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 
32 #include <drm/amdgpu_drm.h>
33 #include <drm/drm_debugfs.h>
34 
35 #include "amdgpu.h"
36 #include "atom.h"
37 #include "amdgpu_trace.h"
38 
39 #define AMDGPU_IB_TEST_TIMEOUT	msecs_to_jiffies(1000)
40 #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT	msecs_to_jiffies(2000)
41 
42 /*
43  * IB
44  * IBs (Indirect Buffers) and areas of GPU accessible memory where
45  * commands are stored.  You can put a pointer to the IB in the
46  * command ring and the hw will fetch the commands from the IB
47  * and execute them.  Generally userspace acceleration drivers
48  * produce command buffers which are send to the kernel and
49  * put in IBs for execution by the requested ring.
50  */
51 
52 /**
53  * amdgpu_ib_get - request an IB (Indirect Buffer)
54  *
55  * @adev: amdgpu_device pointer
56  * @vm: amdgpu_vm pointer
57  * @size: requested IB size
58  * @pool_type: IB pool type (delayed, immediate, direct)
59  * @ib: IB object returned
60  *
61  * Request an IB (all asics).  IBs are allocated using the
62  * suballocator.
63  * Returns 0 on success, error on failure.
64  */
65 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
66 		  unsigned size, enum amdgpu_ib_pool_type pool_type,
67 		  struct amdgpu_ib *ib)
68 {
69 	int r;
70 
71 	if (size) {
72 		r = amdgpu_sa_bo_new(&adev->ib_pools[pool_type],
73 				      &ib->sa_bo, size, 256);
74 		if (r) {
75 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
76 			return r;
77 		}
78 
79 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
80 
81 		if (!vm)
82 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
83 	}
84 
85 	return 0;
86 }
87 
88 /**
89  * amdgpu_ib_free - free an IB (Indirect Buffer)
90  *
91  * @adev: amdgpu_device pointer
92  * @ib: IB object to free
93  * @f: the fence SA bo need wait on for the ib alloation
94  *
95  * Free an IB (all asics).
96  */
97 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
98 		    struct dma_fence *f)
99 {
100 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
101 }
102 
103 /**
104  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
105  *
106  * @ring: ring index the IB is associated with
107  * @num_ibs: number of IBs to schedule
108  * @ibs: IB objects to schedule
109  * @job: job to schedule
110  * @f: fence created during this submission
111  *
112  * Schedule an IB on the associated ring (all asics).
113  * Returns 0 on success, error on failure.
114  *
115  * On SI, there are two parallel engines fed from the primary ring,
116  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
117  * resource descriptors have moved to memory, the CE allows you to
118  * prime the caches while the DE is updating register state so that
119  * the resource descriptors will be already in cache when the draw is
120  * processed.  To accomplish this, the userspace driver submits two
121  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
122  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
123  * to SI there was just a DE IB.
124  */
125 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
126 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
127 		       struct dma_fence **f)
128 {
129 	struct amdgpu_device *adev = ring->adev;
130 	struct amdgpu_ib *ib = &ibs[0];
131 	struct dma_fence *tmp = NULL;
132 	bool skip_preamble, need_ctx_switch;
133 	unsigned patch_offset = ~0;
134 	struct amdgpu_vm *vm;
135 	uint64_t fence_ctx;
136 	uint32_t status = 0, alloc_size;
137 	unsigned fence_flags = 0;
138 	bool secure;
139 
140 	unsigned i;
141 	int r = 0;
142 	bool need_pipe_sync = false;
143 
144 	if (num_ibs == 0)
145 		return -EINVAL;
146 
147 	/* ring tests don't use a job */
148 	if (job) {
149 		vm = job->vm;
150 		fence_ctx = job->base.s_fence ?
151 			job->base.s_fence->scheduled.context : 0;
152 	} else {
153 		vm = NULL;
154 		fence_ctx = 0;
155 	}
156 
157 	if (!ring->sched.ready) {
158 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
159 		return -EINVAL;
160 	}
161 
162 	if (vm && !job->vmid) {
163 		dev_err(adev->dev, "VM IB without ID\n");
164 		return -EINVAL;
165 	}
166 
167 	if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
168 	    (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)) {
169 		dev_err(adev->dev, "secure submissions not supported on compute rings\n");
170 		return -EINVAL;
171 	}
172 
173 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
174 		ring->funcs->emit_ib_size;
175 
176 	r = amdgpu_ring_alloc(ring, alloc_size);
177 	if (r) {
178 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
179 		return r;
180 	}
181 
182 	need_ctx_switch = ring->current_ctx != fence_ctx;
183 	if (ring->funcs->emit_pipeline_sync && job &&
184 	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync)) ||
185 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
186 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
187 		need_pipe_sync = true;
188 
189 		if (tmp)
190 			trace_amdgpu_ib_pipe_sync(job, tmp);
191 
192 		dma_fence_put(tmp);
193 	}
194 
195 	if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
196 		ring->funcs->emit_mem_sync(ring);
197 
198 	if (ring->funcs->emit_wave_limit &&
199 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
200 		ring->funcs->emit_wave_limit(ring, true);
201 
202 	if (ring->funcs->insert_start)
203 		ring->funcs->insert_start(ring);
204 
205 	if (job) {
206 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
207 		if (r) {
208 			amdgpu_ring_undo(ring);
209 			return r;
210 		}
211 	}
212 
213 	if (job && ring->funcs->init_cond_exec)
214 		patch_offset = amdgpu_ring_init_cond_exec(ring);
215 
216 #ifdef CONFIG_X86_64
217 	if (!(adev->flags & AMD_IS_APU))
218 #endif
219 	{
220 		if (ring->funcs->emit_hdp_flush)
221 			amdgpu_ring_emit_hdp_flush(ring);
222 		else
223 			amdgpu_asic_flush_hdp(adev, ring);
224 	}
225 
226 	if (need_ctx_switch)
227 		status |= AMDGPU_HAVE_CTX_SWITCH;
228 
229 	skip_preamble = ring->current_ctx == fence_ctx;
230 	if (job && ring->funcs->emit_cntxcntl) {
231 		status |= job->preamble_status;
232 		status |= job->preemption_status;
233 		amdgpu_ring_emit_cntxcntl(ring, status);
234 	}
235 
236 	/* Setup initial TMZiness and send it off.
237 	 */
238 	secure = false;
239 	if (job && ring->funcs->emit_frame_cntl) {
240 		secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
241 		amdgpu_ring_emit_frame_cntl(ring, true, secure);
242 	}
243 
244 	for (i = 0; i < num_ibs; ++i) {
245 		ib = &ibs[i];
246 
247 		/* drop preamble IBs if we don't have a context switch */
248 		if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
249 		    skip_preamble &&
250 		    !(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
251 		    !amdgpu_mcbp &&
252 		    !amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
253 			continue;
254 
255 		if (job && ring->funcs->emit_frame_cntl) {
256 			if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
257 				amdgpu_ring_emit_frame_cntl(ring, false, secure);
258 				secure = !secure;
259 				amdgpu_ring_emit_frame_cntl(ring, true, secure);
260 			}
261 		}
262 
263 		amdgpu_ring_emit_ib(ring, job, ib, status);
264 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
265 	}
266 
267 	if (job && ring->funcs->emit_frame_cntl)
268 		amdgpu_ring_emit_frame_cntl(ring, false, secure);
269 
270 #ifdef CONFIG_X86_64
271 	if (!(adev->flags & AMD_IS_APU))
272 #endif
273 		amdgpu_asic_invalidate_hdp(adev, ring);
274 
275 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
276 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
277 
278 	/* wrap the last IB with fence */
279 	if (job && job->uf_addr) {
280 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
281 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
282 	}
283 
284 	r = amdgpu_fence_emit(ring, f, fence_flags);
285 	if (r) {
286 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
287 		if (job && job->vmid)
288 			amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
289 		amdgpu_ring_undo(ring);
290 		return r;
291 	}
292 
293 	if (ring->funcs->insert_end)
294 		ring->funcs->insert_end(ring);
295 
296 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
297 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
298 
299 	ring->current_ctx = fence_ctx;
300 	if (vm && ring->funcs->emit_switch_buffer)
301 		amdgpu_ring_emit_switch_buffer(ring);
302 
303 	if (ring->funcs->emit_wave_limit &&
304 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
305 		ring->funcs->emit_wave_limit(ring, false);
306 
307 	amdgpu_ring_commit(ring);
308 	return 0;
309 }
310 
311 /**
312  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
313  *
314  * @adev: amdgpu_device pointer
315  *
316  * Initialize the suballocator to manage a pool of memory
317  * for use as IBs (all asics).
318  * Returns 0 on success, error on failure.
319  */
320 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
321 {
322 	unsigned size;
323 	int r, i;
324 
325 	if (adev->ib_pool_ready)
326 		return 0;
327 
328 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
329 		if (i == AMDGPU_IB_POOL_DIRECT)
330 			size = PAGE_SIZE * 2;
331 		else
332 			size = AMDGPU_IB_POOL_SIZE;
333 
334 		r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
335 					      size, AMDGPU_GPU_PAGE_SIZE,
336 					      AMDGPU_GEM_DOMAIN_GTT);
337 		if (r)
338 			goto error;
339 	}
340 	adev->ib_pool_ready = true;
341 
342 	return 0;
343 
344 error:
345 	while (i--)
346 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
347 	return r;
348 }
349 
350 /**
351  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
352  *
353  * @adev: amdgpu_device pointer
354  *
355  * Tear down the suballocator managing the pool of memory
356  * for use as IBs (all asics).
357  */
358 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
359 {
360 	int i;
361 
362 	if (!adev->ib_pool_ready)
363 		return;
364 
365 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
366 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
367 	adev->ib_pool_ready = false;
368 }
369 
370 /**
371  * amdgpu_ib_ring_tests - test IBs on the rings
372  *
373  * @adev: amdgpu_device pointer
374  *
375  * Test an IB (Indirect Buffer) on each ring.
376  * If the test fails, disable the ring.
377  * Returns 0 on success, error if the primary GFX ring
378  * IB test fails.
379  */
380 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
381 {
382 	long tmo_gfx, tmo_mm;
383 	int r, ret = 0;
384 	unsigned i;
385 
386 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
387 	if (amdgpu_sriov_vf(adev)) {
388 		/* for MM engines in hypervisor side they are not scheduled together
389 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
390 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
391 		 * under SR-IOV should be set to a long time. 8 sec should be enough
392 		 * for the MM comes back to this VF.
393 		 */
394 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
395 	}
396 
397 	if (amdgpu_sriov_runtime(adev)) {
398 		/* for CP & SDMA engines since they are scheduled together so
399 		 * need to make the timeout width enough to cover the time
400 		 * cost waiting for it coming back under RUNTIME only
401 		*/
402 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
403 	} else if (adev->gmc.xgmi.hive_id) {
404 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
405 	}
406 
407 	for (i = 0; i < adev->num_rings; ++i) {
408 		struct amdgpu_ring *ring = adev->rings[i];
409 		long tmo;
410 
411 		/* KIQ rings don't have an IB test because we never submit IBs
412 		 * to them and they have no interrupt support.
413 		 */
414 		if (!ring->sched.ready || !ring->funcs->test_ib)
415 			continue;
416 
417 		/* MM engine need more time */
418 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
419 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
420 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
421 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
422 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
423 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
424 			tmo = tmo_mm;
425 		else
426 			tmo = tmo_gfx;
427 
428 		r = amdgpu_ring_test_ib(ring, tmo);
429 		if (!r) {
430 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
431 				      ring->name);
432 			continue;
433 		}
434 
435 		ring->sched.ready = false;
436 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
437 			  ring->name, r);
438 
439 		if (ring == &adev->gfx.gfx_ring[0]) {
440 			/* oh, oh, that's really bad */
441 			adev->accel_working = false;
442 			return r;
443 
444 		} else {
445 			ret = r;
446 		}
447 	}
448 	return ret;
449 }
450 
451 /*
452  * Debugfs info
453  */
454 #if defined(CONFIG_DEBUG_FS)
455 
456 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
457 {
458 	struct drm_info_node *node = (struct drm_info_node *) m->private;
459 	struct drm_device *dev = node->minor->dev;
460 	struct amdgpu_device *adev = drm_to_adev(dev);
461 
462 	seq_printf(m, "--------------------- DELAYED --------------------- \n");
463 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
464 				     m);
465 	seq_printf(m, "-------------------- IMMEDIATE -------------------- \n");
466 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
467 				     m);
468 	seq_printf(m, "--------------------- DIRECT ---------------------- \n");
469 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
470 
471 	return 0;
472 }
473 
474 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
475 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
476 };
477 
478 #endif
479 
480 int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
481 {
482 #if defined(CONFIG_DEBUG_FS)
483 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list,
484 					ARRAY_SIZE(amdgpu_debugfs_sa_list));
485 #else
486 	return 0;
487 #endif
488 }
489