File: | dev/pci/drm/i915/i915_vma.c |
Warning: | line 1841, column 2 Value stored to 'node' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* |
2 | * Copyright © 2016 Intel Corporation |
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 (including the next |
12 | * paragraph) shall be included in all copies or substantial portions of the |
13 | * Software. |
14 | * |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
18 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
19 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING |
20 | * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
21 | * IN THE SOFTWARE. |
22 | * |
23 | */ |
24 | |
25 | #include <linux/sched/mm.h> |
26 | #include <linux/dma-fence-array.h> |
27 | #include <drm/drm_gem.h> |
28 | |
29 | #include "display/intel_frontbuffer.h" |
30 | #include "gem/i915_gem_lmem.h" |
31 | #include "gem/i915_gem_tiling.h" |
32 | #include "gt/intel_engine.h" |
33 | #include "gt/intel_engine_heartbeat.h" |
34 | #include "gt/intel_gt.h" |
35 | #include "gt/intel_gt_requests.h" |
36 | |
37 | #include "i915_drv.h" |
38 | #include "i915_gem_evict.h" |
39 | #include "i915_sw_fence_work.h" |
40 | #include "i915_trace.h" |
41 | #include "i915_vma.h" |
42 | #include "i915_vma_resource.h" |
43 | |
44 | #include <dev/pci/agpvar.h> |
45 | |
46 | static inline void assert_vma_held_evict(const struct i915_vma *vma) |
47 | { |
48 | /* |
49 | * We may be forced to unbind when the vm is dead, to clean it up. |
50 | * This is the only exception to the requirement of the object lock |
51 | * being held. |
52 | */ |
53 | if (kref_read(&vma->vm->ref)) |
54 | assert_object_held_shared(vma->obj); |
55 | } |
56 | |
57 | static struct pool slab_vmas; |
58 | |
59 | static struct i915_vma *i915_vma_alloc(void) |
60 | { |
61 | #ifdef __linux__ |
62 | return kmem_cache_zalloc(slab_vmas, GFP_KERNEL(0x0001 | 0x0004)); |
63 | #else |
64 | return pool_get(&slab_vmas, PR_WAITOK0x0001 | PR_ZERO0x0008); |
65 | #endif |
66 | } |
67 | |
68 | static void i915_vma_free(struct i915_vma *vma) |
69 | { |
70 | #ifdef __linux__ |
71 | return kmem_cache_free(slab_vmas, vma); |
72 | #else |
73 | pool_put(&slab_vmas, vma); |
74 | #endif |
75 | } |
76 | |
77 | #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM)0 && IS_ENABLED(CONFIG_DRM_DEBUG_MM)0 |
78 | |
79 | #include <linux/stackdepot.h> |
80 | |
81 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) |
82 | { |
83 | char buf[512]; |
84 | |
85 | if (!vma->node.stack) { |
86 | DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",___drm_dbg(((void *)0), DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: unknown owner\n" , vma->node.start, vma->node.size, reason) |
87 | vma->node.start, vma->node.size, reason)___drm_dbg(((void *)0), DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: unknown owner\n" , vma->node.start, vma->node.size, reason); |
88 | return; |
89 | } |
90 | |
91 | stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0); |
92 | DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",___drm_dbg(((void *)0), DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: inserted at %s\n" , vma->node.start, vma->node.size, reason, buf) |
93 | vma->node.start, vma->node.size, reason, buf)___drm_dbg(((void *)0), DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: inserted at %s\n" , vma->node.start, vma->node.size, reason, buf); |
94 | } |
95 | |
96 | #else |
97 | |
98 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) |
99 | { |
100 | } |
101 | |
102 | #endif |
103 | |
104 | static inline struct i915_vma *active_to_vma(struct i915_active *ref) |
105 | { |
106 | return container_of(ref, typeof(struct i915_vma), active)({ const __typeof( ((typeof(struct i915_vma) *)0)->active ) *__mptr = (ref); (typeof(struct i915_vma) *)( (char *)__mptr - __builtin_offsetof(typeof(struct i915_vma), active) );}); |
107 | } |
108 | |
109 | static int __i915_vma_active(struct i915_active *ref) |
110 | { |
111 | return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT2; |
112 | } |
113 | |
114 | static void __i915_vma_retire(struct i915_active *ref) |
115 | { |
116 | i915_vma_put(active_to_vma(ref)); |
117 | } |
118 | |
119 | static struct i915_vma * |
120 | vma_create(struct drm_i915_gem_object *obj, |
121 | struct i915_address_space *vm, |
122 | const struct i915_gtt_view *view) |
123 | { |
124 | struct i915_vma *pos = ERR_PTR(-E2BIG7); |
125 | struct i915_vma *vma; |
126 | struct rb_node *rb, **p; |
127 | int err; |
128 | |
129 | /* The aliasing_ppgtt should never be used directly! */ |
130 | GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm)((void)0); |
131 | |
132 | vma = i915_vma_alloc(); |
133 | if (vma == NULL((void *)0)) |
134 | return ERR_PTR(-ENOMEM12); |
135 | |
136 | vma->ops = &vm->vma_ops; |
137 | vma->obj = obj; |
138 | vma->size = obj->base.size; |
139 | vma->display_alignment = I915_GTT_MIN_ALIGNMENT(1ULL << (12)); |
140 | |
141 | i915_active_init(&vma->active, __i915_vma_active, __i915_vma_retire, 0)do { static struct lock_class_key __mkey; static struct lock_class_key __wkey; __i915_active_init(&vma->active, __i915_vma_active , __i915_vma_retire, 0, &__mkey, &__wkey); } while (0 ); |
142 | |
143 | #ifdef notyet |
144 | /* Declare ourselves safe for use inside shrinkers */ |
145 | if (IS_ENABLED(CONFIG_LOCKDEP)0) { |
146 | fs_reclaim_acquire(GFP_KERNEL(0x0001 | 0x0004)); |
147 | might_lock(&vma->active.mutex); |
148 | fs_reclaim_release(GFP_KERNEL(0x0001 | 0x0004)); |
149 | } |
150 | #endif |
151 | |
152 | INIT_LIST_HEAD(&vma->closed_link); |
153 | INIT_LIST_HEAD(&vma->obj_link); |
154 | RB_CLEAR_NODE(&vma->obj_node)(((&vma->obj_node))->__entry.rbe_parent = (&vma ->obj_node)); |
155 | |
156 | if (view && view->type != I915_GTT_VIEW_NORMAL) { |
157 | vma->gtt_view = *view; |
158 | if (view->type == I915_GTT_VIEW_PARTIAL) { |
159 | GEM_BUG_ON(range_overflows_t(u64,((void)0) |
160 | view->partial.offset,((void)0) |
161 | view->partial.size,((void)0) |
162 | obj->base.size >> PAGE_SHIFT))((void)0); |
163 | vma->size = view->partial.size; |
164 | vma->size <<= PAGE_SHIFT12; |
165 | GEM_BUG_ON(vma->size > obj->base.size)((void)0); |
166 | } else if (view->type == I915_GTT_VIEW_ROTATED) { |
167 | vma->size = intel_rotation_info_size(&view->rotated); |
168 | vma->size <<= PAGE_SHIFT12; |
169 | } else if (view->type == I915_GTT_VIEW_REMAPPED) { |
170 | vma->size = intel_remapped_info_size(&view->remapped); |
171 | vma->size <<= PAGE_SHIFT12; |
172 | } |
173 | } |
174 | |
175 | if (unlikely(vma->size > vm->total)__builtin_expect(!!(vma->size > vm->total), 0)) |
176 | goto err_vma; |
177 | |
178 | GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE))((void)0); |
179 | |
180 | err = mutex_lock_interruptible(&vm->mutex); |
181 | if (err) { |
182 | pos = ERR_PTR(err); |
183 | goto err_vma; |
184 | } |
185 | |
186 | vma->vm = vm; |
187 | list_add_tail(&vma->vm_link, &vm->unbound_list); |
188 | |
189 | spin_lock(&obj->vma.lock)mtx_enter(&obj->vma.lock); |
190 | if (i915_is_ggtt(vm)((vm)->is_ggtt)) { |
191 | if (unlikely(overflows_type(vma->size, u32))__builtin_expect(!!((sizeof(vma->size) > sizeof(u32) && (vma->size) >> (8 * sizeof(u32)))), 0)) |
192 | goto err_unlock; |
193 | |
194 | vma->fence_size = i915_gem_fence_size(vm->i915, vma->size, |
195 | i915_gem_object_get_tiling(obj), |
196 | i915_gem_object_get_stride(obj)); |
197 | if (unlikely(vma->fence_size < vma->size || /* overflow */__builtin_expect(!!(vma->fence_size < vma->size || vma ->fence_size > vm->total), 0) |
198 | vma->fence_size > vm->total)__builtin_expect(!!(vma->fence_size < vma->size || vma ->fence_size > vm->total), 0)) |
199 | goto err_unlock; |
200 | |
201 | GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT))((void)0); |
202 | |
203 | vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size, |
204 | i915_gem_object_get_tiling(obj), |
205 | i915_gem_object_get_stride(obj)); |
206 | GEM_BUG_ON(!is_power_of_2(vma->fence_alignment))((void)0); |
207 | |
208 | __set_bit(I915_VMA_GGTT_BIT13, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
209 | } |
210 | |
211 | rb = NULL((void *)0); |
212 | p = &obj->vma.tree.rb_node; |
213 | while (*p) { |
214 | long cmp; |
215 | |
216 | rb = *p; |
217 | pos = rb_entry(rb, struct i915_vma, obj_node)({ const __typeof( ((struct i915_vma *)0)->obj_node ) *__mptr = (rb); (struct i915_vma *)( (char *)__mptr - __builtin_offsetof (struct i915_vma, obj_node) );}); |
218 | |
219 | /* |
220 | * If the view already exists in the tree, another thread |
221 | * already created a matching vma, so return the older instance |
222 | * and dispose of ours. |
223 | */ |
224 | cmp = i915_vma_compare(pos, vm, view); |
225 | if (cmp < 0) |
226 | p = &rb->rb_right__entry.rbe_right; |
227 | else if (cmp > 0) |
228 | p = &rb->rb_left__entry.rbe_left; |
229 | else |
230 | goto err_unlock; |
231 | } |
232 | rb_link_node(&vma->obj_node, rb, p); |
233 | rb_insert_color(&vma->obj_node, &obj->vma.tree)linux_root_RB_INSERT_COLOR((struct linux_root *)(&obj-> vma.tree), (&vma->obj_node)); |
234 | |
235 | if (i915_vma_is_ggtt(vma)) |
236 | /* |
237 | * We put the GGTT vma at the start of the vma-list, followed |
238 | * by the ppGGTT vma. This allows us to break early when |
239 | * iterating over only the GGTT vma for an object, see |
240 | * for_each_ggtt_vma() |
241 | */ |
242 | list_add(&vma->obj_link, &obj->vma.list); |
243 | else |
244 | list_add_tail(&vma->obj_link, &obj->vma.list); |
245 | |
246 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
247 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
248 | |
249 | return vma; |
250 | |
251 | err_unlock: |
252 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
253 | list_del_init(&vma->vm_link); |
254 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
255 | err_vma: |
256 | i915_vma_free(vma); |
257 | return pos; |
258 | } |
259 | |
260 | static struct i915_vma * |
261 | i915_vma_lookup(struct drm_i915_gem_object *obj, |
262 | struct i915_address_space *vm, |
263 | const struct i915_gtt_view *view) |
264 | { |
265 | struct rb_node *rb; |
266 | |
267 | rb = obj->vma.tree.rb_node; |
268 | while (rb) { |
269 | struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node)({ const __typeof( ((struct i915_vma *)0)->obj_node ) *__mptr = (rb); (struct i915_vma *)( (char *)__mptr - __builtin_offsetof (struct i915_vma, obj_node) );}); |
270 | long cmp; |
271 | |
272 | cmp = i915_vma_compare(vma, vm, view); |
273 | if (cmp == 0) |
274 | return vma; |
275 | |
276 | if (cmp < 0) |
277 | rb = rb->rb_right__entry.rbe_right; |
278 | else |
279 | rb = rb->rb_left__entry.rbe_left; |
280 | } |
281 | |
282 | return NULL((void *)0); |
283 | } |
284 | |
285 | /** |
286 | * i915_vma_instance - return the singleton instance of the VMA |
287 | * @obj: parent &struct drm_i915_gem_object to be mapped |
288 | * @vm: address space in which the mapping is located |
289 | * @view: additional mapping requirements |
290 | * |
291 | * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with |
292 | * the same @view characteristics. If a match is not found, one is created. |
293 | * Once created, the VMA is kept until either the object is freed, or the |
294 | * address space is closed. |
295 | * |
296 | * Returns the vma, or an error pointer. |
297 | */ |
298 | struct i915_vma * |
299 | i915_vma_instance(struct drm_i915_gem_object *obj, |
300 | struct i915_address_space *vm, |
301 | const struct i915_gtt_view *view) |
302 | { |
303 | struct i915_vma *vma; |
304 | |
305 | GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm))((void)0); |
306 | GEM_BUG_ON(!kref_read(&vm->ref))((void)0); |
307 | |
308 | spin_lock(&obj->vma.lock)mtx_enter(&obj->vma.lock); |
309 | vma = i915_vma_lookup(obj, vm, view); |
310 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
311 | |
312 | /* vma_create() will resolve the race if another creates the vma */ |
313 | if (unlikely(!vma)__builtin_expect(!!(!vma), 0)) |
314 | vma = vma_create(obj, vm, view); |
315 | |
316 | GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view))((void)0); |
317 | return vma; |
318 | } |
319 | |
320 | struct i915_vma_work { |
321 | struct dma_fence_work base; |
322 | struct i915_address_space *vm; |
323 | struct i915_vm_pt_stash stash; |
324 | struct i915_vma_resource *vma_res; |
325 | struct drm_i915_gem_object *obj; |
326 | struct i915_sw_dma_fence_cb cb; |
327 | enum i915_cache_level cache_level; |
328 | unsigned int flags; |
329 | }; |
330 | |
331 | static void __vma_bind(struct dma_fence_work *work) |
332 | { |
333 | struct i915_vma_work *vw = container_of(work, typeof(*vw), base)({ const __typeof( ((typeof(*vw) *)0)->base ) *__mptr = (work ); (typeof(*vw) *)( (char *)__mptr - __builtin_offsetof(typeof (*vw), base) );}); |
334 | struct i915_vma_resource *vma_res = vw->vma_res; |
335 | |
336 | /* |
337 | * We are about the bind the object, which must mean we have already |
338 | * signaled the work to potentially clear/move the pages underneath. If |
339 | * something went wrong at that stage then the object should have |
340 | * unknown_state set, in which case we need to skip the bind. |
341 | */ |
342 | if (i915_gem_object_has_unknown_state(vw->obj)) |
343 | return; |
344 | |
345 | vma_res->ops->bind_vma(vma_res->vm, &vw->stash, |
346 | vma_res, vw->cache_level, vw->flags); |
347 | } |
348 | |
349 | static void __vma_release(struct dma_fence_work *work) |
350 | { |
351 | struct i915_vma_work *vw = container_of(work, typeof(*vw), base)({ const __typeof( ((typeof(*vw) *)0)->base ) *__mptr = (work ); (typeof(*vw) *)( (char *)__mptr - __builtin_offsetof(typeof (*vw), base) );}); |
352 | |
353 | if (vw->obj) |
354 | i915_gem_object_put(vw->obj); |
355 | |
356 | i915_vm_free_pt_stash(vw->vm, &vw->stash); |
357 | if (vw->vma_res) |
358 | i915_vma_resource_put(vw->vma_res); |
359 | } |
360 | |
361 | static const struct dma_fence_work_ops bind_ops = { |
362 | .name = "bind", |
363 | .work = __vma_bind, |
364 | .release = __vma_release, |
365 | }; |
366 | |
367 | struct i915_vma_work *i915_vma_work(void) |
368 | { |
369 | struct i915_vma_work *vw; |
370 | |
371 | vw = kzalloc(sizeof(*vw), GFP_KERNEL(0x0001 | 0x0004)); |
372 | if (!vw) |
373 | return NULL((void *)0); |
374 | |
375 | dma_fence_work_init(&vw->base, &bind_ops); |
376 | vw->base.dma.error = -EAGAIN35; /* disable the worker by default */ |
377 | |
378 | return vw; |
379 | } |
380 | |
381 | int i915_vma_wait_for_bind(struct i915_vma *vma) |
382 | { |
383 | int err = 0; |
384 | |
385 | if (rcu_access_pointer(vma->active.excl.fence)(vma->active.excl.fence)) { |
386 | struct dma_fence *fence; |
387 | |
388 | rcu_read_lock(); |
389 | fence = dma_fence_get_rcu_safe(&vma->active.excl.fence); |
390 | rcu_read_unlock(); |
391 | if (fence) { |
392 | err = dma_fence_wait(fence, true1); |
393 | dma_fence_put(fence); |
394 | } |
395 | } |
396 | |
397 | return err; |
398 | } |
399 | |
400 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)0 |
401 | static int i915_vma_verify_bind_complete(struct i915_vma *vma)0 |
402 | { |
403 | struct dma_fence *fence = i915_active_fence_get(&vma->active.excl); |
404 | int err; |
405 | |
406 | if (!fence) |
407 | return 0; |
408 | |
409 | if (dma_fence_is_signaled(fence)) |
410 | err = fence->error; |
411 | else |
412 | err = -EBUSY16; |
413 | |
414 | dma_fence_put(fence); |
415 | |
416 | return err; |
417 | } |
418 | #else |
419 | #define i915_vma_verify_bind_complete(_vma)0 0 |
420 | #endif |
421 | |
422 | I915_SELFTEST_EXPORTstatic void |
423 | i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res, |
424 | struct i915_vma *vma) |
425 | { |
426 | struct drm_i915_gem_object *obj = vma->obj; |
427 | |
428 | i915_vma_resource_init(vma_res, vma->vm, vma->pages, &vma->page_sizes, |
429 | obj->mm.rsgt, i915_gem_object_is_readonly(obj), |
430 | i915_gem_object_is_lmem(obj), obj->mm.region, |
431 | vma->ops, vma->private, vma->node.start, |
432 | vma->node.size, vma->size); |
433 | } |
434 | |
435 | /** |
436 | * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. |
437 | * @vma: VMA to map |
438 | * @cache_level: mapping cache level |
439 | * @flags: flags like global or local mapping |
440 | * @work: preallocated worker for allocating and binding the PTE |
441 | * @vma_res: pointer to a preallocated vma resource. The resource is either |
442 | * consumed or freed. |
443 | * |
444 | * DMA addresses are taken from the scatter-gather table of this object (or of |
445 | * this VMA in case of non-default GGTT views) and PTE entries set up. |
446 | * Note that DMA addresses are also the only part of the SG table we care about. |
447 | */ |
448 | int i915_vma_bind(struct i915_vma *vma, |
449 | enum i915_cache_level cache_level, |
450 | u32 flags, |
451 | struct i915_vma_work *work, |
452 | struct i915_vma_resource *vma_res) |
453 | { |
454 | u32 bind_flags; |
455 | u32 vma_flags; |
456 | int ret; |
457 | |
458 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
459 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node))((void)0); |
460 | GEM_BUG_ON(vma->size > vma->node.size)((void)0); |
461 | |
462 | if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,({ ((void)0); 0; }) |
463 | vma->node.size,({ ((void)0); 0; }) |
464 | vma->vm->total))({ ((void)0); 0; })) { |
465 | i915_vma_resource_free(vma_res); |
466 | return -ENODEV19; |
467 | } |
468 | |
469 | if (GEM_DEBUG_WARN_ON(!flags)({ ((void)0); 0; })) { |
470 | i915_vma_resource_free(vma_res); |
471 | return -EINVAL22; |
472 | } |
473 | |
474 | bind_flags = flags; |
475 | bind_flags &= I915_VMA_GLOBAL_BIND((int)(1UL << (10))) | I915_VMA_LOCAL_BIND((int)(1UL << (11))); |
476 | |
477 | vma_flags = atomic_read(&vma->flags)({ typeof(*(&vma->flags)) __tmp = *(volatile typeof(*( &vma->flags)) *)&(*(&vma->flags)); membar_datadep_consumer (); __tmp; }); |
478 | vma_flags &= I915_VMA_GLOBAL_BIND((int)(1UL << (10))) | I915_VMA_LOCAL_BIND((int)(1UL << (11))); |
479 | |
480 | bind_flags &= ~vma_flags; |
481 | if (bind_flags == 0) { |
482 | i915_vma_resource_free(vma_res); |
483 | return 0; |
484 | } |
485 | |
486 | GEM_BUG_ON(!atomic_read(&vma->pages_count))((void)0); |
487 | |
488 | /* Wait for or await async unbinds touching our range */ |
489 | if (work && bind_flags & vma->vm->bind_async_flags) |
490 | ret = i915_vma_resource_bind_dep_await(vma->vm, |
491 | &work->base.chain, |
492 | vma->node.start, |
493 | vma->node.size, |
494 | true1, |
495 | GFP_NOWAIT0x0002 | |
496 | __GFP_RETRY_MAYFAIL0 | |
497 | __GFP_NOWARN0); |
498 | else |
499 | ret = i915_vma_resource_bind_dep_sync(vma->vm, vma->node.start, |
500 | vma->node.size, true1); |
501 | if (ret) { |
502 | i915_vma_resource_free(vma_res); |
503 | return ret; |
504 | } |
505 | |
506 | if (vma->resource || !vma_res) { |
507 | /* Rebinding with an additional I915_VMA_*_BIND */ |
508 | GEM_WARN_ON(!vma_flags)({ __builtin_expect(!!(!!(!vma_flags)), 0); }); |
509 | i915_vma_resource_free(vma_res); |
510 | } else { |
511 | i915_vma_resource_init_from_vma(vma_res, vma); |
512 | vma->resource = vma_res; |
513 | } |
514 | trace_i915_vma_bind(vma, bind_flags); |
515 | if (work && bind_flags & vma->vm->bind_async_flags) { |
516 | struct dma_fence *prev; |
517 | |
518 | work->vma_res = i915_vma_resource_get(vma->resource); |
519 | work->cache_level = cache_level; |
520 | work->flags = bind_flags; |
521 | |
522 | /* |
523 | * Note we only want to chain up to the migration fence on |
524 | * the pages (not the object itself). As we don't track that, |
525 | * yet, we have to use the exclusive fence instead. |
526 | * |
527 | * Also note that we do not want to track the async vma as |
528 | * part of the obj->resv->excl_fence as it only affects |
529 | * execution and not content or object's backing store lifetime. |
530 | */ |
531 | prev = i915_active_set_exclusive(&vma->active, &work->base.dma); |
532 | if (prev) { |
533 | __i915_sw_fence_await_dma_fence(&work->base.chain, |
534 | prev, |
535 | &work->cb); |
536 | dma_fence_put(prev); |
537 | } |
538 | |
539 | work->base.dma.error = 0; /* enable the queue_work() */ |
540 | work->obj = i915_gem_object_get(vma->obj); |
541 | } else { |
542 | ret = i915_gem_object_wait_moving_fence(vma->obj, true1); |
543 | if (ret) { |
544 | i915_vma_resource_free(vma->resource); |
545 | vma->resource = NULL((void *)0); |
546 | |
547 | return ret; |
548 | } |
549 | vma->ops->bind_vma(vma->vm, NULL((void *)0), vma->resource, cache_level, |
550 | bind_flags); |
551 | } |
552 | |
553 | atomic_or(bind_flags, &vma->flags)x86_atomic_setbits_u32(&vma->flags, bind_flags); |
554 | return 0; |
555 | } |
556 | |
557 | void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) |
558 | { |
559 | void __iomem *ptr; |
560 | int err; |
561 | |
562 | if (WARN_ON_ONCE(vma->obj->flags & I915_BO_ALLOC_GPU_ONLY)({ static int __warned; int __ret = !!(vma->obj->flags & (1UL << (6))); if (__ret && !__warned) { printf ("WARNING %s failed at %s:%d\n", "vma->obj->flags & (1UL << (6))" , "/usr/src/sys/dev/pci/drm/i915/i915_vma.c", 562); __warned = 1; } __builtin_expect(!!(__ret), 0); })) |
563 | return IOMEM_ERR_PTR(-EINVAL22); |
564 | |
565 | GEM_BUG_ON(!i915_vma_is_ggtt(vma))((void)0); |
566 | GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))((void)0); |
567 | GEM_BUG_ON(i915_vma_verify_bind_complete(vma))((void)0); |
568 | |
569 | ptr = READ_ONCE(vma->iomap)({ typeof(vma->iomap) __tmp = *(volatile typeof(vma->iomap ) *)&(vma->iomap); membar_datadep_consumer(); __tmp; } ); |
570 | if (ptr == NULL((void *)0)) { |
571 | /* |
572 | * TODO: consider just using i915_gem_object_pin_map() for lmem |
573 | * instead, which already supports mapping non-contiguous chunks |
574 | * of pages, that way we can also drop the |
575 | * I915_BO_ALLOC_CONTIGUOUS when allocating the object. |
576 | */ |
577 | if (i915_gem_object_is_lmem(vma->obj)) { |
578 | ptr = i915_gem_object_lmem_io_map(vma->obj, 0, |
579 | vma->obj->base.size); |
580 | } else if (i915_vma_is_map_and_fenceable(vma)) { |
581 | #ifdef __linux__ |
582 | ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap, |
583 | vma->node.start, |
584 | vma->node.size); |
585 | #else |
586 | { |
587 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->vm->i915; |
588 | err = agp_map_subregion(dev_priv->agph, vma->node.start, |
589 | vma->node.size, &vma->bsh); |
590 | if (err) { |
591 | err = -err; |
592 | goto err; |
593 | } |
594 | ptr = bus_space_vaddr(dev_priv->bst, vma->bsh)((dev_priv->bst)->vaddr((vma->bsh))); |
595 | } |
596 | #endif |
597 | } else { |
598 | ptr = (void __iomem *) |
599 | i915_gem_object_pin_map(vma->obj, I915_MAP_WC); |
600 | if (IS_ERR(ptr)) { |
601 | err = PTR_ERR(ptr); |
602 | goto err; |
603 | } |
604 | ptr = page_pack_bits(ptr, 1)({ unsigned long __bits = (1); ((void)0); ((typeof(ptr))((unsigned long)(ptr) | __bits)); }); |
605 | } |
606 | |
607 | if (ptr == NULL((void *)0)) { |
608 | err = -ENOMEM12; |
609 | goto err; |
610 | } |
611 | |
612 | if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))__builtin_expect(!!(__sync_val_compare_and_swap(&vma-> iomap, ((void *)0), ptr)), 0)) { |
613 | if (page_unmask_bits(ptr)((unsigned long)(ptr) & ((1UL << (12)) - 1))) |
614 | __i915_gem_object_release_map(vma->obj); |
615 | #ifdef __linux__ |
616 | else |
617 | io_mapping_unmap(ptr); |
618 | #endif |
619 | ptr = vma->iomap; |
620 | } |
621 | } |
622 | |
623 | __i915_vma_pin(vma); |
624 | |
625 | err = i915_vma_pin_fence(vma); |
626 | if (err) |
627 | goto err_unpin; |
628 | |
629 | i915_vma_set_ggtt_write(vma); |
630 | |
631 | /* NB Access through the GTT requires the device to be awake. */ |
632 | return page_mask_bits(ptr)({ unsigned long __v = (unsigned long)(ptr); (typeof(ptr))(__v & -(1UL << (12))); }); |
633 | |
634 | err_unpin: |
635 | __i915_vma_unpin(vma); |
636 | err: |
637 | return IOMEM_ERR_PTR(err); |
638 | } |
639 | |
640 | void i915_vma_flush_writes(struct i915_vma *vma) |
641 | { |
642 | if (i915_vma_unset_ggtt_write(vma)) |
643 | intel_gt_flush_ggtt_writes(vma->vm->gt); |
644 | } |
645 | |
646 | void i915_vma_unpin_iomap(struct i915_vma *vma) |
647 | { |
648 | GEM_BUG_ON(vma->iomap == NULL)((void)0); |
649 | |
650 | /* XXX We keep the mapping until __i915_vma_unbind()/evict() */ |
651 | |
652 | i915_vma_flush_writes(vma); |
653 | |
654 | i915_vma_unpin_fence(vma); |
655 | i915_vma_unpin(vma); |
656 | } |
657 | |
658 | void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) |
659 | { |
660 | struct i915_vma *vma; |
661 | struct drm_i915_gem_object *obj; |
662 | |
663 | vma = fetch_and_zero(p_vma)({ typeof(*p_vma) __T = *(p_vma); *(p_vma) = (typeof(*p_vma)) 0; __T; }); |
664 | if (!vma) |
665 | return; |
666 | |
667 | obj = vma->obj; |
668 | GEM_BUG_ON(!obj)((void)0); |
669 | |
670 | i915_vma_unpin(vma); |
671 | |
672 | if (flags & I915_VMA_RELEASE_MAP(1UL << (0))) |
673 | i915_gem_object_unpin_map(obj); |
674 | |
675 | i915_gem_object_put(obj); |
676 | } |
677 | |
678 | bool_Bool i915_vma_misplaced(const struct i915_vma *vma, |
679 | u64 size, u64 alignment, u64 flags) |
680 | { |
681 | if (!drm_mm_node_allocated(&vma->node)) |
682 | return false0; |
683 | |
684 | if (test_bit(I915_VMA_ERROR_BIT12, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags))) |
685 | return true1; |
686 | |
687 | if (vma->node.size < size) |
688 | return true1; |
689 | |
690 | GEM_BUG_ON(alignment && !is_power_of_2(alignment))((void)0); |
691 | if (alignment && !IS_ALIGNED(vma->node.start, alignment)(((vma->node.start) & ((alignment) - 1)) == 0)) |
692 | return true1; |
693 | |
694 | if (flags & PIN_MAPPABLE(1ULL << (3)) && !i915_vma_is_map_and_fenceable(vma)) |
695 | return true1; |
696 | |
697 | if (flags & PIN_OFFSET_BIAS(1ULL << (6)) && |
698 | vma->node.start < (flags & PIN_OFFSET_MASK-(1ULL << (12)))) |
699 | return true1; |
700 | |
701 | if (flags & PIN_OFFSET_FIXED(1ULL << (7)) && |
702 | vma->node.start != (flags & PIN_OFFSET_MASK-(1ULL << (12)))) |
703 | return true1; |
704 | |
705 | return false0; |
706 | } |
707 | |
708 | void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) |
709 | { |
710 | bool_Bool mappable, fenceable; |
711 | |
712 | GEM_BUG_ON(!i915_vma_is_ggtt(vma))((void)0); |
713 | GEM_BUG_ON(!vma->fence_size)((void)0); |
714 | |
715 | fenceable = (vma->node.size >= vma->fence_size && |
716 | IS_ALIGNED(vma->node.start, vma->fence_alignment)(((vma->node.start) & ((vma->fence_alignment) - 1)) == 0)); |
717 | |
718 | mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end; |
719 | |
720 | if (mappable && fenceable) |
721 | set_bit(I915_VMA_CAN_FENCE_BIT14, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
722 | else |
723 | clear_bit(I915_VMA_CAN_FENCE_BIT14, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
724 | } |
725 | |
726 | bool_Bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) |
727 | { |
728 | struct drm_mm_node *node = &vma->node; |
729 | struct drm_mm_node *other; |
730 | |
731 | /* |
732 | * On some machines we have to be careful when putting differing types |
733 | * of snoopable memory together to avoid the prefetcher crossing memory |
734 | * domains and dying. During vm initialisation, we decide whether or not |
735 | * these constraints apply and set the drm_mm.color_adjust |
736 | * appropriately. |
737 | */ |
738 | if (!i915_vm_has_cache_coloring(vma->vm)) |
739 | return true1; |
740 | |
741 | /* Only valid to be called on an already inserted vma */ |
742 | GEM_BUG_ON(!drm_mm_node_allocated(node))((void)0); |
743 | GEM_BUG_ON(list_empty(&node->node_list))((void)0); |
744 | |
745 | other = list_prev_entry(node, node_list)({ const __typeof( ((typeof(*(node)) *)0)->node_list ) *__mptr = (((node)->node_list.prev)); (typeof(*(node)) *)( (char * )__mptr - __builtin_offsetof(typeof(*(node)), node_list) );}); |
746 | if (i915_node_color_differs(other, color) && |
747 | !drm_mm_hole_follows(other)) |
748 | return false0; |
749 | |
750 | other = list_next_entry(node, node_list)({ const __typeof( ((typeof(*(node)) *)0)->node_list ) *__mptr = (((node)->node_list.next)); (typeof(*(node)) *)( (char * )__mptr - __builtin_offsetof(typeof(*(node)), node_list) );}); |
751 | if (i915_node_color_differs(other, color) && |
752 | !drm_mm_hole_follows(node)) |
753 | return false0; |
754 | |
755 | return true1; |
756 | } |
757 | |
758 | /** |
759 | * i915_vma_insert - finds a slot for the vma in its address space |
760 | * @vma: the vma |
761 | * @size: requested size in bytes (can be larger than the VMA) |
762 | * @alignment: required alignment |
763 | * @flags: mask of PIN_* flags to use |
764 | * |
765 | * First we try to allocate some free space that meets the requirements for |
766 | * the VMA. Failiing that, if the flags permit, it will evict an old VMA, |
767 | * preferrably the oldest idle entry to make room for the new VMA. |
768 | * |
769 | * Returns: |
770 | * 0 on success, negative error code otherwise. |
771 | */ |
772 | static int |
773 | i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
774 | u64 size, u64 alignment, u64 flags) |
775 | { |
776 | unsigned long color; |
777 | u64 start, end; |
778 | int ret; |
779 | |
780 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND))((void)0); |
781 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node))((void)0); |
782 | |
783 | size = max(size, vma->size)(((size)>(vma->size))?(size):(vma->size)); |
784 | alignment = max(alignment, vma->display_alignment)(((alignment)>(vma->display_alignment))?(alignment):(vma ->display_alignment)); |
785 | if (flags & PIN_MAPPABLE(1ULL << (3))) { |
786 | size = max_t(typeof(size), size, vma->fence_size)({ typeof(size) __max_a = (size); typeof(size) __max_b = (vma ->fence_size); __max_a > __max_b ? __max_a : __max_b; } ); |
787 | alignment = max_t(typeof(alignment),({ typeof(alignment) __max_a = (alignment); typeof(alignment) __max_b = (vma->fence_alignment); __max_a > __max_b ? __max_a : __max_b; }) |
788 | alignment, vma->fence_alignment)({ typeof(alignment) __max_a = (alignment); typeof(alignment) __max_b = (vma->fence_alignment); __max_a > __max_b ? __max_a : __max_b; }); |
789 | } |
790 | |
791 | GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE))((void)0); |
792 | GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT))((void)0); |
793 | GEM_BUG_ON(!is_power_of_2(alignment))((void)0); |
794 | |
795 | start = flags & PIN_OFFSET_BIAS(1ULL << (6)) ? flags & PIN_OFFSET_MASK-(1ULL << (12)) : 0; |
796 | GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE))((void)0); |
797 | |
798 | end = vma->vm->total; |
799 | if (flags & PIN_MAPPABLE(1ULL << (3))) |
800 | end = min_t(u64, end, i915_vm_to_ggtt(vma->vm)->mappable_end)({ u64 __min_a = (end); u64 __min_b = (i915_vm_to_ggtt(vma-> vm)->mappable_end); __min_a < __min_b ? __min_a : __min_b ; }); |
801 | if (flags & PIN_ZONE_4G(1ULL << (4))) |
802 | end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE)({ u64 __min_a = (end); u64 __min_b = ((1ULL << 32) - ( 1ULL << (12))); __min_a < __min_b ? __min_a : __min_b ; }); |
803 | GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE))((void)0); |
804 | |
805 | alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj))(((alignment)>(i915_vm_obj_min_alignment(vma->vm, vma-> obj)))?(alignment):(i915_vm_obj_min_alignment(vma->vm, vma ->obj))); |
806 | /* |
807 | * for compact-pt we round up the reservation to prevent |
808 | * any smaller pages being used within the same PDE |
809 | */ |
810 | if (NEEDS_COMPACT_PT(vma->vm->i915)((&(vma->vm->i915)->__info)->needs_compact_pt )) |
811 | size = round_up(size, alignment)((((size) + ((alignment) - 1)) / (alignment)) * (alignment)); |
812 | |
813 | /* If binding the object/GGTT view requires more space than the entire |
814 | * aperture has, reject it early before evicting everything in a vain |
815 | * attempt to find space. |
816 | */ |
817 | if (size > end) { |
818 | DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",___drm_dbg(((void *)0), DRM_UT_CORE, "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n" , size, flags & (1ULL << (3)) ? "mappable" : "total" , end) |
819 | size, flags & PIN_MAPPABLE ? "mappable" : "total",___drm_dbg(((void *)0), DRM_UT_CORE, "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n" , size, flags & (1ULL << (3)) ? "mappable" : "total" , end) |
820 | end)___drm_dbg(((void *)0), DRM_UT_CORE, "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n" , size, flags & (1ULL << (3)) ? "mappable" : "total" , end); |
821 | return -ENOSPC28; |
822 | } |
823 | |
824 | color = 0; |
825 | |
826 | if (i915_vm_has_cache_coloring(vma->vm)) |
827 | color = vma->obj->cache_level; |
828 | |
829 | if (flags & PIN_OFFSET_FIXED(1ULL << (7))) { |
830 | u64 offset = flags & PIN_OFFSET_MASK-(1ULL << (12)); |
831 | if (!IS_ALIGNED(offset, alignment)(((offset) & ((alignment) - 1)) == 0) || |
832 | range_overflows(offset, size, end)({ typeof(offset) start__ = (offset); typeof(size) size__ = ( size); typeof(end) max__ = (end); (void)(&start__ == & size__); (void)(&start__ == &max__); start__ >= max__ || size__ > max__ - start__; })) |
833 | return -EINVAL22; |
834 | |
835 | ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node, |
836 | size, offset, color, |
837 | flags); |
838 | if (ret) |
839 | return ret; |
840 | } else { |
841 | /* |
842 | * We only support huge gtt pages through the 48b PPGTT, |
843 | * however we also don't want to force any alignment for |
844 | * objects which need to be tightly packed into the low 32bits. |
845 | * |
846 | * Note that we assume that GGTT are limited to 4GiB for the |
847 | * forseeable future. See also i915_ggtt_offset(). |
848 | */ |
849 | if (upper_32_bits(end - 1)((u32)(((end - 1) >> 16) >> 16)) && |
850 | vma->page_sizes.sg > I915_GTT_PAGE_SIZE(1ULL << (12))) { |
851 | /* |
852 | * We can't mix 64K and 4K PTEs in the same page-table |
853 | * (2M block), and so to avoid the ugliness and |
854 | * complexity of coloring we opt for just aligning 64K |
855 | * objects to 2M. |
856 | */ |
857 | u64 page_alignment = |
858 | rounddown_pow_of_two(vma->page_sizes.sg | |
859 | I915_GTT_PAGE_SIZE_2M(1ULL << (21))); |
860 | |
861 | /* |
862 | * Check we don't expand for the limited Global GTT |
863 | * (mappable aperture is even more precious!). This |
864 | * also checks that we exclude the aliasing-ppgtt. |
865 | */ |
866 | GEM_BUG_ON(i915_vma_is_ggtt(vma))((void)0); |
867 | |
868 | alignment = max(alignment, page_alignment)(((alignment)>(page_alignment))?(alignment):(page_alignment )); |
869 | |
870 | if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K(1ULL << (16))) |
871 | size = round_up(size, I915_GTT_PAGE_SIZE_2M)((((size) + (((1ULL << (21))) - 1)) / ((1ULL << ( 21)))) * ((1ULL << (21)))); |
872 | } |
873 | |
874 | ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node, |
875 | size, alignment, color, |
876 | start, end, flags); |
877 | if (ret) |
878 | return ret; |
879 | |
880 | GEM_BUG_ON(vma->node.start < start)((void)0); |
881 | GEM_BUG_ON(vma->node.start + vma->node.size > end)((void)0); |
882 | } |
883 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node))((void)0); |
884 | GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color))((void)0); |
885 | |
886 | list_move_tail(&vma->vm_link, &vma->vm->bound_list); |
887 | |
888 | return 0; |
889 | } |
890 | |
891 | static void |
892 | i915_vma_detach(struct i915_vma *vma) |
893 | { |
894 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node))((void)0); |
895 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND))((void)0); |
896 | |
897 | /* |
898 | * And finally now the object is completely decoupled from this |
899 | * vma, we can drop its hold on the backing storage and allow |
900 | * it to be reaped by the shrinker. |
901 | */ |
902 | list_move_tail(&vma->vm_link, &vma->vm->unbound_list); |
903 | } |
904 | |
905 | static bool_Bool try_qad_pin(struct i915_vma *vma, unsigned int flags) |
906 | { |
907 | unsigned int bound; |
908 | |
909 | bound = atomic_read(&vma->flags)({ typeof(*(&vma->flags)) __tmp = *(volatile typeof(*( &vma->flags)) *)&(*(&vma->flags)); membar_datadep_consumer (); __tmp; }); |
910 | |
911 | if (flags & PIN_VALIDATE(1ULL << (8))) { |
912 | flags &= I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))); |
913 | |
914 | return (flags & bound) == flags; |
915 | } |
916 | |
917 | /* with the lock mandatory for unbind, we don't race here */ |
918 | flags &= I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))); |
919 | do { |
920 | if (unlikely(flags & ~bound)__builtin_expect(!!(flags & ~bound), 0)) |
921 | return false0; |
922 | |
923 | if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))__builtin_expect(!!(bound & (0x200 | ((int)(1UL << ( 12))))), 0)) |
924 | return false0; |
925 | |
926 | GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0)((void)0); |
927 | } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); |
928 | |
929 | return true1; |
930 | } |
931 | |
932 | static struct scatterlist * |
933 | rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, |
934 | unsigned int width, unsigned int height, |
935 | unsigned int src_stride, unsigned int dst_stride, |
936 | struct sg_table *st, struct scatterlist *sg) |
937 | { |
938 | unsigned int column, row; |
939 | unsigned int src_idx; |
940 | |
941 | for (column = 0; column < width; column++) { |
942 | unsigned int left; |
943 | |
944 | src_idx = src_stride * (height - 1) + column + offset; |
945 | for (row = 0; row < height; row++) { |
946 | st->nents++; |
947 | /* |
948 | * We don't need the pages, but need to initialize |
949 | * the entries so the sg list can be happily traversed. |
950 | * The only thing we need are DMA addresses. |
951 | */ |
952 | sg_set_page(sg, NULL((void *)0), I915_GTT_PAGE_SIZE(1ULL << (12)), 0); |
953 | sg_dma_address(sg)((sg)->dma_address) = |
954 | i915_gem_object_get_dma_address(obj, src_idx); |
955 | sg_dma_len(sg)((sg)->length) = I915_GTT_PAGE_SIZE(1ULL << (12)); |
956 | sg = sg_next(sg); |
957 | src_idx -= src_stride; |
958 | } |
959 | |
960 | left = (dst_stride - height) * I915_GTT_PAGE_SIZE(1ULL << (12)); |
961 | |
962 | if (!left) |
963 | continue; |
964 | |
965 | st->nents++; |
966 | |
967 | /* |
968 | * The DE ignores the PTEs for the padding tiles, the sg entry |
969 | * here is just a conenience to indicate how many padding PTEs |
970 | * to insert at this spot. |
971 | */ |
972 | sg_set_page(sg, NULL((void *)0), left, 0); |
973 | sg_dma_address(sg)((sg)->dma_address) = 0; |
974 | sg_dma_len(sg)((sg)->length) = left; |
975 | sg = sg_next(sg); |
976 | } |
977 | |
978 | return sg; |
979 | } |
980 | |
981 | static noinline__attribute__((__noinline__)) struct sg_table * |
982 | intel_rotate_pages(struct intel_rotation_info *rot_info, |
983 | struct drm_i915_gem_object *obj) |
984 | { |
985 | unsigned int size = intel_rotation_info_size(rot_info); |
986 | struct drm_i915_privateinteldrm_softc *i915 = to_i915(obj->base.dev); |
987 | struct sg_table *st; |
988 | struct scatterlist *sg; |
989 | int ret = -ENOMEM12; |
990 | int i; |
991 | |
992 | /* Allocate target SG list. */ |
993 | st = kmalloc(sizeof(*st), GFP_KERNEL(0x0001 | 0x0004)); |
994 | if (!st) |
995 | goto err_st_alloc; |
996 | |
997 | ret = sg_alloc_table(st, size, GFP_KERNEL(0x0001 | 0x0004)); |
998 | if (ret) |
999 | goto err_sg_alloc; |
1000 | |
1001 | st->nents = 0; |
1002 | sg = st->sgl; |
1003 | |
1004 | for (i = 0 ; i < ARRAY_SIZE(rot_info->plane)(sizeof((rot_info->plane)) / sizeof((rot_info->plane)[0 ])); i++) |
1005 | sg = rotate_pages(obj, rot_info->plane[i].offset, |
1006 | rot_info->plane[i].width, rot_info->plane[i].height, |
1007 | rot_info->plane[i].src_stride, |
1008 | rot_info->plane[i].dst_stride, |
1009 | st, sg); |
1010 | |
1011 | return st; |
1012 | |
1013 | err_sg_alloc: |
1014 | kfree(st); |
1015 | err_st_alloc: |
1016 | |
1017 | drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n" , obj->base.size, rot_info->plane[0].width, rot_info-> plane[0].height, size) |
1018 | obj->base.size, rot_info->plane[0].width,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n" , obj->base.size, rot_info->plane[0].width, rot_info-> plane[0].height, size) |
1019 | rot_info->plane[0].height, size)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n" , obj->base.size, rot_info->plane[0].width, rot_info-> plane[0].height, size); |
1020 | |
1021 | return ERR_PTR(ret); |
1022 | } |
1023 | |
1024 | static struct scatterlist * |
1025 | add_padding_pages(unsigned int count, |
1026 | struct sg_table *st, struct scatterlist *sg) |
1027 | { |
1028 | st->nents++; |
1029 | |
1030 | /* |
1031 | * The DE ignores the PTEs for the padding tiles, the sg entry |
1032 | * here is just a convenience to indicate how many padding PTEs |
1033 | * to insert at this spot. |
1034 | */ |
1035 | sg_set_page(sg, NULL((void *)0), count * I915_GTT_PAGE_SIZE(1ULL << (12)), 0); |
1036 | sg_dma_address(sg)((sg)->dma_address) = 0; |
1037 | sg_dma_len(sg)((sg)->length) = count * I915_GTT_PAGE_SIZE(1ULL << (12)); |
1038 | sg = sg_next(sg); |
1039 | |
1040 | return sg; |
1041 | } |
1042 | |
1043 | static struct scatterlist * |
1044 | remap_tiled_color_plane_pages(struct drm_i915_gem_object *obj, |
1045 | unsigned int offset, unsigned int alignment_pad, |
1046 | unsigned int width, unsigned int height, |
1047 | unsigned int src_stride, unsigned int dst_stride, |
1048 | struct sg_table *st, struct scatterlist *sg, |
1049 | unsigned int *gtt_offset) |
1050 | { |
1051 | unsigned int row; |
1052 | |
1053 | if (!width || !height) |
1054 | return sg; |
1055 | |
1056 | if (alignment_pad) |
1057 | sg = add_padding_pages(alignment_pad, st, sg); |
1058 | |
1059 | for (row = 0; row < height; row++) { |
1060 | unsigned int left = width * I915_GTT_PAGE_SIZE(1ULL << (12)); |
1061 | |
1062 | while (left) { |
1063 | dma_addr_t addr; |
1064 | unsigned int length; |
1065 | |
1066 | /* |
1067 | * We don't need the pages, but need to initialize |
1068 | * the entries so the sg list can be happily traversed. |
1069 | * The only thing we need are DMA addresses. |
1070 | */ |
1071 | |
1072 | addr = i915_gem_object_get_dma_address_len(obj, offset, &length); |
1073 | |
1074 | length = min(left, length)(((left)<(length))?(left):(length)); |
1075 | |
1076 | st->nents++; |
1077 | |
1078 | sg_set_page(sg, NULL((void *)0), length, 0); |
1079 | sg_dma_address(sg)((sg)->dma_address) = addr; |
1080 | sg_dma_len(sg)((sg)->length) = length; |
1081 | sg = sg_next(sg); |
1082 | |
1083 | offset += length / I915_GTT_PAGE_SIZE(1ULL << (12)); |
1084 | left -= length; |
1085 | } |
1086 | |
1087 | offset += src_stride - width; |
1088 | |
1089 | left = (dst_stride - width) * I915_GTT_PAGE_SIZE(1ULL << (12)); |
1090 | |
1091 | if (!left) |
1092 | continue; |
1093 | |
1094 | sg = add_padding_pages(left >> PAGE_SHIFT12, st, sg); |
1095 | } |
1096 | |
1097 | *gtt_offset += alignment_pad + dst_stride * height; |
1098 | |
1099 | return sg; |
1100 | } |
1101 | |
1102 | static struct scatterlist * |
1103 | remap_contiguous_pages(struct drm_i915_gem_object *obj, |
1104 | unsigned int obj_offset, |
1105 | unsigned int count, |
1106 | struct sg_table *st, struct scatterlist *sg) |
1107 | { |
1108 | struct scatterlist *iter; |
1109 | unsigned int offset; |
1110 | |
1111 | iter = i915_gem_object_get_sg_dma(obj, obj_offset, &offset); |
1112 | GEM_BUG_ON(!iter)((void)0); |
1113 | |
1114 | do { |
1115 | unsigned int len; |
1116 | |
1117 | len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),(((((iter)->length) - (offset << 12))<(count << 12))?(((iter)->length) - (offset << 12)):(count << 12)) |
1118 | count << PAGE_SHIFT)(((((iter)->length) - (offset << 12))<(count << 12))?(((iter)->length) - (offset << 12)):(count << 12)); |
1119 | sg_set_page(sg, NULL((void *)0), len, 0); |
1120 | sg_dma_address(sg)((sg)->dma_address) = |
1121 | sg_dma_address(iter)((iter)->dma_address) + (offset << PAGE_SHIFT12); |
1122 | sg_dma_len(sg)((sg)->length) = len; |
1123 | |
1124 | st->nents++; |
1125 | count -= len >> PAGE_SHIFT12; |
1126 | if (count == 0) |
1127 | return sg; |
1128 | |
1129 | sg = __sg_next(sg); |
1130 | iter = __sg_next(iter); |
1131 | offset = 0; |
1132 | } while (1); |
1133 | } |
1134 | |
1135 | static struct scatterlist * |
1136 | remap_linear_color_plane_pages(struct drm_i915_gem_object *obj, |
1137 | unsigned int obj_offset, unsigned int alignment_pad, |
1138 | unsigned int size, |
1139 | struct sg_table *st, struct scatterlist *sg, |
1140 | unsigned int *gtt_offset) |
1141 | { |
1142 | if (!size) |
1143 | return sg; |
1144 | |
1145 | if (alignment_pad) |
1146 | sg = add_padding_pages(alignment_pad, st, sg); |
1147 | |
1148 | sg = remap_contiguous_pages(obj, obj_offset, size, st, sg); |
1149 | sg = sg_next(sg); |
1150 | |
1151 | *gtt_offset += alignment_pad + size; |
1152 | |
1153 | return sg; |
1154 | } |
1155 | |
1156 | static struct scatterlist * |
1157 | remap_color_plane_pages(const struct intel_remapped_info *rem_info, |
1158 | struct drm_i915_gem_object *obj, |
1159 | int color_plane, |
1160 | struct sg_table *st, struct scatterlist *sg, |
1161 | unsigned int *gtt_offset) |
1162 | { |
1163 | unsigned int alignment_pad = 0; |
1164 | |
1165 | if (rem_info->plane_alignment) |
1166 | alignment_pad = roundup2(*gtt_offset, rem_info->plane_alignment)(((*gtt_offset) + ((rem_info->plane_alignment) - 1)) & (~((__typeof(*gtt_offset))(rem_info->plane_alignment) - 1 ))) - *gtt_offset; |
1167 | |
1168 | if (rem_info->plane[color_plane].linear) |
1169 | sg = remap_linear_color_plane_pages(obj, |
1170 | rem_info->plane[color_plane].offset, |
1171 | alignment_pad, |
1172 | rem_info->plane[color_plane].size, |
1173 | st, sg, |
1174 | gtt_offset); |
1175 | |
1176 | else |
1177 | sg = remap_tiled_color_plane_pages(obj, |
1178 | rem_info->plane[color_plane].offset, |
1179 | alignment_pad, |
1180 | rem_info->plane[color_plane].width, |
1181 | rem_info->plane[color_plane].height, |
1182 | rem_info->plane[color_plane].src_stride, |
1183 | rem_info->plane[color_plane].dst_stride, |
1184 | st, sg, |
1185 | gtt_offset); |
1186 | |
1187 | return sg; |
1188 | } |
1189 | |
1190 | static noinline__attribute__((__noinline__)) struct sg_table * |
1191 | intel_remap_pages(struct intel_remapped_info *rem_info, |
1192 | struct drm_i915_gem_object *obj) |
1193 | { |
1194 | unsigned int size = intel_remapped_info_size(rem_info); |
1195 | struct drm_i915_privateinteldrm_softc *i915 = to_i915(obj->base.dev); |
1196 | struct sg_table *st; |
1197 | struct scatterlist *sg; |
1198 | unsigned int gtt_offset = 0; |
1199 | int ret = -ENOMEM12; |
1200 | int i; |
1201 | |
1202 | /* Allocate target SG list. */ |
1203 | st = kmalloc(sizeof(*st), GFP_KERNEL(0x0001 | 0x0004)); |
1204 | if (!st) |
1205 | goto err_st_alloc; |
1206 | |
1207 | ret = sg_alloc_table(st, size, GFP_KERNEL(0x0001 | 0x0004)); |
1208 | if (ret) |
1209 | goto err_sg_alloc; |
1210 | |
1211 | st->nents = 0; |
1212 | sg = st->sgl; |
1213 | |
1214 | for (i = 0 ; i < ARRAY_SIZE(rem_info->plane)(sizeof((rem_info->plane)) / sizeof((rem_info->plane)[0 ])); i++) |
1215 | sg = remap_color_plane_pages(rem_info, obj, i, st, sg, >t_offset); |
1216 | |
1217 | i915_sg_trim(st); |
1218 | |
1219 | return st; |
1220 | |
1221 | err_sg_alloc: |
1222 | kfree(st); |
1223 | err_st_alloc: |
1224 | |
1225 | drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n" , obj->base.size, rem_info->plane[0].width, rem_info-> plane[0].height, size) |
1226 | obj->base.size, rem_info->plane[0].width,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n" , obj->base.size, rem_info->plane[0].width, rem_info-> plane[0].height, size) |
1227 | rem_info->plane[0].height, size)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n" , obj->base.size, rem_info->plane[0].width, rem_info-> plane[0].height, size); |
1228 | |
1229 | return ERR_PTR(ret); |
1230 | } |
1231 | |
1232 | static noinline__attribute__((__noinline__)) struct sg_table * |
1233 | intel_partial_pages(const struct i915_gtt_view *view, |
1234 | struct drm_i915_gem_object *obj) |
1235 | { |
1236 | struct sg_table *st; |
1237 | struct scatterlist *sg; |
1238 | unsigned int count = view->partial.size; |
1239 | int ret = -ENOMEM12; |
1240 | |
1241 | st = kmalloc(sizeof(*st), GFP_KERNEL(0x0001 | 0x0004)); |
1242 | if (!st) |
1243 | goto err_st_alloc; |
1244 | |
1245 | ret = sg_alloc_table(st, count, GFP_KERNEL(0x0001 | 0x0004)); |
1246 | if (ret) |
1247 | goto err_sg_alloc; |
1248 | |
1249 | st->nents = 0; |
1250 | |
1251 | sg = remap_contiguous_pages(obj, view->partial.offset, count, st, st->sgl); |
1252 | |
1253 | sg_mark_end(sg); |
1254 | i915_sg_trim(st); /* Drop any unused tail entries. */ |
1255 | |
1256 | return st; |
1257 | |
1258 | err_sg_alloc: |
1259 | kfree(st); |
1260 | err_st_alloc: |
1261 | return ERR_PTR(ret); |
1262 | } |
1263 | |
1264 | static int |
1265 | __i915_vma_get_pages(struct i915_vma *vma) |
1266 | { |
1267 | struct sg_table *pages; |
1268 | |
1269 | /* |
1270 | * The vma->pages are only valid within the lifespan of the borrowed |
1271 | * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so |
1272 | * must be the vma->pages. A simple rule is that vma->pages must only |
1273 | * be accessed when the obj->mm.pages are pinned. |
1274 | */ |
1275 | GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj))((void)0); |
1276 | |
1277 | switch (vma->gtt_view.type) { |
1278 | default: |
1279 | GEM_BUG_ON(vma->gtt_view.type)((void)0); |
1280 | fallthroughdo {} while (0); |
1281 | case I915_GTT_VIEW_NORMAL: |
1282 | pages = vma->obj->mm.pages; |
1283 | break; |
1284 | |
1285 | case I915_GTT_VIEW_ROTATED: |
1286 | pages = |
1287 | intel_rotate_pages(&vma->gtt_view.rotated, vma->obj); |
1288 | break; |
1289 | |
1290 | case I915_GTT_VIEW_REMAPPED: |
1291 | pages = |
1292 | intel_remap_pages(&vma->gtt_view.remapped, vma->obj); |
1293 | break; |
1294 | |
1295 | case I915_GTT_VIEW_PARTIAL: |
1296 | pages = intel_partial_pages(&vma->gtt_view, vma->obj); |
1297 | break; |
1298 | } |
1299 | |
1300 | if (IS_ERR(pages)) { |
1301 | drm_err(&vma->vm->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to get pages for VMA view type %u (%ld)!\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , vma-> gtt_view.type, PTR_ERR(pages)) |
1302 | "Failed to get pages for VMA view type %u (%ld)!\n",printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to get pages for VMA view type %u (%ld)!\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , vma-> gtt_view.type, PTR_ERR(pages)) |
1303 | vma->gtt_view.type, PTR_ERR(pages))printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to get pages for VMA view type %u (%ld)!\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , vma-> gtt_view.type, PTR_ERR(pages)); |
1304 | return PTR_ERR(pages); |
1305 | } |
1306 | |
1307 | vma->pages = pages; |
1308 | |
1309 | return 0; |
1310 | } |
1311 | |
1312 | I915_SELFTEST_EXPORTstatic int i915_vma_get_pages(struct i915_vma *vma) |
1313 | { |
1314 | int err; |
1315 | |
1316 | if (atomic_add_unless(&vma->pages_count, 1, 0)) |
1317 | return 0; |
1318 | |
1319 | err = i915_gem_object_pin_pages(vma->obj); |
1320 | if (err) |
1321 | return err; |
1322 | |
1323 | err = __i915_vma_get_pages(vma); |
1324 | if (err) |
1325 | goto err_unpin; |
1326 | |
1327 | vma->page_sizes = vma->obj->mm.page_sizes; |
1328 | atomic_inc(&vma->pages_count)__sync_fetch_and_add(&vma->pages_count, 1); |
1329 | |
1330 | return 0; |
1331 | |
1332 | err_unpin: |
1333 | __i915_gem_object_unpin_pages(vma->obj); |
1334 | |
1335 | return err; |
1336 | } |
1337 | |
1338 | void vma_invalidate_tlb(struct i915_address_space *vm, u32 *tlb) |
1339 | { |
1340 | /* |
1341 | * Before we release the pages that were bound by this vma, we |
1342 | * must invalidate all the TLBs that may still have a reference |
1343 | * back to our physical address. It only needs to be done once, |
1344 | * so after updating the PTE to point away from the pages, record |
1345 | * the most recent TLB invalidation seqno, and if we have not yet |
1346 | * flushed the TLBs upon release, perform a full invalidation. |
1347 | */ |
1348 | WRITE_ONCE(*tlb, intel_gt_next_invalidate_tlb_full(vm->gt))({ typeof(*tlb) __tmp = (intel_gt_next_invalidate_tlb_full(vm ->gt)); *(volatile typeof(*tlb) *)&(*tlb) = __tmp; __tmp ; }); |
1349 | } |
1350 | |
1351 | static void __vma_put_pages(struct i915_vma *vma, unsigned int count) |
1352 | { |
1353 | /* We allocate under vma_get_pages, so beware the shrinker */ |
1354 | GEM_BUG_ON(atomic_read(&vma->pages_count) < count)((void)0); |
1355 | |
1356 | if (atomic_sub_return(count, &vma->pages_count)__sync_sub_and_fetch(&vma->pages_count, count) == 0) { |
1357 | if (vma->pages != vma->obj->mm.pages) { |
1358 | sg_free_table(vma->pages); |
1359 | kfree(vma->pages); |
1360 | } |
1361 | vma->pages = NULL((void *)0); |
1362 | |
1363 | i915_gem_object_unpin_pages(vma->obj); |
1364 | } |
1365 | } |
1366 | |
1367 | I915_SELFTEST_EXPORTstatic void i915_vma_put_pages(struct i915_vma *vma) |
1368 | { |
1369 | if (atomic_add_unless(&vma->pages_count, -1, 1)) |
1370 | return; |
1371 | |
1372 | __vma_put_pages(vma, 1); |
1373 | } |
1374 | |
1375 | static void vma_unbind_pages(struct i915_vma *vma) |
1376 | { |
1377 | unsigned int count; |
1378 | |
1379 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
1380 | |
1381 | /* The upper portion of pages_count is the number of bindings */ |
1382 | count = atomic_read(&vma->pages_count)({ typeof(*(&vma->pages_count)) __tmp = *(volatile typeof (*(&vma->pages_count)) *)&(*(&vma->pages_count )); membar_datadep_consumer(); __tmp; }); |
1383 | count >>= I915_VMA_PAGES_BIAS24; |
1384 | GEM_BUG_ON(!count)((void)0); |
1385 | |
1386 | __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS24); |
1387 | } |
1388 | |
1389 | int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
1390 | u64 size, u64 alignment, u64 flags) |
1391 | { |
1392 | struct i915_vma_work *work = NULL((void *)0); |
1393 | struct dma_fence *moving = NULL((void *)0); |
1394 | struct i915_vma_resource *vma_res = NULL((void *)0); |
1395 | intel_wakeref_t wakeref = 0; |
1396 | unsigned int bound; |
1397 | int err; |
1398 | |
1399 | assert_vma_held(vma)do { (void)(&((vma)->obj->base.resv)->lock.base) ; } while(0); |
1400 | GEM_BUG_ON(!ww)((void)0); |
1401 | |
1402 | BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND)extern char _ctassert[(!((1ULL << (10)) != ((int)(1UL << (10))))) ? 1 : -1 ] __attribute__((__unused__)); |
1403 | BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND)extern char _ctassert[(!((1ULL << (11)) != ((int)(1UL << (11))))) ? 1 : -1 ] __attribute__((__unused__)); |
1404 | |
1405 | GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)))((void)0); |
1406 | |
1407 | /* First try and grab the pin without rebinding the vma */ |
1408 | if (try_qad_pin(vma, flags)) |
1409 | return 0; |
1410 | |
1411 | err = i915_vma_get_pages(vma); |
1412 | if (err) |
1413 | return err; |
1414 | |
1415 | if (flags & PIN_GLOBAL(1ULL << (10))) |
1416 | wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm); |
1417 | |
1418 | if (flags & vma->vm->bind_async_flags) { |
1419 | /* lock VM */ |
1420 | err = i915_vm_lock_objects(vma->vm, ww); |
1421 | if (err) |
1422 | goto err_rpm; |
1423 | |
1424 | work = i915_vma_work(); |
1425 | if (!work) { |
1426 | err = -ENOMEM12; |
1427 | goto err_rpm; |
1428 | } |
1429 | |
1430 | work->vm = vma->vm; |
1431 | |
1432 | err = i915_gem_object_get_moving_fence(vma->obj, &moving); |
1433 | if (err) |
1434 | goto err_rpm; |
1435 | |
1436 | dma_fence_work_chain(&work->base, moving); |
1437 | |
1438 | /* Allocate enough page directories to used PTE */ |
1439 | if (vma->vm->allocate_va_range) { |
1440 | err = i915_vm_alloc_pt_stash(vma->vm, |
1441 | &work->stash, |
1442 | vma->size); |
1443 | if (err) |
1444 | goto err_fence; |
1445 | |
1446 | err = i915_vm_map_pt_stash(vma->vm, &work->stash); |
1447 | if (err) |
1448 | goto err_fence; |
1449 | } |
1450 | } |
1451 | |
1452 | vma_res = i915_vma_resource_alloc(); |
1453 | if (IS_ERR(vma_res)) { |
1454 | err = PTR_ERR(vma_res); |
1455 | goto err_fence; |
1456 | } |
1457 | |
1458 | /* |
1459 | * Differentiate between user/kernel vma inside the aliasing-ppgtt. |
1460 | * |
1461 | * We conflate the Global GTT with the user's vma when using the |
1462 | * aliasing-ppgtt, but it is still vitally important to try and |
1463 | * keep the use cases distinct. For example, userptr objects are |
1464 | * not allowed inside the Global GTT as that will cause lock |
1465 | * inversions when we have to evict them the mmu_notifier callbacks - |
1466 | * but they are allowed to be part of the user ppGTT which can never |
1467 | * be mapped. As such we try to give the distinct users of the same |
1468 | * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt |
1469 | * and i915_ppgtt separate]. |
1470 | * |
1471 | * NB this may cause us to mask real lock inversions -- while the |
1472 | * code is safe today, lockdep may not be able to spot future |
1473 | * transgressions. |
1474 | */ |
1475 | err = mutex_lock_interruptible_nested(&vma->vm->mutex,mutex_lock_interruptible(&vma->vm->mutex) |
1476 | !(flags & PIN_GLOBAL))mutex_lock_interruptible(&vma->vm->mutex); |
1477 | if (err) |
1478 | goto err_vma_res; |
1479 | |
1480 | /* No more allocations allowed now we hold vm->mutex */ |
1481 | |
1482 | if (unlikely(i915_vma_is_closed(vma))__builtin_expect(!!(i915_vma_is_closed(vma)), 0)) { |
1483 | err = -ENOENT2; |
1484 | goto err_unlock; |
1485 | } |
1486 | |
1487 | bound = atomic_read(&vma->flags)({ typeof(*(&vma->flags)) __tmp = *(volatile typeof(*( &vma->flags)) *)&(*(&vma->flags)); membar_datadep_consumer (); __tmp; }); |
1488 | if (unlikely(bound & I915_VMA_ERROR)__builtin_expect(!!(bound & ((int)(1UL << (12)))), 0 )) { |
1489 | err = -ENOMEM12; |
1490 | goto err_unlock; |
1491 | } |
1492 | |
1493 | if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))__builtin_expect(!!(!((bound + 1) & 0x3ff)), 0)) { |
1494 | err = -EAGAIN35; /* pins are meant to be fairly temporary */ |
1495 | goto err_unlock; |
1496 | } |
1497 | |
1498 | if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))__builtin_expect(!!(!(flags & ~bound & (((int)(1UL << (10))) | ((int)(1UL << (11)))))), 0)) { |
1499 | if (!(flags & PIN_VALIDATE(1ULL << (8)))) |
1500 | __i915_vma_pin(vma); |
1501 | goto err_unlock; |
1502 | } |
1503 | |
1504 | err = i915_active_acquire(&vma->active); |
1505 | if (err) |
1506 | goto err_unlock; |
1507 | |
1508 | if (!(bound & I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))))) { |
1509 | err = i915_vma_insert(vma, ww, size, alignment, flags); |
1510 | if (err) |
1511 | goto err_active; |
1512 | |
1513 | if (i915_is_ggtt(vma->vm)((vma->vm)->is_ggtt)) |
1514 | __i915_vma_set_map_and_fenceable(vma); |
1515 | } |
1516 | |
1517 | GEM_BUG_ON(!vma->pages)((void)0); |
1518 | err = i915_vma_bind(vma, |
1519 | vma->obj->cache_level, |
1520 | flags, work, vma_res); |
1521 | vma_res = NULL((void *)0); |
1522 | if (err) |
1523 | goto err_remove; |
1524 | |
1525 | /* There should only be at most 2 active bindings (user, global) */ |
1526 | GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound)((void)0); |
1527 | atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count)__sync_fetch_and_add(&vma->pages_count, ((1UL << (24)) | 1)); |
1528 | list_move_tail(&vma->vm_link, &vma->vm->bound_list); |
1529 | |
1530 | if (!(flags & PIN_VALIDATE(1ULL << (8)))) { |
1531 | __i915_vma_pin(vma); |
1532 | GEM_BUG_ON(!i915_vma_is_pinned(vma))((void)0); |
1533 | } |
1534 | GEM_BUG_ON(!i915_vma_is_bound(vma, flags))((void)0); |
1535 | GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags))((void)0); |
1536 | |
1537 | err_remove: |
1538 | if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))))) { |
1539 | i915_vma_detach(vma); |
1540 | drm_mm_remove_node(&vma->node); |
1541 | } |
1542 | err_active: |
1543 | i915_active_release(&vma->active); |
1544 | err_unlock: |
1545 | mutex_unlock(&vma->vm->mutex)rw_exit_write(&vma->vm->mutex); |
1546 | err_vma_res: |
1547 | i915_vma_resource_free(vma_res); |
1548 | err_fence: |
1549 | if (work) |
1550 | dma_fence_work_commit_imm(&work->base); |
1551 | err_rpm: |
1552 | if (wakeref) |
1553 | intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); |
1554 | |
1555 | if (moving) |
1556 | dma_fence_put(moving); |
1557 | |
1558 | i915_vma_put_pages(vma); |
1559 | return err; |
1560 | } |
1561 | |
1562 | static void flush_idle_contexts(struct intel_gt *gt) |
1563 | { |
1564 | struct intel_engine_cs *engine; |
1565 | enum intel_engine_id id; |
1566 | |
1567 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
1568 | intel_engine_flush_barriers(engine); |
1569 | |
1570 | intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT(0x7fffffff)); |
1571 | } |
1572 | |
1573 | static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
1574 | u32 align, unsigned int flags) |
1575 | { |
1576 | struct i915_address_space *vm = vma->vm; |
1577 | int err; |
1578 | |
1579 | do { |
1580 | err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL(1ULL << (10))); |
1581 | |
1582 | if (err != -ENOSPC28) { |
1583 | if (!err) { |
1584 | err = i915_vma_wait_for_bind(vma); |
1585 | if (err) |
1586 | i915_vma_unpin(vma); |
1587 | } |
1588 | return err; |
1589 | } |
1590 | |
1591 | /* Unlike i915_vma_pin, we don't take no for an answer! */ |
1592 | flush_idle_contexts(vm->gt); |
1593 | if (mutex_lock_interruptible(&vm->mutex) == 0) { |
1594 | /* |
1595 | * We pass NULL ww here, as we don't want to unbind |
1596 | * locked objects when called from execbuf when pinning |
1597 | * is removed. This would probably regress badly. |
1598 | */ |
1599 | i915_gem_evict_vm(vm, NULL((void *)0), NULL((void *)0)); |
1600 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
1601 | } |
1602 | } while (1); |
1603 | } |
1604 | |
1605 | int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
1606 | u32 align, unsigned int flags) |
1607 | { |
1608 | struct i915_gem_ww_ctx _ww; |
1609 | int err; |
1610 | |
1611 | GEM_BUG_ON(!i915_vma_is_ggtt(vma))((void)0); |
1612 | |
1613 | if (ww) |
1614 | return __i915_ggtt_pin(vma, ww, align, flags); |
1615 | |
1616 | lockdep_assert_not_held(&vma->obj->base.resv->lock.base)do { (void)(&vma->obj->base.resv->lock.base); } while (0); |
1617 | |
1618 | for_i915_gem_ww(&_ww, err, true)for (i915_gem_ww_ctx_init(&_ww, 1), (err) = -11; (err) == -11; (err) = __i915_gem_ww_fini(&_ww, err)) { |
1619 | err = i915_gem_object_lock(vma->obj, &_ww); |
1620 | if (!err) |
1621 | err = __i915_ggtt_pin(vma, &_ww, align, flags); |
1622 | } |
1623 | |
1624 | return err; |
1625 | } |
1626 | |
1627 | static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) |
1628 | { |
1629 | /* |
1630 | * We defer actually closing, unbinding and destroying the VMA until |
1631 | * the next idle point, or if the object is freed in the meantime. By |
1632 | * postponing the unbind, we allow for it to be resurrected by the |
1633 | * client, avoiding the work required to rebind the VMA. This is |
1634 | * advantageous for DRI, where the client/server pass objects |
1635 | * between themselves, temporarily opening a local VMA to the |
1636 | * object, and then closing it again. The same object is then reused |
1637 | * on the next frame (or two, depending on the depth of the swap queue) |
1638 | * causing us to rebind the VMA once more. This ends up being a lot |
1639 | * of wasted work for the steady state. |
1640 | */ |
1641 | GEM_BUG_ON(i915_vma_is_closed(vma))((void)0); |
1642 | list_add(&vma->closed_link, >->closed_vma); |
1643 | } |
1644 | |
1645 | void i915_vma_close(struct i915_vma *vma) |
1646 | { |
1647 | struct intel_gt *gt = vma->vm->gt; |
1648 | unsigned long flags; |
1649 | |
1650 | if (i915_vma_is_ggtt(vma)) |
1651 | return; |
1652 | |
1653 | GEM_BUG_ON(!atomic_read(&vma->open_count))((void)0); |
1654 | if (atomic_dec_and_lock_irqsave(&vma->open_count,atomic_dec_and_lock(&vma->open_count, >->closed_lock ) |
1655 | >->closed_lock,atomic_dec_and_lock(&vma->open_count, >->closed_lock ) |
1656 | flags)atomic_dec_and_lock(&vma->open_count, >->closed_lock )) { |
1657 | __vma_close(vma, gt); |
1658 | spin_unlock_irqrestore(>->closed_lock, flags)do { (void)(flags); mtx_leave(>->closed_lock); } while (0); |
1659 | } |
1660 | } |
1661 | |
1662 | static void __i915_vma_remove_closed(struct i915_vma *vma) |
1663 | { |
1664 | list_del_init(&vma->closed_link); |
1665 | } |
1666 | |
1667 | void i915_vma_reopen(struct i915_vma *vma) |
1668 | { |
1669 | struct intel_gt *gt = vma->vm->gt; |
1670 | |
1671 | spin_lock_irq(>->closed_lock)mtx_enter(>->closed_lock); |
1672 | if (i915_vma_is_closed(vma)) |
1673 | __i915_vma_remove_closed(vma); |
1674 | spin_unlock_irq(>->closed_lock)mtx_leave(>->closed_lock); |
1675 | } |
1676 | |
1677 | static void force_unbind(struct i915_vma *vma) |
1678 | { |
1679 | if (!drm_mm_node_allocated(&vma->node)) |
1680 | return; |
1681 | |
1682 | atomic_and(~I915_VMA_PIN_MASK, &vma->flags)__sync_fetch_and_and(&vma->flags, ~0x3ff); |
1683 | WARN_ON(__i915_vma_unbind(vma))({ int __ret = !!(__i915_vma_unbind(vma)); if (__ret) printf( "WARNING %s failed at %s:%d\n", "__i915_vma_unbind(vma)", "/usr/src/sys/dev/pci/drm/i915/i915_vma.c" , 1683); __builtin_expect(!!(__ret), 0); }); |
1684 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node))((void)0); |
1685 | } |
1686 | |
1687 | static void release_references(struct i915_vma *vma, struct intel_gt *gt, |
1688 | bool_Bool vm_ddestroy) |
1689 | { |
1690 | struct drm_i915_gem_object *obj = vma->obj; |
1691 | |
1692 | GEM_BUG_ON(i915_vma_is_active(vma))((void)0); |
1693 | |
1694 | spin_lock(&obj->vma.lock)mtx_enter(&obj->vma.lock); |
1695 | list_del(&vma->obj_link); |
1696 | if (!RB_EMPTY_NODE(&vma->obj_node)((&vma->obj_node)->__entry.rbe_parent == &vma-> obj_node)) |
1697 | rb_erase(&vma->obj_node, &obj->vma.tree)linux_root_RB_REMOVE((struct linux_root *)(&obj->vma.tree ), (&vma->obj_node)); |
1698 | |
1699 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
1700 | |
1701 | spin_lock_irq(>->closed_lock)mtx_enter(>->closed_lock); |
1702 | __i915_vma_remove_closed(vma); |
1703 | spin_unlock_irq(>->closed_lock)mtx_leave(>->closed_lock); |
1704 | |
1705 | if (vm_ddestroy) |
1706 | i915_vm_resv_put(vma->vm); |
1707 | |
1708 | i915_active_fini(&vma->active); |
1709 | GEM_WARN_ON(vma->resource)({ __builtin_expect(!!(!!(vma->resource)), 0); }); |
1710 | i915_vma_free(vma); |
1711 | } |
1712 | |
1713 | /** |
1714 | * i915_vma_destroy_locked - Remove all weak reference to the vma and put |
1715 | * the initial reference. |
1716 | * |
1717 | * This function should be called when it's decided the vma isn't needed |
1718 | * anymore. The caller must assure that it doesn't race with another lookup |
1719 | * plus destroy, typically by taking an appropriate reference. |
1720 | * |
1721 | * Current callsites are |
1722 | * - __i915_gem_object_pages_fini() |
1723 | * - __i915_vm_close() - Blocks the above function by taking a reference on |
1724 | * the object. |
1725 | * - __i915_vma_parked() - Blocks the above functions by taking a reference |
1726 | * on the vm and a reference on the object. Also takes the object lock so |
1727 | * destruction from __i915_vma_parked() can be blocked by holding the |
1728 | * object lock. Since the object lock is only allowed from within i915 with |
1729 | * an object refcount, holding the object lock also implicitly blocks the |
1730 | * vma freeing from __i915_gem_object_pages_fini(). |
1731 | * |
1732 | * Because of locks taken during destruction, a vma is also guaranteed to |
1733 | * stay alive while the following locks are held if it was looked up while |
1734 | * holding one of the locks: |
1735 | * - vm->mutex |
1736 | * - obj->vma.lock |
1737 | * - gt->closed_lock |
1738 | */ |
1739 | void i915_vma_destroy_locked(struct i915_vma *vma) |
1740 | { |
1741 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
1742 | |
1743 | force_unbind(vma); |
1744 | list_del_init(&vma->vm_link); |
1745 | release_references(vma, vma->vm->gt, false0); |
1746 | } |
1747 | |
1748 | void i915_vma_destroy(struct i915_vma *vma) |
1749 | { |
1750 | struct intel_gt *gt; |
1751 | bool_Bool vm_ddestroy; |
1752 | |
1753 | mutex_lock(&vma->vm->mutex)rw_enter_write(&vma->vm->mutex); |
1754 | force_unbind(vma); |
1755 | list_del_init(&vma->vm_link); |
1756 | vm_ddestroy = vma->vm_ddestroy; |
1757 | vma->vm_ddestroy = false0; |
1758 | |
1759 | /* vma->vm may be freed when releasing vma->vm->mutex. */ |
1760 | gt = vma->vm->gt; |
1761 | mutex_unlock(&vma->vm->mutex)rw_exit_write(&vma->vm->mutex); |
1762 | release_references(vma, gt, vm_ddestroy); |
1763 | } |
1764 | |
1765 | void i915_vma_parked(struct intel_gt *gt) |
1766 | { |
1767 | struct i915_vma *vma, *next; |
1768 | DRM_LIST_HEAD(closed)struct list_head closed = { &(closed), &(closed) }; |
1769 | |
1770 | spin_lock_irq(>->closed_lock)mtx_enter(>->closed_lock); |
1771 | list_for_each_entry_safe(vma, next, >->closed_vma, closed_link)for (vma = ({ const __typeof( ((__typeof(*vma) *)0)->closed_link ) *__mptr = ((>->closed_vma)->next); (__typeof(* vma) *)( (char *)__mptr - __builtin_offsetof(__typeof(*vma), closed_link ) );}), next = ({ const __typeof( ((__typeof(*vma) *)0)->closed_link ) *__mptr = (vma->closed_link.next); (__typeof(*vma) *)( ( char *)__mptr - __builtin_offsetof(__typeof(*vma), closed_link ) );}); &vma->closed_link != (>->closed_vma); vma = next, next = ({ const __typeof( ((__typeof(*next) *)0) ->closed_link ) *__mptr = (next->closed_link.next); (__typeof (*next) *)( (char *)__mptr - __builtin_offsetof(__typeof(*next ), closed_link) );})) { |
1772 | struct drm_i915_gem_object *obj = vma->obj; |
1773 | struct i915_address_space *vm = vma->vm; |
1774 | |
1775 | /* XXX All to avoid keeping a reference on i915_vma itself */ |
1776 | |
1777 | if (!kref_get_unless_zero(&obj->base.refcount)) |
1778 | continue; |
1779 | |
1780 | if (!i915_vm_tryget(vm)) { |
1781 | i915_gem_object_put(obj); |
1782 | continue; |
1783 | } |
1784 | |
1785 | list_move(&vma->closed_link, &closed); |
1786 | } |
1787 | spin_unlock_irq(>->closed_lock)mtx_leave(>->closed_lock); |
1788 | |
1789 | /* As the GT is held idle, no vma can be reopened as we destroy them */ |
1790 | list_for_each_entry_safe(vma, next, &closed, closed_link)for (vma = ({ const __typeof( ((__typeof(*vma) *)0)->closed_link ) *__mptr = ((&closed)->next); (__typeof(*vma) *)( (char *)__mptr - __builtin_offsetof(__typeof(*vma), closed_link) ) ;}), next = ({ const __typeof( ((__typeof(*vma) *)0)->closed_link ) *__mptr = (vma->closed_link.next); (__typeof(*vma) *)( ( char *)__mptr - __builtin_offsetof(__typeof(*vma), closed_link ) );}); &vma->closed_link != (&closed); vma = next , next = ({ const __typeof( ((__typeof(*next) *)0)->closed_link ) *__mptr = (next->closed_link.next); (__typeof(*next) *) ( (char *)__mptr - __builtin_offsetof(__typeof(*next), closed_link ) );})) { |
1791 | struct drm_i915_gem_object *obj = vma->obj; |
1792 | struct i915_address_space *vm = vma->vm; |
1793 | |
1794 | if (i915_gem_object_trylock(obj, NULL((void *)0))) { |
1795 | INIT_LIST_HEAD(&vma->closed_link); |
1796 | i915_vma_destroy(vma); |
1797 | i915_gem_object_unlock(obj); |
1798 | } else { |
1799 | /* back you go.. */ |
1800 | spin_lock_irq(>->closed_lock)mtx_enter(>->closed_lock); |
1801 | list_add(&vma->closed_link, >->closed_vma); |
1802 | spin_unlock_irq(>->closed_lock)mtx_leave(>->closed_lock); |
1803 | } |
1804 | |
1805 | i915_gem_object_put(obj); |
1806 | i915_vm_put(vm); |
1807 | } |
1808 | } |
1809 | |
1810 | static void __i915_vma_iounmap(struct i915_vma *vma) |
1811 | { |
1812 | GEM_BUG_ON(i915_vma_is_pinned(vma))((void)0); |
1813 | |
1814 | if (vma->iomap == NULL((void *)0)) |
1815 | return; |
1816 | |
1817 | if (page_unmask_bits(vma->iomap)((unsigned long)(vma->iomap) & ((1UL << (12)) - 1 ))) |
1818 | __i915_gem_object_release_map(vma->obj); |
1819 | else { |
1820 | #ifdef __linux__ |
1821 | io_mapping_unmap(vma->iomap); |
1822 | #else |
1823 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->vm->i915; |
1824 | agp_unmap_subregion(dev_priv->agph, vma->bsh, vma->node.size); |
1825 | #endif |
1826 | } |
1827 | vma->iomap = NULL((void *)0); |
1828 | } |
1829 | |
1830 | void i915_vma_revoke_mmap(struct i915_vma *vma) |
1831 | { |
1832 | struct drm_vma_offset_node *node; |
1833 | u64 vma_offset; |
1834 | |
1835 | if (!i915_vma_has_userfault(vma)) |
1836 | return; |
1837 | |
1838 | GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma))((void)0); |
1839 | GEM_BUG_ON(!vma->obj->userfault_count)((void)0); |
1840 | |
1841 | node = &vma->mmo->vma_node; |
Value stored to 'node' is never read | |
1842 | vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT12; |
1843 | #ifdef __linux__ |
1844 | unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping, |
1845 | drm_vma_node_offset_addr(node) + vma_offset, |
1846 | vma->size, |
1847 | 1); |
1848 | #else |
1849 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->obj->base.dev->dev_private; |
1850 | struct vm_page *pg; |
1851 | |
1852 | for (pg = &dev_priv->pgs[atop(vma->node.start)((vma->node.start) >> 12)]; |
1853 | pg != &dev_priv->pgs[atop(vma->node.start + vma->size)((vma->node.start + vma->size) >> 12)]; |
1854 | pg++) |
1855 | pmap_page_protect(pg, PROT_NONE0x00); |
1856 | #endif |
1857 | |
1858 | i915_vma_unset_userfault(vma); |
1859 | if (!--vma->obj->userfault_count) |
1860 | list_del(&vma->obj->userfault_link); |
1861 | } |
1862 | |
1863 | static int |
1864 | __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) |
1865 | { |
1866 | return __i915_request_await_exclusive(rq, &vma->active); |
1867 | } |
1868 | |
1869 | static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) |
1870 | { |
1871 | int err; |
1872 | |
1873 | /* Wait for the vma to be bound before we start! */ |
1874 | err = __i915_request_await_bind(rq, vma); |
1875 | if (err) |
1876 | return err; |
1877 | |
1878 | return i915_active_add_request(&vma->active, rq); |
1879 | } |
1880 | |
1881 | int _i915_vma_move_to_active(struct i915_vma *vma, |
1882 | struct i915_request *rq, |
1883 | struct dma_fence *fence, |
1884 | unsigned int flags) |
1885 | { |
1886 | struct drm_i915_gem_object *obj = vma->obj; |
1887 | int err; |
1888 | |
1889 | assert_object_held(obj)do { (void)(&((obj)->base.resv)->lock.base); } while (0); |
1890 | |
1891 | GEM_BUG_ON(!vma->pages)((void)0); |
1892 | |
1893 | err = __i915_vma_move_to_active(vma, rq); |
1894 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
1895 | return err; |
1896 | |
1897 | /* |
1898 | * Reserve fences slot early to prevent an allocation after preparing |
1899 | * the workload and associating fences with dma_resv. |
1900 | */ |
1901 | if (fence && !(flags & __EXEC_OBJECT_NO_RESERVE(1UL << (31)))) { |
1902 | struct dma_fence *curr; |
1903 | int idx; |
1904 | |
1905 | dma_fence_array_for_each(curr, idx, fence)for (idx = 0, curr = dma_fence_array_first(fence); curr != (( void *)0); ++(idx), curr = dma_fence_array_next(fence, idx)) |
1906 | ; |
1907 | err = dma_resv_reserve_fences(vma->obj->base.resv, idx); |
1908 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
1909 | return err; |
1910 | } |
1911 | |
1912 | if (flags & EXEC_OBJECT_WRITE(1<<2)) { |
1913 | struct intel_frontbuffer *front; |
1914 | |
1915 | front = __intel_frontbuffer_get(obj); |
1916 | if (unlikely(front)__builtin_expect(!!(front), 0)) { |
1917 | if (intel_frontbuffer_invalidate(front, ORIGIN_CS)) |
1918 | i915_active_add_request(&front->write, rq); |
1919 | intel_frontbuffer_put(front); |
1920 | } |
1921 | } |
1922 | |
1923 | if (fence) { |
1924 | struct dma_fence *curr; |
1925 | enum dma_resv_usage usage; |
1926 | int idx; |
1927 | |
1928 | if (flags & EXEC_OBJECT_WRITE(1<<2)) { |
1929 | usage = DMA_RESV_USAGE_WRITE; |
1930 | obj->write_domain = I915_GEM_DOMAIN_RENDER0x00000002; |
1931 | obj->read_domains = 0; |
1932 | } else { |
1933 | usage = DMA_RESV_USAGE_READ; |
1934 | obj->write_domain = 0; |
1935 | } |
1936 | |
1937 | dma_fence_array_for_each(curr, idx, fence)for (idx = 0, curr = dma_fence_array_first(fence); curr != (( void *)0); ++(idx), curr = dma_fence_array_next(fence, idx)) |
1938 | dma_resv_add_fence(vma->obj->base.resv, curr, usage); |
1939 | } |
1940 | |
1941 | if (flags & EXEC_OBJECT_NEEDS_FENCE(1<<0) && vma->fence) |
1942 | i915_active_add_request(&vma->fence->active, rq); |
1943 | |
1944 | obj->read_domains |= I915_GEM_GPU_DOMAINS(0x00000002 | 0x00000004 | 0x00000008 | 0x00000010 | 0x00000020 ); |
1945 | obj->mm.dirty = true1; |
1946 | |
1947 | GEM_BUG_ON(!i915_vma_is_active(vma))((void)0); |
1948 | return 0; |
1949 | } |
1950 | |
1951 | struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool_Bool async) |
1952 | { |
1953 | struct i915_vma_resource *vma_res = vma->resource; |
1954 | struct dma_fence *unbind_fence; |
1955 | |
1956 | GEM_BUG_ON(i915_vma_is_pinned(vma))((void)0); |
1957 | assert_vma_held_evict(vma); |
1958 | |
1959 | if (i915_vma_is_map_and_fenceable(vma)) { |
1960 | /* Force a pagefault for domain tracking on next user access */ |
1961 | i915_vma_revoke_mmap(vma); |
1962 | |
1963 | /* |
1964 | * Check that we have flushed all writes through the GGTT |
1965 | * before the unbind, other due to non-strict nature of those |
1966 | * indirect writes they may end up referencing the GGTT PTE |
1967 | * after the unbind. |
1968 | * |
1969 | * Note that we may be concurrently poking at the GGTT_WRITE |
1970 | * bit from set-domain, as we mark all GGTT vma associated |
1971 | * with an object. We know this is for another vma, as we |
1972 | * are currently unbinding this one -- so if this vma will be |
1973 | * reused, it will be refaulted and have its dirty bit set |
1974 | * before the next write. |
1975 | */ |
1976 | i915_vma_flush_writes(vma); |
1977 | |
1978 | /* release the fence reg _after_ flushing */ |
1979 | i915_vma_revoke_fence(vma); |
1980 | |
1981 | clear_bit(I915_VMA_CAN_FENCE_BIT14, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
1982 | } |
1983 | |
1984 | __i915_vma_iounmap(vma); |
1985 | |
1986 | GEM_BUG_ON(vma->fence)((void)0); |
1987 | GEM_BUG_ON(i915_vma_has_userfault(vma))((void)0); |
1988 | |
1989 | /* Object backend must be async capable. */ |
1990 | GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt)({ __builtin_expect(!!(!!(async && !vma->resource-> bi.pages_rsgt)), 0); }); |
1991 | |
1992 | /* If vm is not open, unbind is a nop. */ |
1993 | vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND((int)(1UL << (10)))) && |
1994 | kref_read(&vma->vm->ref); |
1995 | vma_res->skip_pte_rewrite = !kref_read(&vma->vm->ref) || |
1996 | vma->vm->skip_pte_rewrite; |
1997 | trace_i915_vma_unbind(vma); |
1998 | |
1999 | if (async) |
2000 | unbind_fence = i915_vma_resource_unbind(vma_res, |
2001 | &vma->obj->mm.tlb); |
2002 | else |
2003 | unbind_fence = i915_vma_resource_unbind(vma_res, NULL((void *)0)); |
2004 | |
2005 | vma->resource = NULL((void *)0); |
2006 | |
2007 | atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),__sync_fetch_and_and(&vma->flags, ~((((int)(1UL << (10))) | ((int)(1UL << (11)))) | ((int)(1UL << ( 12))) | ((int)(1UL << (16))))) |
2008 | &vma->flags)__sync_fetch_and_and(&vma->flags, ~((((int)(1UL << (10))) | ((int)(1UL << (11)))) | ((int)(1UL << ( 12))) | ((int)(1UL << (16))))); |
2009 | |
2010 | i915_vma_detach(vma); |
2011 | |
2012 | if (!async) { |
2013 | if (unbind_fence) { |
2014 | dma_fence_wait(unbind_fence, false0); |
2015 | dma_fence_put(unbind_fence); |
2016 | unbind_fence = NULL((void *)0); |
2017 | } |
2018 | vma_invalidate_tlb(vma->vm, &vma->obj->mm.tlb); |
2019 | } |
2020 | |
2021 | /* |
2022 | * Binding itself may not have completed until the unbind fence signals, |
2023 | * so don't drop the pages until that happens, unless the resource is |
2024 | * async_capable. |
2025 | */ |
2026 | |
2027 | vma_unbind_pages(vma); |
2028 | return unbind_fence; |
2029 | } |
2030 | |
2031 | int __i915_vma_unbind(struct i915_vma *vma) |
2032 | { |
2033 | int ret; |
2034 | |
2035 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
2036 | assert_vma_held_evict(vma); |
2037 | |
2038 | if (!drm_mm_node_allocated(&vma->node)) |
2039 | return 0; |
2040 | |
2041 | if (i915_vma_is_pinned(vma)) { |
2042 | vma_print_allocator(vma, "is pinned"); |
2043 | return -EAGAIN35; |
2044 | } |
2045 | |
2046 | /* |
2047 | * After confirming that no one else is pinning this vma, wait for |
2048 | * any laggards who may have crept in during the wait (through |
2049 | * a residual pin skipping the vm->mutex) to complete. |
2050 | */ |
2051 | ret = i915_vma_sync(vma); |
2052 | if (ret) |
2053 | return ret; |
2054 | |
2055 | GEM_BUG_ON(i915_vma_is_active(vma))((void)0); |
2056 | __i915_vma_evict(vma, false0); |
2057 | |
2058 | drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ |
2059 | return 0; |
2060 | } |
2061 | |
2062 | static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma) |
2063 | { |
2064 | struct dma_fence *fence; |
2065 | |
2066 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
2067 | |
2068 | if (!drm_mm_node_allocated(&vma->node)) |
2069 | return NULL((void *)0); |
2070 | |
2071 | if (i915_vma_is_pinned(vma) || |
2072 | &vma->obj->mm.rsgt->table != vma->resource->bi.pages) |
2073 | return ERR_PTR(-EAGAIN35); |
2074 | |
2075 | /* |
2076 | * We probably need to replace this with awaiting the fences of the |
2077 | * object's dma_resv when the vma active goes away. When doing that |
2078 | * we need to be careful to not add the vma_resource unbind fence |
2079 | * immediately to the object's dma_resv, because then unbinding |
2080 | * the next vma from the object, in case there are many, will |
2081 | * actually await the unbinding of the previous vmas, which is |
2082 | * undesirable. |
2083 | */ |
2084 | if (i915_sw_fence_await_active(&vma->resource->chain, &vma->active, |
2085 | I915_ACTIVE_AWAIT_EXCL(1UL << (0)) | |
2086 | I915_ACTIVE_AWAIT_ACTIVE(1UL << (1))) < 0) { |
2087 | return ERR_PTR(-EBUSY16); |
2088 | } |
2089 | |
2090 | fence = __i915_vma_evict(vma, true1); |
2091 | |
2092 | drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ |
2093 | |
2094 | return fence; |
2095 | } |
2096 | |
2097 | int i915_vma_unbind(struct i915_vma *vma) |
2098 | { |
2099 | struct i915_address_space *vm = vma->vm; |
2100 | intel_wakeref_t wakeref = 0; |
2101 | int err; |
2102 | |
2103 | assert_object_held_shared(vma->obj); |
2104 | |
2105 | /* Optimistic wait before taking the mutex */ |
2106 | err = i915_vma_sync(vma); |
2107 | if (err) |
2108 | return err; |
2109 | |
2110 | if (!drm_mm_node_allocated(&vma->node)) |
2111 | return 0; |
2112 | |
2113 | if (i915_vma_is_pinned(vma)) { |
2114 | vma_print_allocator(vma, "is pinned"); |
2115 | return -EAGAIN35; |
2116 | } |
2117 | |
2118 | if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND((int)(1UL << (10))))) |
2119 | /* XXX not always required: nop_clear_range */ |
2120 | wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); |
2121 | |
2122 | err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref)mutex_lock_interruptible(&vma->vm->mutex); |
2123 | if (err) |
2124 | goto out_rpm; |
2125 | |
2126 | err = __i915_vma_unbind(vma); |
2127 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
2128 | |
2129 | out_rpm: |
2130 | if (wakeref) |
2131 | intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); |
2132 | return err; |
2133 | } |
2134 | |
2135 | int i915_vma_unbind_async(struct i915_vma *vma, bool_Bool trylock_vm) |
2136 | { |
2137 | struct drm_i915_gem_object *obj = vma->obj; |
2138 | struct i915_address_space *vm = vma->vm; |
2139 | intel_wakeref_t wakeref = 0; |
2140 | struct dma_fence *fence; |
2141 | int err; |
2142 | |
2143 | /* |
2144 | * We need the dma-resv lock since we add the |
2145 | * unbind fence to the dma-resv object. |
2146 | */ |
2147 | assert_object_held(obj)do { (void)(&((obj)->base.resv)->lock.base); } while (0); |
2148 | |
2149 | if (!drm_mm_node_allocated(&vma->node)) |
2150 | return 0; |
2151 | |
2152 | if (i915_vma_is_pinned(vma)) { |
2153 | vma_print_allocator(vma, "is pinned"); |
2154 | return -EAGAIN35; |
2155 | } |
2156 | |
2157 | if (!obj->mm.rsgt) |
2158 | return -EBUSY16; |
2159 | |
2160 | err = dma_resv_reserve_fences(obj->base.resv, 2); |
2161 | if (err) |
2162 | return -EBUSY16; |
2163 | |
2164 | /* |
2165 | * It would be great if we could grab this wakeref from the |
2166 | * async unbind work if needed, but we can't because it uses |
2167 | * kmalloc and it's in the dma-fence signalling critical path. |
2168 | */ |
2169 | if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND((int)(1UL << (10))))) |
2170 | wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); |
2171 | |
2172 | if (trylock_vm && !mutex_trylock(&vm->mutex)(rw_enter(&vm->mutex, 0x0001UL | 0x0040UL) == 0)) { |
2173 | err = -EBUSY16; |
2174 | goto out_rpm; |
2175 | } else if (!trylock_vm) { |
2176 | err = mutex_lock_interruptible_nested(&vm->mutex, !wakeref)mutex_lock_interruptible(&vm->mutex); |
2177 | if (err) |
2178 | goto out_rpm; |
2179 | } |
2180 | |
2181 | fence = __i915_vma_unbind_async(vma); |
2182 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
2183 | if (IS_ERR_OR_NULL(fence)) { |
2184 | err = PTR_ERR_OR_ZERO(fence); |
2185 | goto out_rpm; |
2186 | } |
2187 | |
2188 | dma_resv_add_fence(obj->base.resv, fence, DMA_RESV_USAGE_READ); |
2189 | dma_fence_put(fence); |
2190 | |
2191 | out_rpm: |
2192 | if (wakeref) |
2193 | intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); |
2194 | return err; |
2195 | } |
2196 | |
2197 | int i915_vma_unbind_unlocked(struct i915_vma *vma) |
2198 | { |
2199 | int err; |
2200 | |
2201 | i915_gem_object_lock(vma->obj, NULL((void *)0)); |
2202 | err = i915_vma_unbind(vma); |
2203 | i915_gem_object_unlock(vma->obj); |
2204 | |
2205 | return err; |
2206 | } |
2207 | |
2208 | struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) |
2209 | { |
2210 | i915_gem_object_make_unshrinkable(vma->obj); |
2211 | return vma; |
2212 | } |
2213 | |
2214 | void i915_vma_make_shrinkable(struct i915_vma *vma) |
2215 | { |
2216 | i915_gem_object_make_shrinkable(vma->obj); |
2217 | } |
2218 | |
2219 | void i915_vma_make_purgeable(struct i915_vma *vma) |
2220 | { |
2221 | i915_gem_object_make_purgeable(vma->obj); |
2222 | } |
2223 | |
2224 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)0 |
2225 | #include "selftests/i915_vma.c" |
2226 | #endif |
2227 | |
2228 | void i915_vma_module_exit(void) |
2229 | { |
2230 | #ifdef __linux__ |
2231 | kmem_cache_destroy(slab_vmas); |
2232 | #else |
2233 | pool_destroy(&slab_vmas); |
2234 | #endif |
2235 | } |
2236 | |
2237 | int __init i915_vma_module_init(void) |
2238 | { |
2239 | #ifdef __linux__ |
2240 | slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); |
2241 | if (!slab_vmas) |
2242 | return -ENOMEM12; |
2243 | #else |
2244 | pool_init(&slab_vmas, sizeof(struct i915_vma), |
2245 | CACHELINESIZE64, IPL_NONE0x0, 0, "drmvma", NULL((void *)0)); |
2246 | #endif |
2247 | |
2248 | return 0; |
2249 | } |