1 /*
2  * Copyright 2015 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * Authors: monk liu <monk.liu@amd.com>
23  */
24 
25 #include <drm/drmP.h>
26 #include "amdgpu.h"
27 
28 int amdgpu_ctx_init(struct amdgpu_device *adev, bool kernel,
29 		    struct amdgpu_ctx *ctx)
30 {
31 	unsigned i, j;
32 	int r;
33 
34 	memset(ctx, 0, sizeof(*ctx));
35 	ctx->adev = adev;
36 	kref_init(&ctx->refcount);
37 	spin_lock_init(&ctx->ring_lock);
38 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
39 		ctx->rings[i].sequence = 1;
40 
41 	if (amdgpu_enable_scheduler) {
42 		/* create context entity for each ring */
43 		for (i = 0; i < adev->num_rings; i++) {
44 			struct amd_sched_rq *rq;
45 			if (kernel)
46 				rq = &adev->rings[i]->sched.kernel_rq;
47 			else
48 				rq = &adev->rings[i]->sched.sched_rq;
49 			r = amd_sched_entity_init(&adev->rings[i]->sched,
50 						  &ctx->rings[i].entity,
51 						  rq, amdgpu_sched_jobs);
52 			if (r)
53 				break;
54 		}
55 
56 		if (i < adev->num_rings) {
57 			for (j = 0; j < i; j++)
58 				amd_sched_entity_fini(&adev->rings[j]->sched,
59 						      &ctx->rings[j].entity);
60 			kfree(ctx);
61 			return r;
62 		}
63 	}
64 	return 0;
65 }
66 
67 void amdgpu_ctx_fini(struct amdgpu_ctx *ctx)
68 {
69 	struct amdgpu_device *adev = ctx->adev;
70 	unsigned i, j;
71 
72 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
73 		for (j = 0; j < AMDGPU_CTX_MAX_CS_PENDING; ++j)
74 			fence_put(ctx->rings[i].fences[j]);
75 
76 	if (amdgpu_enable_scheduler) {
77 		for (i = 0; i < adev->num_rings; i++)
78 			amd_sched_entity_fini(&adev->rings[i]->sched,
79 					      &ctx->rings[i].entity);
80 	}
81 }
82 
83 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
84 			    struct amdgpu_fpriv *fpriv,
85 			    uint32_t *id)
86 {
87 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
88 	struct amdgpu_ctx *ctx;
89 	int r;
90 
91 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
92 	if (!ctx)
93 		return -ENOMEM;
94 
95 	mutex_lock(&mgr->lock);
96 	r = idr_alloc(&mgr->ctx_handles, ctx, 1, 0, GFP_KERNEL);
97 	if (r < 0) {
98 		mutex_unlock(&mgr->lock);
99 		kfree(ctx);
100 		return r;
101 	}
102 	*id = (uint32_t)r;
103 	r = amdgpu_ctx_init(adev, false, ctx);
104 	mutex_unlock(&mgr->lock);
105 
106 	return r;
107 }
108 
109 static void amdgpu_ctx_do_release(struct kref *ref)
110 {
111 	struct amdgpu_ctx *ctx;
112 
113 	ctx = container_of(ref, struct amdgpu_ctx, refcount);
114 
115 	amdgpu_ctx_fini(ctx);
116 
117 	kfree(ctx);
118 }
119 
120 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
121 {
122 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
123 	struct amdgpu_ctx *ctx;
124 
125 	mutex_lock(&mgr->lock);
126 	ctx = idr_find(&mgr->ctx_handles, id);
127 	if (ctx) {
128 		idr_remove(&mgr->ctx_handles, id);
129 		kref_put(&ctx->refcount, amdgpu_ctx_do_release);
130 		mutex_unlock(&mgr->lock);
131 		return 0;
132 	}
133 	mutex_unlock(&mgr->lock);
134 	return -EINVAL;
135 }
136 
137 static int amdgpu_ctx_query(struct amdgpu_device *adev,
138 			    struct amdgpu_fpriv *fpriv, uint32_t id,
139 			    union drm_amdgpu_ctx_out *out)
140 {
141 	struct amdgpu_ctx *ctx;
142 	struct amdgpu_ctx_mgr *mgr;
143 	unsigned reset_counter;
144 
145 	if (!fpriv)
146 		return -EINVAL;
147 
148 	mgr = &fpriv->ctx_mgr;
149 	mutex_lock(&mgr->lock);
150 	ctx = idr_find(&mgr->ctx_handles, id);
151 	if (!ctx) {
152 		mutex_unlock(&mgr->lock);
153 		return -EINVAL;
154 	}
155 
156 	/* TODO: these two are always zero */
157 	out->state.flags = 0x0;
158 	out->state.hangs = 0x0;
159 
160 	/* determine if a GPU reset has occured since the last call */
161 	reset_counter = atomic_read(&adev->gpu_reset_counter);
162 	/* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
163 	if (ctx->reset_counter == reset_counter)
164 		out->state.reset_status = AMDGPU_CTX_NO_RESET;
165 	else
166 		out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
167 	ctx->reset_counter = reset_counter;
168 
169 	mutex_unlock(&mgr->lock);
170 	return 0;
171 }
172 
173 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
174 		     struct drm_file *filp)
175 {
176 	int r;
177 	uint32_t id;
178 
179 	union drm_amdgpu_ctx *args = data;
180 	struct amdgpu_device *adev = dev->dev_private;
181 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
182 
183 	r = 0;
184 	id = args->in.ctx_id;
185 
186 	switch (args->in.op) {
187 		case AMDGPU_CTX_OP_ALLOC_CTX:
188 			r = amdgpu_ctx_alloc(adev, fpriv, &id);
189 			args->out.alloc.ctx_id = id;
190 			break;
191 		case AMDGPU_CTX_OP_FREE_CTX:
192 			r = amdgpu_ctx_free(fpriv, id);
193 			break;
194 		case AMDGPU_CTX_OP_QUERY_STATE:
195 			r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
196 			break;
197 		default:
198 			return -EINVAL;
199 	}
200 
201 	return r;
202 }
203 
204 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
205 {
206 	struct amdgpu_ctx *ctx;
207 	struct amdgpu_ctx_mgr *mgr;
208 
209 	if (!fpriv)
210 		return NULL;
211 
212 	mgr = &fpriv->ctx_mgr;
213 
214 	mutex_lock(&mgr->lock);
215 	ctx = idr_find(&mgr->ctx_handles, id);
216 	if (ctx)
217 		kref_get(&ctx->refcount);
218 	mutex_unlock(&mgr->lock);
219 	return ctx;
220 }
221 
222 int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
223 {
224 	if (ctx == NULL)
225 		return -EINVAL;
226 
227 	kref_put(&ctx->refcount, amdgpu_ctx_do_release);
228 	return 0;
229 }
230 
231 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
232 			      struct fence *fence)
233 {
234 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
235 	uint64_t seq = cring->sequence;
236 	unsigned idx = 0;
237 	struct fence *other = NULL;
238 
239 	idx = seq % AMDGPU_CTX_MAX_CS_PENDING;
240 	other = cring->fences[idx];
241 	if (other) {
242 		signed long r;
243 		r = fence_wait_timeout(other, false, MAX_SCHEDULE_TIMEOUT);
244 		if (r < 0)
245 			DRM_ERROR("Error (%ld) waiting for fence!\n", r);
246 	}
247 
248 	fence_get(fence);
249 
250 	spin_lock(&ctx->ring_lock);
251 	cring->fences[idx] = fence;
252 	cring->sequence++;
253 	spin_unlock(&ctx->ring_lock);
254 
255 	fence_put(other);
256 
257 	return seq;
258 }
259 
260 struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
261 				   struct amdgpu_ring *ring, uint64_t seq)
262 {
263 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
264 	struct fence *fence;
265 
266 	spin_lock(&ctx->ring_lock);
267 
268 	if (seq >= cring->sequence) {
269 		spin_unlock(&ctx->ring_lock);
270 		return ERR_PTR(-EINVAL);
271 	}
272 
273 
274 	if (seq + AMDGPU_CTX_MAX_CS_PENDING < cring->sequence) {
275 		spin_unlock(&ctx->ring_lock);
276 		return NULL;
277 	}
278 
279 	fence = fence_get(cring->fences[seq % AMDGPU_CTX_MAX_CS_PENDING]);
280 	spin_unlock(&ctx->ring_lock);
281 
282 	return fence;
283 }
284 
285 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
286 {
287 	mutex_init(&mgr->lock);
288 	idr_init(&mgr->ctx_handles);
289 }
290 
291 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
292 {
293 	struct amdgpu_ctx *ctx;
294 	struct idr *idp;
295 	uint32_t id;
296 
297 	idp = &mgr->ctx_handles;
298 
299 	idr_for_each_entry(idp, ctx, id) {
300 		if (kref_put(&ctx->refcount, amdgpu_ctx_do_release) != 1)
301 			DRM_ERROR("ctx %p is still alive\n", ctx);
302 	}
303 
304 	idr_destroy(&mgr->ctx_handles);
305 	mutex_destroy(&mgr->lock);
306 }
307