1 /**************************************************************************
2  *
3  * Copyright © 2017 VMware, Inc., Palo Alto, CA., 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 #include "vmwgfx_drv.h"
29 
30 /*
31  * Template that implements find_first_diff() for a generic
32  * unsigned integer type. @size and return value are in bytes.
33  */
34 #define VMW_FIND_FIRST_DIFF(_type)			 \
35 static size_t vmw_find_first_diff_ ## _type		 \
36 	(const _type * dst, const _type * src, size_t size)\
37 {							 \
38 	size_t i;					 \
39 							 \
40 	for (i = 0; i < size; i += sizeof(_type)) {	 \
41 		if (*dst++ != *src++)			 \
42 			break;				 \
43 	}						 \
44 							 \
45 	return i;					 \
46 }
47 
48 
49 /*
50  * Template that implements find_last_diff() for a generic
51  * unsigned integer type. Pointers point to the item following the
52  * *end* of the area to be examined. @size and return value are in
53  * bytes.
54  */
55 #define VMW_FIND_LAST_DIFF(_type)					\
56 static ssize_t vmw_find_last_diff_ ## _type(				\
57 	const _type * dst, const _type * src, size_t size)		\
58 {									\
59 	while (size) {							\
60 		if (*--dst != *--src)					\
61 			break;						\
62 									\
63 		size -= sizeof(_type);					\
64 	}								\
65 	return size;							\
66 }
67 
68 
69 /*
70  * Instantiate find diff functions for relevant unsigned integer sizes,
71  * assuming that wider integers are faster (including aligning) up to the
72  * architecture native width, which is assumed to be 32 bit unless
73  * CONFIG_64BIT is defined.
74  */
75 VMW_FIND_FIRST_DIFF(u8);
76 VMW_FIND_LAST_DIFF(u8);
77 
78 VMW_FIND_FIRST_DIFF(u16);
79 VMW_FIND_LAST_DIFF(u16);
80 
81 VMW_FIND_FIRST_DIFF(u32);
82 VMW_FIND_LAST_DIFF(u32);
83 
84 #ifdef CONFIG_64BIT
85 VMW_FIND_FIRST_DIFF(u64);
86 VMW_FIND_LAST_DIFF(u64);
87 #endif
88 
89 
90 /* We use size aligned copies. This computes (addr - align(addr)) */
91 #define SPILL(_var, _type) ((unsigned long) _var & (sizeof(_type) - 1))
92 
93 
94 /*
95  * Template to compute find_first_diff() for a certain integer type
96  * including a head copy for alignment, and adjustment of parameters
97  * for tail find or increased resolution find using an unsigned integer find
98  * of smaller width. If finding is complete, and resolution is sufficient,
99  * the macro executes a return statement. Otherwise it falls through.
100  */
101 #define VMW_TRY_FIND_FIRST_DIFF(_type)					\
102 do {									\
103 	unsigned int spill = SPILL(dst, _type);				\
104 	size_t diff_offs;						\
105 									\
106 	if (spill && spill == SPILL(src, _type) &&			\
107 	    sizeof(_type) - spill <= size) {				\
108 		spill = sizeof(_type) - spill;				\
109 		diff_offs = vmw_find_first_diff_u8(dst, src, spill);	\
110 		if (diff_offs < spill)					\
111 			return round_down(offset + diff_offs, granularity); \
112 									\
113 		dst += spill;						\
114 		src += spill;						\
115 		size -= spill;						\
116 		offset += spill;					\
117 		spill = 0;						\
118 	}								\
119 	if (!spill && !SPILL(src, _type)) {				\
120 		size_t to_copy = size &	 ~(sizeof(_type) - 1);		\
121 									\
122 		diff_offs = vmw_find_first_diff_ ## _type		\
123 			((_type *) dst, (_type *) src, to_copy);	\
124 		if (diff_offs >= size || granularity == sizeof(_type))	\
125 			return (offset + diff_offs);			\
126 									\
127 		dst += diff_offs;					\
128 		src += diff_offs;					\
129 		size -= diff_offs;					\
130 		offset += diff_offs;					\
131 	}								\
132 } while (0)								\
133 
134 
135 /**
136  * vmw_find_first_diff - find the first difference between dst and src
137  *
138  * @dst: The destination address
139  * @src: The source address
140  * @size: Number of bytes to compare
141  * @granularity: The granularity needed for the return value in bytes.
142  * return: The offset from find start where the first difference was
143  * encountered in bytes. If no difference was found, the function returns
144  * a value >= @size.
145  */
146 static size_t vmw_find_first_diff(const u8 *dst, const u8 *src, size_t size,
147 				  size_t granularity)
148 {
149 	size_t offset = 0;
150 
151 	/*
152 	 * Try finding with large integers if alignment allows, or we can
153 	 * fix it. Fall through if we need better resolution or alignment
154 	 * was bad.
155 	 */
156 #ifdef CONFIG_64BIT
157 	VMW_TRY_FIND_FIRST_DIFF(u64);
158 #endif
159 	VMW_TRY_FIND_FIRST_DIFF(u32);
160 	VMW_TRY_FIND_FIRST_DIFF(u16);
161 
162 	return round_down(offset + vmw_find_first_diff_u8(dst, src, size),
163 			  granularity);
164 }
165 
166 
167 /*
168  * Template to compute find_last_diff() for a certain integer type
169  * including a tail copy for alignment, and adjustment of parameters
170  * for head find or increased resolution find using an unsigned integer find
171  * of smaller width. If finding is complete, and resolution is sufficient,
172  * the macro executes a return statement. Otherwise it falls through.
173  */
174 #define VMW_TRY_FIND_LAST_DIFF(_type)					\
175 do {									\
176 	unsigned int spill = SPILL(dst, _type);				\
177 	ssize_t location;						\
178 	ssize_t diff_offs;						\
179 									\
180 	if (spill && spill <= size && spill == SPILL(src, _type)) {	\
181 		diff_offs = vmw_find_last_diff_u8(dst, src, spill);	\
182 		if (diff_offs) {					\
183 			location = size - spill + diff_offs - 1;	\
184 			return round_down(location, granularity);	\
185 		}							\
186 									\
187 		dst -= spill;						\
188 		src -= spill;						\
189 		size -= spill;						\
190 		spill = 0;						\
191 	}								\
192 	if (!spill && !SPILL(src, _type)) {				\
193 		size_t to_copy = round_down(size, sizeof(_type));	\
194 									\
195 		diff_offs = vmw_find_last_diff_ ## _type		\
196 			((_type *) dst, (_type *) src, to_copy);	\
197 		location = size - to_copy + diff_offs - sizeof(_type);	\
198 		if (location < 0 || granularity == sizeof(_type))	\
199 			return location;				\
200 									\
201 		dst -= to_copy - diff_offs;				\
202 		src -= to_copy - diff_offs;				\
203 		size -= to_copy - diff_offs;				\
204 	}								\
205 } while (0)
206 
207 
208 /**
209  * vmw_find_last_diff - find the last difference between dst and src
210  *
211  * @dst: The destination address
212  * @src: The source address
213  * @size: Number of bytes to compare
214  * @granularity: The granularity needed for the return value in bytes.
215  * return: The offset from find start where the last difference was
216  * encountered in bytes, or a negative value if no difference was found.
217  */
218 static ssize_t vmw_find_last_diff(const u8 *dst, const u8 *src, size_t size,
219 				  size_t granularity)
220 {
221 	dst += size;
222 	src += size;
223 
224 #ifdef CONFIG_64BIT
225 	VMW_TRY_FIND_LAST_DIFF(u64);
226 #endif
227 	VMW_TRY_FIND_LAST_DIFF(u32);
228 	VMW_TRY_FIND_LAST_DIFF(u16);
229 
230 	return round_down(vmw_find_last_diff_u8(dst, src, size) - 1,
231 			  granularity);
232 }
233 
234 
235 /**
236  * vmw_memcpy - A wrapper around kernel memcpy with allowing to plug it into a
237  * struct vmw_diff_cpy.
238  *
239  * @diff: The struct vmw_diff_cpy closure argument (unused).
240  * @dest: The copy destination.
241  * @src: The copy source.
242  * @n: Number of bytes to copy.
243  */
244 void vmw_memcpy(struct vmw_diff_cpy *diff, u8 *dest, const u8 *src, size_t n)
245 {
246 	memcpy(dest, src, n);
247 }
248 
249 
250 /**
251  * vmw_adjust_rect - Adjust rectangle coordinates for newly found difference
252  *
253  * @diff: The struct vmw_diff_cpy used to track the modified bounding box.
254  * @diff_offs: The offset from @diff->line_offset where the difference was
255  * found.
256  */
257 static void vmw_adjust_rect(struct vmw_diff_cpy *diff, size_t diff_offs)
258 {
259 	size_t offs = (diff_offs + diff->line_offset) / diff->cpp;
260 	struct drm_rect *rect = &diff->rect;
261 
262 	rect->x1 = min_t(int, rect->x1, offs);
263 	rect->x2 = max_t(int, rect->x2, offs + 1);
264 	rect->y1 = min_t(int, rect->y1, diff->line);
265 	rect->y2 = max_t(int, rect->y2, diff->line + 1);
266 }
267 
268 /**
269  * vmw_diff_memcpy - memcpy that creates a bounding box of modified content.
270  *
271  * @diff: The struct vmw_diff_cpy used to track the modified bounding box.
272  * @dest: The copy destination.
273  * @src: The copy source.
274  * @n: Number of bytes to copy.
275  *
276  * In order to correctly track the modified content, the field @diff->line must
277  * be pre-loaded with the current line number, the field @diff->line_offset must
278  * be pre-loaded with the line offset in bytes where the copy starts, and
279  * finally the field @diff->cpp need to be preloaded with the number of bytes
280  * per unit in the horizontal direction of the area we're examining.
281  * Typically bytes per pixel.
282  * This is needed to know the needed granularity of the difference computing
283  * operations. A higher cpp generally leads to faster execution at the cost of
284  * bounding box width precision.
285  */
286 void vmw_diff_memcpy(struct vmw_diff_cpy *diff, u8 *dest, const u8 *src,
287 		     size_t n)
288 {
289 	ssize_t csize, byte_len;
290 
291 	if (WARN_ON_ONCE(round_down(n, diff->cpp) != n))
292 		return;
293 
294 	/* TODO: Possibly use a single vmw_find_first_diff per line? */
295 	csize = vmw_find_first_diff(dest, src, n, diff->cpp);
296 	if (csize < n) {
297 		vmw_adjust_rect(diff, csize);
298 		byte_len = diff->cpp;
299 
300 		/*
301 		 * Starting from where first difference was found, find
302 		 * location of last difference, and then copy.
303 		 */
304 		diff->line_offset += csize;
305 		dest += csize;
306 		src += csize;
307 		n -= csize;
308 		csize = vmw_find_last_diff(dest, src, n, diff->cpp);
309 		if (csize >= 0) {
310 			byte_len += csize;
311 			vmw_adjust_rect(diff, csize);
312 		}
313 		memcpy(dest, src, byte_len);
314 	}
315 	diff->line_offset += n;
316 }
317 
318 /**
319  * struct vmw_bo_blit_line_data - Convenience argument to vmw_bo_cpu_blit_line
320  *
321  * @mapped_dst: Already mapped destination page index in @dst_pages.
322  * @dst_addr: Kernel virtual address of mapped destination page.
323  * @dst_pages: Array of destination bo pages.
324  * @dst_num_pages: Number of destination bo pages.
325  * @dst_prot: Destination bo page protection.
326  * @mapped_src: Already mapped source page index in @dst_pages.
327  * @src_addr: Kernel virtual address of mapped source page.
328  * @src_pages: Array of source bo pages.
329  * @src_num_pages: Number of source bo pages.
330  * @src_prot: Source bo page protection.
331  * @diff: Struct vmw_diff_cpy, in the end forwarded to the memcpy routine.
332  */
333 struct vmw_bo_blit_line_data {
334 	u32 mapped_dst;
335 	u8 *dst_addr;
336 	struct page **dst_pages;
337 	u32 dst_num_pages;
338 	pgprot_t dst_prot;
339 	u32 mapped_src;
340 	u8 *src_addr;
341 	struct page **src_pages;
342 	u32 src_num_pages;
343 	pgprot_t src_prot;
344 	struct vmw_diff_cpy *diff;
345 };
346 
347 /**
348  * vmw_bo_cpu_blit_line - Blit part of a line from one bo to another.
349  *
350  * @d: Blit data as described above.
351  * @dst_offset: Destination copy start offset from start of bo.
352  * @src_offset: Source copy start offset from start of bo.
353  * @bytes_to_copy: Number of bytes to copy in this line.
354  */
355 static int vmw_bo_cpu_blit_line(struct vmw_bo_blit_line_data *d,
356 				u32 dst_offset,
357 				u32 src_offset,
358 				u32 bytes_to_copy)
359 {
360 	struct vmw_diff_cpy *diff = d->diff;
361 
362 	while (bytes_to_copy) {
363 		u32 copy_size = bytes_to_copy;
364 		u32 dst_page = dst_offset >> PAGE_SHIFT;
365 		u32 src_page = src_offset >> PAGE_SHIFT;
366 		u32 dst_page_offset = dst_offset & ~PAGE_MASK;
367 		u32 src_page_offset = src_offset & ~PAGE_MASK;
368 		bool unmap_dst = d->dst_addr && dst_page != d->mapped_dst;
369 		bool unmap_src = d->src_addr && (src_page != d->mapped_src ||
370 						 unmap_dst);
371 
372 		copy_size = min_t(u32, copy_size, PAGE_SIZE - dst_page_offset);
373 		copy_size = min_t(u32, copy_size, PAGE_SIZE - src_page_offset);
374 
375 		if (unmap_src) {
376 			ttm_kunmap_atomic_prot(d->src_addr, d->src_prot);
377 			d->src_addr = NULL;
378 		}
379 
380 		if (unmap_dst) {
381 			ttm_kunmap_atomic_prot(d->dst_addr, d->dst_prot);
382 			d->dst_addr = NULL;
383 		}
384 
385 		if (!d->dst_addr) {
386 			if (WARN_ON_ONCE(dst_page >= d->dst_num_pages))
387 				return -EINVAL;
388 
389 			d->dst_addr =
390 				ttm_kmap_atomic_prot(d->dst_pages[dst_page],
391 						     d->dst_prot);
392 			if (!d->dst_addr)
393 				return -ENOMEM;
394 
395 			d->mapped_dst = dst_page;
396 		}
397 
398 		if (!d->src_addr) {
399 			if (WARN_ON_ONCE(src_page >= d->src_num_pages))
400 				return -EINVAL;
401 
402 			d->src_addr =
403 				ttm_kmap_atomic_prot(d->src_pages[src_page],
404 						     d->src_prot);
405 			if (!d->src_addr)
406 				return -ENOMEM;
407 
408 			d->mapped_src = src_page;
409 		}
410 		diff->do_cpy(diff, d->dst_addr + dst_page_offset,
411 			     d->src_addr + src_page_offset, copy_size);
412 
413 		bytes_to_copy -= copy_size;
414 		dst_offset += copy_size;
415 		src_offset += copy_size;
416 	}
417 
418 	return 0;
419 }
420 
421 /**
422  * ttm_bo_cpu_blit - in-kernel cpu blit.
423  *
424  * @dst: Destination buffer object.
425  * @dst_offset: Destination offset of blit start in bytes.
426  * @dst_stride: Destination stride in bytes.
427  * @src: Source buffer object.
428  * @src_offset: Source offset of blit start in bytes.
429  * @src_stride: Source stride in bytes.
430  * @w: Width of blit.
431  * @h: Height of blit.
432  * return: Zero on success. Negative error value on failure. Will print out
433  * kernel warnings on caller bugs.
434  *
435  * Performs a CPU blit from one buffer object to another avoiding a full
436  * bo vmap which may exhaust- or fragment vmalloc space.
437  * On supported architectures (x86), we're using kmap_atomic which avoids
438  * cross-processor TLB- and cache flushes and may, on non-HIGHMEM systems
439  * reference already set-up mappings.
440  *
441  * Neither of the buffer objects may be placed in PCI memory
442  * (Fixed memory in TTM terminology) when using this function.
443  */
444 int vmw_bo_cpu_blit(struct ttm_buffer_object *dst,
445 		    u32 dst_offset, u32 dst_stride,
446 		    struct ttm_buffer_object *src,
447 		    u32 src_offset, u32 src_stride,
448 		    u32 w, u32 h,
449 		    struct vmw_diff_cpy *diff)
450 {
451 	struct ttm_operation_ctx ctx = {
452 		.interruptible = false,
453 		.no_wait_gpu = false
454 	};
455 	u32 j, initial_line = dst_offset / dst_stride;
456 	struct vmw_bo_blit_line_data d;
457 	int ret = 0;
458 
459 	/* Buffer objects need to be either pinned or reserved: */
460 	if (!(dst->mem.placement & TTM_PL_FLAG_NO_EVICT))
461 		lockdep_assert_held(&dst->resv->lock.base);
462 	if (!(src->mem.placement & TTM_PL_FLAG_NO_EVICT))
463 		lockdep_assert_held(&src->resv->lock.base);
464 
465 	if (dst->ttm->state == tt_unpopulated) {
466 		ret = dst->ttm->bdev->driver->ttm_tt_populate(dst->ttm, &ctx);
467 		if (ret)
468 			return ret;
469 	}
470 
471 	if (src->ttm->state == tt_unpopulated) {
472 		ret = src->ttm->bdev->driver->ttm_tt_populate(src->ttm, &ctx);
473 		if (ret)
474 			return ret;
475 	}
476 
477 	d.mapped_dst = 0;
478 	d.mapped_src = 0;
479 	d.dst_addr = NULL;
480 	d.src_addr = NULL;
481 	d.dst_pages = dst->ttm->pages;
482 	d.src_pages = src->ttm->pages;
483 	d.dst_num_pages = dst->num_pages;
484 	d.src_num_pages = src->num_pages;
485 	d.dst_prot = ttm_io_prot(dst->mem.placement, PAGE_KERNEL);
486 	d.src_prot = ttm_io_prot(src->mem.placement, PAGE_KERNEL);
487 	d.diff = diff;
488 
489 	for (j = 0; j < h; ++j) {
490 		diff->line = j + initial_line;
491 		diff->line_offset = dst_offset % dst_stride;
492 		ret = vmw_bo_cpu_blit_line(&d, dst_offset, src_offset, w);
493 		if (ret)
494 			goto out;
495 
496 		dst_offset += dst_stride;
497 		src_offset += src_stride;
498 	}
499 out:
500 	if (d.src_addr)
501 		ttm_kunmap_atomic_prot(d.src_addr, d.src_prot);
502 	if (d.dst_addr)
503 		ttm_kunmap_atomic_prot(d.dst_addr, d.dst_prot);
504 
505 	return ret;
506 }
507