| File: | dev/pci/drm/i915/i915_vma.c |
| Warning: | line 1217, column 2 Value stored to 'vma_offset' 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 <drm/drm_gem.h> |
| 27 | |
| 28 | #include "display/intel_frontbuffer.h" |
| 29 | |
| 30 | #include "gt/intel_engine.h" |
| 31 | #include "gt/intel_engine_heartbeat.h" |
| 32 | #include "gt/intel_gt.h" |
| 33 | #include "gt/intel_gt_requests.h" |
| 34 | |
| 35 | #include "i915_drv.h" |
| 36 | #include "i915_globals.h" |
| 37 | #include "i915_sw_fence_work.h" |
| 38 | #include "i915_trace.h" |
| 39 | #include "i915_vma.h" |
| 40 | |
| 41 | #include <dev/pci/agpvar.h> |
| 42 | |
| 43 | static struct i915_global_vma { |
| 44 | struct i915_global base; |
| 45 | #ifdef __linux__ |
| 46 | struct kmem_cache *slab_vmas; |
| 47 | #else |
| 48 | struct pool slab_vmas; |
| 49 | #endif |
| 50 | } global; |
| 51 | |
| 52 | struct i915_vma *i915_vma_alloc(void) |
| 53 | { |
| 54 | #ifdef __linux__ |
| 55 | return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL(0x0001 | 0x0004)); |
| 56 | #else |
| 57 | return pool_get(&global.slab_vmas, PR_WAITOK0x0001 | PR_ZERO0x0008); |
| 58 | #endif |
| 59 | } |
| 60 | |
| 61 | void i915_vma_free(struct i915_vma *vma) |
| 62 | { |
| 63 | #ifdef __linux__ |
| 64 | return kmem_cache_free(global.slab_vmas, vma); |
| 65 | #else |
| 66 | pool_put(&global.slab_vmas, vma); |
| 67 | #endif |
| 68 | } |
| 69 | |
| 70 | #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM)0 && IS_ENABLED(CONFIG_DRM_DEBUG_MM)0 |
| 71 | |
| 72 | #include <linux/stackdepot.h> |
| 73 | |
| 74 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) |
| 75 | { |
| 76 | unsigned long *entries; |
| 77 | unsigned int nr_entries; |
| 78 | char buf[512]; |
| 79 | |
| 80 | if (!vma->node.stack) { |
| 81 | DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",__drm_dbg(DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: unknown owner\n" , vma->node.start, vma->node.size, reason) |
| 82 | vma->node.start, vma->node.size, reason)__drm_dbg(DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: unknown owner\n" , vma->node.start, vma->node.size, reason); |
| 83 | return; |
| 84 | } |
| 85 | |
| 86 | nr_entries = stack_depot_fetch(vma->node.stack, &entries); |
| 87 | stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0); |
| 88 | DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",__drm_dbg(DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: inserted at %s\n" , vma->node.start, vma->node.size, reason, buf) |
| 89 | vma->node.start, vma->node.size, reason, buf)__drm_dbg(DRM_UT_DRIVER, "vma.node [%08llx + %08llx] %s: inserted at %s\n" , vma->node.start, vma->node.size, reason, buf); |
| 90 | } |
| 91 | |
| 92 | #else |
| 93 | |
| 94 | static void vma_print_allocator(struct i915_vma *vma, const char *reason) |
| 95 | { |
| 96 | } |
| 97 | |
| 98 | #endif |
| 99 | |
| 100 | static inline struct i915_vma *active_to_vma(struct i915_active *ref) |
| 101 | { |
| 102 | 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) );}); |
| 103 | } |
| 104 | |
| 105 | static int __i915_vma_active(struct i915_active *ref) |
| 106 | { |
| 107 | return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT2; |
| 108 | } |
| 109 | |
| 110 | __i915_active_call__attribute__((__aligned__(4))) |
| 111 | static void __i915_vma_retire(struct i915_active *ref) |
| 112 | { |
| 113 | i915_vma_put(active_to_vma(ref)); |
| 114 | } |
| 115 | |
| 116 | static struct i915_vma * |
| 117 | vma_create(struct drm_i915_gem_object *obj, |
| 118 | struct i915_address_space *vm, |
| 119 | const struct i915_ggtt_view *view) |
| 120 | { |
| 121 | struct i915_vma *pos = ERR_PTR(-E2BIG7); |
| 122 | struct i915_vma *vma; |
| 123 | struct rb_node *rb, **p; |
| 124 | |
| 125 | /* The aliasing_ppgtt should never be used directly! */ |
| 126 | GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm)((void)0); |
| 127 | |
| 128 | vma = i915_vma_alloc(); |
| 129 | if (vma == NULL((void *)0)) |
| 130 | return ERR_PTR(-ENOMEM12); |
| 131 | |
| 132 | kref_init(&vma->ref); |
| 133 | rw_init(&vma->pages_mutex, "vmapg")_rw_init_flags(&vma->pages_mutex, "vmapg", 0, ((void * )0)); |
| 134 | vma->vm = i915_vm_get(vm); |
| 135 | vma->ops = &vm->vma_ops; |
| 136 | vma->obj = obj; |
| 137 | vma->resv = obj->base.resv; |
| 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)do { static struct lock_class_key __mkey; static struct lock_class_key __wkey; __i915_active_init(&vma->active, __i915_vma_active , __i915_vma_retire, &__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 | |
| 154 | if (view && view->type != I915_GGTT_VIEW_NORMAL) { |
| 155 | vma->ggtt_view = *view; |
| 156 | if (view->type == I915_GGTT_VIEW_PARTIAL) { |
| 157 | GEM_BUG_ON(range_overflows_t(u64,((void)0) |
| 158 | view->partial.offset,((void)0) |
| 159 | view->partial.size,((void)0) |
| 160 | obj->base.size >> PAGE_SHIFT))((void)0); |
| 161 | vma->size = view->partial.size; |
| 162 | vma->size <<= PAGE_SHIFT12; |
| 163 | GEM_BUG_ON(vma->size > obj->base.size)((void)0); |
| 164 | } else if (view->type == I915_GGTT_VIEW_ROTATED) { |
| 165 | vma->size = intel_rotation_info_size(&view->rotated); |
| 166 | vma->size <<= PAGE_SHIFT12; |
| 167 | } else if (view->type == I915_GGTT_VIEW_REMAPPED) { |
| 168 | vma->size = intel_remapped_info_size(&view->remapped); |
| 169 | vma->size <<= PAGE_SHIFT12; |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | if (unlikely(vma->size > vm->total)__builtin_expect(!!(vma->size > vm->total), 0)) |
| 174 | goto err_vma; |
| 175 | |
| 176 | GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE))((void)0); |
| 177 | |
| 178 | spin_lock(&obj->vma.lock)mtx_enter(&obj->vma.lock); |
| 179 | |
| 180 | if (i915_is_ggtt(vm)((vm)->is_ggtt)) { |
| 181 | if (unlikely(overflows_type(vma->size, u32))__builtin_expect(!!((sizeof(vma->size) > sizeof(u32) && (vma->size) >> (8 * sizeof(u32)))), 0)) |
| 182 | goto err_unlock; |
| 183 | |
| 184 | vma->fence_size = i915_gem_fence_size(vm->i915, vma->size, |
| 185 | i915_gem_object_get_tiling(obj), |
| 186 | i915_gem_object_get_stride(obj)); |
| 187 | if (unlikely(vma->fence_size < vma->size || /* overflow */__builtin_expect(!!(vma->fence_size < vma->size || vma ->fence_size > vm->total), 0) |
| 188 | vma->fence_size > vm->total)__builtin_expect(!!(vma->fence_size < vma->size || vma ->fence_size > vm->total), 0)) |
| 189 | goto err_unlock; |
| 190 | |
| 191 | GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT))((void)0); |
| 192 | |
| 193 | vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size, |
| 194 | i915_gem_object_get_tiling(obj), |
| 195 | i915_gem_object_get_stride(obj)); |
| 196 | GEM_BUG_ON(!is_power_of_2(vma->fence_alignment))((void)0); |
| 197 | |
| 198 | __set_bit(I915_VMA_GGTT_BIT14, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
| 199 | } |
| 200 | |
| 201 | rb = NULL((void *)0); |
| 202 | p = &obj->vma.tree.rb_node; |
| 203 | while (*p) { |
| 204 | long cmp; |
| 205 | |
| 206 | rb = *p; |
| 207 | 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) );}); |
| 208 | |
| 209 | /* |
| 210 | * If the view already exists in the tree, another thread |
| 211 | * already created a matching vma, so return the older instance |
| 212 | * and dispose of ours. |
| 213 | */ |
| 214 | cmp = i915_vma_compare(pos, vm, view); |
| 215 | if (cmp < 0) |
| 216 | p = &rb->rb_right__entry.rbe_right; |
| 217 | else if (cmp > 0) |
| 218 | p = &rb->rb_left__entry.rbe_left; |
| 219 | else |
| 220 | goto err_unlock; |
| 221 | } |
| 222 | rb_link_node(&vma->obj_node, rb, p); |
| 223 | rb_insert_color(&vma->obj_node, &obj->vma.tree)linux_root_RB_INSERT_COLOR((struct linux_root *)(&obj-> vma.tree), (&vma->obj_node)); |
| 224 | |
| 225 | if (i915_vma_is_ggtt(vma)) |
| 226 | /* |
| 227 | * We put the GGTT vma at the start of the vma-list, followed |
| 228 | * by the ppGGTT vma. This allows us to break early when |
| 229 | * iterating over only the GGTT vma for an object, see |
| 230 | * for_each_ggtt_vma() |
| 231 | */ |
| 232 | list_add(&vma->obj_link, &obj->vma.list); |
| 233 | else |
| 234 | list_add_tail(&vma->obj_link, &obj->vma.list); |
| 235 | |
| 236 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
| 237 | |
| 238 | return vma; |
| 239 | |
| 240 | err_unlock: |
| 241 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
| 242 | err_vma: |
| 243 | i915_vm_put(vm); |
| 244 | i915_vma_free(vma); |
| 245 | return pos; |
| 246 | } |
| 247 | |
| 248 | static struct i915_vma * |
| 249 | vma_lookup(struct drm_i915_gem_object *obj, |
| 250 | struct i915_address_space *vm, |
| 251 | const struct i915_ggtt_view *view) |
| 252 | { |
| 253 | struct rb_node *rb; |
| 254 | |
| 255 | rb = obj->vma.tree.rb_node; |
| 256 | while (rb) { |
| 257 | 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) );}); |
| 258 | long cmp; |
| 259 | |
| 260 | cmp = i915_vma_compare(vma, vm, view); |
| 261 | if (cmp == 0) |
| 262 | return vma; |
| 263 | |
| 264 | if (cmp < 0) |
| 265 | rb = rb->rb_right__entry.rbe_right; |
| 266 | else |
| 267 | rb = rb->rb_left__entry.rbe_left; |
| 268 | } |
| 269 | |
| 270 | return NULL((void *)0); |
| 271 | } |
| 272 | |
| 273 | /** |
| 274 | * i915_vma_instance - return the singleton instance of the VMA |
| 275 | * @obj: parent &struct drm_i915_gem_object to be mapped |
| 276 | * @vm: address space in which the mapping is located |
| 277 | * @view: additional mapping requirements |
| 278 | * |
| 279 | * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with |
| 280 | * the same @view characteristics. If a match is not found, one is created. |
| 281 | * Once created, the VMA is kept until either the object is freed, or the |
| 282 | * address space is closed. |
| 283 | * |
| 284 | * Returns the vma, or an error pointer. |
| 285 | */ |
| 286 | struct i915_vma * |
| 287 | i915_vma_instance(struct drm_i915_gem_object *obj, |
| 288 | struct i915_address_space *vm, |
| 289 | const struct i915_ggtt_view *view) |
| 290 | { |
| 291 | struct i915_vma *vma; |
| 292 | |
| 293 | GEM_BUG_ON(view && !i915_is_ggtt(vm))((void)0); |
| 294 | GEM_BUG_ON(!atomic_read(&vm->open))((void)0); |
| 295 | |
| 296 | spin_lock(&obj->vma.lock)mtx_enter(&obj->vma.lock); |
| 297 | vma = vma_lookup(obj, vm, view); |
| 298 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
| 299 | |
| 300 | /* vma_create() will resolve the race if another creates the vma */ |
| 301 | if (unlikely(!vma)__builtin_expect(!!(!vma), 0)) |
| 302 | vma = vma_create(obj, vm, view); |
| 303 | |
| 304 | GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view))((void)0); |
| 305 | return vma; |
| 306 | } |
| 307 | |
| 308 | struct i915_vma_work { |
| 309 | struct dma_fence_work base; |
| 310 | struct i915_address_space *vm; |
| 311 | struct i915_vm_pt_stash stash; |
| 312 | struct i915_vma *vma; |
| 313 | struct drm_i915_gem_object *pinned; |
| 314 | struct i915_sw_dma_fence_cb cb; |
| 315 | enum i915_cache_level cache_level; |
| 316 | unsigned int flags; |
| 317 | }; |
| 318 | |
| 319 | static int __vma_bind(struct dma_fence_work *work) |
| 320 | { |
| 321 | 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) );}); |
| 322 | struct i915_vma *vma = vw->vma; |
| 323 | |
| 324 | vma->ops->bind_vma(vw->vm, &vw->stash, |
| 325 | vma, vw->cache_level, vw->flags); |
| 326 | return 0; |
| 327 | } |
| 328 | |
| 329 | static void __vma_release(struct dma_fence_work *work) |
| 330 | { |
| 331 | 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) );}); |
| 332 | |
| 333 | if (vw->pinned) { |
| 334 | __i915_gem_object_unpin_pages(vw->pinned); |
| 335 | i915_gem_object_put(vw->pinned); |
| 336 | } |
| 337 | |
| 338 | i915_vm_free_pt_stash(vw->vm, &vw->stash); |
| 339 | i915_vm_put(vw->vm); |
| 340 | } |
| 341 | |
| 342 | static const struct dma_fence_work_ops bind_ops = { |
| 343 | .name = "bind", |
| 344 | .work = __vma_bind, |
| 345 | .release = __vma_release, |
| 346 | }; |
| 347 | |
| 348 | struct i915_vma_work *i915_vma_work(void) |
| 349 | { |
| 350 | struct i915_vma_work *vw; |
| 351 | |
| 352 | vw = kzalloc(sizeof(*vw), GFP_KERNEL(0x0001 | 0x0004)); |
| 353 | if (!vw) |
| 354 | return NULL((void *)0); |
| 355 | |
| 356 | dma_fence_work_init(&vw->base, &bind_ops); |
| 357 | vw->base.dma.error = -EAGAIN35; /* disable the worker by default */ |
| 358 | |
| 359 | return vw; |
| 360 | } |
| 361 | |
| 362 | int i915_vma_wait_for_bind(struct i915_vma *vma) |
| 363 | { |
| 364 | int err = 0; |
| 365 | |
| 366 | if (rcu_access_pointer(vma->active.excl.fence)(vma->active.excl.fence)) { |
| 367 | struct dma_fence *fence; |
| 368 | |
| 369 | rcu_read_lock(); |
| 370 | fence = dma_fence_get_rcu_safe(&vma->active.excl.fence); |
| 371 | rcu_read_unlock(); |
| 372 | if (fence) { |
| 373 | err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT(0x7fffffff)); |
| 374 | dma_fence_put(fence); |
| 375 | } |
| 376 | } |
| 377 | |
| 378 | return err; |
| 379 | } |
| 380 | |
| 381 | /** |
| 382 | * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. |
| 383 | * @vma: VMA to map |
| 384 | * @cache_level: mapping cache level |
| 385 | * @flags: flags like global or local mapping |
| 386 | * @work: preallocated worker for allocating and binding the PTE |
| 387 | * |
| 388 | * DMA addresses are taken from the scatter-gather table of this object (or of |
| 389 | * this VMA in case of non-default GGTT views) and PTE entries set up. |
| 390 | * Note that DMA addresses are also the only part of the SG table we care about. |
| 391 | */ |
| 392 | int i915_vma_bind(struct i915_vma *vma, |
| 393 | enum i915_cache_level cache_level, |
| 394 | u32 flags, |
| 395 | struct i915_vma_work *work) |
| 396 | { |
| 397 | u32 bind_flags; |
| 398 | u32 vma_flags; |
| 399 | |
| 400 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node))((void)0); |
| 401 | GEM_BUG_ON(vma->size > vma->node.size)((void)0); |
| 402 | |
| 403 | if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,({ ((void)0); 0; }) |
| 404 | vma->node.size,({ ((void)0); 0; }) |
| 405 | vma->vm->total))({ ((void)0); 0; })) |
| 406 | return -ENODEV19; |
| 407 | |
| 408 | if (GEM_DEBUG_WARN_ON(!flags)({ ((void)0); 0; })) |
| 409 | return -EINVAL22; |
| 410 | |
| 411 | bind_flags = flags; |
| 412 | bind_flags &= I915_VMA_GLOBAL_BIND((int)(1UL << (10))) | I915_VMA_LOCAL_BIND((int)(1UL << (11))); |
| 413 | |
| 414 | vma_flags = atomic_read(&vma->flags)({ typeof(*(&vma->flags)) __tmp = *(volatile typeof(*( &vma->flags)) *)&(*(&vma->flags)); membar_datadep_consumer (); __tmp; }); |
| 415 | vma_flags &= I915_VMA_GLOBAL_BIND((int)(1UL << (10))) | I915_VMA_LOCAL_BIND((int)(1UL << (11))); |
| 416 | |
| 417 | bind_flags &= ~vma_flags; |
| 418 | if (bind_flags == 0) |
| 419 | return 0; |
| 420 | |
| 421 | GEM_BUG_ON(!vma->pages)((void)0); |
| 422 | |
| 423 | trace_i915_vma_bind(vma, bind_flags); |
| 424 | if (work && bind_flags & vma->vm->bind_async_flags) { |
| 425 | struct dma_fence *prev; |
| 426 | |
| 427 | work->vma = vma; |
| 428 | work->cache_level = cache_level; |
| 429 | work->flags = bind_flags; |
| 430 | |
| 431 | /* |
| 432 | * Note we only want to chain up to the migration fence on |
| 433 | * the pages (not the object itself). As we don't track that, |
| 434 | * yet, we have to use the exclusive fence instead. |
| 435 | * |
| 436 | * Also note that we do not want to track the async vma as |
| 437 | * part of the obj->resv->excl_fence as it only affects |
| 438 | * execution and not content or object's backing store lifetime. |
| 439 | */ |
| 440 | prev = i915_active_set_exclusive(&vma->active, &work->base.dma); |
| 441 | if (prev) { |
| 442 | __i915_sw_fence_await_dma_fence(&work->base.chain, |
| 443 | prev, |
| 444 | &work->cb); |
| 445 | dma_fence_put(prev); |
| 446 | } |
| 447 | |
| 448 | work->base.dma.error = 0; /* enable the queue_work() */ |
| 449 | |
| 450 | if (vma->obj) { |
| 451 | __i915_gem_object_pin_pages(vma->obj); |
| 452 | work->pinned = i915_gem_object_get(vma->obj); |
| 453 | } |
| 454 | } else { |
| 455 | vma->ops->bind_vma(vma->vm, NULL((void *)0), vma, cache_level, bind_flags); |
| 456 | } |
| 457 | |
| 458 | atomic_or(bind_flags, &vma->flags)x86_atomic_setbits_u32(&vma->flags, bind_flags); |
| 459 | return 0; |
| 460 | } |
| 461 | |
| 462 | void __iomem *i915_vma_pin_iomap(struct i915_vma *vma) |
| 463 | { |
| 464 | void __iomem *ptr; |
| 465 | int err; |
| 466 | |
| 467 | if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))({ __builtin_expect(!!(!!(!i915_vma_is_map_and_fenceable(vma) )), 0); })) { |
| 468 | err = -ENODEV19; |
| 469 | goto err; |
| 470 | } |
| 471 | |
| 472 | GEM_BUG_ON(!i915_vma_is_ggtt(vma))((void)0); |
| 473 | GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))((void)0); |
| 474 | |
| 475 | ptr = READ_ONCE(vma->iomap)({ typeof(vma->iomap) __tmp = *(volatile typeof(vma->iomap ) *)&(vma->iomap); membar_datadep_consumer(); __tmp; } ); |
| 476 | if (ptr == NULL((void *)0)) { |
| 477 | #ifdef __linux__ |
| 478 | ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap, |
| 479 | vma->node.start, |
| 480 | vma->node.size); |
| 481 | #else |
| 482 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->vm->i915; |
| 483 | err = agp_map_subregion(dev_priv->agph, vma->node.start, |
| 484 | vma->node.size, &vma->bsh); |
| 485 | if (err) { |
| 486 | err = -err; |
| 487 | goto err; |
| 488 | } |
| 489 | ptr = bus_space_vaddr(dev_priv->bst, vma->bsh)((dev_priv->bst)->vaddr((vma->bsh))); |
| 490 | #endif |
| 491 | if (ptr == NULL((void *)0)) { |
| 492 | err = -ENOMEM12; |
| 493 | goto err; |
| 494 | } |
| 495 | |
| 496 | if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))__builtin_expect(!!(__sync_val_compare_and_swap(&vma-> iomap, ((void *)0), ptr)), 0)) { |
| 497 | #ifdef __linux__ |
| 498 | io_mapping_unmap(ptr); |
| 499 | #endif |
| 500 | ptr = vma->iomap; |
| 501 | } |
| 502 | } |
| 503 | |
| 504 | __i915_vma_pin(vma); |
| 505 | |
| 506 | err = i915_vma_pin_fence(vma); |
| 507 | if (err) |
| 508 | goto err_unpin; |
| 509 | |
| 510 | i915_vma_set_ggtt_write(vma); |
| 511 | |
| 512 | /* NB Access through the GTT requires the device to be awake. */ |
| 513 | return ptr; |
| 514 | |
| 515 | err_unpin: |
| 516 | __i915_vma_unpin(vma); |
| 517 | err: |
| 518 | return IO_ERR_PTR(err)((void *)ERR_PTR(err)); |
| 519 | } |
| 520 | |
| 521 | void i915_vma_flush_writes(struct i915_vma *vma) |
| 522 | { |
| 523 | if (i915_vma_unset_ggtt_write(vma)) |
| 524 | intel_gt_flush_ggtt_writes(vma->vm->gt); |
| 525 | } |
| 526 | |
| 527 | void i915_vma_unpin_iomap(struct i915_vma *vma) |
| 528 | { |
| 529 | GEM_BUG_ON(vma->iomap == NULL)((void)0); |
| 530 | |
| 531 | i915_vma_flush_writes(vma); |
| 532 | |
| 533 | i915_vma_unpin_fence(vma); |
| 534 | i915_vma_unpin(vma); |
| 535 | } |
| 536 | |
| 537 | void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags) |
| 538 | { |
| 539 | struct i915_vma *vma; |
| 540 | struct drm_i915_gem_object *obj; |
| 541 | |
| 542 | vma = fetch_and_zero(p_vma)({ typeof(*p_vma) __T = *(p_vma); *(p_vma) = (typeof(*p_vma)) 0; __T; }); |
| 543 | if (!vma) |
| 544 | return; |
| 545 | |
| 546 | obj = vma->obj; |
| 547 | GEM_BUG_ON(!obj)((void)0); |
| 548 | |
| 549 | i915_vma_unpin(vma); |
| 550 | |
| 551 | if (flags & I915_VMA_RELEASE_MAP(1UL << (0))) |
| 552 | i915_gem_object_unpin_map(obj); |
| 553 | |
| 554 | i915_gem_object_put(obj); |
| 555 | } |
| 556 | |
| 557 | bool_Bool i915_vma_misplaced(const struct i915_vma *vma, |
| 558 | u64 size, u64 alignment, u64 flags) |
| 559 | { |
| 560 | if (!drm_mm_node_allocated(&vma->node)) |
| 561 | return false0; |
| 562 | |
| 563 | if (test_bit(I915_VMA_ERROR_BIT13, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags))) |
| 564 | return true1; |
| 565 | |
| 566 | if (vma->node.size < size) |
| 567 | return true1; |
| 568 | |
| 569 | GEM_BUG_ON(alignment && !is_power_of_2(alignment))((void)0); |
| 570 | if (alignment && !IS_ALIGNED(vma->node.start, alignment)(((vma->node.start) & ((alignment) - 1)) == 0)) |
| 571 | return true1; |
| 572 | |
| 573 | if (flags & PIN_MAPPABLE(1ULL << (3)) && !i915_vma_is_map_and_fenceable(vma)) |
| 574 | return true1; |
| 575 | |
| 576 | if (flags & PIN_OFFSET_BIAS(1ULL << (6)) && |
| 577 | vma->node.start < (flags & PIN_OFFSET_MASK-(1ULL << (12)))) |
| 578 | return true1; |
| 579 | |
| 580 | if (flags & PIN_OFFSET_FIXED(1ULL << (7)) && |
| 581 | vma->node.start != (flags & PIN_OFFSET_MASK-(1ULL << (12)))) |
| 582 | return true1; |
| 583 | |
| 584 | return false0; |
| 585 | } |
| 586 | |
| 587 | void __i915_vma_set_map_and_fenceable(struct i915_vma *vma) |
| 588 | { |
| 589 | bool_Bool mappable, fenceable; |
| 590 | |
| 591 | GEM_BUG_ON(!i915_vma_is_ggtt(vma))((void)0); |
| 592 | GEM_BUG_ON(!vma->fence_size)((void)0); |
| 593 | |
| 594 | fenceable = (vma->node.size >= vma->fence_size && |
| 595 | IS_ALIGNED(vma->node.start, vma->fence_alignment)(((vma->node.start) & ((vma->fence_alignment) - 1)) == 0)); |
| 596 | |
| 597 | mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end; |
| 598 | |
| 599 | if (mappable && fenceable) |
| 600 | set_bit(I915_VMA_CAN_FENCE_BIT15, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
| 601 | else |
| 602 | clear_bit(I915_VMA_CAN_FENCE_BIT15, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
| 603 | } |
| 604 | |
| 605 | bool_Bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) |
| 606 | { |
| 607 | struct drm_mm_node *node = &vma->node; |
| 608 | struct drm_mm_node *other; |
| 609 | |
| 610 | /* |
| 611 | * On some machines we have to be careful when putting differing types |
| 612 | * of snoopable memory together to avoid the prefetcher crossing memory |
| 613 | * domains and dying. During vm initialisation, we decide whether or not |
| 614 | * these constraints apply and set the drm_mm.color_adjust |
| 615 | * appropriately. |
| 616 | */ |
| 617 | if (!i915_vm_has_cache_coloring(vma->vm)) |
| 618 | return true1; |
| 619 | |
| 620 | /* Only valid to be called on an already inserted vma */ |
| 621 | GEM_BUG_ON(!drm_mm_node_allocated(node))((void)0); |
| 622 | GEM_BUG_ON(list_empty(&node->node_list))((void)0); |
| 623 | |
| 624 | 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) );}); |
| 625 | if (i915_node_color_differs(other, color) && |
| 626 | !drm_mm_hole_follows(other)) |
| 627 | return false0; |
| 628 | |
| 629 | 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) );}); |
| 630 | if (i915_node_color_differs(other, color) && |
| 631 | !drm_mm_hole_follows(node)) |
| 632 | return false0; |
| 633 | |
| 634 | return true1; |
| 635 | } |
| 636 | |
| 637 | /** |
| 638 | * i915_vma_insert - finds a slot for the vma in its address space |
| 639 | * @vma: the vma |
| 640 | * @size: requested size in bytes (can be larger than the VMA) |
| 641 | * @alignment: required alignment |
| 642 | * @flags: mask of PIN_* flags to use |
| 643 | * |
| 644 | * First we try to allocate some free space that meets the requirements for |
| 645 | * the VMA. Failiing that, if the flags permit, it will evict an old VMA, |
| 646 | * preferrably the oldest idle entry to make room for the new VMA. |
| 647 | * |
| 648 | * Returns: |
| 649 | * 0 on success, negative error code otherwise. |
| 650 | */ |
| 651 | static int |
| 652 | i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) |
| 653 | { |
| 654 | unsigned long color; |
| 655 | u64 start, end; |
| 656 | int ret; |
| 657 | |
| 658 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND))((void)0); |
| 659 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node))((void)0); |
| 660 | |
| 661 | size = max(size, vma->size)(((size)>(vma->size))?(size):(vma->size)); |
| 662 | alignment = max(alignment, vma->display_alignment)(((alignment)>(vma->display_alignment))?(alignment):(vma ->display_alignment)); |
| 663 | if (flags & PIN_MAPPABLE(1ULL << (3))) { |
| 664 | 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; } ); |
| 665 | 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; }) |
| 666 | alignment, vma->fence_alignment)({ typeof(alignment) __max_a = (alignment); typeof(alignment) __max_b = (vma->fence_alignment); __max_a > __max_b ? __max_a : __max_b; }); |
| 667 | } |
| 668 | |
| 669 | GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE))((void)0); |
| 670 | GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT))((void)0); |
| 671 | GEM_BUG_ON(!is_power_of_2(alignment))((void)0); |
| 672 | |
| 673 | start = flags & PIN_OFFSET_BIAS(1ULL << (6)) ? flags & PIN_OFFSET_MASK-(1ULL << (12)) : 0; |
| 674 | GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE))((void)0); |
| 675 | |
| 676 | end = vma->vm->total; |
| 677 | if (flags & PIN_MAPPABLE(1ULL << (3))) |
| 678 | 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 ; }); |
| 679 | if (flags & PIN_ZONE_4G(1ULL << (4))) |
| 680 | 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 ; }); |
| 681 | GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE))((void)0); |
| 682 | |
| 683 | /* If binding the object/GGTT view requires more space than the entire |
| 684 | * aperture has, reject it early before evicting everything in a vain |
| 685 | * attempt to find space. |
| 686 | */ |
| 687 | if (size > end) { |
| 688 | DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",__drm_dbg(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) |
| 689 | size, flags & PIN_MAPPABLE ? "mappable" : "total",__drm_dbg(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) |
| 690 | end)__drm_dbg(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); |
| 691 | return -ENOSPC28; |
| 692 | } |
| 693 | |
| 694 | color = 0; |
| 695 | if (vma->obj && i915_vm_has_cache_coloring(vma->vm)) |
| 696 | color = vma->obj->cache_level; |
| 697 | |
| 698 | if (flags & PIN_OFFSET_FIXED(1ULL << (7))) { |
| 699 | u64 offset = flags & PIN_OFFSET_MASK-(1ULL << (12)); |
| 700 | if (!IS_ALIGNED(offset, alignment)(((offset) & ((alignment) - 1)) == 0) || |
| 701 | 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__; })) |
| 702 | return -EINVAL22; |
| 703 | |
| 704 | ret = i915_gem_gtt_reserve(vma->vm, &vma->node, |
| 705 | size, offset, color, |
| 706 | flags); |
| 707 | if (ret) |
| 708 | return ret; |
| 709 | } else { |
| 710 | /* |
| 711 | * We only support huge gtt pages through the 48b PPGTT, |
| 712 | * however we also don't want to force any alignment for |
| 713 | * objects which need to be tightly packed into the low 32bits. |
| 714 | * |
| 715 | * Note that we assume that GGTT are limited to 4GiB for the |
| 716 | * forseeable future. See also i915_ggtt_offset(). |
| 717 | */ |
| 718 | if (upper_32_bits(end - 1)((u32)(((end - 1) >> 16) >> 16)) && |
| 719 | vma->page_sizes.sg > I915_GTT_PAGE_SIZE(1ULL << (12))) { |
| 720 | /* |
| 721 | * We can't mix 64K and 4K PTEs in the same page-table |
| 722 | * (2M block), and so to avoid the ugliness and |
| 723 | * complexity of coloring we opt for just aligning 64K |
| 724 | * objects to 2M. |
| 725 | */ |
| 726 | u64 page_alignment = |
| 727 | rounddown_pow_of_two(vma->page_sizes.sg | |
| 728 | I915_GTT_PAGE_SIZE_2M(1ULL << (21))); |
| 729 | |
| 730 | /* |
| 731 | * Check we don't expand for the limited Global GTT |
| 732 | * (mappable aperture is even more precious!). This |
| 733 | * also checks that we exclude the aliasing-ppgtt. |
| 734 | */ |
| 735 | GEM_BUG_ON(i915_vma_is_ggtt(vma))((void)0); |
| 736 | |
| 737 | alignment = max(alignment, page_alignment)(((alignment)>(page_alignment))?(alignment):(page_alignment )); |
| 738 | |
| 739 | if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K(1ULL << (16))) |
| 740 | size = round_up(size, I915_GTT_PAGE_SIZE_2M)((((size) + (((1ULL << (21))) - 1)) / ((1ULL << ( 21)))) * ((1ULL << (21)))); |
| 741 | } |
| 742 | |
| 743 | ret = i915_gem_gtt_insert(vma->vm, &vma->node, |
| 744 | size, alignment, color, |
| 745 | start, end, flags); |
| 746 | if (ret) |
| 747 | return ret; |
| 748 | |
| 749 | GEM_BUG_ON(vma->node.start < start)((void)0); |
| 750 | GEM_BUG_ON(vma->node.start + vma->node.size > end)((void)0); |
| 751 | } |
| 752 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node))((void)0); |
| 753 | GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color))((void)0); |
| 754 | |
| 755 | list_add_tail(&vma->vm_link, &vma->vm->bound_list); |
| 756 | |
| 757 | return 0; |
| 758 | } |
| 759 | |
| 760 | static void |
| 761 | i915_vma_detach(struct i915_vma *vma) |
| 762 | { |
| 763 | GEM_BUG_ON(!drm_mm_node_allocated(&vma->node))((void)0); |
| 764 | GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND))((void)0); |
| 765 | |
| 766 | /* |
| 767 | * And finally now the object is completely decoupled from this |
| 768 | * vma, we can drop its hold on the backing storage and allow |
| 769 | * it to be reaped by the shrinker. |
| 770 | */ |
| 771 | list_del(&vma->vm_link); |
| 772 | } |
| 773 | |
| 774 | static bool_Bool try_qad_pin(struct i915_vma *vma, unsigned int flags) |
| 775 | { |
| 776 | unsigned int bound; |
| 777 | bool_Bool pinned = true1; |
| 778 | |
| 779 | bound = atomic_read(&vma->flags)({ typeof(*(&vma->flags)) __tmp = *(volatile typeof(*( &vma->flags)) *)&(*(&vma->flags)); membar_datadep_consumer (); __tmp; }); |
| 780 | do { |
| 781 | if (unlikely(flags & ~bound)__builtin_expect(!!(flags & ~bound), 0)) |
| 782 | return false0; |
| 783 | |
| 784 | if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))__builtin_expect(!!(bound & (0x200 | ((int)(1UL << ( 13))))), 0)) |
| 785 | return false0; |
| 786 | |
| 787 | if (!(bound & I915_VMA_PIN_MASK0x3ff)) |
| 788 | goto unpinned; |
| 789 | |
| 790 | GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0)((void)0); |
| 791 | } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); |
| 792 | |
| 793 | return true1; |
| 794 | |
| 795 | unpinned: |
| 796 | /* |
| 797 | * If pin_count==0, but we are bound, check under the lock to avoid |
| 798 | * racing with a concurrent i915_vma_unbind(). |
| 799 | */ |
| 800 | mutex_lock(&vma->vm->mutex)rw_enter_write(&vma->vm->mutex); |
| 801 | do { |
| 802 | if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))__builtin_expect(!!(bound & (0x200 | ((int)(1UL << ( 13))))), 0)) { |
| 803 | pinned = false0; |
| 804 | break; |
| 805 | } |
| 806 | |
| 807 | if (unlikely(flags & ~bound)__builtin_expect(!!(flags & ~bound), 0)) { |
| 808 | pinned = false0; |
| 809 | break; |
| 810 | } |
| 811 | } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); |
| 812 | mutex_unlock(&vma->vm->mutex)rw_exit_write(&vma->vm->mutex); |
| 813 | |
| 814 | return pinned; |
| 815 | } |
| 816 | |
| 817 | static int vma_get_pages(struct i915_vma *vma) |
| 818 | { |
| 819 | int err = 0; |
| 820 | |
| 821 | if (atomic_add_unless(&vma->pages_count, 1, 0)) |
| 822 | return 0; |
| 823 | |
| 824 | /* Allocations ahoy! */ |
| 825 | if (mutex_lock_interruptible(&vma->pages_mutex)) |
| 826 | return -EINTR4; |
| 827 | |
| 828 | if (!atomic_read(&vma->pages_count)({ typeof(*(&vma->pages_count)) __tmp = *(volatile typeof (*(&vma->pages_count)) *)&(*(&vma->pages_count )); membar_datadep_consumer(); __tmp; })) { |
| 829 | if (vma->obj) { |
| 830 | err = i915_gem_object_pin_pages(vma->obj); |
| 831 | if (err) |
| 832 | goto unlock; |
| 833 | } |
| 834 | |
| 835 | err = vma->ops->set_pages(vma); |
| 836 | if (err) { |
| 837 | if (vma->obj) |
| 838 | i915_gem_object_unpin_pages(vma->obj); |
| 839 | goto unlock; |
| 840 | } |
| 841 | } |
| 842 | atomic_inc(&vma->pages_count)__sync_fetch_and_add(&vma->pages_count, 1); |
| 843 | |
| 844 | unlock: |
| 845 | mutex_unlock(&vma->pages_mutex)rw_exit_write(&vma->pages_mutex); |
| 846 | |
| 847 | return err; |
| 848 | } |
| 849 | |
| 850 | static void __vma_put_pages(struct i915_vma *vma, unsigned int count) |
| 851 | { |
| 852 | /* We allocate under vma_get_pages, so beware the shrinker */ |
| 853 | mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING)rw_enter_write(&vma->pages_mutex); |
| 854 | GEM_BUG_ON(atomic_read(&vma->pages_count) < count)((void)0); |
| 855 | if (atomic_sub_return(count, &vma->pages_count)__sync_sub_and_fetch(&vma->pages_count, count) == 0) { |
| 856 | vma->ops->clear_pages(vma); |
| 857 | GEM_BUG_ON(vma->pages)((void)0); |
| 858 | if (vma->obj) |
| 859 | i915_gem_object_unpin_pages(vma->obj); |
| 860 | } |
| 861 | mutex_unlock(&vma->pages_mutex)rw_exit_write(&vma->pages_mutex); |
| 862 | } |
| 863 | |
| 864 | static void vma_put_pages(struct i915_vma *vma) |
| 865 | { |
| 866 | if (atomic_add_unless(&vma->pages_count, -1, 1)) |
| 867 | return; |
| 868 | |
| 869 | __vma_put_pages(vma, 1); |
| 870 | } |
| 871 | |
| 872 | static void vma_unbind_pages(struct i915_vma *vma) |
| 873 | { |
| 874 | unsigned int count; |
| 875 | |
| 876 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
| 877 | |
| 878 | /* The upper portion of pages_count is the number of bindings */ |
| 879 | count = atomic_read(&vma->pages_count)({ typeof(*(&vma->pages_count)) __tmp = *(volatile typeof (*(&vma->pages_count)) *)&(*(&vma->pages_count )); membar_datadep_consumer(); __tmp; }); |
| 880 | count >>= I915_VMA_PAGES_BIAS24; |
| 881 | GEM_BUG_ON(!count)((void)0); |
| 882 | |
| 883 | __vma_put_pages(vma, count | count << I915_VMA_PAGES_BIAS24); |
| 884 | } |
| 885 | |
| 886 | int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
| 887 | u64 size, u64 alignment, u64 flags) |
| 888 | { |
| 889 | struct i915_vma_work *work = NULL((void *)0); |
| 890 | intel_wakeref_t wakeref = 0; |
| 891 | unsigned int bound; |
| 892 | int err; |
| 893 | |
| 894 | #ifdef CONFIG_PROVE_LOCKING |
| 895 | if (debug_locks && lockdep_is_held(&vma->vm->i915->drm.struct_mutex)) |
| 896 | WARN_ON(!ww)({ int __ret = !!((!ww)); if (__ret) printf("%s", "WARN_ON(" "!ww" ")"); __builtin_expect(!!(__ret), 0); }); |
| 897 | #endif |
| 898 | |
| 899 | BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND)extern char _ctassert[(!((1ULL << (10)) != ((int)(1UL << (10))))) ? 1 : -1 ] __attribute__((__unused__)); |
| 900 | BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND)extern char _ctassert[(!((1ULL << (11)) != ((int)(1UL << (11))))) ? 1 : -1 ] __attribute__((__unused__)); |
| 901 | |
| 902 | GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)))((void)0); |
| 903 | |
| 904 | /* First try and grab the pin without rebinding the vma */ |
| 905 | if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))))) |
| 906 | return 0; |
| 907 | |
| 908 | err = vma_get_pages(vma); |
| 909 | if (err) |
| 910 | return err; |
| 911 | |
| 912 | if (flags & PIN_GLOBAL(1ULL << (10))) |
| 913 | wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm); |
| 914 | |
| 915 | if (flags & vma->vm->bind_async_flags) { |
| 916 | work = i915_vma_work(); |
| 917 | if (!work) { |
| 918 | err = -ENOMEM12; |
| 919 | goto err_rpm; |
| 920 | } |
| 921 | |
| 922 | work->vm = i915_vm_get(vma->vm); |
| 923 | |
| 924 | /* Allocate enough page directories to used PTE */ |
| 925 | if (vma->vm->allocate_va_range) { |
| 926 | err = i915_vm_alloc_pt_stash(vma->vm, |
| 927 | &work->stash, |
| 928 | vma->size); |
| 929 | if (err) |
| 930 | goto err_fence; |
| 931 | |
| 932 | err = i915_vm_pin_pt_stash(vma->vm, |
| 933 | &work->stash); |
| 934 | if (err) |
| 935 | goto err_fence; |
| 936 | } |
| 937 | } |
| 938 | |
| 939 | /* |
| 940 | * Differentiate between user/kernel vma inside the aliasing-ppgtt. |
| 941 | * |
| 942 | * We conflate the Global GTT with the user's vma when using the |
| 943 | * aliasing-ppgtt, but it is still vitally important to try and |
| 944 | * keep the use cases distinct. For example, userptr objects are |
| 945 | * not allowed inside the Global GTT as that will cause lock |
| 946 | * inversions when we have to evict them the mmu_notifier callbacks - |
| 947 | * but they are allowed to be part of the user ppGTT which can never |
| 948 | * be mapped. As such we try to give the distinct users of the same |
| 949 | * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt |
| 950 | * and i915_ppgtt separate]. |
| 951 | * |
| 952 | * NB this may cause us to mask real lock inversions -- while the |
| 953 | * code is safe today, lockdep may not be able to spot future |
| 954 | * transgressions. |
| 955 | */ |
| 956 | err = mutex_lock_interruptible_nested(&vma->vm->mutex,mutex_lock_interruptible(&vma->vm->mutex) |
| 957 | !(flags & PIN_GLOBAL))mutex_lock_interruptible(&vma->vm->mutex); |
| 958 | if (err) |
| 959 | goto err_fence; |
| 960 | |
| 961 | /* No more allocations allowed now we hold vm->mutex */ |
| 962 | |
| 963 | if (unlikely(i915_vma_is_closed(vma))__builtin_expect(!!(i915_vma_is_closed(vma)), 0)) { |
| 964 | err = -ENOENT2; |
| 965 | goto err_unlock; |
| 966 | } |
| 967 | |
| 968 | bound = atomic_read(&vma->flags)({ typeof(*(&vma->flags)) __tmp = *(volatile typeof(*( &vma->flags)) *)&(*(&vma->flags)); membar_datadep_consumer (); __tmp; }); |
| 969 | if (unlikely(bound & I915_VMA_ERROR)__builtin_expect(!!(bound & ((int)(1UL << (13)))), 0 )) { |
| 970 | err = -ENOMEM12; |
| 971 | goto err_unlock; |
| 972 | } |
| 973 | |
| 974 | if (unlikely(!((bound + 1) & I915_VMA_PIN_MASK))__builtin_expect(!!(!((bound + 1) & 0x3ff)), 0)) { |
| 975 | err = -EAGAIN35; /* pins are meant to be fairly temporary */ |
| 976 | goto err_unlock; |
| 977 | } |
| 978 | |
| 979 | if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))__builtin_expect(!!(!(flags & ~bound & (((int)(1UL << (10))) | ((int)(1UL << (11)))))), 0)) { |
| 980 | __i915_vma_pin(vma); |
| 981 | goto err_unlock; |
| 982 | } |
| 983 | |
| 984 | err = i915_active_acquire(&vma->active); |
| 985 | if (err) |
| 986 | goto err_unlock; |
| 987 | |
| 988 | if (!(bound & I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))))) { |
| 989 | err = i915_vma_insert(vma, size, alignment, flags); |
| 990 | if (err) |
| 991 | goto err_active; |
| 992 | |
| 993 | if (i915_is_ggtt(vma->vm)((vma->vm)->is_ggtt)) |
| 994 | __i915_vma_set_map_and_fenceable(vma); |
| 995 | } |
| 996 | |
| 997 | GEM_BUG_ON(!vma->pages)((void)0); |
| 998 | err = i915_vma_bind(vma, |
| 999 | vma->obj ? vma->obj->cache_level : 0, |
| 1000 | flags, work); |
| 1001 | if (err) |
| 1002 | goto err_remove; |
| 1003 | |
| 1004 | /* There should only be at most 2 active bindings (user, global) */ |
| 1005 | GEM_BUG_ON(bound + I915_VMA_PAGES_ACTIVE < bound)((void)0); |
| 1006 | atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count)__sync_fetch_and_add(&vma->pages_count, ((1UL << (24)) | 1)); |
| 1007 | list_move_tail(&vma->vm_link, &vma->vm->bound_list); |
| 1008 | |
| 1009 | __i915_vma_pin(vma); |
| 1010 | GEM_BUG_ON(!i915_vma_is_pinned(vma))((void)0); |
| 1011 | GEM_BUG_ON(!i915_vma_is_bound(vma, flags))((void)0); |
| 1012 | GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags))((void)0); |
| 1013 | |
| 1014 | err_remove: |
| 1015 | if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK(((int)(1UL << (10))) | ((int)(1UL << (11)))))) { |
| 1016 | i915_vma_detach(vma); |
| 1017 | drm_mm_remove_node(&vma->node); |
| 1018 | } |
| 1019 | err_active: |
| 1020 | i915_active_release(&vma->active); |
| 1021 | err_unlock: |
| 1022 | mutex_unlock(&vma->vm->mutex)rw_exit_write(&vma->vm->mutex); |
| 1023 | err_fence: |
| 1024 | if (work) |
| 1025 | dma_fence_work_commit_imm(&work->base); |
| 1026 | err_rpm: |
| 1027 | if (wakeref) |
| 1028 | intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); |
| 1029 | vma_put_pages(vma); |
| 1030 | return err; |
| 1031 | } |
| 1032 | |
| 1033 | static void flush_idle_contexts(struct intel_gt *gt) |
| 1034 | { |
| 1035 | struct intel_engine_cs *engine; |
| 1036 | enum intel_engine_id id; |
| 1037 | |
| 1038 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 1039 | intel_engine_flush_barriers(engine); |
| 1040 | |
| 1041 | intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT(0x7fffffff)); |
| 1042 | } |
| 1043 | |
| 1044 | int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, |
| 1045 | u32 align, unsigned int flags) |
| 1046 | { |
| 1047 | struct i915_address_space *vm = vma->vm; |
| 1048 | int err; |
| 1049 | |
| 1050 | GEM_BUG_ON(!i915_vma_is_ggtt(vma))((void)0); |
| 1051 | |
| 1052 | do { |
| 1053 | err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL(1ULL << (10))); |
| 1054 | if (err != -ENOSPC28) { |
| 1055 | if (!err) { |
| 1056 | err = i915_vma_wait_for_bind(vma); |
| 1057 | if (err) |
| 1058 | i915_vma_unpin(vma); |
| 1059 | } |
| 1060 | return err; |
| 1061 | } |
| 1062 | |
| 1063 | /* Unlike i915_vma_pin, we don't take no for an answer! */ |
| 1064 | flush_idle_contexts(vm->gt); |
| 1065 | if (mutex_lock_interruptible(&vm->mutex) == 0) { |
| 1066 | i915_gem_evict_vm(vm); |
| 1067 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
| 1068 | } |
| 1069 | } while (1); |
| 1070 | } |
| 1071 | |
| 1072 | static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) |
| 1073 | { |
| 1074 | /* |
| 1075 | * We defer actually closing, unbinding and destroying the VMA until |
| 1076 | * the next idle point, or if the object is freed in the meantime. By |
| 1077 | * postponing the unbind, we allow for it to be resurrected by the |
| 1078 | * client, avoiding the work required to rebind the VMA. This is |
| 1079 | * advantageous for DRI, where the client/server pass objects |
| 1080 | * between themselves, temporarily opening a local VMA to the |
| 1081 | * object, and then closing it again. The same object is then reused |
| 1082 | * on the next frame (or two, depending on the depth of the swap queue) |
| 1083 | * causing us to rebind the VMA once more. This ends up being a lot |
| 1084 | * of wasted work for the steady state. |
| 1085 | */ |
| 1086 | GEM_BUG_ON(i915_vma_is_closed(vma))((void)0); |
| 1087 | list_add(&vma->closed_link, >->closed_vma); |
| 1088 | } |
| 1089 | |
| 1090 | void i915_vma_close(struct i915_vma *vma) |
| 1091 | { |
| 1092 | struct intel_gt *gt = vma->vm->gt; |
| 1093 | unsigned long flags; |
| 1094 | |
| 1095 | if (i915_vma_is_ggtt(vma)) |
| 1096 | return; |
| 1097 | |
| 1098 | GEM_BUG_ON(!atomic_read(&vma->open_count))((void)0); |
| 1099 | if (atomic_dec_and_lock_irqsave(&vma->open_count,atomic_dec_and_lock(&vma->open_count, >->closed_lock ) |
| 1100 | >->closed_lock,atomic_dec_and_lock(&vma->open_count, >->closed_lock ) |
| 1101 | flags)atomic_dec_and_lock(&vma->open_count, >->closed_lock )) { |
| 1102 | __vma_close(vma, gt); |
| 1103 | spin_unlock_irqrestore(>->closed_lock, flags)do { (void)(flags); mtx_leave(>->closed_lock); } while (0); |
| 1104 | } |
| 1105 | } |
| 1106 | |
| 1107 | static void __i915_vma_remove_closed(struct i915_vma *vma) |
| 1108 | { |
| 1109 | struct intel_gt *gt = vma->vm->gt; |
| 1110 | |
| 1111 | spin_lock_irq(>->closed_lock)mtx_enter(>->closed_lock); |
| 1112 | list_del_init(&vma->closed_link); |
| 1113 | spin_unlock_irq(>->closed_lock)mtx_leave(>->closed_lock); |
| 1114 | } |
| 1115 | |
| 1116 | void i915_vma_reopen(struct i915_vma *vma) |
| 1117 | { |
| 1118 | if (i915_vma_is_closed(vma)) |
| 1119 | __i915_vma_remove_closed(vma); |
| 1120 | } |
| 1121 | |
| 1122 | void i915_vma_release(struct kref *ref) |
| 1123 | { |
| 1124 | struct i915_vma *vma = container_of(ref, typeof(*vma), ref)({ const __typeof( ((typeof(*vma) *)0)->ref ) *__mptr = (ref ); (typeof(*vma) *)( (char *)__mptr - __builtin_offsetof(typeof (*vma), ref) );}); |
| 1125 | |
| 1126 | if (drm_mm_node_allocated(&vma->node)) { |
| 1127 | mutex_lock(&vma->vm->mutex)rw_enter_write(&vma->vm->mutex); |
| 1128 | atomic_and(~I915_VMA_PIN_MASK, &vma->flags)__sync_fetch_and_and(&vma->flags, ~0x3ff); |
| 1129 | WARN_ON(__i915_vma_unbind(vma))({ int __ret = !!((__i915_vma_unbind(vma))); if (__ret) printf ("%s", "WARN_ON(" "__i915_vma_unbind(vma)" ")"); __builtin_expect (!!(__ret), 0); }); |
| 1130 | mutex_unlock(&vma->vm->mutex)rw_exit_write(&vma->vm->mutex); |
| 1131 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node))((void)0); |
| 1132 | } |
| 1133 | GEM_BUG_ON(i915_vma_is_active(vma))((void)0); |
| 1134 | |
| 1135 | if (vma->obj) { |
| 1136 | struct drm_i915_gem_object *obj = vma->obj; |
| 1137 | |
| 1138 | spin_lock(&obj->vma.lock)mtx_enter(&obj->vma.lock); |
| 1139 | list_del(&vma->obj_link); |
| 1140 | if (!RB_EMPTY_NODE(&vma->obj_node)((&vma->obj_node)->__entry.rbe_parent == &vma-> obj_node)) |
| 1141 | rb_erase(&vma->obj_node, &obj->vma.tree)linux_root_RB_REMOVE((struct linux_root *)(&obj->vma.tree ), (&vma->obj_node)); |
| 1142 | spin_unlock(&obj->vma.lock)mtx_leave(&obj->vma.lock); |
| 1143 | } |
| 1144 | |
| 1145 | __i915_vma_remove_closed(vma); |
| 1146 | i915_vm_put(vma->vm); |
| 1147 | |
| 1148 | i915_active_fini(&vma->active); |
| 1149 | i915_vma_free(vma); |
| 1150 | } |
| 1151 | |
| 1152 | void i915_vma_parked(struct intel_gt *gt) |
| 1153 | { |
| 1154 | struct i915_vma *vma, *next; |
| 1155 | DRM_LIST_HEAD(closed)struct list_head closed = { &(closed), &(closed) }; |
| 1156 | |
| 1157 | spin_lock_irq(>->closed_lock)mtx_enter(>->closed_lock); |
| 1158 | 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) );})) { |
| 1159 | struct drm_i915_gem_object *obj = vma->obj; |
| 1160 | struct i915_address_space *vm = vma->vm; |
| 1161 | |
| 1162 | /* XXX All to avoid keeping a reference on i915_vma itself */ |
| 1163 | |
| 1164 | if (!kref_get_unless_zero(&obj->base.refcount)) |
| 1165 | continue; |
| 1166 | |
| 1167 | if (!i915_vm_tryopen(vm)) { |
| 1168 | i915_gem_object_put(obj); |
| 1169 | continue; |
| 1170 | } |
| 1171 | |
| 1172 | list_move(&vma->closed_link, &closed); |
| 1173 | } |
| 1174 | spin_unlock_irq(>->closed_lock)mtx_leave(>->closed_lock); |
| 1175 | |
| 1176 | /* As the GT is held idle, no vma can be reopened as we destroy them */ |
| 1177 | 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 ) );})) { |
| 1178 | struct drm_i915_gem_object *obj = vma->obj; |
| 1179 | struct i915_address_space *vm = vma->vm; |
| 1180 | |
| 1181 | INIT_LIST_HEAD(&vma->closed_link); |
| 1182 | __i915_vma_put(vma); |
| 1183 | |
| 1184 | i915_gem_object_put(obj); |
| 1185 | i915_vm_close(vm); |
| 1186 | } |
| 1187 | } |
| 1188 | |
| 1189 | static void __i915_vma_iounmap(struct i915_vma *vma) |
| 1190 | { |
| 1191 | GEM_BUG_ON(i915_vma_is_pinned(vma))((void)0); |
| 1192 | |
| 1193 | if (vma->iomap == NULL((void *)0)) |
| 1194 | return; |
| 1195 | |
| 1196 | #ifdef __linux__ |
| 1197 | io_mapping_unmap(vma->iomap); |
| 1198 | #else |
| 1199 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->vm->i915; |
| 1200 | agp_unmap_subregion(dev_priv->agph, vma->bsh, vma->node.size); |
| 1201 | #endif |
| 1202 | vma->iomap = NULL((void *)0); |
| 1203 | } |
| 1204 | |
| 1205 | void i915_vma_revoke_mmap(struct i915_vma *vma) |
| 1206 | { |
| 1207 | struct drm_vma_offset_node *node; |
| 1208 | u64 vma_offset; |
| 1209 | |
| 1210 | if (!i915_vma_has_userfault(vma)) |
| 1211 | return; |
| 1212 | |
| 1213 | GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma))((void)0); |
| 1214 | GEM_BUG_ON(!vma->obj->userfault_count)((void)0); |
| 1215 | |
| 1216 | node = &vma->mmo->vma_node; |
| 1217 | vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT12; |
Value stored to 'vma_offset' is never read | |
| 1218 | #ifdef __linux__ |
| 1219 | unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping, |
| 1220 | drm_vma_node_offset_addr(node) + vma_offset, |
| 1221 | vma->size, |
| 1222 | 1); |
| 1223 | #else |
| 1224 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->obj->base.dev->dev_private; |
| 1225 | struct vm_page *pg; |
| 1226 | |
| 1227 | for (pg = &dev_priv->pgs[atop(vma->node.start)((vma->node.start) >> 12)]; |
| 1228 | pg != &dev_priv->pgs[atop(vma->node.start + vma->size)((vma->node.start + vma->size) >> 12)]; |
| 1229 | pg++) |
| 1230 | pmap_page_protect(pg, PROT_NONE0x00); |
| 1231 | #endif |
| 1232 | |
| 1233 | i915_vma_unset_userfault(vma); |
| 1234 | if (!--vma->obj->userfault_count) |
| 1235 | list_del(&vma->obj->userfault_link); |
| 1236 | } |
| 1237 | |
| 1238 | static int |
| 1239 | __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) |
| 1240 | { |
| 1241 | return __i915_request_await_exclusive(rq, &vma->active); |
| 1242 | } |
| 1243 | |
| 1244 | int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) |
| 1245 | { |
| 1246 | int err; |
| 1247 | |
| 1248 | GEM_BUG_ON(!i915_vma_is_pinned(vma))((void)0); |
| 1249 | |
| 1250 | /* Wait for the vma to be bound before we start! */ |
| 1251 | err = __i915_request_await_bind(rq, vma); |
| 1252 | if (err) |
| 1253 | return err; |
| 1254 | |
| 1255 | return i915_active_add_request(&vma->active, rq); |
| 1256 | } |
| 1257 | |
| 1258 | int i915_vma_move_to_active(struct i915_vma *vma, |
| 1259 | struct i915_request *rq, |
| 1260 | unsigned int flags) |
| 1261 | { |
| 1262 | struct drm_i915_gem_object *obj = vma->obj; |
| 1263 | int err; |
| 1264 | |
| 1265 | assert_object_held(obj)do { (void)(&((obj)->base.resv)->lock.base); } while (0); |
| 1266 | |
| 1267 | err = __i915_vma_move_to_active(vma, rq); |
| 1268 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 1269 | return err; |
| 1270 | |
| 1271 | if (flags & EXEC_OBJECT_WRITE(1<<2)) { |
| 1272 | struct intel_frontbuffer *front; |
| 1273 | |
| 1274 | front = __intel_frontbuffer_get(obj); |
| 1275 | if (unlikely(front)__builtin_expect(!!(front), 0)) { |
| 1276 | if (intel_frontbuffer_invalidate(front, ORIGIN_CS)) |
| 1277 | i915_active_add_request(&front->write, rq); |
| 1278 | intel_frontbuffer_put(front); |
| 1279 | } |
| 1280 | |
| 1281 | dma_resv_add_excl_fence(vma->resv, &rq->fence); |
| 1282 | obj->write_domain = I915_GEM_DOMAIN_RENDER0x00000002; |
| 1283 | obj->read_domains = 0; |
| 1284 | } else { |
| 1285 | err = dma_resv_reserve_shared(vma->resv, 1); |
| 1286 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 1287 | return err; |
| 1288 | |
| 1289 | dma_resv_add_shared_fence(vma->resv, &rq->fence); |
| 1290 | obj->write_domain = 0; |
| 1291 | } |
| 1292 | |
| 1293 | if (flags & EXEC_OBJECT_NEEDS_FENCE(1<<0) && vma->fence) |
| 1294 | i915_active_add_request(&vma->fence->active, rq); |
| 1295 | |
| 1296 | obj->read_domains |= I915_GEM_GPU_DOMAINS(0x00000002 | 0x00000004 | 0x00000008 | 0x00000010 | 0x00000020 ); |
| 1297 | obj->mm.dirty = true1; |
| 1298 | |
| 1299 | GEM_BUG_ON(!i915_vma_is_active(vma))((void)0); |
| 1300 | return 0; |
| 1301 | } |
| 1302 | |
| 1303 | void __i915_vma_evict(struct i915_vma *vma) |
| 1304 | { |
| 1305 | GEM_BUG_ON(i915_vma_is_pinned(vma))((void)0); |
| 1306 | |
| 1307 | if (i915_vma_is_map_and_fenceable(vma)) { |
| 1308 | /* Force a pagefault for domain tracking on next user access */ |
| 1309 | i915_vma_revoke_mmap(vma); |
| 1310 | |
| 1311 | /* |
| 1312 | * Check that we have flushed all writes through the GGTT |
| 1313 | * before the unbind, other due to non-strict nature of those |
| 1314 | * indirect writes they may end up referencing the GGTT PTE |
| 1315 | * after the unbind. |
| 1316 | * |
| 1317 | * Note that we may be concurrently poking at the GGTT_WRITE |
| 1318 | * bit from set-domain, as we mark all GGTT vma associated |
| 1319 | * with an object. We know this is for another vma, as we |
| 1320 | * are currently unbinding this one -- so if this vma will be |
| 1321 | * reused, it will be refaulted and have its dirty bit set |
| 1322 | * before the next write. |
| 1323 | */ |
| 1324 | i915_vma_flush_writes(vma); |
| 1325 | |
| 1326 | /* release the fence reg _after_ flushing */ |
| 1327 | i915_vma_revoke_fence(vma); |
| 1328 | |
| 1329 | __i915_vma_iounmap(vma); |
| 1330 | clear_bit(I915_VMA_CAN_FENCE_BIT15, __i915_vma_flags(vma)((unsigned long *)&(vma)->flags)); |
| 1331 | } |
| 1332 | GEM_BUG_ON(vma->fence)((void)0); |
| 1333 | GEM_BUG_ON(i915_vma_has_userfault(vma))((void)0); |
| 1334 | |
| 1335 | if (likely(atomic_read(&vma->vm->open))__builtin_expect(!!(({ typeof(*(&vma->vm->open)) __tmp = *(volatile typeof(*(&vma->vm->open)) *)&(*(& vma->vm->open)); membar_datadep_consumer(); __tmp; })), 1)) { |
| 1336 | trace_i915_vma_unbind(vma); |
| 1337 | vma->ops->unbind_vma(vma->vm, vma); |
| 1338 | } |
| 1339 | 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 << ( 13))) | ((int)(1UL << (17))))) |
| 1340 | &vma->flags)__sync_fetch_and_and(&vma->flags, ~((((int)(1UL << (10))) | ((int)(1UL << (11)))) | ((int)(1UL << ( 13))) | ((int)(1UL << (17))))); |
| 1341 | |
| 1342 | i915_vma_detach(vma); |
| 1343 | vma_unbind_pages(vma); |
| 1344 | } |
| 1345 | |
| 1346 | int __i915_vma_unbind(struct i915_vma *vma) |
| 1347 | { |
| 1348 | int ret; |
| 1349 | |
| 1350 | lockdep_assert_held(&vma->vm->mutex)do { (void)(&vma->vm->mutex); } while(0); |
| 1351 | |
| 1352 | if (!drm_mm_node_allocated(&vma->node)) |
| 1353 | return 0; |
| 1354 | |
| 1355 | if (i915_vma_is_pinned(vma)) { |
| 1356 | vma_print_allocator(vma, "is pinned"); |
| 1357 | return -EAGAIN35; |
| 1358 | } |
| 1359 | |
| 1360 | /* |
| 1361 | * After confirming that no one else is pinning this vma, wait for |
| 1362 | * any laggards who may have crept in during the wait (through |
| 1363 | * a residual pin skipping the vm->mutex) to complete. |
| 1364 | */ |
| 1365 | ret = i915_vma_sync(vma); |
| 1366 | if (ret) |
| 1367 | return ret; |
| 1368 | |
| 1369 | GEM_BUG_ON(i915_vma_is_active(vma))((void)0); |
| 1370 | __i915_vma_evict(vma); |
| 1371 | |
| 1372 | drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */ |
| 1373 | return 0; |
| 1374 | } |
| 1375 | |
| 1376 | int i915_vma_unbind(struct i915_vma *vma) |
| 1377 | { |
| 1378 | struct i915_address_space *vm = vma->vm; |
| 1379 | intel_wakeref_t wakeref = 0; |
| 1380 | int err; |
| 1381 | |
| 1382 | /* Optimistic wait before taking the mutex */ |
| 1383 | err = i915_vma_sync(vma); |
| 1384 | if (err) |
| 1385 | return err; |
| 1386 | |
| 1387 | if (!drm_mm_node_allocated(&vma->node)) |
| 1388 | return 0; |
| 1389 | |
| 1390 | if (i915_vma_is_pinned(vma)) { |
| 1391 | vma_print_allocator(vma, "is pinned"); |
| 1392 | return -EAGAIN35; |
| 1393 | } |
| 1394 | |
| 1395 | if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND((int)(1UL << (10))))) |
| 1396 | /* XXX not always required: nop_clear_range */ |
| 1397 | wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm); |
| 1398 | |
| 1399 | err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref)mutex_lock_interruptible(&vma->vm->mutex); |
| 1400 | if (err) |
| 1401 | goto out_rpm; |
| 1402 | |
| 1403 | err = __i915_vma_unbind(vma); |
| 1404 | mutex_unlock(&vm->mutex)rw_exit_write(&vm->mutex); |
| 1405 | |
| 1406 | out_rpm: |
| 1407 | if (wakeref) |
| 1408 | intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref); |
| 1409 | return err; |
| 1410 | } |
| 1411 | |
| 1412 | struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) |
| 1413 | { |
| 1414 | i915_gem_object_make_unshrinkable(vma->obj); |
| 1415 | return vma; |
| 1416 | } |
| 1417 | |
| 1418 | void i915_vma_make_shrinkable(struct i915_vma *vma) |
| 1419 | { |
| 1420 | i915_gem_object_make_shrinkable(vma->obj); |
| 1421 | } |
| 1422 | |
| 1423 | void i915_vma_make_purgeable(struct i915_vma *vma) |
| 1424 | { |
| 1425 | i915_gem_object_make_purgeable(vma->obj); |
| 1426 | } |
| 1427 | |
| 1428 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)0 |
| 1429 | #include "selftests/i915_vma.c" |
| 1430 | #endif |
| 1431 | |
| 1432 | static void i915_global_vma_shrink(void) |
| 1433 | { |
| 1434 | #ifdef notyet |
| 1435 | kmem_cache_shrink(global.slab_vmas); |
| 1436 | #endif |
| 1437 | } |
| 1438 | |
| 1439 | static void i915_global_vma_exit(void) |
| 1440 | { |
| 1441 | #ifdef __linux__ |
| 1442 | kmem_cache_destroy(global.slab_vmas); |
| 1443 | #else |
| 1444 | pool_destroy(&global.slab_vmas); |
| 1445 | #endif |
| 1446 | } |
| 1447 | |
| 1448 | static struct i915_global_vma global = { { |
| 1449 | .shrink = i915_global_vma_shrink, |
| 1450 | .exit = i915_global_vma_exit, |
| 1451 | } }; |
| 1452 | |
| 1453 | int __init i915_global_vma_init(void) |
| 1454 | { |
| 1455 | #ifdef __linux__ |
| 1456 | global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); |
| 1457 | if (!global.slab_vmas) |
| 1458 | return -ENOMEM12; |
| 1459 | #else |
| 1460 | pool_init(&global.slab_vmas, sizeof(struct i915_vma), |
| 1461 | CACHELINESIZE64, IPL_NONE0x0, 0, "drmvma", NULL((void *)0)); |
| 1462 | #endif |
| 1463 | |
| 1464 | i915_global_register(&global.base); |
| 1465 | return 0; |
| 1466 | } |