xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision 110e6f26af80dfd90b6e5c645b1aed7228aa580d)
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 	ib->vm_id = 0;
79 
80 	return 0;
81 }
82 
83 /**
84  * amdgpu_ib_free - free an IB (Indirect Buffer)
85  *
86  * @adev: amdgpu_device pointer
87  * @ib: IB object to free
88  * @f: the fence SA bo need wait on for the ib alloation
89  *
90  * Free an IB (all asics).
91  */
92 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib, struct fence *f)
93 {
94 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
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  * @f: fence created during this submission
104  *
105  * Schedule an IB on the associated ring (all asics).
106  * Returns 0 on success, error on failure.
107  *
108  * On SI, there are two parallel engines fed from the primary ring,
109  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
110  * resource descriptors have moved to memory, the CE allows you to
111  * prime the caches while the DE is updating register state so that
112  * the resource descriptors will be already in cache when the draw is
113  * processed.  To accomplish this, the userspace driver submits two
114  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
115  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
116  * to SI there was just a DE IB.
117  */
118 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
119 		       struct amdgpu_ib *ibs, struct fence *last_vm_update,
120 		       struct fence **f)
121 {
122 	struct amdgpu_device *adev = ring->adev;
123 	struct amdgpu_ib *ib = &ibs[0];
124 	struct amdgpu_ctx *ctx, *old_ctx;
125 	struct amdgpu_vm *vm;
126 	struct fence *hwf;
127 	unsigned i, patch_offset = ~0;
128 
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->vm_id) {
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 (ring->type == AMDGPU_RING_TYPE_SDMA && ring->funcs->init_cond_exec)
154 		patch_offset = amdgpu_ring_init_cond_exec(ring);
155 
156 	if (vm) {
157 		/* do context switch */
158 		r = amdgpu_vm_flush(ring, ib->vm_id, ib->vm_pd_addr,
159 				    ib->gds_base, ib->gds_size,
160 				    ib->gws_base, ib->gws_size,
161 				    ib->oa_base, ib->oa_size);
162 		if (r) {
163 			amdgpu_ring_undo(ring);
164 			return r;
165 		}
166 
167 		if (ring->funcs->emit_hdp_flush)
168 			amdgpu_ring_emit_hdp_flush(ring);
169 	}
170 
171 	/* always set cond_exec_polling to CONTINUE */
172 	*ring->cond_exe_cpu_addr = 1;
173 
174 	old_ctx = ring->current_ctx;
175 	for (i = 0; i < num_ibs; ++i) {
176 		ib = &ibs[i];
177 
178 		if (ib->ctx != ctx || ib->vm != vm) {
179 			ring->current_ctx = old_ctx;
180 			if (ib->vm_id)
181 				amdgpu_vm_reset_id(adev, ib->vm_id);
182 			amdgpu_ring_undo(ring);
183 			return -EINVAL;
184 		}
185 		amdgpu_ring_emit_ib(ring, ib);
186 		ring->current_ctx = ctx;
187 	}
188 
189 	if (vm) {
190 		if (ring->funcs->emit_hdp_invalidate)
191 			amdgpu_ring_emit_hdp_invalidate(ring);
192 	}
193 
194 	r = amdgpu_fence_emit(ring, &hwf);
195 	if (r) {
196 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
197 		ring->current_ctx = old_ctx;
198 		if (ib->vm_id)
199 			amdgpu_vm_reset_id(adev, ib->vm_id);
200 		amdgpu_ring_undo(ring);
201 		return r;
202 	}
203 
204 	/* wrap the last IB with fence */
205 	if (ib->user) {
206 		uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
207 		addr += ib->user->offset;
208 		amdgpu_ring_emit_fence(ring, addr, ib->sequence,
209 				       AMDGPU_FENCE_FLAG_64BIT);
210 	}
211 
212 	if (f)
213 		*f = fence_get(hwf);
214 
215 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
216 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
217 
218 	amdgpu_ring_commit(ring);
219 	return 0;
220 }
221 
222 /**
223  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
224  *
225  * @adev: amdgpu_device pointer
226  *
227  * Initialize the suballocator to manage a pool of memory
228  * for use as IBs (all asics).
229  * Returns 0 on success, error on failure.
230  */
231 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
232 {
233 	int r;
234 
235 	if (adev->ib_pool_ready) {
236 		return 0;
237 	}
238 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
239 				      AMDGPU_IB_POOL_SIZE*64*1024,
240 				      AMDGPU_GPU_PAGE_SIZE,
241 				      AMDGPU_GEM_DOMAIN_GTT);
242 	if (r) {
243 		return r;
244 	}
245 
246 	r = amdgpu_sa_bo_manager_start(adev, &adev->ring_tmp_bo);
247 	if (r) {
248 		return r;
249 	}
250 
251 	adev->ib_pool_ready = true;
252 	if (amdgpu_debugfs_sa_init(adev)) {
253 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
254 	}
255 	return 0;
256 }
257 
258 /**
259  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
260  *
261  * @adev: amdgpu_device pointer
262  *
263  * Tear down the suballocator managing the pool of memory
264  * for use as IBs (all asics).
265  */
266 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
267 {
268 	if (adev->ib_pool_ready) {
269 		amdgpu_sa_bo_manager_suspend(adev, &adev->ring_tmp_bo);
270 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
271 		adev->ib_pool_ready = false;
272 	}
273 }
274 
275 /**
276  * amdgpu_ib_ring_tests - test IBs on the rings
277  *
278  * @adev: amdgpu_device pointer
279  *
280  * Test an IB (Indirect Buffer) on each ring.
281  * If the test fails, disable the ring.
282  * Returns 0 on success, error if the primary GFX ring
283  * IB test fails.
284  */
285 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
286 {
287 	unsigned i;
288 	int r;
289 
290 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
291 		struct amdgpu_ring *ring = adev->rings[i];
292 
293 		if (!ring || !ring->ready)
294 			continue;
295 
296 		r = amdgpu_ring_test_ib(ring);
297 		if (r) {
298 			ring->ready = false;
299 
300 			if (ring == &adev->gfx.gfx_ring[0]) {
301 				/* oh, oh, that's really bad */
302 				DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
303 				adev->accel_working = false;
304 				return r;
305 
306 			} else {
307 				/* still not good, but we can live with it */
308 				DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
309 			}
310 		}
311 	}
312 	return 0;
313 }
314 
315 /*
316  * Debugfs info
317  */
318 #if defined(CONFIG_DEBUG_FS)
319 
320 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
321 {
322 	struct drm_info_node *node = (struct drm_info_node *) m->private;
323 	struct drm_device *dev = node->minor->dev;
324 	struct amdgpu_device *adev = dev->dev_private;
325 
326 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
327 
328 	return 0;
329 
330 }
331 
332 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
333 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
334 };
335 
336 #endif
337 
338 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
339 {
340 #if defined(CONFIG_DEBUG_FS)
341 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
342 #else
343 	return 0;
344 #endif
345 }
346