1/*
  2 * Copyright 2011 Tilera Corporation. All Rights Reserved.
  3 *
  4 *   This program is free software; you can redistribute it and/or
  5 *   modify it under the terms of the GNU General Public License
  6 *   as published by the Free Software Foundation, version 2.
  7 *
  8 *   This program is distributed in the hope that it will be useful, but
  9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
 10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 11 *   NON INFRINGEMENT.  See the GNU General Public License for
 12 *   more details.
 13 */
 14
 15#include <linux/types.h>
 16#include <linux/string.h>
 17#include <linux/module.h>
 18#define __memcpy memcpy
 19/* EXPORT_SYMBOL() is in arch/tile/lib/exports.c since this should be asm. */
 20
 21/* Must be 8 bytes in size. */
 22#define word_t uint64_t
 23
 24#if CHIP_L2_LINE_SIZE() != 64 && CHIP_L2_LINE_SIZE() != 128
 25#error "Assumes 64 or 128 byte line size"
 26#endif
 27
 28/* How many cache lines ahead should we prefetch? */
 29#define PREFETCH_LINES_AHEAD 3
 30
 31/*
 32 * Provide "base versions" of load and store for the normal code path.
 33 * The kernel provides other versions for userspace copies.
 34 */
 35#define ST(p, v) (*(p) = (v))
 36#define LD(p) (*(p))
 37
 38#ifndef USERCOPY_FUNC
 39#define ST1 ST
 40#define ST2 ST
 41#define ST4 ST
 42#define ST8 ST
 43#define LD1 LD
 44#define LD2 LD
 45#define LD4 LD
 46#define LD8 LD
 47#define RETVAL dstv
 48void *memcpy(void *__restrict dstv, const void *__restrict srcv, size_t n)
 49#else
 50/*
 51 * Special kernel version will provide implementation of the LDn/STn
 52 * macros to return a count of uncopied bytes due to mm fault.
 53 */
 54#define RETVAL 0
 55int USERCOPY_FUNC(void *__restrict dstv, const void *__restrict srcv, size_t n)
 56#endif
 57{
 58	char *__restrict dst1 = (char *)dstv;
 59	const char *__restrict src1 = (const char *)srcv;
 60	const char *__restrict src1_end;
 61	const char *__restrict prefetch;
 62	word_t *__restrict dst8;    /* 8-byte pointer to destination memory. */
 63	word_t final; /* Final bytes to write to trailing word, if any */
 64	long i;
 65
 66	if (n < 16) {
 67		for (; n; n--)
 68			ST1(dst1++, LD1(src1++));
 69		return RETVAL;
 70	}
 71
 72	/*
 73	 * Locate the end of source memory we will copy.  Don't
 74	 * prefetch past this.
 75	 */
 76	src1_end = src1 + n - 1;
 77
 78	/* Prefetch ahead a few cache lines, but not past the end. */
 79	prefetch = src1;
 80	for (i = 0; i < PREFETCH_LINES_AHEAD; i++) {
 81		__insn_prefetch(prefetch);
 82		prefetch += CHIP_L2_LINE_SIZE();
 83		prefetch = (prefetch > src1_end) ? prefetch : src1;
 84	}
 85
 86	/* Copy bytes until dst is word-aligned. */
 87	for (; (uintptr_t)dst1 & (sizeof(word_t) - 1); n--)
 88		ST1(dst1++, LD1(src1++));
 89
 90	/* 8-byte pointer to destination memory. */
 91	dst8 = (word_t *)dst1;
 92
 93	if (__builtin_expect((uintptr_t)src1 & (sizeof(word_t) - 1), 0)) {
 94		/*
 95		 * Misaligned copy.  Copy 8 bytes at a time, but don't
 96		 * bother with other fanciness.
 97		 *
 98		 * TODO: Consider prefetching and using wh64 as well.
 99		 */
100
101		/* Create an aligned src8. */
102		const word_t *__restrict src8 =
103			(const word_t *)((uintptr_t)src1 & -sizeof(word_t));
104		word_t b;
105
106		word_t a = LD8(src8++);
107		for (; n >= sizeof(word_t); n -= sizeof(word_t)) {
108			b = LD8(src8++);
109			a = __insn_dblalign(a, b, src1);
110			ST8(dst8++, a);
111			a = b;
112		}
113
114		if (n == 0)
115			return RETVAL;
116
117		b = ((const char *)src8 <= src1_end) ? *src8 : 0;
118
119		/*
120		 * Final source bytes to write to trailing partial
121		 * word, if any.
122		 */
123		final = __insn_dblalign(a, b, src1);
124	} else {
125		/* Aligned copy. */
126
127		const word_t* __restrict src8 = (const word_t *)src1;
128
129		/* src8 and dst8 are both word-aligned. */
130		if (n >= CHIP_L2_LINE_SIZE()) {
131			/* Copy until 'dst' is cache-line-aligned. */
132			for (; (uintptr_t)dst8 & (CHIP_L2_LINE_SIZE() - 1);
133			     n -= sizeof(word_t))
134				ST8(dst8++, LD8(src8++));
135
136			for (; n >= CHIP_L2_LINE_SIZE(); ) {
137				__insn_wh64(dst8);
138
139				/*
140				 * Prefetch and advance to next line
141				 * to prefetch, but don't go past the end
142				 */
143				__insn_prefetch(prefetch);
144				prefetch += CHIP_L2_LINE_SIZE();
145				prefetch = (prefetch > src1_end) ? prefetch :
146					(const char *)src8;
147
148				/*
149				 * Copy an entire cache line.  Manually
150				 * unrolled to avoid idiosyncracies of
151				 * compiler unrolling.
152				 */
153#define COPY_WORD(offset) ({ ST8(dst8+offset, LD8(src8+offset)); n -= 8; })
154				COPY_WORD(0);
155				COPY_WORD(1);
156				COPY_WORD(2);
157				COPY_WORD(3);
158				COPY_WORD(4);
159				COPY_WORD(5);
160				COPY_WORD(6);
161				COPY_WORD(7);
162#if CHIP_L2_LINE_SIZE() == 128
163				COPY_WORD(8);
164				COPY_WORD(9);
165				COPY_WORD(10);
166				COPY_WORD(11);
167				COPY_WORD(12);
168				COPY_WORD(13);
169				COPY_WORD(14);
170				COPY_WORD(15);
171#elif CHIP_L2_LINE_SIZE() != 64
172# error Fix code that assumes particular L2 cache line sizes
173#endif
174
175				dst8 += CHIP_L2_LINE_SIZE() / sizeof(word_t);
176				src8 += CHIP_L2_LINE_SIZE() / sizeof(word_t);
177			}
178		}
179
180		for (; n >= sizeof(word_t); n -= sizeof(word_t))
181			ST8(dst8++, LD8(src8++));
182
183		if (__builtin_expect(n == 0, 1))
184			return RETVAL;
185
186		final = LD8(src8);
187	}
188
189	/* n != 0 if we get here.  Write out any trailing bytes. */
190	dst1 = (char *)dst8;
191	if (n & 4) {
192		ST4((uint32_t *)dst1, final);
193		dst1 += 4;
194		final >>= 32;
195		n &= 3;
196	}
197	if (n & 2) {
198		ST2((uint16_t *)dst1, final);
199		dst1 += 2;
200		final >>= 16;
201		n &= 1;
202	}
203	if (n)
204		ST1((uint8_t *)dst1, final);
205
206	return RETVAL;
207}
208
209
210#ifdef USERCOPY_FUNC
211#undef ST1
212#undef ST2
213#undef ST4
214#undef ST8
215#undef LD1
216#undef LD2
217#undef LD4
218#undef LD8
219#undef USERCOPY_FUNC
220#endif