1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2019 Intel Corporation
5  */
6 
7 #include "gem/i915_gem_context.h"
8 #include "gem/i915_gem_pm.h"
9 
10 #include "i915_drv.h"
11 #include "i915_globals.h"
12 
13 #include "intel_context.h"
14 #include "intel_engine.h"
15 #include "intel_engine_pm.h"
16 
17 static struct i915_global_context {
18 	struct i915_global base;
19 	struct kmem_cache *slab_ce;
20 } global;
21 
22 static struct intel_context *intel_context_alloc(void)
23 {
24 	return kmem_cache_zalloc(global.slab_ce, GFP_KERNEL);
25 }
26 
27 void intel_context_free(struct intel_context *ce)
28 {
29 	kmem_cache_free(global.slab_ce, ce);
30 }
31 
32 struct intel_context *
33 intel_context_create(struct i915_gem_context *ctx,
34 		     struct intel_engine_cs *engine)
35 {
36 	struct intel_context *ce;
37 
38 	ce = intel_context_alloc();
39 	if (!ce)
40 		return ERR_PTR(-ENOMEM);
41 
42 	intel_context_init(ce, ctx, engine);
43 	return ce;
44 }
45 
46 int __intel_context_do_pin(struct intel_context *ce)
47 {
48 	int err;
49 
50 	if (mutex_lock_interruptible(&ce->pin_mutex))
51 		return -EINTR;
52 
53 	if (likely(!atomic_read(&ce->pin_count))) {
54 		intel_wakeref_t wakeref;
55 
56 		if (unlikely(!test_bit(CONTEXT_ALLOC_BIT, &ce->flags))) {
57 			err = ce->ops->alloc(ce);
58 			if (unlikely(err))
59 				goto err;
60 
61 			__set_bit(CONTEXT_ALLOC_BIT, &ce->flags);
62 		}
63 
64 		err = 0;
65 		with_intel_runtime_pm(&ce->engine->i915->runtime_pm, wakeref)
66 			err = ce->ops->pin(ce);
67 		if (err)
68 			goto err;
69 
70 		GEM_TRACE("%s context:%llx pin ring:{head:%04x, tail:%04x}\n",
71 			  ce->engine->name, ce->timeline->fence_context,
72 			  ce->ring->head, ce->ring->tail);
73 
74 		i915_gem_context_get(ce->gem_context); /* for ctx->ppgtt */
75 
76 		smp_mb__before_atomic(); /* flush pin before it is visible */
77 	}
78 
79 	atomic_inc(&ce->pin_count);
80 	GEM_BUG_ON(!intel_context_is_pinned(ce)); /* no overflow! */
81 
82 	mutex_unlock(&ce->pin_mutex);
83 	return 0;
84 
85 err:
86 	mutex_unlock(&ce->pin_mutex);
87 	return err;
88 }
89 
90 void intel_context_unpin(struct intel_context *ce)
91 {
92 	if (likely(atomic_add_unless(&ce->pin_count, -1, 1)))
93 		return;
94 
95 	/* We may be called from inside intel_context_pin() to evict another */
96 	intel_context_get(ce);
97 	mutex_lock_nested(&ce->pin_mutex, SINGLE_DEPTH_NESTING);
98 
99 	if (likely(atomic_dec_and_test(&ce->pin_count))) {
100 		GEM_TRACE("%s context:%llx retire\n",
101 			  ce->engine->name, ce->timeline->fence_context);
102 
103 		ce->ops->unpin(ce);
104 
105 		i915_gem_context_put(ce->gem_context);
106 		intel_context_active_release(ce);
107 	}
108 
109 	mutex_unlock(&ce->pin_mutex);
110 	intel_context_put(ce);
111 }
112 
113 static int __context_pin_state(struct i915_vma *vma)
114 {
115 	u64 flags;
116 	int err;
117 
118 	flags = i915_ggtt_pin_bias(vma) | PIN_OFFSET_BIAS;
119 	flags |= PIN_HIGH | PIN_GLOBAL;
120 
121 	err = i915_vma_pin(vma, 0, 0, flags);
122 	if (err)
123 		return err;
124 
125 	/*
126 	 * And mark it as a globally pinned object to let the shrinker know
127 	 * it cannot reclaim the object until we release it.
128 	 */
129 	i915_vma_make_unshrinkable(vma);
130 	vma->obj->mm.dirty = true;
131 
132 	return 0;
133 }
134 
135 static void __context_unpin_state(struct i915_vma *vma)
136 {
137 	i915_vma_make_shrinkable(vma);
138 	__i915_vma_unpin(vma);
139 }
140 
141 __i915_active_call
142 static void __intel_context_retire(struct i915_active *active)
143 {
144 	struct intel_context *ce = container_of(active, typeof(*ce), active);
145 
146 	GEM_TRACE("%s context:%llx retire\n",
147 		  ce->engine->name, ce->timeline->fence_context);
148 
149 	if (ce->state)
150 		__context_unpin_state(ce->state);
151 
152 	intel_timeline_unpin(ce->timeline);
153 	intel_ring_unpin(ce->ring);
154 
155 	intel_context_put(ce);
156 }
157 
158 static int __intel_context_active(struct i915_active *active)
159 {
160 	struct intel_context *ce = container_of(active, typeof(*ce), active);
161 	int err;
162 
163 	intel_context_get(ce);
164 
165 	err = intel_ring_pin(ce->ring);
166 	if (err)
167 		goto err_put;
168 
169 	err = intel_timeline_pin(ce->timeline);
170 	if (err)
171 		goto err_ring;
172 
173 	if (!ce->state)
174 		return 0;
175 
176 	err = __context_pin_state(ce->state);
177 	if (err)
178 		goto err_timeline;
179 
180 	return 0;
181 
182 err_timeline:
183 	intel_timeline_unpin(ce->timeline);
184 err_ring:
185 	intel_ring_unpin(ce->ring);
186 err_put:
187 	intel_context_put(ce);
188 	return err;
189 }
190 
191 int intel_context_active_acquire(struct intel_context *ce)
192 {
193 	int err;
194 
195 	err = i915_active_acquire(&ce->active);
196 	if (err)
197 		return err;
198 
199 	/* Preallocate tracking nodes */
200 	if (!i915_gem_context_is_kernel(ce->gem_context)) {
201 		err = i915_active_acquire_preallocate_barrier(&ce->active,
202 							      ce->engine);
203 		if (err) {
204 			i915_active_release(&ce->active);
205 			return err;
206 		}
207 	}
208 
209 	return 0;
210 }
211 
212 void intel_context_active_release(struct intel_context *ce)
213 {
214 	/* Nodes preallocated in intel_context_active() */
215 	i915_active_acquire_barrier(&ce->active);
216 	i915_active_release(&ce->active);
217 }
218 
219 void
220 intel_context_init(struct intel_context *ce,
221 		   struct i915_gem_context *ctx,
222 		   struct intel_engine_cs *engine)
223 {
224 	struct i915_address_space *vm;
225 
226 	GEM_BUG_ON(!engine->cops);
227 
228 	kref_init(&ce->ref);
229 
230 	ce->gem_context = ctx;
231 	rcu_read_lock();
232 	vm = rcu_dereference(ctx->vm);
233 	if (vm)
234 		ce->vm = i915_vm_get(vm);
235 	else
236 		ce->vm = i915_vm_get(&engine->gt->ggtt->vm);
237 	rcu_read_unlock();
238 	if (ctx->timeline)
239 		ce->timeline = intel_timeline_get(ctx->timeline);
240 
241 	ce->engine = engine;
242 	ce->ops = engine->cops;
243 	ce->sseu = engine->sseu;
244 	ce->ring = __intel_context_ring_size(SZ_16K);
245 
246 	INIT_LIST_HEAD(&ce->signal_link);
247 	INIT_LIST_HEAD(&ce->signals);
248 
249 	mutex_init(&ce->pin_mutex);
250 
251 	i915_active_init(&ce->active,
252 			 __intel_context_active, __intel_context_retire);
253 }
254 
255 void intel_context_fini(struct intel_context *ce)
256 {
257 	if (ce->timeline)
258 		intel_timeline_put(ce->timeline);
259 	i915_vm_put(ce->vm);
260 
261 	mutex_destroy(&ce->pin_mutex);
262 	i915_active_fini(&ce->active);
263 }
264 
265 static void i915_global_context_shrink(void)
266 {
267 	kmem_cache_shrink(global.slab_ce);
268 }
269 
270 static void i915_global_context_exit(void)
271 {
272 	kmem_cache_destroy(global.slab_ce);
273 }
274 
275 static struct i915_global_context global = { {
276 	.shrink = i915_global_context_shrink,
277 	.exit = i915_global_context_exit,
278 } };
279 
280 int __init i915_global_context_init(void)
281 {
282 	global.slab_ce = KMEM_CACHE(intel_context, SLAB_HWCACHE_ALIGN);
283 	if (!global.slab_ce)
284 		return -ENOMEM;
285 
286 	i915_global_register(&global.base);
287 	return 0;
288 }
289 
290 void intel_context_enter_engine(struct intel_context *ce)
291 {
292 	intel_engine_pm_get(ce->engine);
293 	intel_timeline_enter(ce->timeline);
294 }
295 
296 void intel_context_exit_engine(struct intel_context *ce)
297 {
298 	intel_timeline_exit(ce->timeline);
299 	intel_engine_pm_put(ce->engine);
300 }
301 
302 int intel_context_prepare_remote_request(struct intel_context *ce,
303 					 struct i915_request *rq)
304 {
305 	struct intel_timeline *tl = ce->timeline;
306 	int err;
307 
308 	/* Only suitable for use in remotely modifying this context */
309 	GEM_BUG_ON(rq->hw_context == ce);
310 
311 	if (rcu_access_pointer(rq->timeline) != tl) { /* timeline sharing! */
312 		err = mutex_lock_interruptible_nested(&tl->mutex,
313 						      SINGLE_DEPTH_NESTING);
314 		if (err)
315 			return err;
316 
317 		/* Queue this switch after current activity by this context. */
318 		err = i915_active_fence_set(&tl->last_request, rq);
319 		mutex_unlock(&tl->mutex);
320 		if (err)
321 			return err;
322 	}
323 
324 	/*
325 	 * Guarantee context image and the timeline remains pinned until the
326 	 * modifying request is retired by setting the ce activity tracker.
327 	 *
328 	 * But we only need to take one pin on the account of it. Or in other
329 	 * words transfer the pinned ce object to tracked active request.
330 	 */
331 	GEM_BUG_ON(i915_active_is_idle(&ce->active));
332 	return i915_active_add_request(&ce->active, rq);
333 }
334 
335 struct i915_request *intel_context_create_request(struct intel_context *ce)
336 {
337 	struct i915_request *rq;
338 	int err;
339 
340 	err = intel_context_pin(ce);
341 	if (unlikely(err))
342 		return ERR_PTR(err);
343 
344 	rq = i915_request_create(ce);
345 	intel_context_unpin(ce);
346 
347 	return rq;
348 }
349 
350 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
351 #include "selftest_context.c"
352 #endif
353