_emit_hdp_flush(ring);
182 	}
183 
184 	old_ctx = ring->current_ctx;
185 	for (i = 0; i < num_ibs; ++i) {
186 		ib = &ibs[i];
187 
188 		if (ib->ring != ring || ib->ctx != ctx || ib->vm != vm) {
189 			ring->current_ctx = old_ctx;
190 			amdgpu_ring_unlock_undo(ring);
191 			return -EINVAL;
192 		}
193 		amdgpu_ring_emit_ib(ring, ib);
194 		ring->current_ctx = ctx;
195 	}
196 
197 	r = amdgpu_fence_emit(ring, owner, &ib->fence);
198 	if (r) {
199 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
200 		ring->current_ctx = old_ctx;
201 		amdgpu_ring_unlock_undo(ring);
202 		return r;
203 	}
204 
205 	if (!amdgpu_enable_scheduler && ib->ctx)
206 		ib->sequence = amdgpu_ctx_add_fence(ib->ctx, ring,
207 						    &ib->fence->base);
208 
209 	/* wrap the last IB with fence */
210 	if (ib->user) {
211 		uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
212 		addr += ib->user->offset;
213 		amdgpu_ring_emit_fence(ring, addr, ib->sequence,
214 				       AMDGPU_FENCE_FLAG_64BIT);
215 	}
216 
217 	if (ib->vm)
218 		amdgpu_vm_fence(adev, ib->vm, ib->fence);
219 
220 	amdgpu_ring_unlock_commit(ring);
221 	return 0;
222 }
223 
224 /**
225  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
226  *
227  * @adev: amdgpu_device pointer
228  *
229  * Initialize the suballocator to manage a pool of memory
230  * for use as IBs (all asics).
231  * Returns 0 on success, error on failure.
232  */
233 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
234 {
235 	int r;
236 
237 	if (adev->ib_pool_ready) {
238 		return 0;
239 	}
240 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
241 				      AMDGPU_IB_POOL_SIZE*64*1024,
242 				      AMDGPU_GPU_PAGE_SIZE,
243 				      AMDGPU_GEM_DOMAIN_GTT);
244 	if (r) {
245 		return r;
246 	}
247 
248 	r = amdgpu_sa_bo_manager_start(adev, &adev->ring_tmp_bo);
249 	if (r) {
250 		return r;
251 	}
252 
253 	adev->ib_pool_ready = true;
254 	if (amdgpu_debugfs_sa_init(adev)) {
255 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
256 	}
257 	return 0;
258 }
259 
260 /**
261  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
262  *
263  * @adev: amdgpu_device pointer
264  *
265  * Tear down the suballocator managing the pool of memory
266  * for use as IBs (all asics).
267  */
268 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
269 {
270 	if (adev->ib_pool_ready) {
271 		amdgpu_sa_bo_manager_suspend(adev, &adev->ring_tmp_bo);
272 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
273 		adev->ib_pool_ready = false;
274 	}
275 }
276 
277 /**
278  * amdgpu_ib_ring_tests - test IBs on the rings
279  *
280  * @adev: amdgpu_device pointer
281  *
282  * Test an IB (Indirect Buffer) on each ring.
283  * If the test fails, disable the ring.
284  * Returns 0 on success, error if the primary GFX ring
285  * IB test fails.
286  */
287 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
288 {
289 	unsigned i;
290 	int r;
291 
292 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
293 		struct amdgpu_ring *ring = adev->rings[i];
294 
295 		if (!ring || !ring->ready)
296 			continue;
297 
298 		r = amdgpu_ring_test_ib(ring);
299 		if (r) {
300 			ring->ready = false;
301 			adev->needs_reset = false;
302 
303 			if (ring == &adev->gfx.gfx_ring[0]) {
304 				/* oh, oh, that's really bad */
305 				DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
306 				adev->accel_working = false;
307 				return r;
308 
309 			} else {
310 				/* still not good, but we can live with it */
311 				DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
312 			}
313 		}
314 	}
315 	return 0;
316 }
317 
318 /*
319  * Debugfs info
320  */
321 #if defined(CONFIG_DEBUG_FS)
322 
323 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
324 {
325 	struct drm_info_node *node = (struct drm_info_node *) m->private;
326 	struct drm_device *dev = node->minor->dev;
327 	struct amdgpu_device *adev = dev->dev_private;
328 
329 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
330 
331 	return 0;
332 
333 }
334 
335 static struct drm_info_list amdgpu_debugfs_sa_list[] = {
336 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
337 };
338 
339 #endif
340 
341 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
342 {
343 #if defined(CONFIG_DEBUG_FS)
344 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
345 #else
346 	return 0;
347 #endif
348 }
349