xref: /linux/drivers/gpu/drm/drm_cache.c (revision bb78c145f7f08fda40bcec397939b01be4366c51)
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 #include <linux/cc_platform.h>
31 #include <linux/export.h>
32 #include <linux/highmem.h>
33 #include <linux/ioport.h>
34 #include <linux/iosys-map.h>
35 #include <xen/xen.h>
36 
37 #include <drm/drm_cache.h>
38 
39 /* A small bounce buffer that fits on the stack. */
40 #define MEMCPY_BOUNCE_SIZE 128
41 
42 #if defined(CONFIG_X86)
43 #include <asm/smp.h>
44 
45 /*
46  * clflushopt is an unordered instruction which needs fencing with mfence or
47  * sfence to avoid ordering issues.  For drm_clflush_page this fencing happens
48  * in the caller.
49  */
50 static void
drm_clflush_page(struct page * page)51 drm_clflush_page(struct page *page)
52 {
53 	uint8_t *page_virtual;
54 	unsigned int i;
55 	const int size = boot_cpu_data.x86_clflush_size;
56 
57 	if (unlikely(page == NULL))
58 		return;
59 
60 	page_virtual = kmap_atomic(page);
61 	for (i = 0; i < PAGE_SIZE; i += size)
62 		clflushopt(page_virtual + i);
63 	kunmap_atomic(page_virtual);
64 }
65 
drm_cache_flush_clflush(struct page * pages[],unsigned long num_pages)66 static void drm_cache_flush_clflush(struct page *pages[],
67 				    unsigned long num_pages)
68 {
69 	unsigned long i;
70 
71 	mb(); /*Full memory barrier used before so that CLFLUSH is ordered*/
72 	for (i = 0; i < num_pages; i++)
73 		drm_clflush_page(*pages++);
74 	mb(); /*Also used after CLFLUSH so that all cache is flushed*/
75 }
76 #endif
77 
78 /**
79  * drm_clflush_pages - Flush dcache lines of a set of pages.
80  * @pages: List of pages to be flushed.
81  * @num_pages: Number of pages in the array.
82  *
83  * Flush every data cache line entry that points to an address belonging
84  * to a page in the array.
85  */
86 void
drm_clflush_pages(struct page * pages[],unsigned long num_pages)87 drm_clflush_pages(struct page *pages[], unsigned long num_pages)
88 {
89 
90 #if defined(CONFIG_X86)
91 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
92 		drm_cache_flush_clflush(pages, num_pages);
93 		return;
94 	}
95 
96 	wbinvd_on_all_cpus();
97 
98 #elif defined(__powerpc__)
99 	unsigned long i;
100 
101 	for (i = 0; i < num_pages; i++) {
102 		struct page *page = pages[i];
103 		void *page_virtual;
104 
105 		if (unlikely(page == NULL))
106 			continue;
107 
108 		page_virtual = kmap_atomic(page);
109 		flush_dcache_range((unsigned long)page_virtual,
110 				   (unsigned long)page_virtual + PAGE_SIZE);
111 		kunmap_atomic(page_virtual);
112 	}
113 #else
114 	WARN_ONCE(1, "Architecture has no drm_cache.c support\n");
115 #endif
116 }
117 EXPORT_SYMBOL(drm_clflush_pages);
118 
119 /**
120  * drm_clflush_sg - Flush dcache lines pointing to a scather-gather.
121  * @st: struct sg_table.
122  *
123  * Flush every data cache line entry that points to an address in the
124  * sg.
125  */
126 void
drm_clflush_sg(struct sg_table * st)127 drm_clflush_sg(struct sg_table *st)
128 {
129 #if defined(CONFIG_X86)
130 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
131 		struct sg_page_iter sg_iter;
132 
133 		mb(); /*CLFLUSH is ordered only by using memory barriers*/
134 		for_each_sgtable_page(st, &sg_iter, 0)
135 			drm_clflush_page(sg_page_iter_page(&sg_iter));
136 		mb(); /*Make sure that all cache line entry is flushed*/
137 
138 		return;
139 	}
140 
141 	wbinvd_on_all_cpus();
142 #else
143 	WARN_ONCE(1, "Architecture has no drm_cache.c support\n");
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
drm_clflush_virt_range(void * addr,unsigned long length)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 every data cache line entry is flushed*/
170 		return;
171 	}
172 
173 	wbinvd_on_all_cpus();
174 #else
175 	WARN_ONCE(1, "Architecture has no drm_cache.c support\n");
176 #endif
177 }
178 EXPORT_SYMBOL(drm_clflush_virt_range);
179 
drm_need_swiotlb(int dma_bits)180 bool drm_need_swiotlb(int dma_bits)
181 {
182 	struct resource *tmp;
183 	resource_size_t max_iomem = 0;
184 
185 	/*
186 	 * Xen paravirtual hosts require swiotlb regardless of requested dma
187 	 * transfer size.
188 	 *
189 	 * NOTE: Really, what it requires is use of the dma_alloc_coherent
190 	 *       allocator used in ttm_dma_populate() instead of
191 	 *       ttm_populate_and_map_pages(), which bounce buffers so much in
192 	 *       Xen it leads to swiotlb buffer exhaustion.
193 	 */
194 	if (xen_pv_domain())
195 		return true;
196 
197 	/*
198 	 * Enforce dma_alloc_coherent when memory encryption is active as well
199 	 * for the same reasons as for Xen paravirtual hosts.
200 	 */
201 	if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
202 		return true;
203 
204 	for (tmp = iomem_resource.child; tmp; tmp = tmp->sibling)
205 		max_iomem = max(max_iomem,  tmp->end);
206 
207 	return max_iomem > ((u64)1 << dma_bits);
208 }
209 EXPORT_SYMBOL(drm_need_swiotlb);
210 
memcpy_fallback(struct iosys_map * dst,const struct iosys_map * src,unsigned long len)211 static void memcpy_fallback(struct iosys_map *dst,
212 			    const struct iosys_map *src,
213 			    unsigned long len)
214 {
215 	if (!dst->is_iomem && !src->is_iomem) {
216 		memcpy(dst->vaddr, src->vaddr, len);
217 	} else if (!src->is_iomem) {
218 		iosys_map_memcpy_to(dst, 0, src->vaddr, len);
219 	} else if (!dst->is_iomem) {
220 		memcpy_fromio(dst->vaddr, src->vaddr_iomem, len);
221 	} else {
222 		/*
223 		 * Bounce size is not performance tuned, but using a
224 		 * bounce buffer like this is significantly faster than
225 		 * resorting to ioreadxx() + iowritexx().
226 		 */
227 		char bounce[MEMCPY_BOUNCE_SIZE];
228 		void __iomem *_src = src->vaddr_iomem;
229 		void __iomem *_dst = dst->vaddr_iomem;
230 
231 		while (len >= MEMCPY_BOUNCE_SIZE) {
232 			memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
233 			memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
234 			_src += MEMCPY_BOUNCE_SIZE;
235 			_dst += MEMCPY_BOUNCE_SIZE;
236 			len -= MEMCPY_BOUNCE_SIZE;
237 		}
238 		if (len) {
239 			memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
240 			memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
241 		}
242 	}
243 }
244 
245 #ifdef CONFIG_X86
246 
247 static DEFINE_STATIC_KEY_FALSE(has_movntdqa);
248 
__memcpy_ntdqa(void * dst,const void * src,unsigned long len)249 static void __memcpy_ntdqa(void *dst, const void *src, unsigned long len)
250 {
251 	kernel_fpu_begin();
252 
253 	while (len >= 4) {
254 		asm("movntdqa	(%0), %%xmm0\n"
255 		    "movntdqa 16(%0), %%xmm1\n"
256 		    "movntdqa 32(%0), %%xmm2\n"
257 		    "movntdqa 48(%0), %%xmm3\n"
258 		    "movaps %%xmm0,   (%1)\n"
259 		    "movaps %%xmm1, 16(%1)\n"
260 		    "movaps %%xmm2, 32(%1)\n"
261 		    "movaps %%xmm3, 48(%1)\n"
262 		    :: "r" (src), "r" (dst) : "memory");
263 		src += 64;
264 		dst += 64;
265 		len -= 4;
266 	}
267 	while (len--) {
268 		asm("movntdqa (%0), %%xmm0\n"
269 		    "movaps %%xmm0, (%1)\n"
270 		    :: "r" (src), "r" (dst) : "memory");
271 		src += 16;
272 		dst += 16;
273 	}
274 
275 	kernel_fpu_end();
276 }
277 
278 /*
279  * __drm_memcpy_from_wc copies @len bytes from @src to @dst using
280  * non-temporal instructions where available. Note that all arguments
281  * (@src, @dst) must be aligned to 16 bytes and @len must be a multiple
282  * of 16.
283  */
__drm_memcpy_from_wc(void * dst,const void * src,unsigned long len)284 static void __drm_memcpy_from_wc(void *dst, const void *src, unsigned long len)
285 {
286 	if (unlikely(((unsigned long)dst | (unsigned long)src | len) & 15))
287 		memcpy(dst, src, len);
288 	else if (likely(len))
289 		__memcpy_ntdqa(dst, src, len >> 4);
290 }
291 
292 /**
293  * drm_memcpy_from_wc - Perform the fastest available memcpy from a source
294  * that may be WC.
295  * @dst: The destination pointer
296  * @src: The source pointer
297  * @len: The size of the area o transfer in bytes
298  *
299  * Tries an arch optimized memcpy for prefetching reading out of a WC region,
300  * and if no such beast is available, falls back to a normal memcpy.
301  */
drm_memcpy_from_wc(struct iosys_map * dst,const struct iosys_map * src,unsigned long len)302 void drm_memcpy_from_wc(struct iosys_map *dst,
303 			const struct iosys_map *src,
304 			unsigned long len)
305 {
306 	if (WARN_ON(in_interrupt())) {
307 		memcpy_fallback(dst, src, len);
308 		return;
309 	}
310 
311 	if (static_branch_likely(&has_movntdqa)) {
312 		__drm_memcpy_from_wc(dst->is_iomem ?
313 				     (void __force *)dst->vaddr_iomem :
314 				     dst->vaddr,
315 				     src->is_iomem ?
316 				     (void const __force *)src->vaddr_iomem :
317 				     src->vaddr,
318 				     len);
319 		return;
320 	}
321 
322 	memcpy_fallback(dst, src, len);
323 }
324 EXPORT_SYMBOL(drm_memcpy_from_wc);
325 
326 /*
327  * drm_memcpy_init_early - One time initialization of the WC memcpy code
328  */
drm_memcpy_init_early(void)329 void drm_memcpy_init_early(void)
330 {
331 	/*
332 	 * Some hypervisors (e.g. KVM) don't support VEX-prefix instructions
333 	 * emulation. So don't enable movntdqa in hypervisor guest.
334 	 */
335 	if (static_cpu_has(X86_FEATURE_XMM4_1) &&
336 	    !boot_cpu_has(X86_FEATURE_HYPERVISOR))
337 		static_branch_enable(&has_movntdqa);
338 }
339 #else
drm_memcpy_from_wc(struct iosys_map * dst,const struct iosys_map * src,unsigned long len)340 void drm_memcpy_from_wc(struct iosys_map *dst,
341 			const struct iosys_map *src,
342 			unsigned long len)
343 {
344 	WARN_ON(in_interrupt());
345 
346 	memcpy_fallback(dst, src, len);
347 }
348 EXPORT_SYMBOL(drm_memcpy_from_wc);
349 
drm_memcpy_init_early(void)350 void drm_memcpy_init_early(void)
351 {
352 }
353 #endif /* CONFIG_X86 */
354