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 
36 #define AMDGPU_IB_TEST_TIMEOUT	msecs_to_jiffies(1000)
37 
38 /*
39  * IB
40  * IBs (Indirect Buffers) and areas of GPU accessible memory where
41  * commands are stored.  You can put a pointer to the IB in the
42  * command ring and the hw will fetch the commands from the IB
43  * and execute them.  Generally userspace acceleration drivers
44  * produce command buffers which are send to the kernel and
45  * put in IBs for execution by the requested ring.
46  */
47 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev);
48 
49 /**
50  * amdgpu_ib_get - request an IB (Indirect Buffer)
51  *
52  * @ring: ring index the IB is associated with
53  * @size: requested IB size
54  * @ib: IB object returned
55  *
56  * Request an IB (all asics).  IBs are allocated using the
57  * suballocator.
58  * Returns 0 on success, error on failure.
59  */
60 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
61 		  unsigned size, struct amdgpu_ib *ib)
62 {
63 	int r;
64 
65 	if (size) {
66 		r = amdgpu_sa_bo_new(&adev->ring_tmp_bo,
67 				      &ib->sa_bo, size, 256);
68 		if (r) {
69 			dev_err(adev->dev, "failed to get a new IB (%d)\n", r);
70 			return r;
71 		}
72 
73 		ib->ptr = amdgpu_sa_bo_cpu_addr(ib->sa_bo);
74 
75 		if (!vm)
76 			ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
77 	}
78 
79 	return 0;
80 }
81 
82 /**
83  * amdgpu_ib_free - free an IB (Indirect Buffer)
84  *
85  * @adev: amdgpu_device pointer
86  * @ib: IB object to free
87  * @f: the fence SA bo need wait on for the ib alloation
88  *
89  * Free an IB (all asics).
90  */
91 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
92 		    struct dma_fence *f)
93 {
94 	amdgpu_sa_bo_free(adev, &ib->sa_bo, f);
95 }
96 
97 /**
98  * amdgpu_ib_schedule - schedule an IB (Indirect Buffer) on the ring
99  *
100  * @adev: amdgpu_device pointer
101  * @num_ibs: number of IBs to schedule
102  * @ibs: IB objects to schedule
103  * @f: fence created during this submission
104  *
105  * Schedule an IB on the associated ring (all asics).
106  * Returns 0 on success, error on failure.
107  *
108  * On SI, there are two parallel engines fed from the primary ring,
109  * the CE (Constant Engine) and the DE (Drawing Engine).  Since
110  * resource descriptors have moved to memory, the CE allows you to
111  * prime the caches while the DE is updating register state so that
112  * the resource descriptors will be already in cache when the draw is
113  * processed.  To accomplish this, the userspace driver submits two
114  * IBs, one for the CE and one for the DE.  If there is a CE IB (called
115  * a CONST_IB), it will be put on the ring prior to the DE IB.  Prior
116  * to SI there was just a DE IB.
117  */
118 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
119 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
120 		       struct dma_fence **f)
121 {
122 	struct amdgpu_device *adev = ring->adev;
123 	struct amdgpu_ib *ib = &ibs[0];
124 	bool skip_preamble, need_ctx_switch;
125 	unsigned patch_offset = ~0;
126 	struct amdgpu_vm *vm;
127 	uint64_t fence_ctx;
128 	uint32_t status = 0, alloc_size;
129 
130 	unsigned i;
131 	int r = 0;
132 
133 	if (num_ibs == 0)
134 		return -EINVAL;
135 
136 	/* ring tests don't use a job */
137 	if (job) {
138 		vm = job->vm;
139 		fence_ctx = job->fence_ctx;
140 	} else {
141 		vm = NULL;
142 		fence_ctx = 0;
143 	}
144 
145 	if (!ring->ready) {
146 		dev_err(adev->dev, "couldn't schedule ib on ring <%s>\n", ring->name);
147 		return -EINVAL;
148 	}
149 
150 	if (vm && !job->vm_id) {
151 		dev_err(adev->dev, "VM IB without ID\n");
152 		return -EINVAL;
153 	}
154 
155 	alloc_size = ring->funcs->emit_frame_size + num_ibs *
156 		ring->funcs->emit_ib_size;
157 
158 	r = amdgpu_ring_alloc(ring, alloc_size);
159 	if (r) {
160 		dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
161 		return r;
162 	}
163 
164 	if (ring->funcs->init_cond_exec)
165 		patch_offset = amdgpu_ring_init_cond_exec(ring);
166 
167 	if (vm) {
168 		r = amdgpu_vm_flush(ring, job);
169 		if (r) {
170 			amdgpu_ring_undo(ring);
171 			return r;
172 		}
173 	}
174 
175 	if (ring->funcs->emit_hdp_flush)
176 		amdgpu_ring_emit_hdp_flush(ring);
177 
178 	skip_preamble = ring->current_ctx == fence_ctx;
179 	need_ctx_switch = ring->current_ctx != fence_ctx;
180 	if (job && ring->funcs->emit_cntxcntl) {
181 		if (need_ctx_switch)
182 			status |= AMDGPU_HAVE_CTX_SWITCH;
183 		status |= job->preamble_status;
184 
185 		if (vm)
186 			status |= AMDGPU_VM_DOMAIN;
187 		amdgpu_ring_emit_cntxcntl(ring, status);
188 	}
189 
190 	for (i = 0; i < num_ibs; ++i) {
191 		ib = &ibs[i];
192 
193 		/* drop preamble IBs if we don't have a context switch */
194 		if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) &&
195 			skip_preamble &&
196 			!(status & AMDGPU_PREAMBLE_IB_PRESENT_FIRST) &&
197 			!amdgpu_sriov_vf(adev)) /* for SRIOV preemption, Preamble CE ib must be inserted anyway */
198 			continue;
199 
200 		amdgpu_ring_emit_ib(ring, ib, job ? job->vm_id : 0,
201 				    need_ctx_switch);
202 		need_ctx_switch = false;
203 	}
204 
205 	if (ring->funcs->emit_hdp_invalidate)
206 		amdgpu_ring_emit_hdp_invalidate(ring);
207 
208 	r = amdgpu_fence_emit(ring, f);
209 	if (r) {
210 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
211 		if (job && job->vm_id)
212 			amdgpu_vm_reset_id(adev, job->vm_id);
213 		amdgpu_ring_undo(ring);
214 		return r;
215 	}
216 
217 	/* wrap the last IB with fence */
218 	if (job && job->uf_addr) {
219 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
220 				       AMDGPU_FENCE_FLAG_64BIT);
221 	}
222 
223 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)
224 		amdgpu_ring_patch_cond_exec(ring, patch_offset);
225 
226 	ring->current_ctx = fence_ctx;
227 	if (vm && ring->funcs->emit_switch_buffer)
228 		amdgpu_ring_emit_switch_buffer(ring);
229 	amdgpu_ring_commit(ring);
230 	return 0;
231 }
232 
233 /**
234  * amdgpu_ib_pool_init - Init the IB (Indirect Buffer) pool
235  *
236  * @adev: amdgpu_device pointer
237  *
238  * Initialize the suballocator to manage a pool of memory
239  * for use as IBs (all asics).
240  * Returns 0 on success, error on failure.
241  */
242 int amdgpu_ib_pool_init(struct amdgpu_device *adev)
243 {
244 	int r;
245 
246 	if (adev->ib_pool_ready) {
247 		return 0;
248 	}
249 	r = amdgpu_sa_bo_manager_init(adev, &adev->ring_tmp_bo,
250 				      AMDGPU_IB_POOL_SIZE*64*1024,
251 				      AMDGPU_GPU_PAGE_SIZE,
252 				      AMDGPU_GEM_DOMAIN_GTT);
253 	if (r) {
254 		return r;
255 	}
256 
257 	r = amdgpu_sa_bo_manager_start(adev, &adev->ring_tmp_bo);
258 	if (r) {
259 		return r;
260 	}
261 
262 	adev->ib_pool_ready = true;
263 	if (amdgpu_debugfs_sa_init(adev)) {
264 		dev_err(adev->dev, "failed to register debugfs file for SA\n");
265 	}
266 	return 0;
267 }
268 
269 /**
270  * amdgpu_ib_pool_fini - Free the IB (Indirect Buffer) pool
271  *
272  * @adev: amdgpu_device pointer
273  *
274  * Tear down the suballocator managing the pool of memory
275  * for use as IBs (all asics).
276  */
277 void amdgpu_ib_pool_fini(struct amdgpu_device *adev)
278 {
279 	if (adev->ib_pool_ready) {
280 		amdgpu_sa_bo_manager_suspend(adev, &adev->ring_tmp_bo);
281 		amdgpu_sa_bo_manager_fini(adev, &adev->ring_tmp_bo);
282 		adev->ib_pool_ready = false;
283 	}
284 }
285 
286 /**
287  * amdgpu_ib_ring_tests - test IBs on the rings
288  *
289  * @adev: amdgpu_device pointer
290  *
291  * Test an IB (Indirect Buffer) on each ring.
292  * If the test fails, disable the ring.
293  * Returns 0 on success, error if the primary GFX ring
294  * IB test fails.
295  */
296 int amdgpu_ib_ring_tests(struct amdgpu_device *adev)
297 {
298 	unsigned i;
299 	int r, ret = 0;
300 
301 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
302 		struct amdgpu_ring *ring = adev->rings[i];
303 
304 		if (!ring || !ring->ready)
305 			continue;
306 
307 		r = amdgpu_ring_test_ib(ring, AMDGPU_IB_TEST_TIMEOUT);
308 		if (r) {
309 			ring->ready = false;
310 
311 			if (ring == &adev->gfx.gfx_ring[0]) {
312 				/* oh, oh, that's really bad */
313 				DRM_ERROR("amdgpu: failed testing IB on GFX ring (%d).\n", r);
314 				adev->accel_working = false;
315 				return r;
316 
317 			} else {
318 				/* still not good, but we can live with it */
319 				DRM_ERROR("amdgpu: failed testing IB on ring %d (%d).\n", i, r);
320 				ret = r;
321 			}
322 		}
323 	}
324 	return ret;
325 }
326 
327 /*
328  * Debugfs info
329  */
330 #if defined(CONFIG_DEBUG_FS)
331 
332 static int amdgpu_debugfs_sa_info(struct seq_file *m, void *data)
333 {
334 	struct drm_info_node *node = (struct drm_info_node *) m->private;
335 	struct drm_device *dev = node->minor->dev;
336 	struct amdgpu_device *adev = dev->dev_private;
337 
338 	amdgpu_sa_bo_dump_debug_info(&adev->ring_tmp_bo, m);
339 
340 	return 0;
341 
342 }
343 
344 static const struct drm_info_list amdgpu_debugfs_sa_list[] = {
345 	{"amdgpu_sa_info", &amdgpu_debugfs_sa_info, 0, NULL},
346 };
347 
348 #endif
349 
350 static int amdgpu_debugfs_sa_init(struct amdgpu_device *adev)
351 {
352 #if defined(CONFIG_DEBUG_FS)
353 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_sa_list, 1);
354 #else
355 	return 0;
356 #endif
357 }
358