1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * dma-fence-util: misc functions for dma_fence objects
4  *
5  * Copyright (C) 2022 Advanced Micro Devices, Inc.
6  * Authors:
7  *	Christian König <christian.koenig@amd.com>
8  */
9 
10 #include <linux/dma-fence.h>
11 #include <linux/dma-fence-array.h>
12 #include <linux/dma-fence-chain.h>
13 #include <linux/dma-fence-unwrap.h>
14 #include <linux/slab.h>
15 
16 /* Internal helper to start new array iteration, don't use directly */
17 static struct dma_fence *
18 __dma_fence_unwrap_array(struct dma_fence_unwrap *cursor)
19 {
20 	cursor->array = dma_fence_chain_contained(cursor->chain);
21 	cursor->index = 0;
22 	return dma_fence_array_first(cursor->array);
23 }
24 
25 /**
26  * dma_fence_unwrap_first - return the first fence from fence containers
27  * @head: the entrypoint into the containers
28  * @cursor: current position inside the containers
29  *
30  * Unwraps potential dma_fence_chain/dma_fence_array containers and return the
31  * first fence.
32  */
33 struct dma_fence *dma_fence_unwrap_first(struct dma_fence *head,
34 					 struct dma_fence_unwrap *cursor)
35 {
36 	cursor->chain = dma_fence_get(head);
37 	return __dma_fence_unwrap_array(cursor);
38 }
39 EXPORT_SYMBOL_GPL(dma_fence_unwrap_first);
40 
41 /**
42  * dma_fence_unwrap_next - return the next fence from a fence containers
43  * @cursor: current position inside the containers
44  *
45  * Continue unwrapping the dma_fence_chain/dma_fence_array containers and return
46  * the next fence from them.
47  */
48 struct dma_fence *dma_fence_unwrap_next(struct dma_fence_unwrap *cursor)
49 {
50 	struct dma_fence *tmp;
51 
52 	++cursor->index;
53 	tmp = dma_fence_array_next(cursor->array, cursor->index);
54 	if (tmp)
55 		return tmp;
56 
57 	cursor->chain = dma_fence_chain_walk(cursor->chain);
58 	return __dma_fence_unwrap_array(cursor);
59 }
60 EXPORT_SYMBOL_GPL(dma_fence_unwrap_next);
61 
62 /* Implementation for the dma_fence_merge() marco, don't use directly */
63 struct dma_fence *__dma_fence_unwrap_merge(unsigned int num_fences,
64 					   struct dma_fence **fences,
65 					   struct dma_fence_unwrap *iter)
66 {
67 	struct dma_fence_array *result;
68 	struct dma_fence *tmp, **array;
69 	ktime_t timestamp;
70 	unsigned int i;
71 	size_t count;
72 
73 	count = 0;
74 	timestamp = ns_to_ktime(0);
75 	for (i = 0; i < num_fences; ++i) {
76 		dma_fence_unwrap_for_each(tmp, &iter[i], fences[i]) {
77 			if (!dma_fence_is_signaled(tmp)) {
78 				++count;
79 			} else {
80 				ktime_t t = dma_fence_timestamp(tmp);
81 
82 				if (ktime_after(t, timestamp))
83 					timestamp = t;
84 			}
85 		}
86 	}
87 
88 	/*
89 	 * If we couldn't find a pending fence just return a private signaled
90 	 * fence with the timestamp of the last signaled one.
91 	 */
92 	if (count == 0)
93 		return dma_fence_allocate_private_stub(timestamp);
94 
95 	array = kmalloc_array(count, sizeof(*array), GFP_KERNEL);
96 	if (!array)
97 		return NULL;
98 
99 	/*
100 	 * This trashes the input fence array and uses it as position for the
101 	 * following merge loop. This works because the dma_fence_merge()
102 	 * wrapper macro is creating this temporary array on the stack together
103 	 * with the iterators.
104 	 */
105 	for (i = 0; i < num_fences; ++i)
106 		fences[i] = dma_fence_unwrap_first(fences[i], &iter[i]);
107 
108 	count = 0;
109 	do {
110 		unsigned int sel;
111 
112 restart:
113 		tmp = NULL;
114 		for (i = 0; i < num_fences; ++i) {
115 			struct dma_fence *next;
116 
117 			while (fences[i] && dma_fence_is_signaled(fences[i]))
118 				fences[i] = dma_fence_unwrap_next(&iter[i]);
119 
120 			next = fences[i];
121 			if (!next)
122 				continue;
123 
124 			/*
125 			 * We can't guarantee that inpute fences are ordered by
126 			 * context, but it is still quite likely when this
127 			 * function is used multiple times. So attempt to order
128 			 * the fences by context as we pass over them and merge
129 			 * fences with the same context.
130 			 */
131 			if (!tmp || tmp->context > next->context) {
132 				tmp = next;
133 				sel = i;
134 
135 			} else if (tmp->context < next->context) {
136 				continue;
137 
138 			} else if (dma_fence_is_later(tmp, next)) {
139 				fences[i] = dma_fence_unwrap_next(&iter[i]);
140 				goto restart;
141 			} else {
142 				fences[sel] = dma_fence_unwrap_next(&iter[sel]);
143 				goto restart;
144 			}
145 		}
146 
147 		if (tmp) {
148 			array[count++] = dma_fence_get(tmp);
149 			fences[sel] = dma_fence_unwrap_next(&iter[sel]);
150 		}
151 	} while (tmp);
152 
153 	if (count == 0) {
154 		tmp = dma_fence_allocate_private_stub(ktime_get());
155 		goto return_tmp;
156 	}
157 
158 	if (count == 1) {
159 		tmp = array[0];
160 		goto return_tmp;
161 	}
162 
163 	result = dma_fence_array_create(count, array,
164 					dma_fence_context_alloc(1),
165 					1, false);
166 	if (!result) {
167 		tmp = NULL;
168 		goto return_tmp;
169 	}
170 	return &result->base;
171 
172 return_tmp:
173 	kfree(array);
174 	return tmp;
175 }
176 EXPORT_SYMBOL_GPL(__dma_fence_unwrap_merge);
177