1// SPDX-License-Identifier: MIT
2/*
3 * Copyright © 2016 Intel Corporation
4 */
5
6#include <drm/drm_cache.h>
7
8#include "i915_config.h"
9#include "i915_drv.h"
10#include "i915_gem_clflush.h"
11#include "i915_gem_object_frontbuffer.h"
12#include "i915_sw_fence_work.h"
13#include "i915_trace.h"
14
15struct clflush {
16 struct dma_fence_work base;
17 struct drm_i915_gem_object *obj;
18};
19
20static void __do_clflush(struct drm_i915_gem_object *obj)
21{
22 GEM_BUG_ON(!i915_gem_object_has_pages(obj));
23 drm_clflush_sg(st: obj->mm.pages);
24
25 i915_gem_object_flush_frontbuffer(obj, origin: ORIGIN_CPU);
26}
27
28static void clflush_work(struct dma_fence_work *base)
29{
30 struct clflush *clflush = container_of(base, typeof(*clflush), base);
31
32 __do_clflush(obj: clflush->obj);
33}
34
35static void clflush_release(struct dma_fence_work *base)
36{
37 struct clflush *clflush = container_of(base, typeof(*clflush), base);
38
39 i915_gem_object_unpin_pages(obj: clflush->obj);
40 i915_gem_object_put(obj: clflush->obj);
41}
42
43static const struct dma_fence_work_ops clflush_ops = {
44 .name = "clflush",
45 .work = clflush_work,
46 .release = clflush_release,
47};
48
49static struct clflush *clflush_work_create(struct drm_i915_gem_object *obj)
50{
51 struct clflush *clflush;
52
53 GEM_BUG_ON(!obj->cache_dirty);
54
55 clflush = kmalloc(sizeof(*clflush), GFP_KERNEL);
56 if (!clflush)
57 return NULL;
58
59 if (__i915_gem_object_get_pages(obj) < 0) {
60 kfree(objp: clflush);
61 return NULL;
62 }
63
64 dma_fence_work_init(f: &clflush->base, ops: &clflush_ops);
65 clflush->obj = i915_gem_object_get(obj); /* obj <-> clflush cycle */
66
67 return clflush;
68}
69
70bool i915_gem_clflush_object(struct drm_i915_gem_object *obj,
71 unsigned int flags)
72{
73 struct drm_i915_private *i915 = to_i915(dev: obj->base.dev);
74 struct clflush *clflush;
75
76 assert_object_held(obj);
77
78 if (IS_DGFX(i915)) {
79 WARN_ON_ONCE(obj->cache_dirty);
80 return false;
81 }
82
83 /*
84 * Stolen memory is always coherent with the GPU as it is explicitly
85 * marked as wc by the system, or the system is cache-coherent.
86 * Similarly, we only access struct pages through the CPU cache, so
87 * anything not backed by physical memory we consider to be always
88 * coherent and not need clflushing.
89 */
90 if (!i915_gem_object_has_struct_page(obj)) {
91 obj->cache_dirty = false;
92 return false;
93 }
94
95 /* If the GPU is snooping the contents of the CPU cache,
96 * we do not need to manually clear the CPU cache lines. However,
97 * the caches are only snooped when the render cache is
98 * flushed/invalidated. As we always have to emit invalidations
99 * and flushes when moving into and out of the RENDER domain, correct
100 * snooping behaviour occurs naturally as the result of our domain
101 * tracking.
102 */
103 if (!(flags & I915_CLFLUSH_FORCE) &&
104 obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)
105 return false;
106
107 trace_i915_gem_object_clflush(obj);
108
109 clflush = NULL;
110 if (!(flags & I915_CLFLUSH_SYNC) &&
111 dma_resv_reserve_fences(obj: obj->base.resv, num_fences: 1) == 0)
112 clflush = clflush_work_create(obj);
113 if (clflush) {
114 i915_sw_fence_await_reservation(fence: &clflush->base.chain,
115 resv: obj->base.resv, write: true,
116 timeout: i915_fence_timeout(i915),
117 I915_FENCE_GFP);
118 dma_resv_add_fence(obj: obj->base.resv, fence: &clflush->base.dma,
119 usage: DMA_RESV_USAGE_KERNEL);
120 dma_fence_work_commit(f: &clflush->base);
121 /*
122 * We must have successfully populated the pages(since we are
123 * holding a pin on the pages as per the flush worker) to reach
124 * this point, which must mean we have already done the required
125 * flush-on-acquire, hence resetting cache_dirty here should be
126 * safe.
127 */
128 obj->cache_dirty = false;
129 } else if (obj->mm.pages) {
130 __do_clflush(obj);
131 obj->cache_dirty = false;
132 } else {
133 GEM_BUG_ON(obj->write_domain != I915_GEM_DOMAIN_CPU);
134 }
135
136 return true;
137}
138