projects / openwrt / staging / xback.git / blob
? search:


430059cddf0005962f98df8b72c7cb252713fb84
[openwrt/staging/xback.git] /
1 From e4d81adf1f2c84b229901cddb403d00010524b28 Mon Sep 17 00:00:00 2001
2 From: popcornmix <popcornmix@gmail.com>
3 Date: Mon, 28 Nov 2016 16:50:04 +0000
4 Subject: [PATCH] Improve __copy_to_user and __copy_from_user
5 performance
6
7 Provide a __copy_from_user that uses memcpy. On BCM2708, use
8 optimised memcpy/memmove/memcmp/memset implementations.
9
10 arch/arm: Add mmiocpy/set aliases for memcpy/set
11
12 See: https://github.com/raspberrypi/linux/issues/1082
13
14 copy_from_user: CPU_SW_DOMAIN_PAN compatibility
15
16 The downstream copy_from_user acceleration must also play nice with
17 CONFIG_CPU_SW_DOMAIN_PAN.
18
19 See: https://github.com/raspberrypi/linux/issues/1381
20
21 Signed-off-by: Phil Elwell <phil@raspberrypi.org>
22 ---
23 arch/arm/include/asm/string.h | 5 +
24 arch/arm/include/asm/uaccess.h | 3 +
25 arch/arm/lib/Makefile | 14 +-
26 arch/arm/lib/arm-mem.h | 159 +++++++++
27 arch/arm/lib/copy_from_user.S | 4 +-
28 arch/arm/lib/exports_rpi.c | 37 +++
29 arch/arm/lib/memcmp_rpi.S | 285 ++++++++++++++++
30 arch/arm/lib/memcpy_rpi.S | 61 ++++
31 arch/arm/lib/memcpymove.h | 506 +++++++++++++++++++++++++++++
32 arch/arm/lib/memmove_rpi.S | 61 ++++
33 arch/arm/lib/memset_rpi.S | 128 ++++++++
34 arch/arm/lib/uaccess_with_memcpy.c | 120 ++++++-
35 arch/arm/mach-bcm/Kconfig | 7 +
36 13 files changed, 1385 insertions(+), 5 deletions(-)
37 create mode 100644 arch/arm/lib/arm-mem.h
38 create mode 100644 arch/arm/lib/exports_rpi.c
39 create mode 100644 arch/arm/lib/memcmp_rpi.S
40 create mode 100644 arch/arm/lib/memcpy_rpi.S
41 create mode 100644 arch/arm/lib/memcpymove.h
42 create mode 100644 arch/arm/lib/memmove_rpi.S
43 create mode 100644 arch/arm/lib/memset_rpi.S
44
45 --- a/arch/arm/include/asm/string.h
46 +++ b/arch/arm/include/asm/string.h
47 @@ -39,4 +39,9 @@ static inline void *memset64(uint64_t *p
48 return __memset64(p, v, n * 8, v >> 32);
49 }
50
51 +#ifdef CONFIG_BCM2835_FAST_MEMCPY
52 +#define __HAVE_ARCH_MEMCMP
53 +extern int memcmp(const void *, const void *, size_t);
54 +#endif
55 +
56 #endif
57 --- a/arch/arm/include/asm/uaccess.h
58 +++ b/arch/arm/include/asm/uaccess.h
59 @@ -514,6 +514,9 @@ do { \
60 extern unsigned long __must_check
61 arm_copy_from_user(void *to, const void __user *from, unsigned long n);
62
63 +extern unsigned long __must_check
64 +__copy_from_user_std(void *to, const void __user *from, unsigned long n);
65 +
66 static inline unsigned long __must_check
67 raw_copy_from_user(void *to, const void __user *from, unsigned long n)
68 {
69 --- a/arch/arm/lib/Makefile
70 +++ b/arch/arm/lib/Makefile
71 @@ -7,8 +7,8 @@
72
73 lib-y := backtrace.o changebit.o csumipv6.o csumpartial.o \
74 csumpartialcopy.o csumpartialcopyuser.o clearbit.o \
75 - delay.o delay-loop.o findbit.o memchr.o memcpy.o \
76 - memmove.o memset.o setbit.o \
77 + delay.o delay-loop.o findbit.o memchr.o \
78 + setbit.o \
79 strchr.o strrchr.o \
80 testchangebit.o testclearbit.o testsetbit.o \
81 ashldi3.o ashrdi3.o lshrdi3.o muldi3.o \
82 @@ -19,6 +19,16 @@ lib-y := backtrace.o changebit.o csumip
83 mmu-y := clear_user.o copy_page.o getuser.o putuser.o \
84 copy_from_user.o copy_to_user.o
85
86 +# Choose optimised implementations for Raspberry Pi
87 +ifeq ($(CONFIG_BCM2835_FAST_MEMCPY),y)
88 + CFLAGS_uaccess_with_memcpy.o += -DCOPY_FROM_USER_THRESHOLD=1600
89 + CFLAGS_uaccess_with_memcpy.o += -DCOPY_TO_USER_THRESHOLD=672
90 + obj-$(CONFIG_MODULES) += exports_rpi.o
91 + lib-y += memcpy_rpi.o memmove_rpi.o memset_rpi.o memcmp_rpi.o
92 +else
93 + lib-y += memcpy.o memmove.o memset.o
94 +endif
95 +
96 # using lib_ here won't override already available weak symbols
97 obj-$(CONFIG_UACCESS_WITH_MEMCPY) += uaccess_with_memcpy.o
98
99 --- /dev/null
100 +++ b/arch/arm/lib/arm-mem.h
101 @@ -0,0 +1,159 @@
102 +/*
103 +Copyright (c) 2013, Raspberry Pi Foundation
104 +Copyright (c) 2013, RISC OS Open Ltd
105 +All rights reserved.
106 +
107 +Redistribution and use in source and binary forms, with or without
108 +modification, are permitted provided that the following conditions are met:
109 + * Redistributions of source code must retain the above copyright
110 + notice, this list of conditions and the following disclaimer.
111 + * Redistributions in binary form must reproduce the above copyright
112 + notice, this list of conditions and the following disclaimer in the
113 + documentation and/or other materials provided with the distribution.
114 + * Neither the name of the copyright holder nor the
115 + names of its contributors may be used to endorse or promote products
116 + derived from this software without specific prior written permission.
117 +
118 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
119 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
120 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
121 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
122 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
123 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
124 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
125 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
126 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
127 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
128 +*/
129 +
130 +.macro myfunc fname
131 + .func fname
132 + .global fname
133 +fname:
134 +.endm
135 +
136 +.macro preload_leading_step1 backwards, ptr, base
137 +/* If the destination is already 16-byte aligned, then we need to preload
138 + * between 0 and prefetch_distance (inclusive) cache lines ahead so there
139 + * are no gaps when the inner loop starts.
140 + */
141 + .if backwards
142 + sub ptr, base, #1
143 + bic ptr, ptr, #31
144 + .else
145 + bic ptr, base, #31
146 + .endif
147 + .set OFFSET, 0
148 + .rept prefetch_distance+1
149 + pld [ptr, #OFFSET]
150 + .if backwards
151 + .set OFFSET, OFFSET-32
152 + .else
153 + .set OFFSET, OFFSET+32
154 + .endif
155 + .endr
156 +.endm
157 +
158 +.macro preload_leading_step2 backwards, ptr, base, leading_bytes, tmp
159 +/* However, if the destination is not 16-byte aligned, we may need to
160 + * preload one more cache line than that. The question we need to ask is:
161 + * are the leading bytes more than the amount by which the source
162 + * pointer will be rounded down for preloading, and if so, by how many
163 + * cache lines?
164 + */
165 + .if backwards
166 +/* Here we compare against how many bytes we are into the
167 + * cache line, counting down from the highest such address.
168 + * Effectively, we want to calculate
169 + * leading_bytes = dst&15
170 + * cacheline_offset = 31-((src-leading_bytes-1)&31)
171 + * extra_needed = leading_bytes - cacheline_offset
172 + * and test if extra_needed is <= 0, or rearranging:
173 + * leading_bytes + (src-leading_bytes-1)&31 <= 31
174 + */
175 + mov tmp, base, lsl #32-5
176 + sbc tmp, tmp, leading_bytes, lsl #32-5
177 + adds tmp, tmp, leading_bytes, lsl #32-5
178 + bcc 61f
179 + pld [ptr, #-32*(prefetch_distance+1)]
180 + .else
181 +/* Effectively, we want to calculate
182 + * leading_bytes = (-dst)&15
183 + * cacheline_offset = (src+leading_bytes)&31
184 + * extra_needed = leading_bytes - cacheline_offset
185 + * and test if extra_needed is <= 0.
186 + */
187 + mov tmp, base, lsl #32-5
188 + add tmp, tmp, leading_bytes, lsl #32-5
189 + rsbs tmp, tmp, leading_bytes, lsl #32-5
190 + bls 61f
191 + pld [ptr, #32*(prefetch_distance+1)]
192 + .endif
193 +61:
194 +.endm
195 +
196 +.macro preload_trailing backwards, base, remain, tmp
197 + /* We need either 0, 1 or 2 extra preloads */
198 + .if backwards
199 + rsb tmp, base, #0
200 + mov tmp, tmp, lsl #32-5
201 + .else
202 + mov tmp, base, lsl #32-5
203 + .endif
204 + adds tmp, tmp, remain, lsl #32-5
205 + adceqs tmp, tmp, #0
206 + /* The instruction above has two effects: ensures Z is only
207 + * set if C was clear (so Z indicates that both shifted quantities
208 + * were 0), and clears C if Z was set (so C indicates that the sum
209 + * of the shifted quantities was greater and not equal to 32) */
210 + beq 82f
211 + .if backwards
212 + sub tmp, base, #1
213 + bic tmp, tmp, #31
214 + .else
215 + bic tmp, base, #31
216 + .endif
217 + bcc 81f
218 + .if backwards
219 + pld [tmp, #-32*(prefetch_distance+1)]
220 +81:
221 + pld [tmp, #-32*prefetch_distance]
222 + .else
223 + pld [tmp, #32*(prefetch_distance+2)]
224 +81:
225 + pld [tmp, #32*(prefetch_distance+1)]
226 + .endif
227 +82:
228 +.endm
229 +
230 +.macro preload_all backwards, narrow_case, shift, base, remain, tmp0, tmp1
231 + .if backwards
232 + sub tmp0, base, #1
233 + bic tmp0, tmp0, #31
234 + pld [tmp0]
235 + sub tmp1, base, remain, lsl #shift
236 + .else
237 + bic tmp0, base, #31
238 + pld [tmp0]
239 + add tmp1, base, remain, lsl #shift
240 + sub tmp1, tmp1, #1
241 + .endif
242 + bic tmp1, tmp1, #31
243 + cmp tmp1, tmp0
244 + beq 92f
245 + .if narrow_case
246 + /* In this case, all the data fits in either 1 or 2 cache lines */
247 + pld [tmp1]
248 + .else
249 +91:
250 + .if backwards
251 + sub tmp0, tmp0, #32
252 + .else
253 + add tmp0, tmp0, #32
254 + .endif
255 + cmp tmp0, tmp1
256 + pld [tmp0]
257 + bne 91b
258 + .endif
259 +92:
260 +.endm
261 --- a/arch/arm/lib/copy_from_user.S
262 +++ b/arch/arm/lib/copy_from_user.S
263 @@ -89,7 +89,8 @@
264
265 .text
266
267 -ENTRY(arm_copy_from_user)
268 +ENTRY(__copy_from_user_std)
269 +WEAK(arm_copy_from_user)
270 #ifdef CONFIG_CPU_SPECTRE
271 get_thread_info r3
272 ldr r3, [r3, #TI_ADDR_LIMIT]
273 @@ -99,6 +100,7 @@ ENTRY(arm_copy_from_user)
274 #include "copy_template.S"
275
276 ENDPROC(arm_copy_from_user)
277 +ENDPROC(__copy_from_user_std)
278
279 .pushsection .fixup,"ax"
280 .align 0
281 --- /dev/null
282 +++ b/arch/arm/lib/exports_rpi.c
283 @@ -0,0 +1,37 @@
284 +/**
285 + * Copyright (c) 2014, Raspberry Pi (Trading) Ltd.
286 + *
287 + * Redistribution and use in source and binary forms, with or without
288 + * modification, are permitted provided that the following conditions
289 + * are met:
290 + * 1. Redistributions of source code must retain the above copyright
291 + * notice, this list of conditions, and the following disclaimer,
292 + * without modification.
293 + * 2. Redistributions in binary form must reproduce the above copyright
294 + * notice, this list of conditions and the following disclaimer in the
295 + * documentation and/or other materials provided with the distribution.
296 + * 3. The names of the above-listed copyright holders may not be used
297 + * to endorse or promote products derived from this software without
298 + * specific prior written permission.
299 + *
300 + * ALTERNATIVELY, this software may be distributed under the terms of the
301 + * GNU General Public License ("GPL") version 2, as published by the Free
302 + * Software Foundation.
303 + *
304 + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
305 + * IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
306 + * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
307 + * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
308 + * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
309 + * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
310 + * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
311 + * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
312 + * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
313 + * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
314 + * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
315 + */
316 +
317 +#include <linux/kernel.h>
318 +#include <linux/module.h>
319 +
320 +EXPORT_SYMBOL(memcmp);
321 --- /dev/null
322 +++ b/arch/arm/lib/memcmp_rpi.S
323 @@ -0,0 +1,285 @@
324 +/*
325 +Copyright (c) 2013, Raspberry Pi Foundation
326 +Copyright (c) 2013, RISC OS Open Ltd
327 +All rights reserved.
328 +
329 +Redistribution and use in source and binary forms, with or without
330 +modification, are permitted provided that the following conditions are met:
331 + * Redistributions of source code must retain the above copyright
332 + notice, this list of conditions and the following disclaimer.
333 + * Redistributions in binary form must reproduce the above copyright
334 + notice, this list of conditions and the following disclaimer in the
335 + documentation and/or other materials provided with the distribution.
336 + * Neither the name of the copyright holder nor the
337 + names of its contributors may be used to endorse or promote products
338 + derived from this software without specific prior written permission.
339 +
340 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
341 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
342 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
343 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
344 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
345 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
346 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
347 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
348 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
349 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
350 +*/
351 +
352 +#include <linux/linkage.h>
353 +#include "arm-mem.h"
354 +
355 +/* Prevent the stack from becoming executable */
356 +#if defined(__linux__) && defined(__ELF__)
357 +.section .note.GNU-stack,"",%progbits
358 +#endif
359 +
360 + .text
361 + .arch armv6
362 + .object_arch armv4
363 + .arm
364 + .altmacro
365 + .p2align 2
366 +
367 +.macro memcmp_process_head unaligned
368 + .if unaligned
369 + ldr DAT0, [S_1], #4
370 + ldr DAT1, [S_1], #4
371 + ldr DAT2, [S_1], #4
372 + ldr DAT3, [S_1], #4
373 + .else
374 + ldmia S_1!, {DAT0, DAT1, DAT2, DAT3}
375 + .endif
376 + ldmia S_2!, {DAT4, DAT5, DAT6, DAT7}
377 +.endm
378 +
379 +.macro memcmp_process_tail
380 + cmp DAT0, DAT4
381 + cmpeq DAT1, DAT5
382 + cmpeq DAT2, DAT6
383 + cmpeq DAT3, DAT7
384 + bne 200f
385 +.endm
386 +
387 +.macro memcmp_leading_31bytes
388 + movs DAT0, OFF, lsl #31
389 + ldrmib DAT0, [S_1], #1
390 + ldrcsh DAT1, [S_1], #2
391 + ldrmib DAT4, [S_2], #1
392 + ldrcsh DAT5, [S_2], #2
393 + movpl DAT0, #0
394 + movcc DAT1, #0
395 + movpl DAT4, #0
396 + movcc DAT5, #0
397 + submi N, N, #1
398 + subcs N, N, #2
399 + cmp DAT0, DAT4
400 + cmpeq DAT1, DAT5
401 + bne 200f
402 + movs DAT0, OFF, lsl #29
403 + ldrmi DAT0, [S_1], #4
404 + ldrcs DAT1, [S_1], #4
405 + ldrcs DAT2, [S_1], #4
406 + ldrmi DAT4, [S_2], #4
407 + ldmcsia S_2!, {DAT5, DAT6}
408 + movpl DAT0, #0
409 + movcc DAT1, #0
410 + movcc DAT2, #0
411 + movpl DAT4, #0
412 + movcc DAT5, #0
413 + movcc DAT6, #0
414 + submi N, N, #4
415 + subcs N, N, #8
416 + cmp DAT0, DAT4
417 + cmpeq DAT1, DAT5
418 + cmpeq DAT2, DAT6
419 + bne 200f
420 + tst OFF, #16
421 + beq 105f
422 + memcmp_process_head 1
423 + sub N, N, #16
424 + memcmp_process_tail
425 +105:
426 +.endm
427 +
428 +.macro memcmp_trailing_15bytes unaligned
429 + movs N, N, lsl #29
430 + .if unaligned
431 + ldrcs DAT0, [S_1], #4
432 + ldrcs DAT1, [S_1], #4
433 + .else
434 + ldmcsia S_1!, {DAT0, DAT1}
435 + .endif
436 + ldrmi DAT2, [S_1], #4
437 + ldmcsia S_2!, {DAT4, DAT5}
438 + ldrmi DAT6, [S_2], #4
439 + movcc DAT0, #0
440 + movcc DAT1, #0
441 + movpl DAT2, #0
442 + movcc DAT4, #0
443 + movcc DAT5, #0
444 + movpl DAT6, #0
445 + cmp DAT0, DAT4
446 + cmpeq DAT1, DAT5
447 + cmpeq DAT2, DAT6
448 + bne 200f
449 + movs N, N, lsl #2
450 + ldrcsh DAT0, [S_1], #2
451 + ldrmib DAT1, [S_1]
452 + ldrcsh DAT4, [S_2], #2
453 + ldrmib DAT5, [S_2]
454 + movcc DAT0, #0
455 + movpl DAT1, #0
456 + movcc DAT4, #0
457 + movpl DAT5, #0
458 + cmp DAT0, DAT4
459 + cmpeq DAT1, DAT5
460 + bne 200f
461 +.endm
462 +
463 +.macro memcmp_long_inner_loop unaligned
464 +110:
465 + memcmp_process_head unaligned
466 + pld [S_2, #prefetch_distance*32 + 16]
467 + memcmp_process_tail
468 + memcmp_process_head unaligned
469 + pld [S_1, OFF]
470 + memcmp_process_tail
471 + subs N, N, #32
472 + bhs 110b
473 + /* Just before the final (prefetch_distance+1) 32-byte blocks,
474 + * deal with final preloads */
475 + preload_trailing 0, S_1, N, DAT0
476 + preload_trailing 0, S_2, N, DAT0
477 + add N, N, #(prefetch_distance+2)*32 - 16
478 +120:
479 + memcmp_process_head unaligned
480 + memcmp_process_tail
481 + subs N, N, #16
482 + bhs 120b
483 + /* Trailing words and bytes */
484 + tst N, #15
485 + beq 199f
486 + memcmp_trailing_15bytes unaligned
487 +199: /* Reached end without detecting a difference */
488 + mov a1, #0
489 + setend le
490 + pop {DAT1-DAT6, pc}
491 +.endm
492 +
493 +.macro memcmp_short_inner_loop unaligned
494 + subs N, N, #16 /* simplifies inner loop termination */
495 + blo 122f
496 +120:
497 + memcmp_process_head unaligned
498 + memcmp_process_tail
499 + subs N, N, #16
500 + bhs 120b
501 +122: /* Trailing words and bytes */
502 + tst N, #15
503 + beq 199f
504 + memcmp_trailing_15bytes unaligned
505 +199: /* Reached end without detecting a difference */
506 + mov a1, #0
507 + setend le
508 + pop {DAT1-DAT6, pc}
509 +.endm
510 +
511 +/*
512 + * int memcmp(const void *s1, const void *s2, size_t n);
513 + * On entry:
514 + * a1 = pointer to buffer 1
515 + * a2 = pointer to buffer 2
516 + * a3 = number of bytes to compare (as unsigned chars)
517 + * On exit:
518 + * a1 = >0/=0/<0 if s1 >/=/< s2
519 + */
520 +
521 +.set prefetch_distance, 2
522 +
523 +ENTRY(memcmp)
524 + S_1 .req a1
525 + S_2 .req a2
526 + N .req a3
527 + DAT0 .req a4
528 + DAT1 .req v1
529 + DAT2 .req v2
530 + DAT3 .req v3
531 + DAT4 .req v4
532 + DAT5 .req v5
533 + DAT6 .req v6
534 + DAT7 .req ip
535 + OFF .req lr
536 +
537 + push {DAT1-DAT6, lr}
538 + setend be /* lowest-addressed bytes are most significant */
539 +
540 + /* To preload ahead as we go, we need at least (prefetch_distance+2) 32-byte blocks */
541 + cmp N, #(prefetch_distance+3)*32 - 1
542 + blo 170f
543 +
544 + /* Long case */
545 + /* Adjust N so that the decrement instruction can also test for
546 + * inner loop termination. We want it to stop when there are
547 + * (prefetch_distance+1) complete blocks to go. */
548 + sub N, N, #(prefetch_distance+2)*32
549 + preload_leading_step1 0, DAT0, S_1
550 + preload_leading_step1 0, DAT1, S_2
551 + tst S_2, #31
552 + beq 154f
553 + rsb OFF, S_2, #0 /* no need to AND with 15 here */
554 + preload_leading_step2 0, DAT0, S_1, OFF, DAT2
555 + preload_leading_step2 0, DAT1, S_2, OFF, DAT2
556 + memcmp_leading_31bytes
557 +154: /* Second source now cacheline (32-byte) aligned; we have at
558 + * least one prefetch to go. */
559 + /* Prefetch offset is best selected such that it lies in the
560 + * first 8 of each 32 bytes - but it's just as easy to aim for
561 + * the first one */
562 + and OFF, S_1, #31
563 + rsb OFF, OFF, #32*prefetch_distance
564 + tst S_1, #3
565 + bne 140f
566 + memcmp_long_inner_loop 0
567 +140: memcmp_long_inner_loop 1
568 +
569 +170: /* Short case */
570 + teq N, #0
571 + beq 199f
572 + preload_all 0, 0, 0, S_1, N, DAT0, DAT1
573 + preload_all 0, 0, 0, S_2, N, DAT0, DAT1
574 + tst S_2, #3
575 + beq 174f
576 +172: subs N, N, #1
577 + blo 199f
578 + ldrb DAT0, [S_1], #1
579 + ldrb DAT4, [S_2], #1
580 + cmp DAT0, DAT4
581 + bne 200f
582 + tst S_2, #3
583 + bne 172b
584 +174: /* Second source now 4-byte aligned; we have 0 or more bytes to go */
585 + tst S_1, #3
586 + bne 140f
587 + memcmp_short_inner_loop 0
588 +140: memcmp_short_inner_loop 1
589 +
590 +200: /* Difference found: determine sign. */
591 + movhi a1, #1
592 + movlo a1, #-1
593 + setend le
594 + pop {DAT1-DAT6, pc}
595 +
596 + .unreq S_1
597 + .unreq S_2
598 + .unreq N
599 + .unreq DAT0
600 + .unreq DAT1
601 + .unreq DAT2
602 + .unreq DAT3
603 + .unreq DAT4
604 + .unreq DAT5
605 + .unreq DAT6
606 + .unreq DAT7
607 + .unreq OFF
608 +ENDPROC(memcmp)
609 --- /dev/null
610 +++ b/arch/arm/lib/memcpy_rpi.S
611 @@ -0,0 +1,61 @@
612 +/*
613 +Copyright (c) 2013, Raspberry Pi Foundation
614 +Copyright (c) 2013, RISC OS Open Ltd
615 +All rights reserved.
616 +
617 +Redistribution and use in source and binary forms, with or without
618 +modification, are permitted provided that the following conditions are met:
619 + * Redistributions of source code must retain the above copyright
620 + notice, this list of conditions and the following disclaimer.
621 + * Redistributions in binary form must reproduce the above copyright
622 + notice, this list of conditions and the following disclaimer in the
623 + documentation and/or other materials provided with the distribution.
624 + * Neither the name of the copyright holder nor the
625 + names of its contributors may be used to endorse or promote products
626 + derived from this software without specific prior written permission.
627 +
628 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
629 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
630 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
631 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
632 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
633 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
634 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
635 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
636 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
637 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
638 +*/
639 +
640 +#include <linux/linkage.h>
641 +#include "arm-mem.h"
642 +#include "memcpymove.h"
643 +
644 +/* Prevent the stack from becoming executable */
645 +#if defined(__linux__) && defined(__ELF__)
646 +.section .note.GNU-stack,"",%progbits
647 +#endif
648 +
649 + .text
650 + .arch armv6
651 + .object_arch armv4
652 + .arm
653 + .altmacro
654 + .p2align 2
655 +
656 +/*
657 + * void *memcpy(void * restrict s1, const void * restrict s2, size_t n);
658 + * On entry:
659 + * a1 = pointer to destination
660 + * a2 = pointer to source
661 + * a3 = number of bytes to copy
662 + * On exit:
663 + * a1 preserved
664 + */
665 +
666 +.set prefetch_distance, 3
667 +
668 +ENTRY(mmiocpy)
669 +ENTRY(memcpy)
670 + memcpy 0
671 +ENDPROC(memcpy)
672 +ENDPROC(mmiocpy)
673 --- /dev/null
674 +++ b/arch/arm/lib/memcpymove.h
675 @@ -0,0 +1,506 @@
676 +/*
677 +Copyright (c) 2013, Raspberry Pi Foundation
678 +Copyright (c) 2013, RISC OS Open Ltd
679 +All rights reserved.
680 +
681 +Redistribution and use in source and binary forms, with or without
682 +modification, are permitted provided that the following conditions are met:
683 + * Redistributions of source code must retain the above copyright
684 + notice, this list of conditions and the following disclaimer.
685 + * Redistributions in binary form must reproduce the above copyright
686 + notice, this list of conditions and the following disclaimer in the
687 + documentation and/or other materials provided with the distribution.
688 + * Neither the name of the copyright holder nor the
689 + names of its contributors may be used to endorse or promote products
690 + derived from this software without specific prior written permission.
691 +
692 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
693 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
694 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
695 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
696 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
697 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
698 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
699 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
700 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
701 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
702 +*/
703 +
704 +.macro unaligned_words backwards, align, use_pld, words, r0, r1, r2, r3, r4, r5, r6, r7, r8
705 + .if words == 1
706 + .if backwards
707 + mov r1, r0, lsl #32-align*8
708 + ldr r0, [S, #-4]!
709 + orr r1, r1, r0, lsr #align*8
710 + str r1, [D, #-4]!
711 + .else
712 + mov r0, r1, lsr #align*8
713 + ldr r1, [S, #4]!
714 + orr r0, r0, r1, lsl #32-align*8
715 + str r0, [D], #4
716 + .endif
717 + .elseif words == 2
718 + .if backwards
719 + ldr r1, [S, #-4]!
720 + mov r2, r0, lsl #32-align*8
721 + ldr r0, [S, #-4]!
722 + orr r2, r2, r1, lsr #align*8
723 + mov r1, r1, lsl #32-align*8
724 + orr r1, r1, r0, lsr #align*8
725 + stmdb D!, {r1, r2}
726 + .else
727 + ldr r1, [S, #4]!
728 + mov r0, r2, lsr #align*8
729 + ldr r2, [S, #4]!
730 + orr r0, r0, r1, lsl #32-align*8
731 + mov r1, r1, lsr #align*8
732 + orr r1, r1, r2, lsl #32-align*8
733 + stmia D!, {r0, r1}
734 + .endif
735 + .elseif words == 4
736 + .if backwards
737 + ldmdb S!, {r2, r3}
738 + mov r4, r0, lsl #32-align*8
739 + ldmdb S!, {r0, r1}
740 + orr r4, r4, r3, lsr #align*8
741 + mov r3, r3, lsl #32-align*8
742 + orr r3, r3, r2, lsr #align*8
743 + mov r2, r2, lsl #32-align*8
744 + orr r2, r2, r1, lsr #align*8
745 + mov r1, r1, lsl #32-align*8
746 + orr r1, r1, r0, lsr #align*8
747 + stmdb D!, {r1, r2, r3, r4}
748 + .else
749 + ldmib S!, {r1, r2}
750 + mov r0, r4, lsr #align*8
751 + ldmib S!, {r3, r4}
752 + orr r0, r0, r1, lsl #32-align*8
753 + mov r1, r1, lsr #align*8
754 + orr r1, r1, r2, lsl #32-align*8
755 + mov r2, r2, lsr #align*8
756 + orr r2, r2, r3, lsl #32-align*8
757 + mov r3, r3, lsr #align*8
758 + orr r3, r3, r4, lsl #32-align*8
759 + stmia D!, {r0, r1, r2, r3}
760 + .endif
761 + .elseif words == 8
762 + .if backwards
763 + ldmdb S!, {r4, r5, r6, r7}
764 + mov r8, r0, lsl #32-align*8
765 + ldmdb S!, {r0, r1, r2, r3}
766 + .if use_pld
767 + pld [S, OFF]
768 + .endif
769 + orr r8, r8, r7, lsr #align*8
770 + mov r7, r7, lsl #32-align*8
771 + orr r7, r7, r6, lsr #align*8
772 + mov r6, r6, lsl #32-align*8
773 + orr r6, r6, r5, lsr #align*8
774 + mov r5, r5, lsl #32-align*8
775 + orr r5, r5, r4, lsr #align*8
776 + mov r4, r4, lsl #32-align*8
777 + orr r4, r4, r3, lsr #align*8
778 + mov r3, r3, lsl #32-align*8
779 + orr r3, r3, r2, lsr #align*8
780 + mov r2, r2, lsl #32-align*8
781 + orr r2, r2, r1, lsr #align*8
782 + mov r1, r1, lsl #32-align*8
783 + orr r1, r1, r0, lsr #align*8
784 + stmdb D!, {r5, r6, r7, r8}
785 + stmdb D!, {r1, r2, r3, r4}
786 + .else
787 + ldmib S!, {r1, r2, r3, r4}
788 + mov r0, r8, lsr #align*8
789 + ldmib S!, {r5, r6, r7, r8}
790 + .if use_pld
791 + pld [S, OFF]
792 + .endif
793 + orr r0, r0, r1, lsl #32-align*8
794 + mov r1, r1, lsr #align*8
795 + orr r1, r1, r2, lsl #32-align*8
796 + mov r2, r2, lsr #align*8
797 + orr r2, r2, r3, lsl #32-align*8
798 + mov r3, r3, lsr #align*8
799 + orr r3, r3, r4, lsl #32-align*8
800 + mov r4, r4, lsr #align*8
801 + orr r4, r4, r5, lsl #32-align*8
802 + mov r5, r5, lsr #align*8
803 + orr r5, r5, r6, lsl #32-align*8
804 + mov r6, r6, lsr #align*8
805 + orr r6, r6, r7, lsl #32-align*8
806 + mov r7, r7, lsr #align*8
807 + orr r7, r7, r8, lsl #32-align*8
808 + stmia D!, {r0, r1, r2, r3}
809 + stmia D!, {r4, r5, r6, r7}
810 + .endif
811 + .endif
812 +.endm
813 +
814 +.macro memcpy_leading_15bytes backwards, align
815 + movs DAT1, DAT2, lsl #31
816 + sub N, N, DAT2
817 + .if backwards
818 + ldrmib DAT0, [S, #-1]!
819 + ldrcsh DAT1, [S, #-2]!
820 + strmib DAT0, [D, #-1]!
821 + strcsh DAT1, [D, #-2]!
822 + .else
823 + ldrmib DAT0, [S], #1
824 + ldrcsh DAT1, [S], #2
825 + strmib DAT0, [D], #1
826 + strcsh DAT1, [D], #2
827 + .endif
828 + movs DAT1, DAT2, lsl #29
829 + .if backwards
830 + ldrmi DAT0, [S, #-4]!
831 + .if align == 0
832 + ldmcsdb S!, {DAT1, DAT2}
833 + .else
834 + ldrcs DAT2, [S, #-4]!
835 + ldrcs DAT1, [S, #-4]!
836 + .endif
837 + strmi DAT0, [D, #-4]!
838 + stmcsdb D!, {DAT1, DAT2}
839 + .else
840 + ldrmi DAT0, [S], #4
841 + .if align == 0
842 + ldmcsia S!, {DAT1, DAT2}
843 + .else
844 + ldrcs DAT1, [S], #4
845 + ldrcs DAT2, [S], #4
846 + .endif
847 + strmi DAT0, [D], #4
848 + stmcsia D!, {DAT1, DAT2}
849 + .endif
850 +.endm
851 +
852 +.macro memcpy_trailing_15bytes backwards, align
853 + movs N, N, lsl #29
854 + .if backwards
855 + .if align == 0
856 + ldmcsdb S!, {DAT0, DAT1}
857 + .else
858 + ldrcs DAT1, [S, #-4]!
859 + ldrcs DAT0, [S, #-4]!
860 + .endif
861 + ldrmi DAT2, [S, #-4]!
862 + stmcsdb D!, {DAT0, DAT1}
863 + strmi DAT2, [D, #-4]!
864 + .else
865 + .if align == 0
866 + ldmcsia S!, {DAT0, DAT1}
867 + .else
868 + ldrcs DAT0, [S], #4
869 + ldrcs DAT1, [S], #4
870 + .endif
871 + ldrmi DAT2, [S], #4
872 + stmcsia D!, {DAT0, DAT1}
873 + strmi DAT2, [D], #4
874 + .endif
875 + movs N, N, lsl #2
876 + .if backwards
877 + ldrcsh DAT0, [S, #-2]!
878 + ldrmib DAT1, [S, #-1]
879 + strcsh DAT0, [D, #-2]!
880 + strmib DAT1, [D, #-1]
881 + .else
882 + ldrcsh DAT0, [S], #2
883 + ldrmib DAT1, [S]
884 + strcsh DAT0, [D], #2
885 + strmib DAT1, [D]
886 + .endif
887 +.endm
888 +
889 +.macro memcpy_long_inner_loop backwards, align
890 + .if align != 0
891 + .if backwards
892 + ldr DAT0, [S, #-align]!
893 + .else
894 + ldr LAST, [S, #-align]!
895 + .endif
896 + .endif
897 +110:
898 + .if align == 0
899 + .if backwards
900 + ldmdb S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
901 + pld [S, OFF]
902 + stmdb D!, {DAT4, DAT5, DAT6, LAST}
903 + stmdb D!, {DAT0, DAT1, DAT2, DAT3}
904 + .else
905 + ldmia S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
906 + pld [S, OFF]
907 + stmia D!, {DAT0, DAT1, DAT2, DAT3}
908 + stmia D!, {DAT4, DAT5, DAT6, LAST}
909 + .endif
910 + .else
911 + unaligned_words backwards, align, 1, 8, DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, DAT7, LAST
912 + .endif
913 + subs N, N, #32
914 + bhs 110b
915 + /* Just before the final (prefetch_distance+1) 32-byte blocks, deal with final preloads */
916 + preload_trailing backwards, S, N, OFF
917 + add N, N, #(prefetch_distance+2)*32 - 32
918 +120:
919 + .if align == 0
920 + .if backwards
921 + ldmdb S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
922 + stmdb D!, {DAT4, DAT5, DAT6, LAST}
923 + stmdb D!, {DAT0, DAT1, DAT2, DAT3}
924 + .else
925 + ldmia S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
926 + stmia D!, {DAT0, DAT1, DAT2, DAT3}
927 + stmia D!, {DAT4, DAT5, DAT6, LAST}
928 + .endif
929 + .else
930 + unaligned_words backwards, align, 0, 8, DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, DAT7, LAST
931 + .endif
932 + subs N, N, #32
933 + bhs 120b
934 + tst N, #16
935 + .if align == 0
936 + .if backwards
937 + ldmnedb S!, {DAT0, DAT1, DAT2, LAST}
938 + stmnedb D!, {DAT0, DAT1, DAT2, LAST}
939 + .else
940 + ldmneia S!, {DAT0, DAT1, DAT2, LAST}
941 + stmneia D!, {DAT0, DAT1, DAT2, LAST}
942 + .endif
943 + .else
944 + beq 130f
945 + unaligned_words backwards, align, 0, 4, DAT0, DAT1, DAT2, DAT3, LAST
946 +130:
947 + .endif
948 + /* Trailing words and bytes */
949 + tst N, #15
950 + beq 199f
951 + .if align != 0
952 + add S, S, #align
953 + .endif
954 + memcpy_trailing_15bytes backwards, align
955 +199:
956 + pop {DAT3, DAT4, DAT5, DAT6, DAT7}
957 + pop {D, DAT1, DAT2, pc}
958 +.endm
959 +
960 +.macro memcpy_medium_inner_loop backwards, align
961 +120:
962 + .if backwards
963 + .if align == 0
964 + ldmdb S!, {DAT0, DAT1, DAT2, LAST}
965 + .else
966 + ldr LAST, [S, #-4]!
967 + ldr DAT2, [S, #-4]!
968 + ldr DAT1, [S, #-4]!
969 + ldr DAT0, [S, #-4]!
970 + .endif
971 + stmdb D!, {DAT0, DAT1, DAT2, LAST}
972 + .else
973 + .if align == 0
974 + ldmia S!, {DAT0, DAT1, DAT2, LAST}
975 + .else
976 + ldr DAT0, [S], #4
977 + ldr DAT1, [S], #4
978 + ldr DAT2, [S], #4
979 + ldr LAST, [S], #4
980 + .endif
981 + stmia D!, {DAT0, DAT1, DAT2, LAST}
982 + .endif
983 + subs N, N, #16
984 + bhs 120b
985 + /* Trailing words and bytes */
986 + tst N, #15
987 + beq 199f
988 + memcpy_trailing_15bytes backwards, align
989 +199:
990 + pop {D, DAT1, DAT2, pc}
991 +.endm
992 +
993 +.macro memcpy_short_inner_loop backwards, align
994 + tst N, #16
995 + .if backwards
996 + .if align == 0
997 + ldmnedb S!, {DAT0, DAT1, DAT2, LAST}
998 + .else
999 + ldrne LAST, [S, #-4]!
1000 + ldrne DAT2, [S, #-4]!
1001 + ldrne DAT1, [S, #-4]!
1002 + ldrne DAT0, [S, #-4]!
1003 + .endif
1004 + stmnedb D!, {DAT0, DAT1, DAT2, LAST}
1005 + .else
1006 + .if align == 0
1007 + ldmneia S!, {DAT0, DAT1, DAT2, LAST}
1008 + .else
1009 + ldrne DAT0, [S], #4
1010 + ldrne DAT1, [S], #4
1011 + ldrne DAT2, [S], #4
1012 + ldrne LAST, [S], #4
1013 + .endif
1014 + stmneia D!, {DAT0, DAT1, DAT2, LAST}
1015 + .endif
1016 + memcpy_trailing_15bytes backwards, align
1017 +199:
1018 + pop {D, DAT1, DAT2, pc}
1019 +.endm
1020 +
1021 +.macro memcpy backwards
1022 + D .req a1
1023 + S .req a2
1024 + N .req a3
1025 + DAT0 .req a4
1026 + DAT1 .req v1
1027 + DAT2 .req v2
1028 + DAT3 .req v3
1029 + DAT4 .req v4
1030 + DAT5 .req v5
1031 + DAT6 .req v6
1032 + DAT7 .req sl
1033 + LAST .req ip
1034 + OFF .req lr
1035 +
1036 + .cfi_startproc
1037 +
1038 + push {D, DAT1, DAT2, lr}
1039 +
1040 + .cfi_def_cfa_offset 16
1041 + .cfi_rel_offset D, 0
1042 + .cfi_undefined S
1043 + .cfi_undefined N
1044 + .cfi_undefined DAT0
1045 + .cfi_rel_offset DAT1, 4
1046 + .cfi_rel_offset DAT2, 8
1047 + .cfi_undefined LAST
1048 + .cfi_rel_offset lr, 12
1049 +
1050 + .if backwards
1051 + add D, D, N
1052 + add S, S, N
1053 + .endif
1054 +
1055 + /* See if we're guaranteed to have at least one 16-byte aligned 16-byte write */
1056 + cmp N, #31
1057 + blo 170f
1058 + /* To preload ahead as we go, we need at least (prefetch_distance+2) 32-byte blocks */
1059 + cmp N, #(prefetch_distance+3)*32 - 1
1060 + blo 160f
1061 +
1062 + /* Long case */
1063 + push {DAT3, DAT4, DAT5, DAT6, DAT7}
1064 +
1065 + .cfi_def_cfa_offset 36
1066 + .cfi_rel_offset D, 20
1067 + .cfi_rel_offset DAT1, 24
1068 + .cfi_rel_offset DAT2, 28
1069 + .cfi_rel_offset DAT3, 0
1070 + .cfi_rel_offset DAT4, 4
1071 + .cfi_rel_offset DAT5, 8
1072 + .cfi_rel_offset DAT6, 12
1073 + .cfi_rel_offset DAT7, 16
1074 + .cfi_rel_offset lr, 32
1075 +
1076 + /* Adjust N so that the decrement instruction can also test for
1077 + * inner loop termination. We want it to stop when there are
1078 + * (prefetch_distance+1) complete blocks to go. */
1079 + sub N, N, #(prefetch_distance+2)*32
1080 + preload_leading_step1 backwards, DAT0, S
1081 + .if backwards
1082 + /* Bug in GAS: it accepts, but mis-assembles the instruction
1083 + * ands DAT2, D, #60, 2
1084 + * which sets DAT2 to the number of leading bytes until destination is aligned and also clears C (sets borrow)
1085 + */
1086 + .word 0xE210513C
1087 + beq 154f
1088 + .else
1089 + ands DAT2, D, #15
1090 + beq 154f
1091 + rsb DAT2, DAT2, #16 /* number of leading bytes until destination aligned */
1092 + .endif
1093 + preload_leading_step2 backwards, DAT0, S, DAT2, OFF
1094 + memcpy_leading_15bytes backwards, 1
1095 +154: /* Destination now 16-byte aligned; we have at least one prefetch as well as at least one 16-byte output block */
1096 + /* Prefetch offset is best selected such that it lies in the first 8 of each 32 bytes - but it's just as easy to aim for the first one */
1097 + .if backwards
1098 + rsb OFF, S, #3
1099 + and OFF, OFF, #28
1100 + sub OFF, OFF, #32*(prefetch_distance+1)
1101 + .else
1102 + and OFF, S, #28
1103 + rsb OFF, OFF, #32*prefetch_distance
1104 + .endif
1105 + movs DAT0, S, lsl #31
1106 + bhi 157f
1107 + bcs 156f
1108 + bmi 155f
1109 + memcpy_long_inner_loop backwards, 0
1110 +155: memcpy_long_inner_loop backwards, 1
1111 +156: memcpy_long_inner_loop backwards, 2
1112 +157: memcpy_long_inner_loop backwards, 3
1113 +
1114 + .cfi_def_cfa_offset 16
1115 + .cfi_rel_offset D, 0
1116 + .cfi_rel_offset DAT1, 4
1117 + .cfi_rel_offset DAT2, 8
1118 + .cfi_same_value DAT3
1119 + .cfi_same_value DAT4
1120 + .cfi_same_value DAT5
1121 + .cfi_same_value DAT6
1122 + .cfi_same_value DAT7
1123 + .cfi_rel_offset lr, 12
1124 +
1125 +160: /* Medium case */
1126 + preload_all backwards, 0, 0, S, N, DAT2, OFF
1127 + sub N, N, #16 /* simplifies inner loop termination */
1128 + .if backwards
1129 + ands DAT2, D, #15
1130 + beq 164f
1131 + .else
1132 + ands DAT2, D, #15
1133 + beq 164f
1134 + rsb DAT2, DAT2, #16
1135 + .endif
1136 + memcpy_leading_15bytes backwards, align
1137 +164: /* Destination now 16-byte aligned; we have at least one 16-byte output block */
1138 + tst S, #3
1139 + bne 140f
1140 + memcpy_medium_inner_loop backwards, 0
1141 +140: memcpy_medium_inner_loop backwards, 1
1142 +
1143 +170: /* Short case, less than 31 bytes, so no guarantee of at least one 16-byte block */
1144 + teq N, #0
1145 + beq 199f
1146 + preload_all backwards, 1, 0, S, N, DAT2, LAST
1147 + tst D, #3
1148 + beq 174f
1149 +172: subs N, N, #1
1150 + blo 199f
1151 + .if backwards
1152 + ldrb DAT0, [S, #-1]!
1153 + strb DAT0, [D, #-1]!
1154 + .else
1155 + ldrb DAT0, [S], #1
1156 + strb DAT0, [D], #1
1157 + .endif
1158 + tst D, #3
1159 + bne 172b
1160 +174: /* Destination now 4-byte aligned; we have 0 or more output bytes to go */
1161 + tst S, #3
1162 + bne 140f
1163 + memcpy_short_inner_loop backwards, 0
1164 +140: memcpy_short_inner_loop backwards, 1
1165 +
1166 + .cfi_endproc
1167 +
1168 + .unreq D
1169 + .unreq S
1170 + .unreq N
1171 + .unreq DAT0
1172 + .unreq DAT1
1173 + .unreq DAT2
1174 + .unreq DAT3
1175 + .unreq DAT4
1176 + .unreq DAT5
1177 + .unreq DAT6
1178 + .unreq DAT7
1179 + .unreq LAST
1180 + .unreq OFF
1181 +.endm
1182 --- /dev/null
1183 +++ b/arch/arm/lib/memmove_rpi.S
1184 @@ -0,0 +1,61 @@
1185 +/*
1186 +Copyright (c) 2013, Raspberry Pi Foundation
1187 +Copyright (c) 2013, RISC OS Open Ltd
1188 +All rights reserved.
1189 +
1190 +Redistribution and use in source and binary forms, with or without
1191 +modification, are permitted provided that the following conditions are met:
1192 + * Redistributions of source code must retain the above copyright
1193 + notice, this list of conditions and the following disclaimer.
1194 + * Redistributions in binary form must reproduce the above copyright
1195 + notice, this list of conditions and the following disclaimer in the
1196 + documentation and/or other materials provided with the distribution.
1197 + * Neither the name of the copyright holder nor the
1198 + names of its contributors may be used to endorse or promote products
1199 + derived from this software without specific prior written permission.
1200 +
1201 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
1202 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
1203 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
1204 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
1205 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
1206 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
1207 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
1208 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
1209 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
1210 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1211 +*/
1212 +
1213 +#include <linux/linkage.h>
1214 +#include "arm-mem.h"
1215 +#include "memcpymove.h"
1216 +
1217 +/* Prevent the stack from becoming executable */
1218 +#if defined(__linux__) && defined(__ELF__)
1219 +.section .note.GNU-stack,"",%progbits
1220 +#endif
1221 +
1222 + .text
1223 + .arch armv6
1224 + .object_arch armv4
1225 + .arm
1226 + .altmacro
1227 + .p2align 2
1228 +
1229 +/*
1230 + * void *memmove(void *s1, const void *s2, size_t n);
1231 + * On entry:
1232 + * a1 = pointer to destination
1233 + * a2 = pointer to source
1234 + * a3 = number of bytes to copy
1235 + * On exit:
1236 + * a1 preserved
1237 + */
1238 +
1239 +.set prefetch_distance, 3
1240 +
1241 +ENTRY(memmove)
1242 + cmp a2, a1
1243 + bpl memcpy /* pl works even over -1 - 0 and 0x7fffffff - 0x80000000 boundaries */
1244 + memcpy 1
1245 +ENDPROC(memmove)
1246 --- /dev/null
1247 +++ b/arch/arm/lib/memset_rpi.S
1248 @@ -0,0 +1,128 @@
1249 +/*
1250 +Copyright (c) 2013, Raspberry Pi Foundation
1251 +Copyright (c) 2013, RISC OS Open Ltd
1252 +All rights reserved.
1253 +
1254 +Redistribution and use in source and binary forms, with or without
1255 +modification, are permitted provided that the following conditions are met:
1256 + * Redistributions of source code must retain the above copyright
1257 + notice, this list of conditions and the following disclaimer.
1258 + * Redistributions in binary form must reproduce the above copyright
1259 + notice, this list of conditions and the following disclaimer in the
1260 + documentation and/or other materials provided with the distribution.
1261 + * Neither the name of the copyright holder nor the
1262 + names of its contributors may be used to endorse or promote products
1263 + derived from this software without specific prior written permission.
1264 +
1265 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
1266 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
1267 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
1268 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
1269 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
1270 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
1271 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
1272 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
1273 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
1274 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1275 +*/
1276 +
1277 +#include <linux/linkage.h>
1278 +#include "arm-mem.h"
1279 +
1280 +/* Prevent the stack from becoming executable */
1281 +#if defined(__linux__) && defined(__ELF__)
1282 +.section .note.GNU-stack,"",%progbits
1283 +#endif
1284 +
1285 + .text
1286 + .arch armv6
1287 + .object_arch armv4
1288 + .arm
1289 + .altmacro
1290 + .p2align 2
1291 +
1292 +/*
1293 + * void *memset(void *s, int c, size_t n);
1294 + * On entry:
1295 + * a1 = pointer to buffer to fill
1296 + * a2 = byte pattern to fill with (caller-narrowed)
1297 + * a3 = number of bytes to fill
1298 + * On exit:
1299 + * a1 preserved
1300 + */
1301 +ENTRY(mmioset)
1302 +ENTRY(memset)
1303 +ENTRY(__memset32)
1304 +ENTRY(__memset64)
1305 +
1306 + S .req a1
1307 + DAT0 .req a2
1308 + N .req a3
1309 + DAT1 .req a4
1310 + DAT2 .req ip
1311 + DAT3 .req lr
1312 +
1313 + orr DAT0, DAT0, DAT0, lsl #8
1314 + push {S, lr}
1315 + orr DAT0, DAT0, DAT0, lsl #16
1316 + mov DAT1, DAT0
1317 +
1318 + /* See if we're guaranteed to have at least one 16-byte aligned 16-byte write */
1319 + cmp N, #31
1320 + blo 170f
1321 +
1322 +161: sub N, N, #16 /* simplifies inner loop termination */
1323 + /* Leading words and bytes */
1324 + tst S, #15
1325 + beq 164f
1326 + rsb DAT3, S, #0 /* bits 0-3 = number of leading bytes until aligned */
1327 + movs DAT2, DAT3, lsl #31
1328 + submi N, N, #1
1329 + strmib DAT0, [S], #1
1330 + subcs N, N, #2
1331 + strcsh DAT0, [S], #2
1332 + movs DAT2, DAT3, lsl #29
1333 + submi N, N, #4
1334 + strmi DAT0, [S], #4
1335 + subcs N, N, #8
1336 + stmcsia S!, {DAT0, DAT1}
1337 +164: /* Delayed set up of DAT2 and DAT3 so we could use them as scratch registers above */
1338 + mov DAT2, DAT0
1339 + mov DAT3, DAT0
1340 + /* Now the inner loop of 16-byte stores */
1341 +165: stmia S!, {DAT0, DAT1, DAT2, DAT3}
1342 + subs N, N, #16
1343 + bhs 165b
1344 +166: /* Trailing words and bytes */
1345 + movs N, N, lsl #29
1346 + stmcsia S!, {DAT0, DAT1}
1347 + strmi DAT0, [S], #4
1348 + movs N, N, lsl #2
1349 + strcsh DAT0, [S], #2
1350 + strmib DAT0, [S]
1351 +199: pop {S, pc}
1352 +
1353 +170: /* Short case */
1354 + mov DAT2, DAT0
1355 + mov DAT3, DAT0
1356 + tst S, #3
1357 + beq 174f
1358 +172: subs N, N, #1
1359 + blo 199b
1360 + strb DAT0, [S], #1
1361 + tst S, #3
1362 + bne 172b
1363 +174: tst N, #16
1364 + stmneia S!, {DAT0, DAT1, DAT2, DAT3}
1365 + b 166b
1366 +
1367 + .unreq S
1368 + .unreq DAT0
1369 + .unreq N
1370 + .unreq DAT1
1371 + .unreq DAT2
1372 + .unreq DAT3
1373 +ENDPROC(__memset64)
1374 +ENDPROC(__memset32)
1375 +ENDPROC(memset)
1376 +ENDPROC(mmioset)
1377 --- a/arch/arm/lib/uaccess_with_memcpy.c
1378 +++ b/arch/arm/lib/uaccess_with_memcpy.c
1379 @@ -22,6 +22,14 @@
1380 #include <asm/current.h>
1381 #include <asm/page.h>
1382
1383 +#ifndef COPY_FROM_USER_THRESHOLD
1384 +#define COPY_FROM_USER_THRESHOLD 64
1385 +#endif
1386 +
1387 +#ifndef COPY_TO_USER_THRESHOLD
1388 +#define COPY_TO_USER_THRESHOLD 64
1389 +#endif
1390 +
1391 static int
1392 pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
1393 {
1394 @@ -84,7 +92,44 @@ pin_page_for_write(const void __user *_a
1395 return 1;
1396 }
1397
1398 -static unsigned long noinline
1399 +static int
1400 +pin_page_for_read(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
1401 +{
1402 + unsigned long addr = (unsigned long)_addr;
1403 + pgd_t *pgd;
1404 + pmd_t *pmd;
1405 + pte_t *pte;
1406 + pud_t *pud;
1407 + spinlock_t *ptl;
1408 +
1409 + pgd = pgd_offset(current->mm, addr);
1410 + if (unlikely(pgd_none(*pgd) || pgd_bad(*pgd)))
1411 + {
1412 + return 0;
1413 + }
1414 + pud = pud_offset(pgd, addr);
1415 + if (unlikely(pud_none(*pud) || pud_bad(*pud)))
1416 + {
1417 + return 0;
1418 + }
1419 +
1420 + pmd = pmd_offset(pud, addr);
1421 + if (unlikely(pmd_none(*pmd) || pmd_bad(*pmd)))
1422 + return 0;
1423 +
1424 + pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl);
1425 + if (unlikely(!pte_present(*pte) || !pte_young(*pte))) {
1426 + pte_unmap_unlock(pte, ptl);
1427 + return 0;
1428 + }
1429 +
1430 + *ptep = pte;
1431 + *ptlp = ptl;
1432 +
1433 + return 1;
1434 +}
1435 +
1436 +unsigned long noinline
1437 __copy_to_user_memcpy(void __user *to, const void *from, unsigned long n)
1438 {
1439 unsigned long ua_flags;
1440 @@ -137,6 +182,57 @@ out:
1441 return n;
1442 }
1443
1444 +unsigned long noinline
1445 +__copy_from_user_memcpy(void *to, const void __user *from, unsigned long n)
1446 +{
1447 + unsigned long ua_flags;
1448 + int atomic;
1449 +
1450 + if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
1451 + memcpy(to, (const void *)from, n);
1452 + return 0;
1453 + }
1454 +
1455 + /* the mmap semaphore is taken only if not in an atomic context */
1456 + atomic = in_atomic();
1457 +
1458 + if (!atomic)
1459 + down_read(&current->mm->mmap_sem);
1460 + while (n) {
1461 + pte_t *pte;
1462 + spinlock_t *ptl;
1463 + int tocopy;
1464 +
1465 + while (!pin_page_for_read(from, &pte, &ptl)) {
1466 + char temp;
1467 + if (!atomic)
1468 + up_read(&current->mm->mmap_sem);
1469 + if (__get_user(temp, (char __user *)from))
1470 + goto out;
1471 + if (!atomic)
1472 + down_read(&current->mm->mmap_sem);
1473 + }
1474 +
1475 + tocopy = (~(unsigned long)from & ~PAGE_MASK) + 1;
1476 + if (tocopy > n)
1477 + tocopy = n;
1478 +
1479 + ua_flags = uaccess_save_and_enable();
1480 + memcpy(to, (const void *)from, tocopy);
1481 + uaccess_restore(ua_flags);
1482 + to += tocopy;
1483 + from += tocopy;
1484 + n -= tocopy;
1485 +
1486 + pte_unmap_unlock(pte, ptl);
1487 + }
1488 + if (!atomic)
1489 + up_read(&current->mm->mmap_sem);
1490 +
1491 +out:
1492 + return n;
1493 +}
1494 +
1495 unsigned long
1496 arm_copy_to_user(void __user *to, const void *from, unsigned long n)
1497 {
1498 @@ -147,7 +243,7 @@ arm_copy_to_user(void __user *to, const
1499 * With frame pointer disabled, tail call optimization kicks in
1500 * as well making this test almost invisible.
1501 */
1502 - if (n < 64) {
1503 + if (n < COPY_TO_USER_THRESHOLD) {
1504 unsigned long ua_flags = uaccess_save_and_enable();
1505 n = __copy_to_user_std(to, from, n);
1506 uaccess_restore(ua_flags);
1507 @@ -157,6 +253,26 @@ arm_copy_to_user(void __user *to, const
1508 }
1509 return n;
1510 }
1511 +
1512 +unsigned long __must_check
1513 +arm_copy_from_user(void *to, const void __user *from, unsigned long n)
1514 +{
1515 + /*
1516 + * This test is stubbed out of the main function above to keep
1517 + * the overhead for small copies low by avoiding a large
1518 + * register dump on the stack just to reload them right away.
1519 + * With frame pointer disabled, tail call optimization kicks in
1520 + * as well making this test almost invisible.
1521 + */
1522 + if (n < COPY_TO_USER_THRESHOLD) {
1523 + unsigned long ua_flags = uaccess_save_and_enable();
1524 + n = __copy_from_user_std(to, from, n);
1525 + uaccess_restore(ua_flags);
1526 + } else {
1527 + n = __copy_from_user_memcpy(to, from, n);
1528 + }
1529 + return n;
1530 +}
1531
1532 static unsigned long noinline
1533 __clear_user_memset(void __user *addr, unsigned long n)
1534 --- a/arch/arm/mach-bcm/Kconfig
1535 +++ b/arch/arm/mach-bcm/Kconfig
1536 @@ -187,6 +187,13 @@ config ARCH_BCM_53573
1537 The base chip is BCM53573 and there are some packaging modifications
1538 like BCM47189 and BCM47452.
1539
1540 +config BCM2835_FAST_MEMCPY
1541 + bool "Enable optimized __copy_to_user and __copy_from_user"
1542 + depends on ARCH_BCM2835 && ARCH_MULTI_V6
1543 + default y
1544 + help
1545 + Optimized versions of __copy_to_user and __copy_from_user for Pi1.
1546 +
1547 config ARCH_BCM_63XX
1548 bool "Broadcom BCM63xx DSL SoC"
1549 depends on ARCH_MULTI_V7
Staging tree of Koen Vandeputte