1 /* $NetBSD: i915_gem_clflush.c,v 1.5 2021/12/19 11:33:30 riastradh Exp $ */ 2 3 /* 4 * SPDX-License-Identifier: MIT 5 * 6 * Copyright 2016 Intel Corporation 7 */ 8 9 #include <sys/cdefs.h> 10 __KERNEL_RCSID(0, "$NetBSD: i915_gem_clflush.c,v 1.5 2021/12/19 11:33:30 riastradh Exp $"); 11 12 #include "display/intel_frontbuffer.h" 13 14 #include "i915_drv.h" 15 #include "i915_gem_clflush.h" 16 #include "i915_sw_fence_work.h" 17 #include "i915_trace.h" 18 19 struct clflush { 20 struct dma_fence_work base; 21 struct drm_i915_gem_object *obj; 22 }; 23 24 static void __do_clflush(struct drm_i915_gem_object *obj) 25 { 26 GEM_BUG_ON(!i915_gem_object_has_pages(obj)); 27 drm_clflush_sg(obj->mm.pages); 28 29 i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU); 30 } 31 32 static int clflush_work(struct dma_fence_work *base) 33 { 34 struct clflush *clflush = container_of(base, typeof(*clflush), base); 35 struct drm_i915_gem_object *obj = clflush->obj; 36 int err; 37 38 err = i915_gem_object_pin_pages(obj); 39 if (err) 40 return err; 41 42 __do_clflush(obj); 43 i915_gem_object_unpin_pages(obj); 44 45 return 0; 46 } 47 48 static void clflush_release(struct dma_fence_work *base) 49 { 50 struct clflush *clflush = container_of(base, typeof(*clflush), base); 51 52 i915_gem_object_put(clflush->obj); 53 } 54 55 static const struct dma_fence_work_ops clflush_ops = { 56 .name = "clflush", 57 .work = clflush_work, 58 .release = clflush_release, 59 }; 60 61 static struct clflush *clflush_work_create(struct drm_i915_gem_object *obj) 62 { 63 struct clflush *clflush; 64 65 GEM_BUG_ON(!obj->cache_dirty); 66 67 clflush = kmalloc(sizeof(*clflush), GFP_KERNEL); 68 if (!clflush) 69 return NULL; 70 71 dma_fence_work_init(&clflush->base, &clflush_ops); 72 clflush->obj = i915_gem_object_get(obj); /* obj <-> clflush cycle */ 73 74 return clflush; 75 } 76 77 bool i915_gem_clflush_object(struct drm_i915_gem_object *obj, 78 unsigned int flags) 79 { 80 struct clflush *clflush; 81 82 assert_object_held(obj); 83 84 /* 85 * Stolen memory is always coherent with the GPU as it is explicitly 86 * marked as wc by the system, or the system is cache-coherent. 87 * Similarly, we only access struct pages through the CPU cache, so 88 * anything not backed by physical memory we consider to be always 89 * coherent and not need clflushing. 90 */ 91 if (!i915_gem_object_has_struct_page(obj)) { 92 obj->cache_dirty = false; 93 return false; 94 } 95 96 /* If the GPU is snooping the contents of the CPU cache, 97 * we do not need to manually clear the CPU cache lines. However, 98 * the caches are only snooped when the render cache is 99 * flushed/invalidated. As we always have to emit invalidations 100 * and flushes when moving into and out of the RENDER domain, correct 101 * snooping behaviour occurs naturally as the result of our domain 102 * tracking. 103 */ 104 if (!(flags & I915_CLFLUSH_FORCE) && 105 obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) 106 return false; 107 108 trace_i915_gem_object_clflush(obj); 109 110 clflush = NULL; 111 if (!(flags & I915_CLFLUSH_SYNC)) 112 clflush = clflush_work_create(obj); 113 if (clflush) { 114 i915_sw_fence_await_reservation(&clflush->base.chain, 115 obj->base.resv, NULL, true, 116 I915_FENCE_TIMEOUT, 117 I915_FENCE_GFP); 118 dma_resv_add_excl_fence(obj->base.resv, &clflush->base.dma); 119 dma_fence_work_commit(&clflush->base); 120 } else if (obj->mm.pages) { 121 __do_clflush(obj); 122 } else { 123 GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU); 124 } 125 126 obj->cache_dirty = false; 127 return true; 128 } 129