xref: /openbmc/linux/drivers/gpu/drm/drm_cache.c (revision f68f2ff9)
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/dma-buf-map.h>
32 #include <linux/export.h>
33 #include <linux/highmem.h>
34 #include <linux/cc_platform.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
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 
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
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 	if (wbinvd_on_all_cpus())
97 		pr_err("Timed out waiting for cache flush\n");
98 
99 #elif defined(__powerpc__)
100 	unsigned long i;
101 
102 	for (i = 0; i < num_pages; i++) {
103 		struct page *page = pages[i];
104 		void *page_virtual;
105 
106 		if (unlikely(page == NULL))
107 			continue;
108 
109 		page_virtual = kmap_atomic(page);
110 		flush_dcache_range((unsigned long)page_virtual,
111 				   (unsigned long)page_virtual + PAGE_SIZE);
112 		kunmap_atomic(page_virtual);
113 	}
114 #else
115 	pr_err("Architecture has no drm_cache.c support\n");
116 	WARN_ON_ONCE(1);
117 #endif
118 }
119 EXPORT_SYMBOL(drm_clflush_pages);
120 
121 /**
122  * drm_clflush_sg - Flush dcache lines pointing to a scather-gather.
123  * @st: struct sg_table.
124  *
125  * Flush every data cache line entry that points to an address in the
126  * sg.
127  */
128 void
129 drm_clflush_sg(struct sg_table *st)
130 {
131 #if defined(CONFIG_X86)
132 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
133 		struct sg_page_iter sg_iter;
134 
135 		mb(); /*CLFLUSH is ordered only by using memory barriers*/
136 		for_each_sgtable_page(st, &sg_iter, 0)
137 			drm_clflush_page(sg_page_iter_page(&sg_iter));
138 		mb(); /*Make sure that all cache line entry is flushed*/
139 
140 		return;
141 	}
142 
143 	if (wbinvd_on_all_cpus())
144 		pr_err("Timed out waiting for cache flush\n");
145 #else
146 	pr_err("Architecture has no drm_cache.c support\n");
147 	WARN_ON_ONCE(1);
148 #endif
149 }
150 EXPORT_SYMBOL(drm_clflush_sg);
151 
152 /**
153  * drm_clflush_virt_range - Flush dcache lines of a region
154  * @addr: Initial kernel memory address.
155  * @length: Region size.
156  *
157  * Flush every data cache line entry that points to an address in the
158  * region requested.
159  */
160 void
161 drm_clflush_virt_range(void *addr, unsigned long length)
162 {
163 #if defined(CONFIG_X86)
164 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
165 		const int size = boot_cpu_data.x86_clflush_size;
166 		void *end = addr + length;
167 
168 		addr = (void *)(((unsigned long)addr) & -size);
169 		mb(); /*CLFLUSH is only ordered with a full memory barrier*/
170 		for (; addr < end; addr += size)
171 			clflushopt(addr);
172 		clflushopt(end - 1); /* force serialisation */
173 		mb(); /*Ensure that every data cache line entry is flushed*/
174 		return;
175 	}
176 
177 	if (wbinvd_on_all_cpus())
178 		pr_err("Timed out waiting for cache flush\n");
179 #else
180 	pr_err("Architecture has no drm_cache.c support\n");
181 	WARN_ON_ONCE(1);
182 #endif
183 }
184 EXPORT_SYMBOL(drm_clflush_virt_range);
185 
186 bool drm_need_swiotlb(int dma_bits)
187 {
188 	struct resource *tmp;
189 	resource_size_t max_iomem = 0;
190 
191 	/*
192 	 * Xen paravirtual hosts require swiotlb regardless of requested dma
193 	 * transfer size.
194 	 *
195 	 * NOTE: Really, what it requires is use of the dma_alloc_coherent
196 	 *       allocator used in ttm_dma_populate() instead of
197 	 *       ttm_populate_and_map_pages(), which bounce buffers so much in
198 	 *       Xen it leads to swiotlb buffer exhaustion.
199 	 */
200 	if (xen_pv_domain())
201 		return true;
202 
203 	/*
204 	 * Enforce dma_alloc_coherent when memory encryption is active as well
205 	 * for the same reasons as for Xen paravirtual hosts.
206 	 */
207 	if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
208 		return true;
209 
210 	for (tmp = iomem_resource.child; tmp; tmp = tmp->sibling)
211 		max_iomem = max(max_iomem,  tmp->end);
212 
213 	return max_iomem > ((u64)1 << dma_bits);
214 }
215 EXPORT_SYMBOL(drm_need_swiotlb);
216 
217 static void memcpy_fallback(struct dma_buf_map *dst,
218 			    const struct dma_buf_map *src,
219 			    unsigned long len)
220 {
221 	if (!dst->is_iomem && !src->is_iomem) {
222 		memcpy(dst->vaddr, src->vaddr, len);
223 	} else if (!src->is_iomem) {
224 		dma_buf_map_memcpy_to(dst, src->vaddr, len);
225 	} else if (!dst->is_iomem) {
226 		memcpy_fromio(dst->vaddr, src->vaddr_iomem, len);
227 	} else {
228 		/*
229 		 * Bounce size is not performance tuned, but using a
230 		 * bounce buffer like this is significantly faster than
231 		 * resorting to ioreadxx() + iowritexx().
232 		 */
233 		char bounce[MEMCPY_BOUNCE_SIZE];
234 		void __iomem *_src = src->vaddr_iomem;
235 		void __iomem *_dst = dst->vaddr_iomem;
236 
237 		while (len >= MEMCPY_BOUNCE_SIZE) {
238 			memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
239 			memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
240 			_src += MEMCPY_BOUNCE_SIZE;
241 			_dst += MEMCPY_BOUNCE_SIZE;
242 			len -= MEMCPY_BOUNCE_SIZE;
243 		}
244 		if (len) {
245 			memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
246 			memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
247 		}
248 	}
249 }
250 
251 #ifdef CONFIG_X86
252 
253 static DEFINE_STATIC_KEY_FALSE(has_movntdqa);
254 
255 static void __memcpy_ntdqa(void *dst, const void *src, unsigned long len)
256 {
257 	kernel_fpu_begin();
258 
259 	while (len >= 4) {
260 		asm("movntdqa	(%0), %%xmm0\n"
261 		    "movntdqa 16(%0), %%xmm1\n"
262 		    "movntdqa 32(%0), %%xmm2\n"
263 		    "movntdqa 48(%0), %%xmm3\n"
264 		    "movaps %%xmm0,   (%1)\n"
265 		    "movaps %%xmm1, 16(%1)\n"
266 		    "movaps %%xmm2, 32(%1)\n"
267 		    "movaps %%xmm3, 48(%1)\n"
268 		    :: "r" (src), "r" (dst) : "memory");
269 		src += 64;
270 		dst += 64;
271 		len -= 4;
272 	}
273 	while (len--) {
274 		asm("movntdqa (%0), %%xmm0\n"
275 		    "movaps %%xmm0, (%1)\n"
276 		    :: "r" (src), "r" (dst) : "memory");
277 		src += 16;
278 		dst += 16;
279 	}
280 
281 	kernel_fpu_end();
282 }
283 
284 /*
285  * __drm_memcpy_from_wc copies @len bytes from @src to @dst using
286  * non-temporal instructions where available. Note that all arguments
287  * (@src, @dst) must be aligned to 16 bytes and @len must be a multiple
288  * of 16.
289  */
290 static void __drm_memcpy_from_wc(void *dst, const void *src, unsigned long len)
291 {
292 	if (unlikely(((unsigned long)dst | (unsigned long)src | len) & 15))
293 		memcpy(dst, src, len);
294 	else if (likely(len))
295 		__memcpy_ntdqa(dst, src, len >> 4);
296 }
297 
298 /**
299  * drm_memcpy_from_wc - Perform the fastest available memcpy from a source
300  * that may be WC.
301  * @dst: The destination pointer
302  * @src: The source pointer
303  * @len: The size of the area o transfer in bytes
304  *
305  * Tries an arch optimized memcpy for prefetching reading out of a WC region,
306  * and if no such beast is available, falls back to a normal memcpy.
307  */
308 void drm_memcpy_from_wc(struct dma_buf_map *dst,
309 			const struct dma_buf_map *src,
310 			unsigned long len)
311 {
312 	if (WARN_ON(in_interrupt())) {
313 		memcpy_fallback(dst, src, len);
314 		return;
315 	}
316 
317 	if (static_branch_likely(&has_movntdqa)) {
318 		__drm_memcpy_from_wc(dst->is_iomem ?
319 				     (void __force *)dst->vaddr_iomem :
320 				     dst->vaddr,
321 				     src->is_iomem ?
322 				     (void const __force *)src->vaddr_iomem :
323 				     src->vaddr,
324 				     len);
325 		return;
326 	}
327 
328 	memcpy_fallback(dst, src, len);
329 }
330 EXPORT_SYMBOL(drm_memcpy_from_wc);
331 
332 /*
333  * drm_memcpy_init_early - One time initialization of the WC memcpy code
334  */
335 void drm_memcpy_init_early(void)
336 {
337 	/*
338 	 * Some hypervisors (e.g. KVM) don't support VEX-prefix instructions
339 	 * emulation. So don't enable movntdqa in hypervisor guest.
340 	 */
341 	if (static_cpu_has(X86_FEATURE_XMM4_1) &&
342 	    !boot_cpu_has(X86_FEATURE_HYPERVISOR))
343 		static_branch_enable(&has_movntdqa);
344 }
345 #else
346 void drm_memcpy_from_wc(struct dma_buf_map *dst,
347 			const struct dma_buf_map *src,
348 			unsigned long len)
349 {
350 	WARN_ON(in_interrupt());
351 
352 	memcpy_fallback(dst, src, len);
353 }
354 EXPORT_SYMBOL(drm_memcpy_from_wc);
355 
356 void drm_memcpy_init_early(void)
357 {
358 }
359 #endif /* CONFIG_X86 */
360