xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c (revision 4b0aaacee51eb6592a03fdefd5ce97558518e291)
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 #define AMDGPU_IB_TEST_TIMEOUT	msecs_to_jiffies(1000)
37 
38 /*
39  * IB
40  * IBs (Indirect Buffers) and areas of GPU accessible memory where
41  * commands are stored.  You can put a pointer to the IB in the
42  * command ring and the hw will fetch the commands from the IB
43  * and execute them.  Generally userspace acceleration drivers
44  * produce command buffers which are send to the kernel and
45  * put in IBs for execution by the requested ring.
46  */
47 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev);
48 
49 /**
50  * amdgpu_ib_get - request an IB (Indirect Buffer)
51  *
52  * @ring: ring index the IB is associated with
53  * @size: requested IB size
54  * @ib: IB object returned
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 amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
61 		  unsigned size, struct amdgpu_ib *ib)
62 {
63 	int r;
64 
65 	if (size) {
66 		r = amdgpu_sa_bo_new(&adev->ring_tmp_bo,
67 				      &ib->sa_bo, size, 256);
68 		if (r) {
69 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
70 			return r;
71 		}
72 
73 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
74 
75 		if (!vm)
76 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
77 	}
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  * @f: the fence SA bo need wait on for the ib alloation
88  *
89  * Free an IB (all asics).
90  */
91 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
92 		    struct dma_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 amdgpu_job *job,
120 		       struct dma_fence **f)
121 {
122 	struct amdgpu_device *adev = ring->adev;
123 	struct amdgpu_ib *ib = &ibs[0];
124 	struct dma_fence *tmp = NULL;
125 	bool skip_preamble, need_ctx_switch;
126 	unsigned patch_offset = ~0;
127 	struct amdgpu_vm *vm;
128 	uint64_t fence_ctx;
129 	uint32_t status = 0, alloc_size;
130 	unsigned fence_flags = 0;
131 
132 	unsigned i;
133 	int r = 0;
134 	bool need_pipe_sync = false;
135 
136 	if (num_ibs == 0)
137 		return -EINVAL;
138 
139 	/* ring tests don't use a job */
140 	if (job) {
141 		vm = job->vm;
142 		fence_ctx = job->base.s_fence->scheduled.context;
143 	} else {
144 		vm = NULL;
145 		fence_ctx = 0;
146 	}
147 
148 	if (!ring->ready) {
149 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
150 		return -EINVAL;
151 	}
152 
153 	if (vm && !job->vmid) {
154 		dev_err(adev->dev, "VM IB without ID\n");
155 		return -EINVAL;
156 	}
157 
158 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
159 		ring->funcs->emit_ib_size;
160 
161 	r = amdgpu_ring_alloc(ring, alloc_size);
162 	if (r) {
163 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
164 		return r;
165 	}
166 
167 	need_ctx_switch = ring->current_ctx != fence_ctx;
168 	if (ring->funcs->emit_pipeline_sync && job &&
169 	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync, NULL)) ||
170 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
171 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
172 		need_pipe_sync = true;
173 		dma_fence_put(tmp);
174 	}
175 
176 	if (ring->funcs->insert_start)
177 		ring->funcs->insert_start(ring);
178 
179 	if (job) {
180 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
181 		if (r) {
182 			amdgpu_ring_undo(ring);
183 			return r;
184 		}
185 	}
186 
187 	if (job && ring->funcs->init_cond_exec)
188 		patch_offset = amdgpu_ring_init_cond_exec(ring);
189 
190 #ifdef CONFIG_X86_64
191 	if (!(adev->flags & AMD_IS_APU))
192 #endif
193 	{
194 		if (ring->funcs->emit_hdp_flush)
195 			amdgpu_ring_emit_hdp_flush(ring);
196 		else
197 			amdgpu_asic_flush_hdp(adev, ring);
198 	}
199 
200 	skip_preamble = ring->current_ctx == fence_ctx;
201 	if (job && ring->funcs->emit_cntxcntl) {
202 		if (need_ctx_switch)
203 			status |= AMDGPU_HAVE_CTX_SWITCH;
204 		status |= job->preamble_status;
205 
206 		amdgpu_ring_emit_cntxcntl(ring, status);
207 	}
208 
209 	for (i = 0; i < num_ibs; ++i) {
210 		ib = &ibs[i];
211 
212 		/* drop preamble IBs if we don't have a context switch */
213 		if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
214 			skip_preamble &&
215 			!(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
216 			!amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
217 			continue;
218 
219 		amdgpu_ring_emit_ib(ring, ib, job ? job->vmid : 0,
220 				    need_ctx_switch);
221 		need_ctx_switch = false;
222 	}
223 
224 	if (ring->funcs->emit_tmz)
225 		amdgpu_ring_emit_tmz(ring, false);
226 
227 #ifdef CONFIG_X86_64
228 	if (!(adev->flags & AMD_IS_APU))
229 #endif
230 		amdgpu_asic_invalidate_hdp(adev, ring);
231 
232 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
233 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
234 
235 	/* wrap the last IB with fence */
236 	if (job && job->uf_addr) {
237 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
238 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
239 	}
240 
241 	r = amdgpu_fence_emit(ring, f, fence_flags);
242 	if (r) {
243 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
244 		if (job && job->vmid)
245 			amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
246 		amdgpu_ring_undo(ring);
247 		return r;
248 	}
249 
250 	if (ring->funcs->insert_end)
251 		ring->funcs->insert_end(ring);
252 
253 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
254 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
255 
256 	ring->current_ctx = fence_ctx;
257 	if (vm && ring->funcs->emit_switch_buffer)
258 		amdgpu_ring_emit_switch_buffer(ring);
259 	amdgpu_ring_commit(ring);
260 	return 0;
261 }
262 
263 /**
264  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
265  *
266  * @adev: amdgpu_device pointer
267  *
268  * Initialize the suballocator to manage a pool of memory
269  * for use as IBs (all asics).
270  * Returns 0 on success, error on failure.
271  */
272 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
273 {
274 	int r;
275 
276 	if (adev->ib_pool_ready) {
277 		return 0;
278 	}
279 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
280 				      AMDGPU_IB_POOL_SIZE*64*1024,
281 				      AMDGPU_GPU_PAGE_SIZE,
282 				      AMDGPU_GEM_DOMAIN_GTT);
283 	if (r) {
284 		return r;
285 	}
286 
287 	adev->ib_pool_ready = true;
288 	if (amdgpu_debugfs_sa_init(adev)) {
289 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
290 	}
291 	return 0;
292 }
293 
294 /**
295  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
296  *
297  * @adev: amdgpu_device pointer
298  *
299  * Tear down the suballocator managing the pool of memory
300  * for use as IBs (all asics).
301  */
302 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
303 {
304 	if (adev->ib_pool_ready) {
305 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
306 		adev->ib_pool_ready = false;
307 	}
308 }
309 
310 /**
311  * amdgpu_ib_ring_tests - test IBs on the rings
312  *
313  * @adev: amdgpu_device pointer
314  *
315  * Test an IB (Indirect Buffer) on each ring.
316  * If the test fails, disable the ring.
317  * Returns 0 on success, error if the primary GFX ring
318  * IB test fails.
319  */
320 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
321 {
322 	unsigned i;
323 	int r, ret = 0;
324 	long tmo_gfx, tmo_mm;
325 
326 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
327 	if (amdgpu_sriov_vf(adev)) {
328 		/* for MM engines in hypervisor side they are not scheduled together
329 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
330 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
331 		 * under SR-IOV should be set to a long time. 8 sec should be enough
332 		 * for the MM comes back to this VF.
333 		 */
334 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
335 	}
336 
337 	if (amdgpu_sriov_runtime(adev)) {
338 		/* for CP & SDMA engines since they are scheduled together so
339 		 * need to make the timeout width enough to cover the time
340 		 * cost waiting for it coming back under RUNTIME only
341 		*/
342 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
343 	}
344 
345 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
346 		struct amdgpu_ring *ring = adev->rings[i];
347 		long tmo;
348 
349 		if (!ring || !ring->ready)
350 			continue;
351 
352 		/* MM engine need more time */
353 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
354 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
355 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
356 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
357 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
358 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
359 			tmo = tmo_mm;
360 		else
361 			tmo = tmo_gfx;
362 
363 		r = amdgpu_ring_test_ib(ring, tmo);
364 		if (r) {
365 			ring->ready = false;
366 
367 			if (ring == &adev->gfx.gfx_ring[0]) {
368 				/* oh, oh, that's really bad */
369 				DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
370 				adev->accel_working = false;
371 				return r;
372 
373 			} else {
374 				/* still not good, but we can live with it */
375 				DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
376 				ret = r;
377 			}
378 		}
379 	}
380 	return ret;
381 }
382 
383 /*
384  * Debugfs info
385  */
386 #if defined(CONFIG_DEBUG_FS)
387 
388 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
389 {
390 	struct drm_info_node *node = (struct drm_info_node *) m->private;
391 	struct drm_device *dev = node->minor->dev;
392 	struct amdgpu_device *adev = dev->dev_private;
393 
394 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
395 
396 	return 0;
397 
398 }
399 
400 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
401 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
402 };
403 
404 #endif
405 
406 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
407 {
408 #if defined(CONFIG_DEBUG_FS)
409 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
410 #else
411 	return 0;
412 #endif
413 }
414