xref: /openbsd-src/sys/dev/pci/drm/drm_cache.c (revision c1a45aed656e7d5627c30c92421893a76f370ccb)
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/mem_encrypt.h>
35 #ifdef __linux__
36 #include <xen/xen.h>
37 #endif
38 
39 #include <drm/drm_cache.h>
40 
41 /* A small bounce buffer that fits on the stack. */
42 #define MEMCPY_BOUNCE_SIZE 128
43 
44 #if defined(CONFIG_X86)
45 #include <asm/smp.h>
46 
47 /*
48  * clflushopt is an unordered instruction which needs fencing with mfence or
49  * sfence to avoid ordering issues.  For drm_clflush_page this fencing happens
50  * in the caller.
51  */
52 static void
53 drm_clflush_page(struct vm_page *page)
54 {
55 	uint8_t *page_virtual;
56 	unsigned int i;
57 	const int size = curcpu()->ci_cflushsz;
58 
59 	if (unlikely(page == NULL))
60 		return;
61 
62 	page_virtual = kmap_atomic(page);
63 	for (i = 0; i < PAGE_SIZE; i += size)
64 		clflushopt(page_virtual + i);
65 	kunmap_atomic(page_virtual);
66 }
67 
68 static void drm_cache_flush_clflush(struct vm_page *pages[],
69 				    unsigned long num_pages)
70 {
71 	unsigned long i;
72 
73 	mb(); /*Full memory barrier used before so that CLFLUSH is ordered*/
74 	for (i = 0; i < num_pages; i++)
75 		drm_clflush_page(*pages++);
76 	mb(); /*Also used after CLFLUSH so that all cache is flushed*/
77 }
78 #endif
79 
80 /**
81  * drm_clflush_pages - Flush dcache lines of a set of pages.
82  * @pages: List of pages to be flushed.
83  * @num_pages: Number of pages in the array.
84  *
85  * Flush every data cache line entry that points to an address belonging
86  * to a page in the array.
87  */
88 void
89 drm_clflush_pages(struct vm_page *pages[], unsigned long num_pages)
90 {
91 
92 #if defined(CONFIG_X86)
93 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
94 		drm_cache_flush_clflush(pages, num_pages);
95 		return;
96 	}
97 
98 	if (wbinvd_on_all_cpus())
99 		pr_err("Timed out waiting for cache flush\n");
100 
101 #elif defined(__powerpc__) && defined(__linux__)
102 	unsigned long i;
103 
104 	for (i = 0; i < num_pages; i++) {
105 		struct vm_page *page = pages[i];
106 		void *page_virtual;
107 
108 		if (unlikely(page == NULL))
109 			continue;
110 
111 		page_virtual = kmap_atomic(page);
112 		flush_dcache_range((unsigned long)page_virtual,
113 				   (unsigned long)page_virtual + PAGE_SIZE);
114 		kunmap_atomic(page_virtual);
115 	}
116 #else
117 	pr_err("Architecture has no drm_cache.c support\n");
118 	WARN_ON_ONCE(1);
119 #endif
120 }
121 EXPORT_SYMBOL(drm_clflush_pages);
122 
123 /**
124  * drm_clflush_sg - Flush dcache lines pointing to a scather-gather.
125  * @st: struct sg_table.
126  *
127  * Flush every data cache line entry that points to an address in the
128  * sg.
129  */
130 void
131 drm_clflush_sg(struct sg_table *st)
132 {
133 #if defined(CONFIG_X86)
134 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
135 		struct sg_page_iter sg_iter;
136 
137 		mb(); /*CLFLUSH is ordered only by using memory barriers*/
138 		for_each_sgtable_page(st, &sg_iter, 0)
139 			drm_clflush_page(sg_page_iter_page(&sg_iter));
140 		mb(); /*Make sure that all cache line entry is flushed*/
141 
142 		return;
143 	}
144 
145 	if (wbinvd_on_all_cpus())
146 		pr_err("Timed out waiting for cache flush\n");
147 #else
148 	pr_err("Architecture has no drm_cache.c support\n");
149 	WARN_ON_ONCE(1);
150 #endif
151 }
152 EXPORT_SYMBOL(drm_clflush_sg);
153 
154 /**
155  * drm_clflush_virt_range - Flush dcache lines of a region
156  * @addr: Initial kernel memory address.
157  * @length: Region size.
158  *
159  * Flush every data cache line entry that points to an address in the
160  * region requested.
161  */
162 void
163 drm_clflush_virt_range(void *addr, unsigned long length)
164 {
165 #if defined(CONFIG_X86)
166 	if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
167 		const int size = curcpu()->ci_cflushsz;
168 		void *end = addr + length;
169 
170 		addr = (void *)(((unsigned long)addr) & -size);
171 		mb(); /*CLFLUSH is only ordered with a full memory barrier*/
172 		for (; addr < end; addr += size)
173 			clflushopt(addr);
174 		clflushopt(end - 1); /* force serialisation */
175 		mb(); /*Ensure that every data cache line entry is flushed*/
176 		return;
177 	}
178 
179 	if (wbinvd_on_all_cpus())
180 		pr_err("Timed out waiting for cache flush\n");
181 #else
182 	pr_err("Architecture has no drm_cache.c support\n");
183 	WARN_ON_ONCE(1);
184 #endif
185 }
186 EXPORT_SYMBOL(drm_clflush_virt_range);
187 
188 bool drm_need_swiotlb(int dma_bits)
189 {
190 	return false;
191 #ifdef notyet
192 	struct resource *tmp;
193 	resource_size_t max_iomem = 0;
194 
195 	/*
196 	 * Xen paravirtual hosts require swiotlb regardless of requested dma
197 	 * transfer size.
198 	 *
199 	 * NOTE: Really, what it requires is use of the dma_alloc_coherent
200 	 *       allocator used in ttm_dma_populate() instead of
201 	 *       ttm_populate_and_map_pages(), which bounce buffers so much in
202 	 *       Xen it leads to swiotlb buffer exhaustion.
203 	 */
204 	if (xen_pv_domain())
205 		return true;
206 
207 	/*
208 	 * Enforce dma_alloc_coherent when memory encryption is active as well
209 	 * for the same reasons as for Xen paravirtual hosts.
210 	 */
211 	if (mem_encrypt_active())
212 		return true;
213 
214 	for (tmp = iomem_resource.child; tmp; tmp = tmp->sibling)
215 		max_iomem = max(max_iomem,  tmp->end);
216 
217 	return max_iomem > ((u64)1 << dma_bits);
218 #endif
219 }
220 EXPORT_SYMBOL(drm_need_swiotlb);
221 
222 static void memcpy_fallback(struct dma_buf_map *dst,
223 			    const struct dma_buf_map *src,
224 			    unsigned long len)
225 {
226 	if (!dst->is_iomem && !src->is_iomem) {
227 		memcpy(dst->vaddr, src->vaddr, len);
228 	} else if (!src->is_iomem) {
229 		dma_buf_map_memcpy_to(dst, src->vaddr, len);
230 	} else if (!dst->is_iomem) {
231 		memcpy_fromio(dst->vaddr, src->vaddr_iomem, len);
232 	} else {
233 		/*
234 		 * Bounce size is not performance tuned, but using a
235 		 * bounce buffer like this is significantly faster than
236 		 * resorting to ioreadxx() + iowritexx().
237 		 */
238 		char bounce[MEMCPY_BOUNCE_SIZE];
239 		void __iomem *_src = src->vaddr_iomem;
240 		void __iomem *_dst = dst->vaddr_iomem;
241 
242 		while (len >= MEMCPY_BOUNCE_SIZE) {
243 			memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
244 			memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
245 			_src += MEMCPY_BOUNCE_SIZE;
246 			_dst += MEMCPY_BOUNCE_SIZE;
247 			len -= MEMCPY_BOUNCE_SIZE;
248 		}
249 		if (len) {
250 			memcpy_fromio(bounce, _src, MEMCPY_BOUNCE_SIZE);
251 			memcpy_toio(_dst, bounce, MEMCPY_BOUNCE_SIZE);
252 		}
253 	}
254 }
255 
256 #ifdef CONFIG_X86
257 
258 #ifdef __linux__
259 static DEFINE_STATIC_KEY_FALSE(has_movntdqa);
260 #else
261 static int has_movntdqa;
262 
263 #include <asm/fpu/api.h>
264 
265 static inline void
266 static_branch_enable(int *x)
267 {
268 	*x = 1;
269 }
270 
271 static inline int
272 static_branch_likely(int *x)
273 {
274 	return (likely(*x == 1));
275 }
276 
277 #endif
278 
279 static void __memcpy_ntdqa(void *dst, const void *src, unsigned long len)
280 {
281 	kernel_fpu_begin();
282 
283 	while (len >= 4) {
284 		asm("movntdqa	(%0), %%xmm0\n"
285 		    "movntdqa 16(%0), %%xmm1\n"
286 		    "movntdqa 32(%0), %%xmm2\n"
287 		    "movntdqa 48(%0), %%xmm3\n"
288 		    "movaps %%xmm0,   (%1)\n"
289 		    "movaps %%xmm1, 16(%1)\n"
290 		    "movaps %%xmm2, 32(%1)\n"
291 		    "movaps %%xmm3, 48(%1)\n"
292 		    :: "r" (src), "r" (dst) : "memory");
293 		src += 64;
294 		dst += 64;
295 		len -= 4;
296 	}
297 	while (len--) {
298 		asm("movntdqa (%0), %%xmm0\n"
299 		    "movaps %%xmm0, (%1)\n"
300 		    :: "r" (src), "r" (dst) : "memory");
301 		src += 16;
302 		dst += 16;
303 	}
304 
305 	kernel_fpu_end();
306 }
307 
308 /*
309  * __drm_memcpy_from_wc copies @len bytes from @src to @dst using
310  * non-temporal instructions where available. Note that all arguments
311  * (@src, @dst) must be aligned to 16 bytes and @len must be a multiple
312  * of 16.
313  */
314 static void __drm_memcpy_from_wc(void *dst, const void *src, unsigned long len)
315 {
316 	if (unlikely(((unsigned long)dst | (unsigned long)src | len) & 15))
317 		memcpy(dst, src, len);
318 	else if (likely(len))
319 		__memcpy_ntdqa(dst, src, len >> 4);
320 }
321 
322 /**
323  * drm_memcpy_from_wc - Perform the fastest available memcpy from a source
324  * that may be WC.
325  * @dst: The destination pointer
326  * @src: The source pointer
327  * @len: The size of the area o transfer in bytes
328  *
329  * Tries an arch optimized memcpy for prefetching reading out of a WC region,
330  * and if no such beast is available, falls back to a normal memcpy.
331  */
332 void drm_memcpy_from_wc(struct dma_buf_map *dst,
333 			const struct dma_buf_map *src,
334 			unsigned long len)
335 {
336 	if (WARN_ON(in_interrupt())) {
337 		memcpy_fallback(dst, src, len);
338 		return;
339 	}
340 
341 	if (static_branch_likely(&has_movntdqa)) {
342 		__drm_memcpy_from_wc(dst->is_iomem ?
343 				     (void __force *)dst->vaddr_iomem :
344 				     dst->vaddr,
345 				     src->is_iomem ?
346 				     (void const __force *)src->vaddr_iomem :
347 				     src->vaddr,
348 				     len);
349 		return;
350 	}
351 
352 	memcpy_fallback(dst, src, len);
353 }
354 EXPORT_SYMBOL(drm_memcpy_from_wc);
355 
356 /*
357  * drm_memcpy_init_early - One time initialization of the WC memcpy code
358  */
359 void drm_memcpy_init_early(void)
360 {
361 	/*
362 	 * Some hypervisors (e.g. KVM) don't support VEX-prefix instructions
363 	 * emulation. So don't enable movntdqa in hypervisor guest.
364 	 */
365 	if (static_cpu_has(X86_FEATURE_XMM4_1) &&
366 	    !boot_cpu_has(X86_FEATURE_HYPERVISOR))
367 		static_branch_enable(&has_movntdqa);
368 }
369 #else
370 void drm_memcpy_from_wc(struct dma_buf_map *dst,
371 			const struct dma_buf_map *src,
372 			unsigned long len)
373 {
374 	WARN_ON(in_interrupt());
375 
376 	memcpy_fallback(dst, src, len);
377 }
378 EXPORT_SYMBOL(drm_memcpy_from_wc);
379 
380 void drm_memcpy_init_early(void)
381 {
382 }
383 #endif /* CONFIG_X86 */
384