1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2021 Intel Corporation
4 */
5
6 #include <linux/interval_tree_generic.h>
7 #include <linux/sched/mm.h>
8
9 #include "i915_sw_fence.h"
10 #include "i915_vma_resource.h"
11 #include "i915_drv.h"
12 #include "intel_memory_region.h"
13
14 #include "gt/intel_gtt.h"
15
16 static struct kmem_cache *slab_vma_resources;
17
18 /**
19 * DOC:
20 * We use a per-vm interval tree to keep track of vma_resources
21 * scheduled for unbind but not yet unbound. The tree is protected by
22 * the vm mutex, and nodes are removed just after the unbind fence signals.
23 * The removal takes the vm mutex from a kernel thread which we need to
24 * keep in mind so that we don't grab the mutex and try to wait for all
25 * pending unbinds to complete, because that will temporaryily block many
26 * of the workqueue threads, and people will get angry.
27 *
28 * We should consider using a single ordered fence per VM instead but that
29 * requires ordering the unbinds and might introduce unnecessary waiting
30 * for unrelated unbinds. Amount of code will probably be roughly the same
31 * due to the simplicity of using the interval tree interface.
32 *
33 * Another drawback of this interval tree is that the complexity of insertion
34 * and removal of fences increases as O(ln(pending_unbinds)) instead of
35 * O(1) for a single fence without interval tree.
36 */
37 #define VMA_RES_START(_node) ((_node)->start)
38 #define VMA_RES_LAST(_node) ((_node)->start + (_node)->node_size - 1)
39 INTERVAL_TREE_DEFINE(struct i915_vma_resource, rb,
40 u64, __subtree_last,
41 VMA_RES_START, VMA_RES_LAST, static, vma_res_itree);
42
43 /* Callbacks for the unbind dma-fence. */
44
45 /**
46 * i915_vma_resource_alloc - Allocate a vma resource
47 *
48 * Return: A pointer to a cleared struct i915_vma_resource or
49 * a -ENOMEM error pointer if allocation fails.
50 */
i915_vma_resource_alloc(void)51 struct i915_vma_resource *i915_vma_resource_alloc(void)
52 {
53 struct i915_vma_resource *vma_res =
54 kmem_cache_zalloc(slab_vma_resources, GFP_KERNEL);
55
56 return vma_res ? vma_res : ERR_PTR(-ENOMEM);
57 }
58
59 /**
60 * i915_vma_resource_free - Free a vma resource
61 * @vma_res: The vma resource to free.
62 */
i915_vma_resource_free(struct i915_vma_resource * vma_res)63 void i915_vma_resource_free(struct i915_vma_resource *vma_res)
64 {
65 if (vma_res)
66 kmem_cache_free(slab_vma_resources, vma_res);
67 }
68
get_driver_name(struct dma_fence * fence)69 static const char *get_driver_name(struct dma_fence *fence)
70 {
71 return "vma unbind fence";
72 }
73
get_timeline_name(struct dma_fence * fence)74 static const char *get_timeline_name(struct dma_fence *fence)
75 {
76 return "unbound";
77 }
78
unbind_fence_free_rcu(struct rcu_head * head)79 static void unbind_fence_free_rcu(struct rcu_head *head)
80 {
81 struct i915_vma_resource *vma_res =
82 container_of(head, typeof(*vma_res), unbind_fence.rcu);
83
84 i915_vma_resource_free(vma_res);
85 }
86
unbind_fence_release(struct dma_fence * fence)87 static void unbind_fence_release(struct dma_fence *fence)
88 {
89 struct i915_vma_resource *vma_res =
90 container_of(fence, typeof(*vma_res), unbind_fence);
91
92 i915_sw_fence_fini(&vma_res->chain);
93
94 call_rcu(&fence->rcu, unbind_fence_free_rcu);
95 }
96
97 static struct dma_fence_ops unbind_fence_ops = {
98 .get_driver_name = get_driver_name,
99 .get_timeline_name = get_timeline_name,
100 .release = unbind_fence_release,
101 };
102
__i915_vma_resource_unhold(struct i915_vma_resource * vma_res)103 static void __i915_vma_resource_unhold(struct i915_vma_resource *vma_res)
104 {
105 struct i915_address_space *vm;
106
107 if (!refcount_dec_and_test(&vma_res->hold_count))
108 return;
109
110 dma_fence_signal(&vma_res->unbind_fence);
111
112 vm = vma_res->vm;
113 if (vma_res->wakeref)
114 intel_runtime_pm_put(&vm->i915->runtime_pm, vma_res->wakeref);
115
116 vma_res->vm = NULL;
117 if (!RB_EMPTY_NODE(&vma_res->rb)) {
118 mutex_lock(&vm->mutex);
119 vma_res_itree_remove(vma_res, &vm->pending_unbind);
120 mutex_unlock(&vm->mutex);
121 }
122
123 if (vma_res->bi.pages_rsgt)
124 i915_refct_sgt_put(vma_res->bi.pages_rsgt);
125 }
126
127 /**
128 * i915_vma_resource_unhold - Unhold the signaling of the vma resource unbind
129 * fence.
130 * @vma_res: The vma resource.
131 * @lockdep_cookie: The lockdep cookie returned from i915_vma_resource_hold.
132 *
133 * The function may leave a dma_fence critical section.
134 */
i915_vma_resource_unhold(struct i915_vma_resource * vma_res,bool lockdep_cookie)135 void i915_vma_resource_unhold(struct i915_vma_resource *vma_res,
136 bool lockdep_cookie)
137 {
138 dma_fence_end_signalling(lockdep_cookie);
139
140 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
141 unsigned long irq_flags;
142
143 /* Inefficient open-coded might_lock_irqsave() */
144 spin_lock_irqsave(&vma_res->lock, irq_flags);
145 spin_unlock_irqrestore(&vma_res->lock, irq_flags);
146 }
147
148 __i915_vma_resource_unhold(vma_res);
149 }
150
151 /**
152 * i915_vma_resource_hold - Hold the signaling of the vma resource unbind fence.
153 * @vma_res: The vma resource.
154 * @lockdep_cookie: Pointer to a bool serving as a lockdep cooke that should
155 * be given as an argument to the pairing i915_vma_resource_unhold.
156 *
157 * If returning true, the function enters a dma_fence signalling critical
158 * section if not in one already.
159 *
160 * Return: true if holding successful, false if not.
161 */
i915_vma_resource_hold(struct i915_vma_resource * vma_res,bool * lockdep_cookie)162 bool i915_vma_resource_hold(struct i915_vma_resource *vma_res,
163 bool *lockdep_cookie)
164 {
165 bool held = refcount_inc_not_zero(&vma_res->hold_count);
166
167 if (held)
168 *lockdep_cookie = dma_fence_begin_signalling();
169
170 return held;
171 }
172
i915_vma_resource_unbind_work(struct work_struct * work)173 static void i915_vma_resource_unbind_work(struct work_struct *work)
174 {
175 struct i915_vma_resource *vma_res =
176 container_of(work, typeof(*vma_res), work);
177 struct i915_address_space *vm = vma_res->vm;
178 bool lockdep_cookie;
179
180 lockdep_cookie = dma_fence_begin_signalling();
181 if (likely(!vma_res->skip_pte_rewrite))
182 vma_res->ops->unbind_vma(vm, vma_res);
183
184 dma_fence_end_signalling(lockdep_cookie);
185 __i915_vma_resource_unhold(vma_res);
186 i915_vma_resource_put(vma_res);
187 }
188
189 static int
i915_vma_resource_fence_notify(struct i915_sw_fence * fence,enum i915_sw_fence_notify state)190 i915_vma_resource_fence_notify(struct i915_sw_fence *fence,
191 enum i915_sw_fence_notify state)
192 {
193 struct i915_vma_resource *vma_res =
194 container_of(fence, typeof(*vma_res), chain);
195 struct dma_fence *unbind_fence =
196 &vma_res->unbind_fence;
197
198 switch (state) {
199 case FENCE_COMPLETE:
200 dma_fence_get(unbind_fence);
201 if (vma_res->immediate_unbind) {
202 i915_vma_resource_unbind_work(&vma_res->work);
203 } else {
204 INIT_WORK(&vma_res->work, i915_vma_resource_unbind_work);
205 queue_work(system_unbound_wq, &vma_res->work);
206 }
207 break;
208 case FENCE_FREE:
209 i915_vma_resource_put(vma_res);
210 break;
211 }
212
213 return NOTIFY_DONE;
214 }
215
216 /**
217 * i915_vma_resource_unbind - Unbind a vma resource
218 * @vma_res: The vma resource to unbind.
219 *
220 * At this point this function does little more than publish a fence that
221 * signals immediately unless signaling is held back.
222 *
223 * Return: A refcounted pointer to a dma-fence that signals when unbinding is
224 * complete.
225 */
i915_vma_resource_unbind(struct i915_vma_resource * vma_res,u32 * tlb)226 struct dma_fence *i915_vma_resource_unbind(struct i915_vma_resource *vma_res,
227 u32 *tlb)
228 {
229 struct i915_address_space *vm = vma_res->vm;
230
231 vma_res->tlb = tlb;
232
233 /* Reference for the sw fence */
234 i915_vma_resource_get(vma_res);
235
236 /* Caller must already have a wakeref in this case. */
237 if (vma_res->needs_wakeref)
238 vma_res->wakeref = intel_runtime_pm_get_if_in_use(&vm->i915->runtime_pm);
239
240 if (atomic_read(&vma_res->chain.pending) <= 1) {
241 RB_CLEAR_NODE(&vma_res->rb);
242 vma_res->immediate_unbind = 1;
243 } else {
244 vma_res_itree_insert(vma_res, &vma_res->vm->pending_unbind);
245 }
246
247 i915_sw_fence_commit(&vma_res->chain);
248
249 return &vma_res->unbind_fence;
250 }
251
252 /**
253 * __i915_vma_resource_init - Initialize a vma resource.
254 * @vma_res: The vma resource to initialize
255 *
256 * Initializes the private members of a vma resource.
257 */
__i915_vma_resource_init(struct i915_vma_resource * vma_res)258 void __i915_vma_resource_init(struct i915_vma_resource *vma_res)
259 {
260 spin_lock_init(&vma_res->lock);
261 dma_fence_init(&vma_res->unbind_fence, &unbind_fence_ops,
262 &vma_res->lock, 0, 0);
263 refcount_set(&vma_res->hold_count, 1);
264 i915_sw_fence_init(&vma_res->chain, i915_vma_resource_fence_notify);
265 }
266
267 static void
i915_vma_resource_color_adjust_range(struct i915_address_space * vm,u64 * start,u64 * end)268 i915_vma_resource_color_adjust_range(struct i915_address_space *vm,
269 u64 *start,
270 u64 *end)
271 {
272 if (i915_vm_has_cache_coloring(vm)) {
273 if (*start)
274 *start -= I915_GTT_PAGE_SIZE;
275 *end += I915_GTT_PAGE_SIZE;
276 }
277 }
278
279 /**
280 * i915_vma_resource_bind_dep_sync - Wait for / sync all unbinds touching a
281 * certain vm range.
282 * @vm: The vm to look at.
283 * @offset: The range start.
284 * @size: The range size.
285 * @intr: Whether to wait interrubtible.
286 *
287 * The function needs to be called with the vm lock held.
288 *
289 * Return: Zero on success, -ERESTARTSYS if interrupted and @intr==true
290 */
i915_vma_resource_bind_dep_sync(struct i915_address_space * vm,u64 offset,u64 size,bool intr)291 int i915_vma_resource_bind_dep_sync(struct i915_address_space *vm,
292 u64 offset,
293 u64 size,
294 bool intr)
295 {
296 struct i915_vma_resource *node;
297 u64 last = offset + size - 1;
298
299 lockdep_assert_held(&vm->mutex);
300 might_sleep();
301
302 i915_vma_resource_color_adjust_range(vm, &offset, &last);
303 node = vma_res_itree_iter_first(&vm->pending_unbind, offset, last);
304 while (node) {
305 int ret = dma_fence_wait(&node->unbind_fence, intr);
306
307 if (ret)
308 return ret;
309
310 node = vma_res_itree_iter_next(node, offset, last);
311 }
312
313 return 0;
314 }
315
316 /**
317 * i915_vma_resource_bind_dep_sync_all - Wait for / sync all unbinds of a vm,
318 * releasing the vm lock while waiting.
319 * @vm: The vm to look at.
320 *
321 * The function may not be called with the vm lock held.
322 * Typically this is called at vm destruction to finish any pending
323 * unbind operations. The vm mutex is released while waiting to avoid
324 * stalling kernel workqueues trying to grab the mutex.
325 */
i915_vma_resource_bind_dep_sync_all(struct i915_address_space * vm)326 void i915_vma_resource_bind_dep_sync_all(struct i915_address_space *vm)
327 {
328 struct i915_vma_resource *node;
329 struct dma_fence *fence;
330
331 do {
332 fence = NULL;
333 mutex_lock(&vm->mutex);
334 node = vma_res_itree_iter_first(&vm->pending_unbind, 0,
335 U64_MAX);
336 if (node)
337 fence = dma_fence_get_rcu(&node->unbind_fence);
338 mutex_unlock(&vm->mutex);
339
340 if (fence) {
341 /*
342 * The wait makes sure the node eventually removes
343 * itself from the tree.
344 */
345 dma_fence_wait(fence, false);
346 dma_fence_put(fence);
347 }
348 } while (node);
349 }
350
351 /**
352 * i915_vma_resource_bind_dep_await - Have a struct i915_sw_fence await all
353 * pending unbinds in a certain range of a vm.
354 * @vm: The vm to look at.
355 * @sw_fence: The struct i915_sw_fence that will be awaiting the unbinds.
356 * @offset: The range start.
357 * @size: The range size.
358 * @intr: Whether to wait interrubtible.
359 * @gfp: Allocation mode for memory allocations.
360 *
361 * The function makes @sw_fence await all pending unbinds in a certain
362 * vm range before calling the complete notifier. To be able to await
363 * each individual unbind, the function needs to allocate memory using
364 * the @gpf allocation mode. If that fails, the function will instead
365 * wait for the unbind fence to signal, using @intr to judge whether to
366 * wait interruptible or not. Note that @gfp should ideally be selected so
367 * as to avoid any expensive memory allocation stalls and rather fail and
368 * synchronize itself. For now the vm mutex is required when calling this
369 * function with means that @gfp can't call into direct reclaim. In reality
370 * this means that during heavy memory pressure, we will sync in this
371 * function.
372 *
373 * Return: Zero on success, -ERESTARTSYS if interrupted and @intr==true
374 */
i915_vma_resource_bind_dep_await(struct i915_address_space * vm,struct i915_sw_fence * sw_fence,u64 offset,u64 size,bool intr,gfp_t gfp)375 int i915_vma_resource_bind_dep_await(struct i915_address_space *vm,
376 struct i915_sw_fence *sw_fence,
377 u64 offset,
378 u64 size,
379 bool intr,
380 gfp_t gfp)
381 {
382 struct i915_vma_resource *node;
383 u64 last = offset + size - 1;
384
385 lockdep_assert_held(&vm->mutex);
386 might_alloc(gfp);
387 might_sleep();
388
389 i915_vma_resource_color_adjust_range(vm, &offset, &last);
390 node = vma_res_itree_iter_first(&vm->pending_unbind, offset, last);
391 while (node) {
392 int ret;
393
394 ret = i915_sw_fence_await_dma_fence(sw_fence,
395 &node->unbind_fence,
396 0, gfp);
397 if (ret < 0) {
398 ret = dma_fence_wait(&node->unbind_fence, intr);
399 if (ret)
400 return ret;
401 }
402
403 node = vma_res_itree_iter_next(node, offset, last);
404 }
405
406 return 0;
407 }
408
i915_vma_resource_module_exit(void)409 void i915_vma_resource_module_exit(void)
410 {
411 kmem_cache_destroy(slab_vma_resources);
412 }
413
i915_vma_resource_module_init(void)414 int __init i915_vma_resource_module_init(void)
415 {
416 slab_vma_resources = KMEM_CACHE(i915_vma_resource, SLAB_HWCACHE_ALIGN);
417 if (!slab_vma_resources)
418 return -ENOMEM;
419
420 return 0;
421 }
422