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 #include "amdgpu_trace.h"
36 
37 #define AMDGPU_IB_TEST_TIMEOUT	msecs_to_jiffies(1000)
38 
39 /*
40  * IB
41  * IBs (Indirect Buffers) and areas of GPU accessible memory where
42  * commands are stored.  You can put a pointer to the IB in the
43  * command ring and the hw will fetch the commands from the IB
44  * and execute them.  Generally userspace acceleration drivers
45  * produce command buffers which are send to the kernel and
46  * put in IBs for execution by the requested ring.
47  */
48 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev);
49 
50 /**
51  * amdgpu_ib_get - request an IB (Indirect Buffer)
52  *
53  * @ring: ring index the IB is associated with
54  * @size: requested IB size
55  * @ib: IB object returned
56  *
57  * Request an IB (all asics).  IBs are allocated using the
58  * suballocator.
59  * Returns 0 on success, error on failure.
60  */
61 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
62 		  unsigned size, struct amdgpu_ib *ib)
63 {
64 	int r;
65 
66 	if (size) {
67 		r = amdgpu_sa_bo_new(&adev->ring_tmp_bo,
68 				      &ib->sa_bo, size, 256);
69 		if (r) {
70 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
71 			return r;
72 		}
73 
74 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
75 
76 		if (!vm)
77 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
78 	}
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,
93 		    struct dma_fence *f)
94 {
95 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
96 }
97 
98 /**
99  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
100  *
101  * @adev: amdgpu_device pointer
102  * @num_ibs: number of IBs to schedule
103  * @ibs: IB objects to schedule
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, struct amdgpu_job *job,
121 		       struct dma_fence **f)
122 {
123 	struct amdgpu_device *adev = ring->adev;
124 	struct amdgpu_ib *ib = &ibs[0];
125 	struct dma_fence *tmp = NULL;
126 	bool skip_preamble, need_ctx_switch;
127 	unsigned patch_offset = ~0;
128 	struct amdgpu_vm *vm;
129 	uint64_t fence_ctx;
130 	uint32_t status = 0, alloc_size;
131 	unsigned fence_flags = 0;
132 
133 	unsigned i;
134 	int r = 0;
135 	bool need_pipe_sync = false;
136 
137 	if (num_ibs == 0)
138 		return -EINVAL;
139 
140 	/* ring tests don't use a job */
141 	if (job) {
142 		vm = job->vm;
143 		fence_ctx = job->base.s_fence->scheduled.context;
144 	} else {
145 		vm = NULL;
146 		fence_ctx = 0;
147 	}
148 
149 	if (!ring->sched.ready) {
150 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
151 		return -EINVAL;
152 	}
153 
154 	if (vm && !job->vmid) {
155 		dev_err(adev->dev, "VM IB without ID\n");
156 		return -EINVAL;
157 	}
158 
159 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
160 		ring->funcs->emit_ib_size;
161 
162 	r = amdgpu_ring_alloc(ring, alloc_size);
163 	if (r) {
164 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
165 		return r;
166 	}
167 
168 	need_ctx_switch = ring->current_ctx != fence_ctx;
169 	if (ring->funcs->emit_pipeline_sync && job &&
170 	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync, NULL)) ||
171 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
172 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
173 		need_pipe_sync = true;
174 
175 		if (tmp)
176 			trace_amdgpu_ib_pipe_sync(job, tmp);
177 
178 		dma_fence_put(tmp);
179 	}
180 
181 	if (ring->funcs->insert_start)
182 		ring->funcs->insert_start(ring);
183 
184 	if (job) {
185 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
186 		if (r) {
187 			amdgpu_ring_undo(ring);
188 			return r;
189 		}
190 	}
191 
192 	if (job && ring->funcs->init_cond_exec)
193 		patch_offset = amdgpu_ring_init_cond_exec(ring);
194 
195 #ifdef CONFIG_X86_64
196 	if (!(adev->flags & AMD_IS_APU))
197 #endif
198 	{
199 		if (ring->funcs->emit_hdp_flush)
200 			amdgpu_ring_emit_hdp_flush(ring);
201 		else
202 			amdgpu_asic_flush_hdp(adev, ring);
203 	}
204 
205 	skip_preamble = ring->current_ctx == fence_ctx;
206 	if (job && ring->funcs->emit_cntxcntl) {
207 		if (need_ctx_switch)
208 			status |= AMDGPU_HAVE_CTX_SWITCH;
209 		status |= job->preamble_status;
210 
211 		amdgpu_ring_emit_cntxcntl(ring, status);
212 	}
213 
214 	for (i = 0; i < num_ibs; ++i) {
215 		ib = &ibs[i];
216 
217 		/* drop preamble IBs if we don't have a context switch */
218 		if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
219 			skip_preamble &&
220 			!(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
221 			!amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
222 			continue;
223 
224 		amdgpu_ring_emit_ib(ring, job, ib, need_ctx_switch);
225 		need_ctx_switch = false;
226 	}
227 
228 	if (ring->funcs->emit_tmz)
229 		amdgpu_ring_emit_tmz(ring, false);
230 
231 #ifdef CONFIG_X86_64
232 	if (!(adev->flags & AMD_IS_APU))
233 #endif
234 		amdgpu_asic_invalidate_hdp(adev, ring);
235 
236 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
237 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
238 
239 	/* wrap the last IB with fence */
240 	if (job && job->uf_addr) {
241 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
242 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
243 	}
244 
245 	r = amdgpu_fence_emit(ring, f, fence_flags);
246 	if (r) {
247 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
248 		if (job && job->vmid)
249 			amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
250 		amdgpu_ring_undo(ring);
251 		return r;
252 	}
253 
254 	if (ring->funcs->insert_end)
255 		ring->funcs->insert_end(ring);
256 
257 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
258 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
259 
260 	ring->current_ctx = fence_ctx;
261 	if (vm && ring->funcs->emit_switch_buffer)
262 		amdgpu_ring_emit_switch_buffer(ring);
263 	amdgpu_ring_commit(ring);
264 	return 0;
265 }
266 
267 /**
268  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
269  *
270  * @adev: amdgpu_device pointer
271  *
272  * Initialize the suballocator to manage a pool of memory
273  * for use as IBs (all asics).
274  * Returns 0 on success, error on failure.
275  */
276 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
277 {
278 	int r;
279 
280 	if (adev->ib_pool_ready) {
281 		return 0;
282 	}
283 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
284 				      AMDGPU_IB_POOL_SIZE*64*1024,
285 				      AMDGPU_GPU_PAGE_SIZE,
286 				      AMDGPU_GEM_DOMAIN_GTT);
287 	if (r) {
288 		return r;
289 	}
290 
291 	adev->ib_pool_ready = true;
292 	if (amdgpu_debugfs_sa_init(adev)) {
293 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
294 	}
295 	return 0;
296 }
297 
298 /**
299  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
300  *
301  * @adev: amdgpu_device pointer
302  *
303  * Tear down the suballocator managing the pool of memory
304  * for use as IBs (all asics).
305  */
306 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
307 {
308 	if (adev->ib_pool_ready) {
309 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
310 		adev->ib_pool_ready = false;
311 	}
312 }
313 
314 /**
315  * amdgpu_ib_ring_tests - test IBs on the rings
316  *
317  * @adev: amdgpu_device pointer
318  *
319  * Test an IB (Indirect Buffer) on each ring.
320  * If the test fails, disable the ring.
321  * Returns 0 on success, error if the primary GFX ring
322  * IB test fails.
323  */
324 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
325 {
326 	unsigned i;
327 	int r, ret = 0;
328 	long tmo_gfx, tmo_mm;
329 
330 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
331 	if (amdgpu_sriov_vf(adev)) {
332 		/* for MM engines in hypervisor side they are not scheduled together
333 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
334 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
335 		 * under SR-IOV should be set to a long time. 8 sec should be enough
336 		 * for the MM comes back to this VF.
337 		 */
338 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
339 	}
340 
341 	if (amdgpu_sriov_runtime(adev)) {
342 		/* for CP & SDMA engines since they are scheduled together so
343 		 * need to make the timeout width enough to cover the time
344 		 * cost waiting for it coming back under RUNTIME only
345 		*/
346 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
347 	}
348 
349 	for (i = 0; i < adev->num_rings; ++i) {
350 		struct amdgpu_ring *ring = adev->rings[i];
351 		long tmo;
352 
353 		/* KIQ rings don't have an IB test because we never submit IBs
354 		 * to them and they have no interrupt support.
355 		 */
356 		if (!ring->sched.ready || !ring->funcs->test_ib)
357 			continue;
358 
359 		/* MM engine need more time */
360 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
361 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
362 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
363 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
364 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
365 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
366 			tmo = tmo_mm;
367 		else
368 			tmo = tmo_gfx;
369 
370 		r = amdgpu_ring_test_ib(ring, tmo);
371 		if (!r) {
372 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
373 				      ring->name);
374 			continue;
375 		}
376 
377 		ring->sched.ready = false;
378 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
379 			  ring->name, r);
380 
381 		if (ring == &adev->gfx.gfx_ring[0]) {
382 			/* oh, oh, that's really bad */
383 			adev->accel_working = false;
384 			return r;
385 
386 		} else {
387 			ret = r;
388 		}
389 	}
390 	return ret;
391 }
392 
393 /*
394  * Debugfs info
395  */
396 #if defined(CONFIG_DEBUG_FS)
397 
398 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
399 {
400 	struct drm_info_node *node = (struct drm_info_node *) m->private;
401 	struct drm_device *dev = node->minor->dev;
402 	struct amdgpu_device *adev = dev->dev_private;
403 
404 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
405 
406 	return 0;
407 
408 }
409 
410 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
411 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
412 };
413 
414 #endif
415 
416 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
417 {
418 #if defined(CONFIG_DEBUG_FS)
419 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
420 #else
421 	return 0;
422 #endif
423 }
424