drivers/gpu/drm/i915/i915_gem_clflush.c

   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 "i915_drv.h"
  26 #include "intel_frontbuffer.h"
  27 #include "i915_gem_clflush.h"
  28
  29 static DEFINE_SPINLOCK(clflush_lock);
  30
  31 struct clflush {
  32         struct dma_fence dma; /* Must be first for dma_fence_free() */
  33         struct i915_sw_fence wait;
  34         struct work_struct work;
  35         struct drm_i915_gem_object *obj;
  36 };
  37
  38 static const char *i915_clflush_get_driver_name(struct dma_fence *fence)
  39 {
  40         return DRIVER_NAME;
  41 }
  42
  43 static const char *i915_clflush_get_timeline_name(struct dma_fence *fence)
  44 {
  45         return "clflush";
  46 }
  47
  48 static bool i915_clflush_enable_signaling(struct dma_fence *fence)
  49 {
  50         return true;
  51 }
  52
  53 static void i915_clflush_release(struct dma_fence *fence)
  54 {
  55         struct clflush *clflush = container_of(fence, typeof(*clflush), dma);
  56
  57         i915_sw_fence_fini(&clflush->wait);
  58
  59         BUILD_BUG_ON(offsetof(typeof(*clflush), dma));
  60         dma_fence_free(&clflush->dma);
  61 }
  62
  63 static const struct dma_fence_ops i915_clflush_ops = {
  64         .get_driver_name = i915_clflush_get_driver_name,
  65         .get_timeline_name = i915_clflush_get_timeline_name,
  66         .enable_signaling = i915_clflush_enable_signaling,
  67         .wait = dma_fence_default_wait,
  68         .release = i915_clflush_release,
  69 };
  70
  71 static void __i915_do_clflush(struct drm_i915_gem_object *obj)
  72 {
  73         drm_clflush_sg(obj->mm.pages);
  74         intel_fb_obj_flush(obj, ORIGIN_CPU);
  75 }
  76
  77 static void i915_clflush_work(struct work_struct *work)
  78 {
  79         struct clflush *clflush = container_of(work, typeof(*clflush), work);
  80         struct drm_i915_gem_object *obj = clflush->obj;
  81
  82         if (i915_gem_object_pin_pages(obj)) {
  83                 DRM_ERROR("Failed to acquire obj->pages for clflushing\n");
  84                 goto out;
  85         }
  86
  87         __i915_do_clflush(obj);
  88
  89         i915_gem_object_unpin_pages(obj);
  90
  91 out:
  92         i915_gem_object_put(obj);
  93
  94         dma_fence_signal(&clflush->dma);
  95         dma_fence_put(&clflush->dma);
  96 }
  97
  98 static int __i915_sw_fence_call
  99 i915_clflush_notify(struct i915_sw_fence *fence,
 100                     enum i915_sw_fence_notify state)
 101 {
 102         struct clflush *clflush = container_of(fence, typeof(*clflush), wait);
 103
 104         switch (state) {
 105         case FENCE_COMPLETE:
 106                 schedule_work(&clflush->work);
 107                 break;
 108
 109         case FENCE_FREE:
 110                 dma_fence_put(&clflush->dma);
 111                 break;
 112         }
 113
 114         return NOTIFY_DONE;
 115 }
 116
 117 bool i915_gem_clflush_object(struct drm_i915_gem_object *obj,
 118                              unsigned int flags)
 119 {
 120         struct clflush *clflush;
 121
 122         /*
 123          * Stolen memory is always coherent with the GPU as it is explicitly
 124          * marked as wc by the system, or the system is cache-coherent.
 125          * Similarly, we only access struct pages through the CPU cache, so
 126          * anything not backed by physical memory we consider to be always
 127          * coherent and not need clflushing.
 128          */
 129         if (!i915_gem_object_has_struct_page(obj)) {
 130                 obj->cache_dirty = false;
 131                 return false;
 132         }
 133
 134         /* If the GPU is snooping the contents of the CPU cache,
 135          * we do not need to manually clear the CPU cache lines.  However,
 136          * the caches are only snooped when the render cache is
 137          * flushed/invalidated.  As we always have to emit invalidations
 138          * and flushes when moving into and out of the RENDER domain, correct
 139          * snooping behaviour occurs naturally as the result of our domain
 140          * tracking.
 141          */
 142         if (!(flags & I915_CLFLUSH_FORCE) && obj->cache_coherent)
 143                 return false;
 144
 145         trace_i915_gem_object_clflush(obj);
 146
 147         clflush = NULL;
 148         if (!(flags & I915_CLFLUSH_SYNC))
 149                 clflush = kmalloc(sizeof(*clflush), GFP_KERNEL);
 150         if (clflush) {
 151                 GEM_BUG_ON(!obj->cache_dirty);
 152
 153                 dma_fence_init(&clflush->dma,
 154                                &i915_clflush_ops,
 155                                &clflush_lock,
 156                                to_i915(obj->base.dev)->mm.unordered_timeline,
 157                                0);
 158                 i915_sw_fence_init(&clflush->wait, i915_clflush_notify);
 159
 160                 clflush->obj = i915_gem_object_get(obj);
 161                 INIT_WORK(&clflush->work, i915_clflush_work);
 162
 163                 dma_fence_get(&clflush->dma);
 164
 165                 i915_sw_fence_await_reservation(&clflush->wait,
 166                                                 obj->resv, NULL,
 167                                                 true, I915_FENCE_TIMEOUT,
 168                                                 GFP_KERNEL);
 169
 170                 reservation_object_lock(obj->resv, NULL);
 171                 reservation_object_add_excl_fence(obj->resv, &clflush->dma);
 172                 reservation_object_unlock(obj->resv);
 173
 174                 i915_sw_fence_commit(&clflush->wait);
 175         } else if (obj->mm.pages) {
 176                 __i915_do_clflush(obj);
 177         } else {
 178                 GEM_BUG_ON(obj->base.write_domain != I915_GEM_DOMAIN_CPU);
 179         }
 180
 181         obj->cache_dirty = false;
 182         return true;
 183 }