1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2016 Intel Corporation 5 */ 6 7 #include "display/intel_frontbuffer.h" 8 9 #include "i915_drv.h" 10 #include "i915_gem_clflush.h" 11 #include "i915_sw_fence_work.h" 12 #include "i915_trace.h" 13 14 struct clflush { 15 struct dma_fence_work base; 16 struct drm_i915_gem_object *obj; 17 }; 18 19 static void __do_clflush(struct drm_i915_gem_object *obj) 20 { 21 GEM_BUG_ON(!i915_gem_object_has_pages(obj)); 22 drm_clflush_sg(obj->mm.pages); 23 24 i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU); 25 } 26 27 static int clflush_work(struct dma_fence_work *base) 28 { 29 struct clflush *clflush = container_of(base, typeof(*clflush), base); 30 struct drm_i915_gem_object *obj = fetch_and_zero(&clflush->obj); 31 int err; 32 33 err = i915_gem_object_pin_pages(obj); 34 if (err) 35 goto put; 36 37 __do_clflush(obj); 38 i915_gem_object_unpin_pages(obj); 39 40 put: 41 i915_gem_object_put(obj); 42 return err; 43 } 44 45 static void clflush_release(struct dma_fence_work *base) 46 { 47 struct clflush *clflush = container_of(base, typeof(*clflush), base); 48 49 if (clflush->obj) 50 i915_gem_object_put(clflush->obj); 51 } 52 53 static const struct dma_fence_work_ops clflush_ops = { 54 .name = "clflush", 55 .work = clflush_work, 56 .release = clflush_release, 57 }; 58 59 static struct clflush *clflush_work_create(struct drm_i915_gem_object *obj) 60 { 61 struct clflush *clflush; 62 63 GEM_BUG_ON(!obj->cache_dirty); 64 65 clflush = kmalloc(sizeof(*clflush), GFP_KERNEL); 66 if (!clflush) 67 return NULL; 68 69 dma_fence_work_init(&clflush->base, &clflush_ops); 70 clflush->obj = i915_gem_object_get(obj); /* obj <-> clflush cycle */ 71 72 return clflush; 73 } 74 75 bool i915_gem_clflush_object(struct drm_i915_gem_object *obj, 76 unsigned int flags) 77 { 78 struct clflush *clflush; 79 80 assert_object_held(obj); 81 82 /* 83 * Stolen memory is always coherent with the GPU as it is explicitly 84 * marked as wc by the system, or the system is cache-coherent. 85 * Similarly, we only access struct pages through the CPU cache, so 86 * anything not backed by physical memory we consider to be always 87 * coherent and not need clflushing. 88 */ 89 if (!i915_gem_object_has_struct_page(obj)) { 90 obj->cache_dirty = false; 91 return false; 92 } 93 94 /* If the GPU is snooping the contents of the CPU cache, 95 * we do not need to manually clear the CPU cache lines. However, 96 * the caches are only snooped when the render cache is 97 * flushed/invalidated. As we always have to emit invalidations 98 * and flushes when moving into and out of the RENDER domain, correct 99 * snooping behaviour occurs naturally as the result of our domain 100 * tracking. 101 */ 102 if (!(flags & I915_CLFLUSH_FORCE) && 103 obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) 104 return false; 105 106 trace_i915_gem_object_clflush(obj); 107 108 clflush = NULL; 109 if (!(flags & I915_CLFLUSH_SYNC)) 110 clflush = clflush_work_create(obj); 111 if (clflush) { 112 i915_sw_fence_await_reservation(&clflush->base.chain, 113 obj->base.resv, NULL, true, 114 I915_FENCE_TIMEOUT, 115 I915_FENCE_GFP); 116 dma_resv_add_excl_fence(obj->base.resv, &clflush->base.dma); 117 dma_fence_work_commit(&clflush->base); 118 } else if (obj->mm.pages) { 119 __do_clflush(obj); 120 } else { 121 GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU); 122 } 123 124 obj->cache_dirty = false; 125 return true; 126 } 127