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, 256);
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;
140 
141 	unsigned i;
142 	int r = 0;
143 	bool need_pipe_sync = false;
144 
145 	if (num_ibs == 0)
146 		return -EINVAL;
147 
148 	/* ring tests don't use a job */
149 	if (job) {
150 		vm = job->vm;
151 		fence_ctx = job->base.s_fence ?
152 			job->base.s_fence->scheduled.context : 0;
153 	} else {
154 		vm = NULL;
155 		fence_ctx = 0;
156 	}
157 
158 	if (!ring->sched.ready) {
159 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
160 		return -EINVAL;
161 	}
162 
163 	if (vm && !job->vmid) {
164 		dev_err(adev->dev, "VM IB without ID\n");
165 		return -EINVAL;
166 	}
167 
168 	if ((ib->flags & AMDGPU_IB_FLAGS_SECURE) &&
169 	    (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)) {
170 		dev_err(adev->dev, "secure submissions not supported on compute rings\n");
171 		return -EINVAL;
172 	}
173 
174 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
175 		ring->funcs->emit_ib_size;
176 
177 	r = amdgpu_ring_alloc(ring, alloc_size);
178 	if (r) {
179 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
180 		return r;
181 	}
182 
183 	need_ctx_switch = ring->current_ctx != fence_ctx;
184 	if (ring->funcs->emit_pipeline_sync && job &&
185 	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync)) ||
186 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
187 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
188 		need_pipe_sync = true;
189 
190 		if (tmp)
191 			trace_amdgpu_ib_pipe_sync(job, tmp);
192 
193 		dma_fence_put(tmp);
194 	}
195 
196 	if ((ib->flags & AMDGPU_IB_FLAG_EMIT_MEM_SYNC) && ring->funcs->emit_mem_sync)
197 		ring->funcs->emit_mem_sync(ring);
198 
199 	if (ring->funcs->emit_wave_limit &&
200 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
201 		ring->funcs->emit_wave_limit(ring, true);
202 
203 	if (ring->funcs->insert_start)
204 		ring->funcs->insert_start(ring);
205 
206 	if (job) {
207 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
208 		if (r) {
209 			amdgpu_ring_undo(ring);
210 			return r;
211 		}
212 	}
213 
214 	if (job && ring->funcs->init_cond_exec)
215 		patch_offset = amdgpu_ring_init_cond_exec(ring);
216 
217 	amdgpu_device_flush_hdp(adev, ring);
218 
219 	if (need_ctx_switch)
220 		status |= AMDGPU_HAVE_CTX_SWITCH;
221 
222 	if (job && ring->funcs->emit_cntxcntl) {
223 		status |= job->preamble_status;
224 		status |= job->preemption_status;
225 		amdgpu_ring_emit_cntxcntl(ring, status);
226 	}
227 
228 	/* Setup initial TMZiness and send it off.
229 	 */
230 	secure = false;
231 	if (job && ring->funcs->emit_frame_cntl) {
232 		secure = ib->flags & AMDGPU_IB_FLAGS_SECURE;
233 		amdgpu_ring_emit_frame_cntl(ring, true, secure);
234 	}
235 
236 	for (i = 0; i < num_ibs; ++i) {
237 		ib = &ibs[i];
238 
239 		if (job && ring->funcs->emit_frame_cntl) {
240 			if (secure != !!(ib->flags & AMDGPU_IB_FLAGS_SECURE)) {
241 				amdgpu_ring_emit_frame_cntl(ring, false, secure);
242 				secure = !secure;
243 				amdgpu_ring_emit_frame_cntl(ring, true, secure);
244 			}
245 		}
246 
247 		amdgpu_ring_emit_ib(ring, job, ib, status);
248 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
249 	}
250 
251 	if (job && ring->funcs->emit_frame_cntl)
252 		amdgpu_ring_emit_frame_cntl(ring, false, secure);
253 
254 	amdgpu_device_invalidate_hdp(adev, ring);
255 
256 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
257 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
258 
259 	/* wrap the last IB with fence */
260 	if (job && job->uf_addr) {
261 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
262 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
263 	}
264 
265 	r = amdgpu_fence_emit(ring, f, fence_flags);
266 	if (r) {
267 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
268 		if (job && job->vmid)
269 			amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
270 		amdgpu_ring_undo(ring);
271 		return r;
272 	}
273 
274 	if (ring->funcs->insert_end)
275 		ring->funcs->insert_end(ring);
276 
277 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
278 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
279 
280 	ring->current_ctx = fence_ctx;
281 	if (vm && ring->funcs->emit_switch_buffer)
282 		amdgpu_ring_emit_switch_buffer(ring);
283 
284 	if (ring->funcs->emit_wave_limit &&
285 	    ring->hw_prio == AMDGPU_GFX_PIPE_PRIO_HIGH)
286 		ring->funcs->emit_wave_limit(ring, false);
287 
288 	amdgpu_ring_commit(ring);
289 	return 0;
290 }
291 
292 /**
293  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
294  *
295  * @adev: amdgpu_device pointer
296  *
297  * Initialize the suballocator to manage a pool of memory
298  * for use as IBs (all asics).
299  * Returns 0 on success, error on failure.
300  */
301 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
302 {
303 	unsigned size;
304 	int r, i;
305 
306 	if (adev->ib_pool_ready)
307 		return 0;
308 
309 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
310 		if (i == AMDGPU_IB_POOL_DIRECT)
311 			size = PAGE_SIZE * 6;
312 		else
313 			size = AMDGPU_IB_POOL_SIZE;
314 
315 		r = amdgpu_sa_bo_manager_init(adev, &adev->ib_pools[i],
316 					      size, AMDGPU_GPU_PAGE_SIZE,
317 					      AMDGPU_GEM_DOMAIN_GTT);
318 		if (r)
319 			goto error;
320 	}
321 	adev->ib_pool_ready = true;
322 
323 	return 0;
324 
325 error:
326 	while (i--)
327 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
328 	return r;
329 }
330 
331 /**
332  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
333  *
334  * @adev: amdgpu_device pointer
335  *
336  * Tear down the suballocator managing the pool of memory
337  * for use as IBs (all asics).
338  */
339 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
340 {
341 	int i;
342 
343 	if (!adev->ib_pool_ready)
344 		return;
345 
346 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
347 		amdgpu_sa_bo_manager_fini(adev, &adev->ib_pools[i]);
348 	adev->ib_pool_ready = false;
349 }
350 
351 /**
352  * amdgpu_ib_ring_tests - test IBs on the rings
353  *
354  * @adev: amdgpu_device pointer
355  *
356  * Test an IB (Indirect Buffer) on each ring.
357  * If the test fails, disable the ring.
358  * Returns 0 on success, error if the primary GFX ring
359  * IB test fails.
360  */
361 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
362 {
363 	long tmo_gfx, tmo_mm;
364 	int r, ret = 0;
365 	unsigned i;
366 
367 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
368 	if (amdgpu_sriov_vf(adev)) {
369 		/* for MM engines in hypervisor side they are not scheduled together
370 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
371 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
372 		 * under SR-IOV should be set to a long time. 8 sec should be enough
373 		 * for the MM comes back to this VF.
374 		 */
375 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
376 	}
377 
378 	if (amdgpu_sriov_runtime(adev)) {
379 		/* for CP & SDMA engines since they are scheduled together so
380 		 * need to make the timeout width enough to cover the time
381 		 * cost waiting for it coming back under RUNTIME only
382 		*/
383 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
384 	} else if (adev->gmc.xgmi.hive_id) {
385 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
386 	}
387 
388 	for (i = 0; i < adev->num_rings; ++i) {
389 		struct amdgpu_ring *ring = adev->rings[i];
390 		long tmo;
391 
392 		/* KIQ rings don't have an IB test because we never submit IBs
393 		 * to them and they have no interrupt support.
394 		 */
395 		if (!ring->sched.ready || !ring->funcs->test_ib)
396 			continue;
397 
398 		/* MM engine need more time */
399 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
400 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
401 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
402 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
403 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
404 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
405 			tmo = tmo_mm;
406 		else
407 			tmo = tmo_gfx;
408 
409 		r = amdgpu_ring_test_ib(ring, tmo);
410 		if (!r) {
411 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
412 				      ring->name);
413 			continue;
414 		}
415 
416 		ring->sched.ready = false;
417 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
418 			  ring->name, r);
419 
420 		if (ring == &adev->gfx.gfx_ring[0]) {
421 			/* oh, oh, that's really bad */
422 			adev->accel_working = false;
423 			return r;
424 
425 		} else {
426 			ret = r;
427 		}
428 	}
429 	return ret;
430 }
431 
432 /*
433  * Debugfs info
434  */
435 #if defined(CONFIG_DEBUG_FS)
436 
437 static int amdgpu_debugfs_sa_info_show(struct seq_file *m, void *unused)
438 {
439 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
440 
441 	seq_printf(m, "--------------------- DELAYED --------------------- \n");
442 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DELAYED],
443 				     m);
444 	seq_printf(m, "-------------------- IMMEDIATE -------------------- \n");
445 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_IMMEDIATE],
446 				     m);
447 	seq_printf(m, "--------------------- DIRECT ---------------------- \n");
448 	amdgpu_sa_bo_dump_debug_info(&adev->ib_pools[AMDGPU_IB_POOL_DIRECT], m);
449 
450 	return 0;
451 }
452 
453 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_sa_info);
454 
455 #endif
456 
457 void amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
458 {
459 #if defined(CONFIG_DEBUG_FS)
460 	struct drm_minor *minor = adev_to_drm(adev)->primary;
461 	struct dentry *root = minor->debugfs_root;
462 
463 	debugfs_create_file("amdgpu_sa_info", 0444, root, adev,
464 			    &amdgpu_debugfs_sa_info_fops);
465 
466 #endif
467 }
468