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  * @ring: ring index the IB is associated with
56  * @size: requested IB size
57  * @ib: IB object returned
58  *
59  * Request an IB (all asics).  IBs are allocated using the
60  * suballocator.
61  * Returns 0 on success, error on failure.
62  */
63 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
64 		unsigned size,
65 		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->ring_tmp_bo[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 
80 		if (!vm)
81 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
82 	}
83 
84 	return 0;
85 }
86 
87 /**
88  * amdgpu_ib_free - free an IB (Indirect Buffer)
89  *
90  * @adev: amdgpu_device pointer
91  * @ib: IB object to free
92  * @f: the fence SA bo need wait on for the ib alloation
93  *
94  * Free an IB (all asics).
95  */
96 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
97 		    struct dma_fence *f)
98 {
99 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
100 }
101 
102 /**
103  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
104  *
105  * @adev: amdgpu_device pointer
106  * @num_ibs: number of IBs to schedule
107  * @ibs: IB objects to schedule
108  * @f: fence created during this submission
109  *
110  * Schedule an IB on the associated ring (all asics).
111  * Returns 0 on success, error on failure.
112  *
113  * On SI, there are two parallel engines fed from the primary ring,
114  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
115  * resource descriptors have moved to memory, the CE allows you to
116  * prime the caches while the DE is updating register state so that
117  * the resource descriptors will be already in cache when the draw is
118  * processed.  To accomplish this, the userspace driver submits two
119  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
120  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
121  * to SI there was just a DE IB.
122  */
123 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
124 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
125 		       struct dma_fence **f)
126 {
127 	struct amdgpu_device *adev = ring->adev;
128 	struct amdgpu_ib *ib = &ibs[0];
129 	struct dma_fence *tmp = NULL;
130 	bool skip_preamble, need_ctx_switch;
131 	unsigned patch_offset = ~0;
132 	struct amdgpu_vm *vm;
133 	uint64_t fence_ctx;
134 	uint32_t status = 0, alloc_size;
135 	unsigned fence_flags = 0;
136 
137 	unsigned i;
138 	int r = 0;
139 	bool need_pipe_sync = false;
140 
141 	if (num_ibs == 0)
142 		return -EINVAL;
143 
144 	/* ring tests don't use a job */
145 	if (job) {
146 		vm = job->vm;
147 		fence_ctx = job->base.s_fence ?
148 			job->base.s_fence->scheduled.context : 0;
149 	} else {
150 		vm = NULL;
151 		fence_ctx = 0;
152 	}
153 
154 	if (!ring->sched.ready) {
155 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
156 		return -EINVAL;
157 	}
158 
159 	if (vm && !job->vmid) {
160 		dev_err(adev->dev, "VM IB without ID\n");
161 		return -EINVAL;
162 	}
163 
164 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
165 		ring->funcs->emit_ib_size;
166 
167 	r = amdgpu_ring_alloc(ring, alloc_size);
168 	if (r) {
169 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
170 		return r;
171 	}
172 
173 	need_ctx_switch = ring->current_ctx != fence_ctx;
174 	if (ring->funcs->emit_pipeline_sync && job &&
175 	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync, NULL)) ||
176 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
177 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
178 		need_pipe_sync = true;
179 
180 		if (tmp)
181 			trace_amdgpu_ib_pipe_sync(job, tmp);
182 
183 		dma_fence_put(tmp);
184 	}
185 
186 	if (ring->funcs->insert_start)
187 		ring->funcs->insert_start(ring);
188 
189 	if (job) {
190 		r = amdgpu_vm_flush(ring, job, need_pipe_sync);
191 		if (r) {
192 			amdgpu_ring_undo(ring);
193 			return r;
194 		}
195 	}
196 
197 	if (job && ring->funcs->init_cond_exec)
198 		patch_offset = amdgpu_ring_init_cond_exec(ring);
199 
200 #ifdef CONFIG_X86_64
201 	if (!(adev->flags & AMD_IS_APU))
202 #endif
203 	{
204 		if (ring->funcs->emit_hdp_flush)
205 			amdgpu_ring_emit_hdp_flush(ring);
206 		else
207 			amdgpu_asic_flush_hdp(adev, ring);
208 	}
209 
210 	if (need_ctx_switch)
211 		status |= AMDGPU_HAVE_CTX_SWITCH;
212 
213 	skip_preamble = ring->current_ctx == fence_ctx;
214 	if (job && ring->funcs->emit_cntxcntl) {
215 		status |= job->preamble_status;
216 		status |= job->preemption_status;
217 		amdgpu_ring_emit_cntxcntl(ring, status);
218 	}
219 
220 	for (i = 0; i < num_ibs; ++i) {
221 		ib = &ibs[i];
222 
223 		/* drop preamble IBs if we don't have a context switch */
224 		if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
225 		    skip_preamble &&
226 		    !(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
227 		    !amdgpu_mcbp &&
228 		    !amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
229 			continue;
230 
231 		amdgpu_ring_emit_ib(ring, job, ib, status);
232 		status &= ~AMDGPU_HAVE_CTX_SWITCH;
233 	}
234 
235 	if (ring->funcs->emit_tmz)
236 		amdgpu_ring_emit_tmz(ring, false);
237 
238 #ifdef CONFIG_X86_64
239 	if (!(adev->flags & AMD_IS_APU))
240 #endif
241 		amdgpu_asic_invalidate_hdp(adev, ring);
242 
243 	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
244 		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
245 
246 	/* wrap the last IB with fence */
247 	if (job && job->uf_addr) {
248 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
249 				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
250 	}
251 
252 	r = amdgpu_fence_emit(ring, f, fence_flags);
253 	if (r) {
254 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
255 		if (job && job->vmid)
256 			amdgpu_vmid_reset(adev, ring->funcs->vmhub, job->vmid);
257 		amdgpu_ring_undo(ring);
258 		return r;
259 	}
260 
261 	if (ring->funcs->insert_end)
262 		ring->funcs->insert_end(ring);
263 
264 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
265 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
266 
267 	ring->current_ctx = fence_ctx;
268 	if (vm && ring->funcs->emit_switch_buffer)
269 		amdgpu_ring_emit_switch_buffer(ring);
270 	amdgpu_ring_commit(ring);
271 	return 0;
272 }
273 
274 /**
275  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
276  *
277  * @adev: amdgpu_device pointer
278  *
279  * Initialize the suballocator to manage a pool of memory
280  * for use as IBs (all asics).
281  * Returns 0 on success, error on failure.
282  */
283 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
284 {
285 	int r, i;
286 	unsigned size;
287 
288 	if (adev->ib_pool_ready) {
289 		return 0;
290 	}
291 	for (i = 0; i < AMDGPU_IB_POOL_MAX; i++) {
292 		if (i == AMDGPU_IB_POOL_DIRECT)
293 			size = PAGE_SIZE * 2;
294 		else
295 			size = AMDGPU_IB_POOL_SIZE*64*1024;
296 		r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo[i],
297 				size,
298 				AMDGPU_GPU_PAGE_SIZE,
299 				AMDGPU_GEM_DOMAIN_GTT);
300 		if (r) {
301 			for (i--; i >= 0; i--)
302 				amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo[i]);
303 			return r;
304 		}
305 	}
306 	adev->ib_pool_ready = true;
307 
308 	return 0;
309 }
310 
311 /**
312  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
313  *
314  * @adev: amdgpu_device pointer
315  *
316  * Tear down the suballocator managing the pool of memory
317  * for use as IBs (all asics).
318  */
319 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
320 {
321 	int i;
322 
323 	if (adev->ib_pool_ready) {
324 		for (i = 0; i < AMDGPU_IB_POOL_MAX; i++)
325 			amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo[i]);
326 		adev->ib_pool_ready = false;
327 	}
328 }
329 
330 /**
331  * amdgpu_ib_ring_tests - test IBs on the rings
332  *
333  * @adev: amdgpu_device pointer
334  *
335  * Test an IB (Indirect Buffer) on each ring.
336  * If the test fails, disable the ring.
337  * Returns 0 on success, error if the primary GFX ring
338  * IB test fails.
339  */
340 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
341 {
342 	unsigned i;
343 	int r, ret = 0;
344 	long tmo_gfx, tmo_mm;
345 
346 	tmo_mm = tmo_gfx = AMDGPU_IB_TEST_TIMEOUT;
347 	if (amdgpu_sriov_vf(adev)) {
348 		/* for MM engines in hypervisor side they are not scheduled together
349 		 * with CP and SDMA engines, so even in exclusive mode MM engine could
350 		 * still running on other VF thus the IB TEST TIMEOUT for MM engines
351 		 * under SR-IOV should be set to a long time. 8 sec should be enough
352 		 * for the MM comes back to this VF.
353 		 */
354 		tmo_mm = 8 * AMDGPU_IB_TEST_TIMEOUT;
355 	}
356 
357 	if (amdgpu_sriov_runtime(adev)) {
358 		/* for CP & SDMA engines since they are scheduled together so
359 		 * need to make the timeout width enough to cover the time
360 		 * cost waiting for it coming back under RUNTIME only
361 		*/
362 		tmo_gfx = 8 * AMDGPU_IB_TEST_TIMEOUT;
363 	} else if (adev->gmc.xgmi.hive_id) {
364 		tmo_gfx = AMDGPU_IB_TEST_GFX_XGMI_TIMEOUT;
365 	}
366 
367 	for (i = 0; i < adev->num_rings; ++i) {
368 		struct amdgpu_ring *ring = adev->rings[i];
369 		long tmo;
370 
371 		/* KIQ rings don't have an IB test because we never submit IBs
372 		 * to them and they have no interrupt support.
373 		 */
374 		if (!ring->sched.ready || !ring->funcs->test_ib)
375 			continue;
376 
377 		/* MM engine need more time */
378 		if (ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
379 			ring->funcs->type == AMDGPU_RING_TYPE_VCE ||
380 			ring->funcs->type == AMDGPU_RING_TYPE_UVD_ENC ||
381 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_DEC ||
382 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
383 			ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG)
384 			tmo = tmo_mm;
385 		else
386 			tmo = tmo_gfx;
387 
388 		r = amdgpu_ring_test_ib(ring, tmo);
389 		if (!r) {
390 			DRM_DEV_DEBUG(adev->dev, "ib test on %s succeeded\n",
391 				      ring->name);
392 			continue;
393 		}
394 
395 		ring->sched.ready = false;
396 		DRM_DEV_ERROR(adev->dev, "IB test failed on %s (%d).\n",
397 			  ring->name, r);
398 
399 		if (ring == &adev->gfx.gfx_ring[0]) {
400 			/* oh, oh, that's really bad */
401 			adev->accel_working = false;
402 			return r;
403 
404 		} else {
405 			ret = r;
406 		}
407 	}
408 	return ret;
409 }
410 
411 /*
412  * Debugfs info
413  */
414 #if defined(CONFIG_DEBUG_FS)
415 
416 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
417 {
418 	struct drm_info_node *node = (struct drm_info_node *) m->private;
419 	struct drm_device *dev = node->minor->dev;
420 	struct amdgpu_device *adev = dev->dev_private;
421 
422 	seq_printf(m, "-------------------- NORMAL -------------------- \n");
423 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo[AMDGPU_IB_POOL_NORMAL], m);
424 	seq_printf(m, "---------------------- VM ---------------------- \n");
425 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo[AMDGPU_IB_POOL_VM], m);
426 	seq_printf(m, "-------------------- DIRECT--------------------- \n");
427 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo[AMDGPU_IB_POOL_DIRECT], m);
428 
429 	return 0;
430 
431 }
432 
433 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
434 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
435 };
436 
437 #endif
438 
439 int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
440 {
441 #if defined(CONFIG_DEBUG_FS)
442 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
443 #else
444 	return 0;
445 #endif
446 }
447