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 #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT	msecs_to_jiffies(2000)
39 
40 /*
41  * IB
42  * IBs (Indirect Buffers) and areas of GPU accessible memory where
43  * commands are stored.  You can put a pointer to the IB in the
44  * command ring and the hw will fetch the commands from the IB
45  * and execute them.  Generally userspace acceleration drivers
46  * produce command buffers which are send to the kernel and
47  * put in IBs for execution by the requested ring.
48  */
49 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev);
50 
51 /**
52  * amdgpu_ib_get - request an IB (Indirect Buffer)
53  *
54  * @ring: ring index the IB is associated with
55  * @size: requested IB size
56  * @ib: IB object returned
57  *
58  * Request an IB (all asics).  IBs are allocated using the
59  * suballocator.
60  * Returns 0 on success, error on failure.
61  */
62 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
63 		  unsigned size, struct amdgpu_ib *ib)
64 {
65 	int r;
66 
67 	if (size) {
68 		r = amdgpu_sa_bo_new(&adev->ring_tmp_bo,
69 				      &ib->sa_bo, size, 256);
70 		if (r) {
71 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
72 			return r;
73 		}
74 
75 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
76 
77 		if (!vm)
78 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
79 	}
80 
81 	return 0;
82 }
83 
84 /**
85  * amdgpu_ib_free - free an IB (Indirect Buffer)
86  *
87  * @adev: amdgpu_device pointer
88  * @ib: IB object to free
89  * @f: the fence SA bo need wait on for the ib alloation
90  *
91  * Free an IB (all asics).
92  */
93 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
94 		    struct dma_fence *f)
95 {
96 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
97 }
98 
99 /**
100  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
101  *
102  * @adev: amdgpu_device pointer
103  * @num_ibs: number of IBs to schedule
104  * @ibs: IB objects to schedule
105  * @f: fence created during this submission
106  *
107  * Schedule an IB on the associated ring (all asics).
108  * Returns 0 on success, error on failure.
109  *
110  * On SI, there are two parallel engines fed from the primary ring,
111  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
112  * resource descriptors have moved to memory, the CE allows you to
113  * prime the caches while the DE is updating register state so that
114  * the resource descriptors will be already in cache when the draw is
115  * processed.  To accomplish this, the userspace driver submits two
116  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
117  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
118  * to SI there was just a DE IB.
119  */
120 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
121 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
122 		       struct dma_fence **f)
123 {
124 	struct amdgpu_device *adev = ring->adev;
125 	struct amdgpu_ib *ib = &ibs[0];
126 	struct dma_fence *tmp = NULL;
127 	bool skip_preamble, need_ctx_switch;
128 	unsigned patch_offset = ~0;
129 	struct amdgpu_vm *vm;
130 	uint64_t fence_ctx;
131 	uint32_t status = 0, alloc_size;
132 	unsigned fence_flags = 0;
133 
134 	unsigned i;
135 	int r = 0;
136 	bool need_pipe_sync = false;
137 
138 	if (num_ibs == 0)
139 		return -EINVAL;
140 
141 	/* ring tests don't use a job */
142 	if (job) {
143 		vm = job->vm;
144 		fence_ctx = job->base.s_fence->scheduled.context;
145 	} else {
146 		vm = NULL;
147 		fence_ctx = 0;
148 	}
149 
150 	if (!ring->sched.ready) {
151 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
152 		return -EINVAL;
153 	}
154 
155 	if (vm && !job->vmid) {
156 		dev_err(adev->dev, "VM IB without ID\n");
157 		return -EINVAL;
158 	}
159 
160 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
161 		ring->funcs->emit_ib_size;
162 
163 	r = amdgpu_ring_alloc(ring, alloc_size);
164 	if (r) {
165 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
166 		return r;
167 	}
168 
169 	need_ctx_switch = ring->current_ctx != fence_ctx;
170 	if (ring->funcs->emit_pipeline_sync && job &&
171 	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync, NULL)) ||
172 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
173 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
174 		need_pipe_sync = true;
175 
176 		if (tmp)
177 			trace_amdgpu_ib_pipe_sync(job, tmp);
178 
179 		dma_fence_put(tmp);
180 	}
181 
182 	if (ring->funcs->insert_start)
183 		ring->funcs->insert_start(ring);
184 
185 	if (job) {
186 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
187 		if (r) {
188 			amdgpu_ring_undo(ring);
189 			return r;
190 		}
191 	}
192 
193 	if (job && ring->funcs->init_cond_exec)
194 		patch_offset = amdgpu_ring_init_cond_exec(ring);
195 
196 #ifdef CONFIG_X86_64
197 	if (!(adev->flags & AMD_IS_APU))
198 #endif
199 	{
200 		if (ring->funcs->emit_hdp_flush)
201 			amdgpu_ring_emit_hdp_flush(ring);
202 		else
203 			amdgpu_asic_flush_hdp(adev, ring);
204 	}
205 
206 	if (need_ctx_switch)
207 		status |= AMDGPU_HAVE_CTX_SWITCH;
208 
209 	skip_preamble = ring->current_ctx == fence_ctx;
210 	if (job && ring->funcs->emit_cntxcntl) {
211 		status |= job->preamble_status;
212 		status |= job->preemption_status;
213 		amdgpu_ring_emit_cntxcntl(ring, status);
214 	}
215 
216 	for (i = 0; i < num_ibs; ++i) {
217 		ib = &ibs[i];
218 
219 		/* drop preamble IBs if we don't have a context switch */
220 		if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
221 		    skip_preamble &&
222 		    !(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
223 		    !amdgpu_mcbp &&
224 		    !amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
225 			continue;
226 
227 		amdgpu_ring_emit_ib(ring, job, ib, status);
228 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
229 	}
230 
231 	if (ring->funcs->emit_tmz)
232 		amdgpu_ring_emit_tmz(ring, false);
233 
234 #ifdef CONFIG_X86_64
235 	if (!(adev->flags & AMD_IS_APU))
236 #endif
237 		amdgpu_asic_invalidate_hdp(adev, ring);
238 
239 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
240 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
241 
242 	/* wrap the last IB with fence */
243 	if (job && job->uf_addr) {
244 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
245 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
246 	}
247 
248 	r = amdgpu_fence_emit(ring, f, fence_flags);
249 	if (r) {
250 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
251 		if (job && job->vmid)
252 			amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
253 		amdgpu_ring_undo(ring);
254 		return r;
255 	}
256 
257 	if (ring->funcs->insert_end)
258 		ring->funcs->insert_end(ring);
259 
260 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
261 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
262 
263 	ring->current_ctx = fence_ctx;
264 	if (vm && ring->funcs->emit_switch_buffer)
265 		amdgpu_ring_emit_switch_buffer(ring);
266 	amdgpu_ring_commit(ring);
267 	return 0;
268 }
269 
270 /**
271  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
272  *
273  * @adev: amdgpu_device pointer
274  *
275  * Initialize the suballocator to manage a pool of memory
276  * for use as IBs (all asics).
277  * Returns 0 on success, error on failure.
278  */
279 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
280 {
281 	int r;
282 
283 	if (adev->ib_pool_ready) {
284 		return 0;
285 	}
286 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
287 				      AMDGPU_IB_POOL_SIZE*64*1024,
288 				      AMDGPU_GPU_PAGE_SIZE,
289 				      AMDGPU_GEM_DOMAIN_GTT);
290 	if (r) {
291 		return r;
292 	}
293 
294 	adev->ib_pool_ready = true;
295 	if (amdgpu_debugfs_sa_init(adev)) {
296 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
297 	}
298 	return 0;
299 }
300 
301 /**
302  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
303  *
304  * @adev: amdgpu_device pointer
305  *
306  * Tear down the suballocator managing the pool of memory
307  * for use as IBs (all asics).
308  */
309 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
310 {
311 	if (adev->ib_pool_ready) {
312 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
313 		adev->ib_pool_ready = false;
314 	}
315 }
316 
317 /**
318  * amdgpu_ib_ring_tests - test IBs on the rings
319  *
320  * @adev: amdgpu_device pointer
321  *
322  * Test an IB (Indirect Buffer) on each ring.
323  * If the test fails, disable the ring.
324  * Returns 0 on success, error if the primary GFX ring
325  * IB test fails.
326  */
327 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
328 {
329 	unsigned i;
330 	int r, ret = 0;
331 	long tmo_gfx, tmo_mm;
332 
333 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
334 	if (amdgpu_sriov_vf(adev)) {
335 		/* for MM engines in hypervisor side they are not scheduled together
336 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
337 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
338 		 * under SR-IOV should be set to a long time. 8 sec should be enough
339 		 * for the MM comes back to this VF.
340 		 */
341 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
342 	}
343 
344 	if (amdgpu_sriov_runtime(adev)) {
345 		/* for CP & SDMA engines since they are scheduled together so
346 		 * need to make the timeout width enough to cover the time
347 		 * cost waiting for it coming back under RUNTIME only
348 		*/
349 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
350 	} else if (adev->gmc.xgmi.hive_id) {
351 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
352 	}
353 
354 	for (i = 0; i < adev->num_rings; ++i) {
355 		struct amdgpu_ring *ring = adev->rings[i];
356 		long tmo;
357 
358 		/* KIQ rings don't have an IB test because we never submit IBs
359 		 * to them and they have no interrupt support.
360 		 */
361 		if (!ring->sched.ready || !ring->funcs->test_ib)
362 			continue;
363 
364 		/* MM engine need more time */
365 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
366 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
367 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
368 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
369 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
370 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
371 			tmo = tmo_mm;
372 		else
373 			tmo = tmo_gfx;
374 
375 		r = amdgpu_ring_test_ib(ring, tmo);
376 		if (!r) {
377 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
378 				      ring->name);
379 			continue;
380 		}
381 
382 		ring->sched.ready = false;
383 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
384 			  ring->name, r);
385 
386 		if (ring == &adev->gfx.gfx_ring[0]) {
387 			/* oh, oh, that's really bad */
388 			adev->accel_working = false;
389 			return r;
390 
391 		} else {
392 			ret = r;
393 		}
394 	}
395 	return ret;
396 }
397 
398 /*
399  * Debugfs info
400  */
401 #if defined(CONFIG_DEBUG_FS)
402 
403 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
404 {
405 	struct drm_info_node *node = (struct drm_info_node *) m->private;
406 	struct drm_device *dev = node->minor->dev;
407 	struct amdgpu_device *adev = dev->dev_private;
408 
409 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
410 
411 	return 0;
412 
413 }
414 
415 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
416 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
417 };
418 
419 #endif
420 
421 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
422 {
423 #if defined(CONFIG_DEBUG_FS)
424 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
425 #else
426 	return 0;
427 #endif
428 }
429