xref: /openbmc/linux/drivers/gpu/drm/drm_cache.c (revision 413d6ed3)
1 /**************************************************************************
2  *
3  * Copyright (c) 2006-2007 Tungsten Graphics, Inc., Cedar Park, TX., USA
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 /*
28  * Authors: Thomas Hellström <thomas-at-tungstengraphics-dot-com>
29  */
30 
31 #include <linux/export.h>
32 #include <linux/highmem.h>
33 #include <linux/mem_encrypt.h>
34 #include <xen/xen.h>
35 
36 #include <drm/drm_cache.h>
37 
38 #if defined(CONFIG_X86)
39 #include <asm/smp.h>
40 
41 /*
42  * clflushopt is an unordered instruction which needs fencing with mfence or
43  * sfence to avoid ordering issues.  For drm_clflush_page this fencing happens
44  * in the caller.
45  */
46 static void
47 drm_clflush_page(struct page *page)
48 {
49 	uint8_t *page_virtual;
50 	unsigned int i;
51 	const int size = boot_cpu_data.x86_clflush_size;
52 
53 	if (unlikely(page == NULL))
54 		return;
55 
56 	page_virtual = kmap_atomic(page);
57 	for (i = 0; i < PAGE_SIZE; i += size)
58 		clflushopt(page_virtual + i);
59 	kunmap_atomic(page_virtual);
60 }
61 
62 static void drm_cache_flush_clflush(struct page *pages[],
63 				    unsigned long num_pages)
64 {
65 	unsigned long i;
66 
67 	mb(); /*Full memory barrier used before so that CLFLUSH is ordered*/
68 	for (i = 0; i < num_pages; i++)
69 		drm_clflush_page(*pages++);
70 	mb(); /*Also used after CLFLUSH so that all cache is flushed*/
71 }
72 #endif
73 
74 /**
75  * drm_clflush_pages - Flush dcache lines of a set of pages.
76  * @pages: List of pages to be flushed.
77  * @num_pages: Number of pages in the array.
78  *
79  * Flush every data cache line entry that points to an address belonging
80  * to a page in the array.
81  */
82 void
83 drm_clflush_pages(struct page *pages[], unsigned long num_pages)
84 {
85 
86 #if defined(CONFIG_X86)
87 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
88 		drm_cache_flush_clflush(pages, num_pages);
89 		return;
90 	}
91 
92 	if (wbinvd_on_all_cpus())
93 		pr_err("Timed out waiting for cache flush\n");
94 
95 #elif defined(__powerpc__)
96 	unsigned long i;
97 
98 	for (i = 0; i < num_pages; i++) {
99 		struct page *page = pages[i];
100 		void *page_virtual;
101 
102 		if (unlikely(page == NULL))
103 			continue;
104 
105 		page_virtual = kmap_atomic(page);
106 		flush_dcache_range((unsigned long)page_virtual,
107 				   (unsigned long)page_virtual + PAGE_SIZE);
108 		kunmap_atomic(page_virtual);
109 	}
110 #else
111 	pr_err("Architecture has no drm_cache.c support\n");
112 	WARN_ON_ONCE(1);
113 #endif
114 }
115 EXPORT_SYMBOL(drm_clflush_pages);
116 
117 /**
118  * drm_clflush_sg - Flush dcache lines pointing to a scather-gather.
119  * @st: struct sg_table.
120  *
121  * Flush every data cache line entry that points to an address in the
122  * sg.
123  */
124 void
125 drm_clflush_sg(struct sg_table *st)
126 {
127 #if defined(CONFIG_X86)
128 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
129 		struct sg_page_iter sg_iter;
130 
131 		mb(); /*CLFLUSH is ordered only by using memory barriers*/
132 		for_each_sgtable_page(st, &sg_iter, 0)
133 			drm_clflush_page(sg_page_iter_page(&sg_iter));
134 		mb(); /*Make sure that all cache line entry is flushed*/
135 
136 		return;
137 	}
138 
139 	if (wbinvd_on_all_cpus())
140 		pr_err("Timed out waiting for cache flush\n");
141 #else
142 	pr_err("Architecture has no drm_cache.c support\n");
143 	WARN_ON_ONCE(1);
144 #endif
145 }
146 EXPORT_SYMBOL(drm_clflush_sg);
147 
148 /**
149  * drm_clflush_virt_range - Flush dcache lines of a region
150  * @addr: Initial kernel memory address.
151  * @length: Region size.
152  *
153  * Flush every data cache line entry that points to an address in the
154  * region requested.
155  */
156 void
157 drm_clflush_virt_range(void *addr, unsigned long length)
158 {
159 #if defined(CONFIG_X86)
160 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
161 		const int size = boot_cpu_data.x86_clflush_size;
162 		void *end = addr + length;
163 
164 		addr = (void *)(((unsigned long)addr) & -size);
165 		mb(); /*CLFLUSH is only ordered with a full memory barrier*/
166 		for (; addr < end; addr += size)
167 			clflushopt(addr);
168 		clflushopt(end - 1); /* force serialisation */
169 		mb(); /*Ensure that evry data cache line entry is flushed*/
170 		return;
171 	}
172 
173 	if (wbinvd_on_all_cpus())
174 		pr_err("Timed out waiting for cache flush\n");
175 #else
176 	pr_err("Architecture has no drm_cache.c support\n");
177 	WARN_ON_ONCE(1);
178 #endif
179 }
180 EXPORT_SYMBOL(drm_clflush_virt_range);
181 
182 bool drm_need_swiotlb(int dma_bits)
183 {
184 	struct resource *tmp;
185 	resource_size_t max_iomem = 0;
186 
187 	/*
188 	 * Xen paravirtual hosts require swiotlb regardless of requested dma
189 	 * transfer size.
190 	 *
191 	 * NOTE: Really, what it requires is use of the dma_alloc_coherent
192 	 *       allocator used in ttm_dma_populate() instead of
193 	 *       ttm_populate_and_map_pages(), which bounce buffers so much in
194 	 *       Xen it leads to swiotlb buffer exhaustion.
195 	 */
196 	if (xen_pv_domain())
197 		return true;
198 
199 	/*
200 	 * Enforce dma_alloc_coherent when memory encryption is active as well
201 	 * for the same reasons as for Xen paravirtual hosts.
202 	 */
203 	if (mem_encrypt_active())
204 		return true;
205 
206 	for (tmp = iomem_resource.child; tmp; tmp = tmp->sibling)
207 		max_iomem = max(max_iomem,  tmp->end);
208 
209 	return max_iomem > ((u64)1 << dma_bits);
210 }
211 EXPORT_SYMBOL(drm_need_swiotlb);
212