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 
34 #include "amdgpu.h"
35 #include "atom.h"
36 #include "amdgpu_trace.h"
37 
38 #define AMDGPU_IB_TEST_TIMEOUT	msecs_to_jiffies(1000)
39 #define AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT	msecs_to_jiffies(2000)
40 
41 /*
42  * IB
43  * IBs (Indirect Buffers) and areas of GPU accessible memory where
44  * commands are stored.  You can put a pointer to the IB in the
45  * command ring and the hw will fetch the commands from the IB
46  * and execute them.  Generally userspace acceleration drivers
47  * produce command buffers which are send to the kernel and
48  * put in IBs for execution by the requested ring.
49  */
50 
51 /**
52  * amdgpu_ib_get - request an IB (Indirect Buffer)
53  *
54  * @adev: amdgpu_device pointer
55  * @vm: amdgpu_vm pointer
56  * @size: requested IB size
57  * @pool_type: IB pool type (delayed, immediate, direct)
58  * @ib: IB object returned
59  *
60  * Request an IB (all asics).  IBs are allocated using the
61  * suballocator.
62  * Returns 0 on success, error on failure.
63  */
64 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
65 		  unsigned size, enum amdgpu_ib_pool_type pool_type,
66 		  struct amdgpu_ib *ib)
67 {
68 	int r;
69 
70 	if (size) {
71 		r = amdgpu_sa_bo_new(&adev->ib_pools[pool_type],
72 				     &ib->sa_bo, size);
73 		if (r) {
74 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
75 			return r;
76 		}
77 
78 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
79 		/* flush the cache before commit the IB */
80 		ib->flags = AMDGPU_IB_FLAG_EMIT_MEM_SYNC;
81 
82 		if (!vm)
83 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
84 	}
85 
86 	return 0;
87 }
88 
89 /**
90  * amdgpu_ib_free - free an IB (Indirect Buffer)
91  *
92  * @adev: amdgpu_device pointer
93  * @ib: IB object to free
94  * @f: the fence SA bo need wait on for the ib alloation
95  *
96  * Free an IB (all asics).
97  */
98 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
99 		    struct dma_fence *f)
100 {
101 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
102 }
103 
104 /**
105  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
106  *
107  * @ring: ring index the IB is associated with
108  * @num_ibs: number of IBs to schedule
109  * @ibs: IB objects to schedule
110  * @job: job to schedule
111  * @f: fence created during this submission
112  *
113  * Schedule an IB on the associated ring (all asics).
114  * Returns 0 on success, error on failure.
115  *
116  * On SI, there are two parallel engines fed from the primary ring,
117  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
118  * resource descriptors have moved to memory, the CE allows you to
119  * prime the caches while the DE is updating register state so that
120  * the resource descriptors will be already in cache when the draw is
121  * processed.  To accomplish this, the userspace driver submits two
122  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
123  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
124  * to SI there was just a DE IB.
125  */
126 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
127 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
128 		       struct dma_fence **f)
129 {
130 	struct amdgpu_device *adev = ring->adev;
131 	struct amdgpu_ib *ib = &ibs[0];
132 	struct dma_fence *tmp = NULL;
133 	bool need_ctx_switch;
134 	unsigned patch_offset = ~0;
135 	struct amdgpu_vm *vm;
136 	uint64_t fence_ctx;
137 	uint32_t status = 0, alloc_size;
138 	unsigned fence_flags = 0;
139 	bool secure, init_shadow;
140 	u64 shadow_va, csa_va, gds_va;
141 	int vmid = AMDGPU_JOB_GET_VMID(job);
142 
143 	unsigned i;
144 	int r = 0;
145 	bool need_pipe_sync = false;
146 
147 	if (num_ibs == 0)
148 		return -EINVAL;
149 
150 	/* ring tests don't use a job */
151 	if (job) {
152 		vm = job->vm;
153 		fence_ctx = job->base.s_fence ?
154 			job->base.s_fence->scheduled.context : 0;
155 		shadow_va = job->shadow_va;
156 		csa_va = job->csa_va;
157 		gds_va = job->gds_va;
158 		init_shadow = job->init_shadow;
159 	} else {
160 		vm = NULL;
161 		fence_ctx = 0;
162 		shadow_va = 0;
163 		csa_va = 0;
164 		gds_va = 0;
165 		init_shadow = false;
166 	}
167 
168 	if (!ring->sched.ready && !ring->is_mes_queue) {
169 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
170 		return -EINVAL;
171 	}
172 
173 	if (vm && !job->vmid && !ring->is_mes_queue) {
174 		dev_err(adev->dev, "VM IB without ID\n");
175 		return -EINVAL;
176 	}
177 
178 	if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
179 	    (!ring->funcs->secure_submission_supported)) {
180 		dev_err(adev->dev, "secure submissions not supported on ring <%s>\n", ring->name);
181 		return -EINVAL;
182 	}
183 
184 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
185 		ring->funcs->emit_ib_size;
186 
187 	r = amdgpu_ring_alloc(ring, alloc_size);
188 	if (r) {
189 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
190 		return r;
191 	}
192 
193 	need_ctx_switch = ring->current_ctx != fence_ctx;
194 	if (ring->funcs->emit_pipeline_sync && job &&
195 	    ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
196 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
197 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
198 		need_pipe_sync = true;
199 
200 		if (tmp)
201 			trace_amdgpu_ib_pipe_sync(job, tmp);
202 
203 		dma_fence_put(tmp);
204 	}
205 
206 	if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
207 		ring->funcs->emit_mem_sync(ring);
208 
209 	if (ring->funcs->emit_wave_limit &&
210 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
211 		ring->funcs->emit_wave_limit(ring, true);
212 
213 	if (ring->funcs->insert_start)
214 		ring->funcs->insert_start(ring);
215 
216 	if (job) {
217 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
218 		if (r) {
219 			amdgpu_ring_undo(ring);
220 			return r;
221 		}
222 	}
223 
224 	amdgpu_ring_ib_begin(ring);
225 
226 	if (ring->funcs->emit_gfx_shadow)
227 		amdgpu_ring_emit_gfx_shadow(ring, shadow_va, csa_va, gds_va,
228 					    init_shadow, vmid);
229 
230 	if (ring->funcs->init_cond_exec)
231 		patch_offset = amdgpu_ring_init_cond_exec(ring);
232 
233 	amdgpu_device_flush_hdp(adev, ring);
234 
235 	if (need_ctx_switch)
236 		status |= AMDGPU_HAVE_CTX_SWITCH;
237 
238 	if (job && ring->funcs->emit_cntxcntl) {
239 		status |= job->preamble_status;
240 		status |= job->preemption_status;
241 		amdgpu_ring_emit_cntxcntl(ring, status);
242 	}
243 
244 	/* Setup initial TMZiness and send it off.
245 	 */
246 	secure = false;
247 	if (job && ring->funcs->emit_frame_cntl) {
248 		secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
249 		amdgpu_ring_emit_frame_cntl(ring, true, secure);
250 	}
251 
252 	for (i = 0; i < num_ibs; ++i) {
253 		ib = &ibs[i];
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 	amdgpu_device_invalidate_hdp(adev, ring);
271 
272 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
273 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
274 
275 	/* wrap the last IB with fence */
276 	if (job && job->uf_addr) {
277 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
278 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
279 	}
280 
281 	if (ring->funcs->emit_gfx_shadow) {
282 		amdgpu_ring_emit_gfx_shadow(ring, 0, 0, 0, false, 0);
283 
284 		if (ring->funcs->init_cond_exec) {
285 			unsigned ce_offset = ~0;
286 
287 			ce_offset = amdgpu_ring_init_cond_exec(ring);
288 			if (ce_offset != ~0 && ring->funcs->patch_cond_exec)
289 				amdgpu_ring_patch_cond_exec(ring, ce_offset);
290 		}
291 	}
292 
293 	r = amdgpu_fence_emit(ring, f, job, fence_flags);
294 	if (r) {
295 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
296 		if (job && job->vmid)
297 			amdgpu_vmid_reset(adev, ring->vm_hub, job->vmid);
298 		amdgpu_ring_undo(ring);
299 		return r;
300 	}
301 
302 	if (ring->funcs->insert_end)
303 		ring->funcs->insert_end(ring);
304 
305 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
306 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
307 
308 	ring->current_ctx = fence_ctx;
309 	if (vm && ring->funcs->emit_switch_buffer)
310 		amdgpu_ring_emit_switch_buffer(ring);
311 
312 	if (ring->funcs->emit_wave_limit &&
313 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
314 		ring->funcs->emit_wave_limit(ring, false);
315 
316 	amdgpu_ring_ib_end(ring);
317 	amdgpu_ring_commit(ring);
318 	return 0;
319 }
320 
321 /**
322  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
323  *
324  * @adev: amdgpu_device pointer
325  *
326  * Initialize the suballocator to manage a pool of memory
327  * for use as IBs (all asics).
328  * Returns 0 on success, error on failure.
329  */
330 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
331 {
332 	int r, i;
333 
334 	if (adev->ib_pool_ready)
335 		return 0;
336 
337 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
338 		r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
339 					      AMDGPU_IB_POOL_SIZE, 256,
340 					      AMDGPU_GEM_DOMAIN_GTT);
341 		if (r)
342 			goto error;
343 	}
344 	adev->ib_pool_ready = true;
345 
346 	return 0;
347 
348 error:
349 	while (i--)
350 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
351 	return r;
352 }
353 
354 /**
355  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
356  *
357  * @adev: amdgpu_device pointer
358  *
359  * Tear down the suballocator managing the pool of memory
360  * for use as IBs (all asics).
361  */
362 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
363 {
364 	int i;
365 
366 	if (!adev->ib_pool_ready)
367 		return;
368 
369 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
370 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
371 	adev->ib_pool_ready = false;
372 }
373 
374 /**
375  * amdgpu_ib_ring_tests - test IBs on the rings
376  *
377  * @adev: amdgpu_device pointer
378  *
379  * Test an IB (Indirect Buffer) on each ring.
380  * If the test fails, disable the ring.
381  * Returns 0 on success, error if the primary GFX ring
382  * IB test fails.
383  */
384 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
385 {
386 	long tmo_gfx, tmo_mm;
387 	int r, ret = 0;
388 	unsigned i;
389 
390 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
391 	if (amdgpu_sriov_vf(adev)) {
392 		/* for MM engines in hypervisor side they are not scheduled together
393 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
394 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
395 		 * under SR-IOV should be set to a long time. 8 sec should be enough
396 		 * for the MM comes back to this VF.
397 		 */
398 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
399 	}
400 
401 	if (amdgpu_sriov_runtime(adev)) {
402 		/* for CP & SDMA engines since they are scheduled together so
403 		 * need to make the timeout width enough to cover the time
404 		 * cost waiting for it coming back under RUNTIME only
405 		*/
406 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
407 	} else if (adev->gmc.xgmi.hive_id) {
408 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
409 	}
410 
411 	for (i = 0; i < adev->num_rings; ++i) {
412 		struct amdgpu_ring *ring = adev->rings[i];
413 		long tmo;
414 
415 		/* KIQ rings don't have an IB test because we never submit IBs
416 		 * to them and they have no interrupt support.
417 		 */
418 		if (!ring->sched.ready || !ring->funcs->test_ib)
419 			continue;
420 
421 		if (adev->enable_mes &&
422 		    ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
423 			continue;
424 
425 		/* MM engine need more time */
426 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
427 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
428 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
429 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
430 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
431 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
432 			tmo = tmo_mm;
433 		else
434 			tmo = tmo_gfx;
435 
436 		r = amdgpu_ring_test_ib(ring, tmo);
437 		if (!r) {
438 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
439 				      ring->name);
440 			continue;
441 		}
442 
443 		ring->sched.ready = false;
444 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
445 			  ring->name, r);
446 
447 		if (ring == &adev->gfx.gfx_ring[0]) {
448 			/* oh, oh, that's really bad */
449 			adev->accel_working = false;
450 			return r;
451 
452 		} else {
453 			ret = r;
454 		}
455 	}
456 	return ret;
457 }
458 
459 /*
460  * Debugfs info
461  */
462 #if defined(CONFIG_DEBUG_FS)
463 
464 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
465 {
466 	struct amdgpu_device *adev = m->private;
467 
468 	seq_printf(m, "--------------------- DELAYED --------------------- \n");
469 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
470 				     m);
471 	seq_printf(m, "-------------------- IMMEDIATE -------------------- \n");
472 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
473 				     m);
474 	seq_printf(m, "--------------------- DIRECT ---------------------- \n");
475 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
476 
477 	return 0;
478 }
479 
480 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
481 
482 #endif
483 
484 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
485 {
486 #if defined(CONFIG_DEBUG_FS)
487 	struct drm_minor *minor = adev_to_drm(adev)->primary;
488 	struct dentry *root = minor->debugfs_root;
489 
490 	debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
491 			    &amdgpu_debugfs_sa_info_fops);
492 
493 #endif
494 }
495