drivers/gpu/drm/i915/i915_gem.c

   1 /*
   2  * Copyright © 2008-2015 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  * Authors:
  24  *    Eric Anholt <[email protected]>
  25  *
  26  */
  27
  28 #include <linux/dma-fence-array.h>
  29 #include <linux/kthread.h>
  30 #include <linux/dma-resv.h>
  31 #include <linux/shmem_fs.h>
  32 #include <linux/slab.h>
  33 #include <linux/stop_machine.h>
  34 #include <linux/swap.h>
  35 #include <linux/pci.h>
  36 #include <linux/dma-buf.h>
  37 #include <linux/mman.h>
  38
  39 #include <drm/drm_cache.h>
  40 #include <drm/drm_vma_manager.h>
  41
  42 #include "display/intel_display.h"
  43 #include "display/intel_frontbuffer.h"
  44
  45 #include "gem/i915_gem_clflush.h"
  46 #include "gem/i915_gem_context.h"
  47 #include "gem/i915_gem_ioctls.h"
  48 #include "gem/i915_gem_mman.h"
  49 #include "gem/i915_gem_pm.h"
  50 #include "gem/i915_gem_region.h"
  51 #include "gem/i915_gem_userptr.h"
  52 #include "gt/intel_engine_user.h"
  53 #include "gt/intel_gt.h"
  54 #include "gt/intel_gt_pm.h"
  55 #include "gt/intel_workarounds.h"
  56
  57 #include "i915_drv.h"
  58 #include "i915_file_private.h"
  59 #include "i915_trace.h"
  60 #include "i915_vgpu.h"
  61 #include "intel_pm.h"
  62
  63 static int
  64 insert_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node, u32 size)
  65 {
  66         int err;
  67
  68         err = mutex_lock_interruptible(&ggtt->vm.mutex);
  69         if (err)
  70                 return err;
  71
  72         memset(node, 0, sizeof(*node));
  73         err = drm_mm_insert_node_in_range(&ggtt->vm.mm, node,
  74                                           size, 0, I915_COLOR_UNEVICTABLE,
  75                                           0, ggtt->mappable_end,
  76                                           DRM_MM_INSERT_LOW);
  77
  78         mutex_unlock(&ggtt->vm.mutex);
  79
  80         return err;
  81 }
  82
  83 static void
  84 remove_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node)
  85 {
  86         mutex_lock(&ggtt->vm.mutex);
  87         drm_mm_remove_node(node);
  88         mutex_unlock(&ggtt->vm.mutex);
  89 }
  90
  91 int
  92 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
  93                             struct drm_file *file)
  94 {
  95         struct drm_i915_private *i915 = to_i915(dev);
  96         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
  97         struct drm_i915_gem_get_aperture *args = data;
  98         struct i915_vma *vma;
  99         u64 pinned;
 100
 101         if (mutex_lock_interruptible(&ggtt->vm.mutex))
 102                 return -EINTR;
 103
 104         pinned = ggtt->vm.reserved;
 105         list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link)
 106                 if (i915_vma_is_pinned(vma))
 107                         pinned += vma->node.size;
 108
 109         mutex_unlock(&ggtt->vm.mutex);
 110
 111         args->aper_size = ggtt->vm.total;
 112         args->aper_available_size = args->aper_size - pinned;
 113
 114         return 0;
 115 }
 116
 117 int i915_gem_object_unbind(struct drm_i915_gem_object *obj,
 118                            unsigned long flags)
 119 {
 120         struct intel_runtime_pm *rpm = &to_i915(obj->base.dev)->runtime_pm;
 121         bool vm_trylock = !!(flags & I915_GEM_OBJECT_UNBIND_VM_TRYLOCK);
 122         LIST_HEAD(still_in_list);
 123         intel_wakeref_t wakeref;
 124         struct i915_vma *vma;
 125         int ret;
 126
 127         assert_object_held(obj);
 128
 129         if (list_empty(&obj->vma.list))
 130                 return 0;
 131
 132         /*
 133          * As some machines use ACPI to handle runtime-resume callbacks, and
 134          * ACPI is quite kmalloc happy, we cannot resume beneath the vm->mutex
 135          * as they are required by the shrinker. Ergo, we wake the device up
 136          * first just in case.
 137          */
 138         wakeref = intel_runtime_pm_get(rpm);
 139
 140 try_again:
 141         ret = 0;
 142         spin_lock(&obj->vma.lock);
 143         while (!ret && (vma = list_first_entry_or_null(&obj->vma.list,
 144                                                        struct i915_vma,
 145                                                        obj_link))) {
 146                 list_move_tail(&vma->obj_link, &still_in_list);
 147                 if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK))
 148                         continue;
 149
 150                 if (flags & I915_GEM_OBJECT_UNBIND_TEST) {
 151                         ret = -EBUSY;
 152                         break;
 153                 }
 154
 155                 /*
 156                  * Requiring the vm destructor to take the object lock
 157                  * before destroying a vma would help us eliminate the
 158                  * i915_vm_tryget() here, AND thus also the barrier stuff
 159                  * at the end. That's an easy fix, but sleeping locks in
 160                  * a kthread should generally be avoided.
 161                  */
 162                 ret = -EAGAIN;
 163                 if (!i915_vm_tryget(vma->vm))
 164                         break;
 165
 166                 spin_unlock(&obj->vma.lock);
 167
 168                 /*
 169                  * Since i915_vma_parked() takes the object lock
 170                  * before vma destruction, it won't race us here,
 171                  * and destroy the vma from under us.
 172                  */
 173
 174                 ret = -EBUSY;
 175                 if (flags & I915_GEM_OBJECT_UNBIND_ASYNC) {
 176                         assert_object_held(vma->obj);
 177                         ret = i915_vma_unbind_async(vma, vm_trylock);
 178                 }
 179
 180                 if (ret == -EBUSY && (flags & I915_GEM_OBJECT_UNBIND_ACTIVE ||
 181                                       !i915_vma_is_active(vma))) {
 182                         if (vm_trylock) {
 183                                 if (mutex_trylock(&vma->vm->mutex)) {
 184                                         ret = __i915_vma_unbind(vma);
 185                                         mutex_unlock(&vma->vm->mutex);
 186                                 }
 187                         } else {
 188                                 ret = i915_vma_unbind(vma);
 189                         }
 190                 }
 191
 192                 i915_vm_put(vma->vm);
 193                 spin_lock(&obj->vma.lock);
 194         }
 195         list_splice_init(&still_in_list, &obj->vma.list);
 196         spin_unlock(&obj->vma.lock);
 197
 198         if (ret == -EAGAIN && flags & I915_GEM_OBJECT_UNBIND_BARRIER) {
 199                 rcu_barrier(); /* flush the i915_vm_release() */
 200                 goto try_again;
 201         }
 202
 203         intel_runtime_pm_put(rpm, wakeref);
 204
 205         return ret;
 206 }
 207
 208 static int
 209 shmem_pread(struct page *page, int offset, int len, char __user *user_data,
 210             bool needs_clflush)
 211 {
 212         char *vaddr;
 213         int ret;
 214
 215         vaddr = kmap(page);
 216
 217         if (needs_clflush)
 218                 drm_clflush_virt_range(vaddr + offset, len);
 219
 220         ret = __copy_to_user(user_data, vaddr + offset, len);
 221
 222         kunmap(page);
 223
 224         return ret ? -EFAULT : 0;
 225 }
 226
 227 static int
 228 i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
 229                      struct drm_i915_gem_pread *args)
 230 {
 231         unsigned int needs_clflush;
 232         unsigned int idx, offset;
 233         char __user *user_data;
 234         u64 remain;
 235         int ret;
 236
 237         ret = i915_gem_object_lock_interruptible(obj, NULL);
 238         if (ret)
 239                 return ret;
 240
 241         ret = i915_gem_object_pin_pages(obj);
 242         if (ret)
 243                 goto err_unlock;
 244
 245         ret = i915_gem_object_prepare_read(obj, &needs_clflush);
 246         if (ret)
 247                 goto err_unpin;
 248
 249         i915_gem_object_finish_access(obj);
 250         i915_gem_object_unlock(obj);
 251
 252         remain = args->size;
 253         user_data = u64_to_user_ptr(args->data_ptr);
 254         offset = offset_in_page(args->offset);
 255         for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 256                 struct page *page = i915_gem_object_get_page(obj, idx);
 257                 unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 258
 259                 ret = shmem_pread(page, offset, length, user_data,
 260                                   needs_clflush);
 261                 if (ret)
 262                         break;
 263
 264                 remain -= length;
 265                 user_data += length;
 266                 offset = 0;
 267         }
 268
 269         i915_gem_object_unpin_pages(obj);
 270         return ret;
 271
 272 err_unpin:
 273         i915_gem_object_unpin_pages(obj);
 274 err_unlock:
 275         i915_gem_object_unlock(obj);
 276         return ret;
 277 }
 278
 279 static inline bool
 280 gtt_user_read(struct io_mapping *mapping,
 281               loff_t base, int offset,
 282               char __user *user_data, int length)
 283 {
 284         void __iomem *vaddr;
 285         unsigned long unwritten;
 286
 287         /* We can use the cpu mem copy function because this is X86. */
 288         vaddr = io_mapping_map_atomic_wc(mapping, base);
 289         unwritten = __copy_to_user_inatomic(user_data,
 290                                             (void __force *)vaddr + offset,
 291                                             length);
 292         io_mapping_unmap_atomic(vaddr);
 293         if (unwritten) {
 294                 vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 295                 unwritten = copy_to_user(user_data,
 296                                          (void __force *)vaddr + offset,
 297                                          length);
 298                 io_mapping_unmap(vaddr);
 299         }
 300         return unwritten;
 301 }
 302
 303 static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj,
 304                                              struct drm_mm_node *node,
 305                                              bool write)
 306 {
 307         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 308         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 309         struct i915_vma *vma;
 310         struct i915_gem_ww_ctx ww;
 311         int ret;
 312
 313         i915_gem_ww_ctx_init(&ww, true);
 314 retry:
 315         vma = ERR_PTR(-ENODEV);
 316         ret = i915_gem_object_lock(obj, &ww);
 317         if (ret)
 318                 goto err_ww;
 319
 320         ret = i915_gem_object_set_to_gtt_domain(obj, write);
 321         if (ret)
 322                 goto err_ww;
 323
 324         if (!i915_gem_object_is_tiled(obj))
 325                 vma = i915_gem_object_ggtt_pin_ww(obj, &ww, NULL, 0, 0,
 326                                                   PIN_MAPPABLE |
 327                                                   PIN_NONBLOCK /* NOWARN */ |
 328                                                   PIN_NOEVICT);
 329         if (vma == ERR_PTR(-EDEADLK)) {
 330                 ret = -EDEADLK;
 331                 goto err_ww;
 332         } else if (!IS_ERR(vma)) {
 333                 node->start = i915_ggtt_offset(vma);
 334                 node->flags = 0;
 335         } else {
 336                 ret = insert_mappable_node(ggtt, node, PAGE_SIZE);
 337                 if (ret)
 338                         goto err_ww;
 339                 GEM_BUG_ON(!drm_mm_node_allocated(node));
 340                 vma = NULL;
 341         }
 342
 343         ret = i915_gem_object_pin_pages(obj);
 344         if (ret) {
 345                 if (drm_mm_node_allocated(node)) {
 346                         ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
 347                         remove_mappable_node(ggtt, node);
 348                 } else {
 349                         i915_vma_unpin(vma);
 350                 }
 351         }
 352
 353 err_ww:
 354         if (ret == -EDEADLK) {
 355                 ret = i915_gem_ww_ctx_backoff(&ww);
 356                 if (!ret)
 357                         goto retry;
 358         }
 359         i915_gem_ww_ctx_fini(&ww);
 360
 361         return ret ? ERR_PTR(ret) : vma;
 362 }
 363
 364 static void i915_gem_gtt_cleanup(struct drm_i915_gem_object *obj,
 365                                  struct drm_mm_node *node,
 366                                  struct i915_vma *vma)
 367 {
 368         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 369         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 370
 371         i915_gem_object_unpin_pages(obj);
 372         if (drm_mm_node_allocated(node)) {
 373                 ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
 374                 remove_mappable_node(ggtt, node);
 375         } else {
 376                 i915_vma_unpin(vma);
 377         }
 378 }
 379
 380 static int
 381 i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
 382                    const struct drm_i915_gem_pread *args)
 383 {
 384         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 385         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 386         intel_wakeref_t wakeref;
 387         struct drm_mm_node node;
 388         void __user *user_data;
 389         struct i915_vma *vma;
 390         u64 remain, offset;
 391         int ret = 0;
 392
 393         wakeref = intel_runtime_pm_get(&i915->runtime_pm);
 394
 395         vma = i915_gem_gtt_prepare(obj, &node, false);
 396         if (IS_ERR(vma)) {
 397                 ret = PTR_ERR(vma);
 398                 goto out_rpm;
 399         }
 400
 401         user_data = u64_to_user_ptr(args->data_ptr);
 402         remain = args->size;
 403         offset = args->offset;
 404
 405         while (remain > 0) {
 406                 /* Operation in this page
 407                  *
 408                  * page_base = page offset within aperture
 409                  * page_offset = offset within page
 410                  * page_length = bytes to copy for this page
 411                  */
 412                 u32 page_base = node.start;
 413                 unsigned page_offset = offset_in_page(offset);
 414                 unsigned page_length = PAGE_SIZE - page_offset;
 415                 page_length = remain < page_length ? remain : page_length;
 416                 if (drm_mm_node_allocated(&node)) {
 417                         ggtt->vm.insert_page(&ggtt->vm,
 418                                              i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
 419                                              node.start, I915_CACHE_NONE, 0);
 420                 } else {
 421                         page_base += offset & PAGE_MASK;
 422                 }
 423
 424                 if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
 425                                   user_data, page_length)) {
 426                         ret = -EFAULT;
 427                         break;
 428                 }
 429
 430                 remain -= page_length;
 431                 user_data += page_length;
 432                 offset += page_length;
 433         }
 434
 435         i915_gem_gtt_cleanup(obj, &node, vma);
 436 out_rpm:
 437         intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 438         return ret;
 439 }
 440
 441 /**
 442  * Reads data from the object referenced by handle.
 443  * @dev: drm device pointer
 444  * @data: ioctl data blob
 445  * @file: drm file pointer
 446  *
 447  * On error, the contents of *data are undefined.
 448  */
 449 int
 450 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 451                      struct drm_file *file)
 452 {
 453         struct drm_i915_private *i915 = to_i915(dev);
 454         struct drm_i915_gem_pread *args = data;
 455         struct drm_i915_gem_object *obj;
 456         int ret;
 457
 458         /* PREAD is disallowed for all platforms after TGL-LP.  This also
 459          * covers all platforms with local memory.
 460          */
 461         if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
 462                 return -EOPNOTSUPP;
 463
 464         if (args->size == 0)
 465                 return 0;
 466
 467         if (!access_ok(u64_to_user_ptr(args->data_ptr),
 468                        args->size))
 469                 return -EFAULT;
 470
 471         obj = i915_gem_object_lookup(file, args->handle);
 472         if (!obj)
 473                 return -ENOENT;
 474
 475         /* Bounds check source.  */
 476         if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 477                 ret = -EINVAL;
 478                 goto out;
 479         }
 480
 481         trace_i915_gem_object_pread(obj, args->offset, args->size);
 482         ret = -ENODEV;
 483         if (obj->ops->pread)
 484                 ret = obj->ops->pread(obj, args);
 485         if (ret != -ENODEV)
 486                 goto out;
 487
 488         ret = i915_gem_object_wait(obj,
 489                                    I915_WAIT_INTERRUPTIBLE,
 490                                    MAX_SCHEDULE_TIMEOUT);
 491         if (ret)
 492                 goto out;
 493
 494         ret = i915_gem_shmem_pread(obj, args);
 495         if (ret == -EFAULT || ret == -ENODEV)
 496                 ret = i915_gem_gtt_pread(obj, args);
 497
 498 out:
 499         i915_gem_object_put(obj);
 500         return ret;
 501 }
 502
 503 /* This is the fast write path which cannot handle
 504  * page faults in the source data
 505  */
 506
 507 static inline bool
 508 ggtt_write(struct io_mapping *mapping,
 509            loff_t base, int offset,
 510            char __user *user_data, int length)
 511 {
 512         void __iomem *vaddr;
 513         unsigned long unwritten;
 514
 515         /* We can use the cpu mem copy function because this is X86. */
 516         vaddr = io_mapping_map_atomic_wc(mapping, base);
 517         unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
 518                                                       user_data, length);
 519         io_mapping_unmap_atomic(vaddr);
 520         if (unwritten) {
 521                 vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 522                 unwritten = copy_from_user((void __force *)vaddr + offset,
 523                                            user_data, length);
 524                 io_mapping_unmap(vaddr);
 525         }
 526
 527         return unwritten;
 528 }
 529
 530 /**
 531  * This is the fast pwrite path, where we copy the data directly from the
 532  * user into the GTT, uncached.
 533  * @obj: i915 GEM object
 534  * @args: pwrite arguments structure
 535  */
 536 static int
 537 i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 538                          const struct drm_i915_gem_pwrite *args)
 539 {
 540         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 541         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 542         struct intel_runtime_pm *rpm = &i915->runtime_pm;
 543         intel_wakeref_t wakeref;
 544         struct drm_mm_node node;
 545         struct i915_vma *vma;
 546         u64 remain, offset;
 547         void __user *user_data;
 548         int ret = 0;
 549
 550         if (i915_gem_object_has_struct_page(obj)) {
 551                 /*
 552                  * Avoid waking the device up if we can fallback, as
 553                  * waking/resuming is very slow (worst-case 10-100 ms
 554                  * depending on PCI sleeps and our own resume time).
 555                  * This easily dwarfs any performance advantage from
 556                  * using the cache bypass of indirect GGTT access.
 557                  */
 558                 wakeref = intel_runtime_pm_get_if_in_use(rpm);
 559                 if (!wakeref)
 560                         return -EFAULT;
 561         } else {
 562                 /* No backing pages, no fallback, we must force GGTT access */
 563                 wakeref = intel_runtime_pm_get(rpm);
 564         }
 565
 566         vma = i915_gem_gtt_prepare(obj, &node, true);
 567         if (IS_ERR(vma)) {
 568                 ret = PTR_ERR(vma);
 569                 goto out_rpm;
 570         }
 571
 572         i915_gem_object_invalidate_frontbuffer(obj, ORIGIN_CPU);
 573
 574         user_data = u64_to_user_ptr(args->data_ptr);
 575         offset = args->offset;
 576         remain = args->size;
 577         while (remain) {
 578                 /* Operation in this page
 579                  *
 580                  * page_base = page offset within aperture
 581                  * page_offset = offset within page
 582                  * page_length = bytes to copy for this page
 583                  */
 584                 u32 page_base = node.start;
 585                 unsigned int page_offset = offset_in_page(offset);
 586                 unsigned int page_length = PAGE_SIZE - page_offset;
 587                 page_length = remain < page_length ? remain : page_length;
 588                 if (drm_mm_node_allocated(&node)) {
 589                         /* flush the write before we modify the GGTT */
 590                         intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 591                         ggtt->vm.insert_page(&ggtt->vm,
 592                                              i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
 593                                              node.start, I915_CACHE_NONE, 0);
 594                         wmb(); /* flush modifications to the GGTT (insert_page) */
 595                 } else {
 596                         page_base += offset & PAGE_MASK;
 597                 }
 598                 /* If we get a fault while copying data, then (presumably) our
 599                  * source page isn't available.  Return the error and we'll
 600                  * retry in the slow path.
 601                  * If the object is non-shmem backed, we retry again with the
 602                  * path that handles page fault.
 603                  */
 604                 if (ggtt_write(&ggtt->iomap, page_base, page_offset,
 605                                user_data, page_length)) {
 606                         ret = -EFAULT;
 607                         break;
 608                 }
 609
 610                 remain -= page_length;
 611                 user_data += page_length;
 612                 offset += page_length;
 613         }
 614
 615         intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 616         i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 617
 618         i915_gem_gtt_cleanup(obj, &node, vma);
 619 out_rpm:
 620         intel_runtime_pm_put(rpm, wakeref);
 621         return ret;
 622 }
 623
 624 /* Per-page copy function for the shmem pwrite fastpath.
 625  * Flushes invalid cachelines before writing to the target if
 626  * needs_clflush_before is set and flushes out any written cachelines after
 627  * writing if needs_clflush is set.
 628  */
 629 static int
 630 shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
 631              bool needs_clflush_before,
 632              bool needs_clflush_after)
 633 {
 634         char *vaddr;
 635         int ret;
 636
 637         vaddr = kmap(page);
 638
 639         if (needs_clflush_before)
 640                 drm_clflush_virt_range(vaddr + offset, len);
 641
 642         ret = __copy_from_user(vaddr + offset, user_data, len);
 643         if (!ret && needs_clflush_after)
 644                 drm_clflush_virt_range(vaddr + offset, len);
 645
 646         kunmap(page);
 647
 648         return ret ? -EFAULT : 0;
 649 }
 650
 651 static int
 652 i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
 653                       const struct drm_i915_gem_pwrite *args)
 654 {
 655         unsigned int partial_cacheline_write;
 656         unsigned int needs_clflush;
 657         unsigned int offset, idx;
 658         void __user *user_data;
 659         u64 remain;
 660         int ret;
 661
 662         ret = i915_gem_object_lock_interruptible(obj, NULL);
 663         if (ret)
 664                 return ret;
 665
 666         ret = i915_gem_object_pin_pages(obj);
 667         if (ret)
 668                 goto err_unlock;
 669
 670         ret = i915_gem_object_prepare_write(obj, &needs_clflush);
 671         if (ret)
 672                 goto err_unpin;
 673
 674         i915_gem_object_finish_access(obj);
 675         i915_gem_object_unlock(obj);
 676
 677         /* If we don't overwrite a cacheline completely we need to be
 678          * careful to have up-to-date data by first clflushing. Don't
 679          * overcomplicate things and flush the entire patch.
 680          */
 681         partial_cacheline_write = 0;
 682         if (needs_clflush & CLFLUSH_BEFORE)
 683                 partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
 684
 685         user_data = u64_to_user_ptr(args->data_ptr);
 686         remain = args->size;
 687         offset = offset_in_page(args->offset);
 688         for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 689                 struct page *page = i915_gem_object_get_page(obj, idx);
 690                 unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 691
 692                 ret = shmem_pwrite(page, offset, length, user_data,
 693                                    (offset | length) & partial_cacheline_write,
 694                                    needs_clflush & CLFLUSH_AFTER);
 695                 if (ret)
 696                         break;
 697
 698                 remain -= length;
 699                 user_data += length;
 700                 offset = 0;
 701         }
 702
 703         i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 704
 705         i915_gem_object_unpin_pages(obj);
 706         return ret;
 707
 708 err_unpin:
 709         i915_gem_object_unpin_pages(obj);
 710 err_unlock:
 711         i915_gem_object_unlock(obj);
 712         return ret;
 713 }
 714
 715 /**
 716  * Writes data to the object referenced by handle.
 717  * @dev: drm device
 718  * @data: ioctl data blob
 719  * @file: drm file
 720  *
 721  * On error, the contents of the buffer that were to be modified are undefined.
 722  */
 723 int
 724 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 725                       struct drm_file *file)
 726 {
 727         struct drm_i915_private *i915 = to_i915(dev);
 728         struct drm_i915_gem_pwrite *args = data;
 729         struct drm_i915_gem_object *obj;
 730         int ret;
 731
 732         /* PWRITE is disallowed for all platforms after TGL-LP.  This also
 733          * covers all platforms with local memory.
 734          */
 735         if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
 736                 return -EOPNOTSUPP;
 737
 738         if (args->size == 0)
 739                 return 0;
 740
 741         if (!access_ok(u64_to_user_ptr(args->data_ptr), args->size))
 742                 return -EFAULT;
 743
 744         obj = i915_gem_object_lookup(file, args->handle);
 745         if (!obj)
 746                 return -ENOENT;
 747
 748         /* Bounds check destination. */
 749         if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 750                 ret = -EINVAL;
 751                 goto err;
 752         }
 753
 754         /* Writes not allowed into this read-only object */
 755         if (i915_gem_object_is_readonly(obj)) {
 756                 ret = -EINVAL;
 757                 goto err;
 758         }
 759
 760         trace_i915_gem_object_pwrite(obj, args->offset, args->size);
 761
 762         ret = -ENODEV;
 763         if (obj->ops->pwrite)
 764                 ret = obj->ops->pwrite(obj, args);
 765         if (ret != -ENODEV)
 766                 goto err;
 767
 768         ret = i915_gem_object_wait(obj,
 769                                    I915_WAIT_INTERRUPTIBLE |
 770                                    I915_WAIT_ALL,
 771                                    MAX_SCHEDULE_TIMEOUT);
 772         if (ret)
 773                 goto err;
 774
 775         ret = -EFAULT;
 776         /* We can only do the GTT pwrite on untiled buffers, as otherwise
 777          * it would end up going through the fenced access, and we'll get
 778          * different detiling behavior between reading and writing.
 779          * pread/pwrite currently are reading and writing from the CPU
 780          * perspective, requiring manual detiling by the client.
 781          */
 782         if (!i915_gem_object_has_struct_page(obj) ||
 783             i915_gem_cpu_write_needs_clflush(obj))
 784                 /* Note that the gtt paths might fail with non-page-backed user
 785                  * pointers (e.g. gtt mappings when moving data between
 786                  * textures). Fallback to the shmem path in that case.
 787                  */
 788                 ret = i915_gem_gtt_pwrite_fast(obj, args);
 789
 790         if (ret == -EFAULT || ret == -ENOSPC) {
 791                 if (i915_gem_object_has_struct_page(obj))
 792                         ret = i915_gem_shmem_pwrite(obj, args);
 793         }
 794
 795 err:
 796         i915_gem_object_put(obj);
 797         return ret;
 798 }
 799
 800 /**
 801  * Called when user space has done writes to this buffer
 802  * @dev: drm device
 803  * @data: ioctl data blob
 804  * @file: drm file
 805  */
 806 int
 807 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
 808                          struct drm_file *file)
 809 {
 810         struct drm_i915_gem_sw_finish *args = data;
 811         struct drm_i915_gem_object *obj;
 812
 813         obj = i915_gem_object_lookup(file, args->handle);
 814         if (!obj)
 815                 return -ENOENT;
 816
 817         /*
 818          * Proxy objects are barred from CPU access, so there is no
 819          * need to ban sw_finish as it is a nop.
 820          */
 821
 822         /* Pinned buffers may be scanout, so flush the cache */
 823         i915_gem_object_flush_if_display(obj);
 824         i915_gem_object_put(obj);
 825
 826         return 0;
 827 }
 828
 829 void i915_gem_runtime_suspend(struct drm_i915_private *i915)
 830 {
 831         struct drm_i915_gem_object *obj, *on;
 832         int i;
 833
 834         /*
 835          * Only called during RPM suspend. All users of the userfault_list
 836          * must be holding an RPM wakeref to ensure that this can not
 837          * run concurrently with themselves (and use the struct_mutex for
 838          * protection between themselves).
 839          */
 840
 841         list_for_each_entry_safe(obj, on,
 842                                  &to_gt(i915)->ggtt->userfault_list, userfault_link)
 843                 __i915_gem_object_release_mmap_gtt(obj);
 844
 845         /*
 846          * The fence will be lost when the device powers down. If any were
 847          * in use by hardware (i.e. they are pinned), we should not be powering
 848          * down! All other fences will be reacquired by the user upon waking.
 849          */
 850         for (i = 0; i < to_gt(i915)->ggtt->num_fences; i++) {
 851                 struct i915_fence_reg *reg = &to_gt(i915)->ggtt->fence_regs[i];
 852
 853                 /*
 854                  * Ideally we want to assert that the fence register is not
 855                  * live at this point (i.e. that no piece of code will be
 856                  * trying to write through fence + GTT, as that both violates
 857                  * our tracking of activity and associated locking/barriers,
 858                  * but also is illegal given that the hw is powered down).
 859                  *
 860                  * Previously we used reg->pin_count as a "liveness" indicator.
 861                  * That is not sufficient, and we need a more fine-grained
 862                  * tool if we want to have a sanity check here.
 863                  */
 864
 865                 if (!reg->vma)
 866                         continue;
 867
 868                 GEM_BUG_ON(i915_vma_has_userfault(reg->vma));
 869                 reg->dirty = true;
 870         }
 871 }
 872
 873 static void discard_ggtt_vma(struct i915_vma *vma)
 874 {
 875         struct drm_i915_gem_object *obj = vma->obj;
 876
 877         spin_lock(&obj->vma.lock);
 878         if (!RB_EMPTY_NODE(&vma->obj_node)) {
 879                 rb_erase(&vma->obj_node, &obj->vma.tree);
 880                 RB_CLEAR_NODE(&vma->obj_node);
 881         }
 882         spin_unlock(&obj->vma.lock);
 883 }
 884
 885 struct i915_vma *
 886 i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj,
 887                             struct i915_gem_ww_ctx *ww,
 888                             const struct i915_ggtt_view *view,
 889                             u64 size, u64 alignment, u64 flags)
 890 {
 891         struct drm_i915_private *i915 = to_i915(obj->base.dev);
 892         struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 893         struct i915_vma *vma;
 894         int ret;
 895
 896         GEM_WARN_ON(!ww);
 897
 898         if (flags & PIN_MAPPABLE &&
 899             (!view || view->type == I915_GGTT_VIEW_NORMAL)) {
 900                 /*
 901                  * If the required space is larger than the available
 902                  * aperture, we will not able to find a slot for the
 903                  * object and unbinding the object now will be in
 904                  * vain. Worse, doing so may cause us to ping-pong
 905                  * the object in and out of the Global GTT and
 906                  * waste a lot of cycles under the mutex.
 907                  */
 908                 if (obj->base.size > ggtt->mappable_end)
 909                         return ERR_PTR(-E2BIG);
 910
 911                 /*
 912                  * If NONBLOCK is set the caller is optimistically
 913                  * trying to cache the full object within the mappable
 914                  * aperture, and *must* have a fallback in place for
 915                  * situations where we cannot bind the object. We
 916                  * can be a little more lax here and use the fallback
 917                  * more often to avoid costly migrations of ourselves
 918                  * and other objects within the aperture.
 919                  *
 920                  * Half-the-aperture is used as a simple heuristic.
 921                  * More interesting would to do search for a free
 922                  * block prior to making the commitment to unbind.
 923                  * That caters for the self-harm case, and with a
 924                  * little more heuristics (e.g. NOFAULT, NOEVICT)
 925                  * we could try to minimise harm to others.
 926                  */
 927                 if (flags & PIN_NONBLOCK &&
 928                     obj->base.size > ggtt->mappable_end / 2)
 929                         return ERR_PTR(-ENOSPC);
 930         }
 931
 932 new_vma:
 933         vma = i915_vma_instance(obj, &ggtt->vm, view);
 934         if (IS_ERR(vma))
 935                 return vma;
 936
 937         if (i915_vma_misplaced(vma, size, alignment, flags)) {
 938                 if (flags & PIN_NONBLOCK) {
 939                         if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))
 940                                 return ERR_PTR(-ENOSPC);
 941
 942                         /*
 943                          * If this misplaced vma is too big (i.e, at-least
 944                          * half the size of aperture) or hasn't been pinned
 945                          * mappable before, we ignore the misplacement when
 946                          * PIN_NONBLOCK is set in order to avoid the ping-pong
 947                          * issue described above. In other words, we try to
 948                          * avoid the costly operation of unbinding this vma
 949                          * from the GGTT and rebinding it back because there
 950                          * may not be enough space for this vma in the aperture.
 951                          */
 952                         if (flags & PIN_MAPPABLE &&
 953                             (vma->fence_size > ggtt->mappable_end / 2 ||
 954                             !i915_vma_is_map_and_fenceable(vma)))
 955                                 return ERR_PTR(-ENOSPC);
 956                 }
 957
 958                 if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma)) {
 959                         discard_ggtt_vma(vma);
 960                         goto new_vma;
 961                 }
 962
 963                 ret = i915_vma_unbind(vma);
 964                 if (ret)
 965                         return ERR_PTR(ret);
 966         }
 967
 968         ret = i915_vma_pin_ww(vma, ww, size, alignment, flags | PIN_GLOBAL);
 969
 970         if (ret)
 971                 return ERR_PTR(ret);
 972
 973         if (vma->fence && !i915_gem_object_is_tiled(obj)) {
 974                 mutex_lock(&ggtt->vm.mutex);
 975                 i915_vma_revoke_fence(vma);
 976                 mutex_unlock(&ggtt->vm.mutex);
 977         }
 978
 979         ret = i915_vma_wait_for_bind(vma);
 980         if (ret) {
 981                 i915_vma_unpin(vma);
 982                 return ERR_PTR(ret);
 983         }
 984
 985         return vma;
 986 }
 987
 988 struct i915_vma * __must_check
 989 i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
 990                          const struct i915_ggtt_view *view,
 991                          u64 size, u64 alignment, u64 flags)
 992 {
 993         struct i915_gem_ww_ctx ww;
 994         struct i915_vma *ret;
 995         int err;
 996
 997         for_i915_gem_ww(&ww, err, true) {
 998                 err = i915_gem_object_lock(obj, &ww);
 999                 if (err)
1000                         continue;
1001
1002                 ret = i915_gem_object_ggtt_pin_ww(obj, &ww, view, size,
1003                                                   alignment, flags);
1004                 if (IS_ERR(ret))
1005                         err = PTR_ERR(ret);
1006         }
1007
1008         return err ? ERR_PTR(err) : ret;
1009 }
1010
1011 int
1012 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
1013                        struct drm_file *file_priv)
1014 {
1015         struct drm_i915_private *i915 = to_i915(dev);
1016         struct drm_i915_gem_madvise *args = data;
1017         struct drm_i915_gem_object *obj;
1018         int err;
1019
1020         switch (args->madv) {
1021         case I915_MADV_DONTNEED:
1022         case I915_MADV_WILLNEED:
1023             break;
1024         default:
1025             return -EINVAL;
1026         }
1027
1028         obj = i915_gem_object_lookup(file_priv, args->handle);
1029         if (!obj)
1030                 return -ENOENT;
1031
1032         err = i915_gem_object_lock_interruptible(obj, NULL);
1033         if (err)
1034                 goto out;
1035
1036         if (i915_gem_object_has_pages(obj) &&
1037             i915_gem_object_is_tiled(obj) &&
1038             i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
1039                 if (obj->mm.madv == I915_MADV_WILLNEED) {
1040                         GEM_BUG_ON(!i915_gem_object_has_tiling_quirk(obj));
1041                         i915_gem_object_clear_tiling_quirk(obj);
1042                         i915_gem_object_make_shrinkable(obj);
1043                 }
1044                 if (args->madv == I915_MADV_WILLNEED) {
1045                         GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));
1046                         i915_gem_object_make_unshrinkable(obj);
1047                         i915_gem_object_set_tiling_quirk(obj);
1048                 }
1049         }
1050
1051         if (obj->mm.madv != __I915_MADV_PURGED) {
1052                 obj->mm.madv = args->madv;
1053                 if (obj->ops->adjust_lru)
1054                         obj->ops->adjust_lru(obj);
1055         }
1056
1057         if (i915_gem_object_has_pages(obj) ||
1058             i915_gem_object_has_self_managed_shrink_list(obj)) {
1059                 unsigned long flags;
1060
1061                 spin_lock_irqsave(&i915->mm.obj_lock, flags);
1062                 if (!list_empty(&obj->mm.link)) {
1063                         struct list_head *list;
1064
1065                         if (obj->mm.madv != I915_MADV_WILLNEED)
1066                                 list = &i915->mm.purge_list;
1067                         else
1068                                 list = &i915->mm.shrink_list;
1069                         list_move_tail(&obj->mm.link, list);
1070
1071                 }
1072                 spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
1073         }
1074
1075         /* if the object is no longer attached, discard its backing storage */
1076         if (obj->mm.madv == I915_MADV_DONTNEED &&
1077             !i915_gem_object_has_pages(obj))
1078                 i915_gem_object_truncate(obj);
1079
1080         args->retained = obj->mm.madv != __I915_MADV_PURGED;
1081
1082         i915_gem_object_unlock(obj);
1083 out:
1084         i915_gem_object_put(obj);
1085         return err;
1086 }
1087
1088 int i915_gem_init(struct drm_i915_private *dev_priv)
1089 {
1090         int ret;
1091
1092         /* We need to fallback to 4K pages if host doesn't support huge gtt. */
1093         if (intel_vgpu_active(dev_priv) && !intel_vgpu_has_huge_gtt(dev_priv))
1094                 mkwrite_device_info(dev_priv)->page_sizes =
1095                         I915_GTT_PAGE_SIZE_4K;
1096
1097         ret = i915_gem_init_userptr(dev_priv);
1098         if (ret)
1099                 return ret;
1100
1101         intel_uc_fetch_firmwares(&to_gt(dev_priv)->uc);
1102         intel_wopcm_init(&dev_priv->wopcm);
1103
1104         ret = i915_init_ggtt(dev_priv);
1105         if (ret) {
1106                 GEM_BUG_ON(ret == -EIO);
1107                 goto err_unlock;
1108         }
1109
1110         /*
1111          * Despite its name intel_init_clock_gating applies both display
1112          * clock gating workarounds; GT mmio workarounds and the occasional
1113          * GT power context workaround. Worse, sometimes it includes a context
1114          * register workaround which we need to apply before we record the
1115          * default HW state for all contexts.
1116          *
1117          * FIXME: break up the workarounds and apply them at the right time!
1118          */
1119         intel_init_clock_gating(dev_priv);
1120
1121         ret = intel_gt_init(to_gt(dev_priv));
1122         if (ret)
1123                 goto err_unlock;
1124
1125         return 0;
1126
1127         /*
1128          * Unwinding is complicated by that we want to handle -EIO to mean
1129          * disable GPU submission but keep KMS alive. We want to mark the
1130          * HW as irrevisibly wedged, but keep enough state around that the
1131          * driver doesn't explode during runtime.
1132          */
1133 err_unlock:
1134         i915_gem_drain_workqueue(dev_priv);
1135
1136         if (ret != -EIO)
1137                 intel_uc_cleanup_firmwares(&to_gt(dev_priv)->uc);
1138
1139         if (ret == -EIO) {
1140                 /*
1141                  * Allow engines or uC initialisation to fail by marking the GPU
1142                  * as wedged. But we only want to do this when the GPU is angry,
1143                  * for all other failure, such as an allocation failure, bail.
1144                  */
1145                 if (!intel_gt_is_wedged(to_gt(dev_priv))) {
1146                         i915_probe_error(dev_priv,
1147                                          "Failed to initialize GPU, declaring it wedged!\n");
1148                         intel_gt_set_wedged(to_gt(dev_priv));
1149                 }
1150
1151                 /* Minimal basic recovery for KMS */
1152                 ret = i915_ggtt_enable_hw(dev_priv);
1153                 i915_ggtt_resume(to_gt(dev_priv)->ggtt);
1154                 intel_init_clock_gating(dev_priv);
1155         }
1156
1157         i915_gem_drain_freed_objects(dev_priv);
1158
1159         return ret;
1160 }
1161
1162 void i915_gem_driver_register(struct drm_i915_private *i915)
1163 {
1164         i915_gem_driver_register__shrinker(i915);
1165
1166         intel_engines_driver_register(i915);
1167 }
1168
1169 void i915_gem_driver_unregister(struct drm_i915_private *i915)
1170 {
1171         i915_gem_driver_unregister__shrinker(i915);
1172 }
1173
1174 void i915_gem_driver_remove(struct drm_i915_private *dev_priv)
1175 {
1176         intel_wakeref_auto_fini(&to_gt(dev_priv)->ggtt->userfault_wakeref);
1177
1178         i915_gem_suspend_late(dev_priv);
1179         intel_gt_driver_remove(to_gt(dev_priv));
1180         dev_priv->uabi_engines = RB_ROOT;
1181
1182         /* Flush any outstanding unpin_work. */
1183         i915_gem_drain_workqueue(dev_priv);
1184
1185         i915_gem_drain_freed_objects(dev_priv);
1186 }
1187
1188 void i915_gem_driver_release(struct drm_i915_private *dev_priv)
1189 {
1190         intel_gt_driver_release(to_gt(dev_priv));
1191
1192         intel_uc_cleanup_firmwares(&to_gt(dev_priv)->uc);
1193
1194         i915_gem_drain_freed_objects(dev_priv);
1195
1196         drm_WARN_ON(&dev_priv->drm, !list_empty(&dev_priv->gem.contexts.list));
1197 }
1198
1199 static void i915_gem_init__mm(struct drm_i915_private *i915)
1200 {
1201         spin_lock_init(&i915->mm.obj_lock);
1202
1203         init_llist_head(&i915->mm.free_list);
1204
1205         INIT_LIST_HEAD(&i915->mm.purge_list);
1206         INIT_LIST_HEAD(&i915->mm.shrink_list);
1207
1208         i915_gem_init__objects(i915);
1209 }
1210
1211 void i915_gem_init_early(struct drm_i915_private *dev_priv)
1212 {
1213         i915_gem_init__mm(dev_priv);
1214         i915_gem_init__contexts(dev_priv);
1215
1216         spin_lock_init(&dev_priv->fb_tracking.lock);
1217 }
1218
1219 void i915_gem_cleanup_early(struct drm_i915_private *dev_priv)
1220 {
1221         i915_gem_drain_freed_objects(dev_priv);
1222         GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list));
1223         GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count));
1224         drm_WARN_ON(&dev_priv->drm, dev_priv->mm.shrink_count);
1225 }
1226
1227 int i915_gem_open(struct drm_i915_private *i915, struct drm_file *file)
1228 {
1229         struct drm_i915_file_private *file_priv;
1230         struct i915_drm_client *client;
1231         int ret = -ENOMEM;
1232
1233         DRM_DEBUG("\n");
1234
1235         file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
1236         if (!file_priv)
1237                 goto err_alloc;
1238
1239         client = i915_drm_client_add(&i915->clients);
1240         if (IS_ERR(client)) {
1241                 ret = PTR_ERR(client);
1242                 goto err_client;
1243         }
1244
1245         file->driver_priv = file_priv;
1246         file_priv->dev_priv = i915;
1247         file_priv->file = file;
1248         file_priv->client = client;
1249
1250         file_priv->bsd_engine = -1;
1251         file_priv->hang_timestamp = jiffies;
1252
1253         ret = i915_gem_context_open(i915, file);
1254         if (ret)
1255                 goto err_context;
1256
1257         return 0;
1258
1259 err_context:
1260         i915_drm_client_put(client);
1261 err_client:
1262         kfree(file_priv);
1263 err_alloc:
1264         return ret;
1265 }
1266
1267 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1268 #include "selftests/mock_gem_device.c"
1269 #include "selftests/i915_gem.c"
1270 #endif