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


a7ce64ba3811bbe51555648774fa24ac82fec4aa
[openwrt/staging/nbd.git] /
1 From ebd731dd71ec9728a5a87ec1cd695be15828c32c 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 performance
5
6 Provide a __copy_from_user that uses memcpy. On BCM2708, use
7 optimised memcpy/memmove/memcmp/memset implementations.
8
9 arch/arm: Add mmiocpy/set aliases for memcpy/set
10
11 See: https://github.com/raspberrypi/linux/issues/1082
12
13 copy_from_user: CPU_SW_DOMAIN_PAN compatibility
14
15 The downstream copy_from_user acceleration must also play nice with
16 CONFIG_CPU_SW_DOMAIN_PAN.
17
18 See: https://github.com/raspberrypi/linux/issues/1381
19
20 Signed-off-by: Phil Elwell <phil@raspberrypi.org>
21 ---
22 arch/arm/include/asm/string.h | 5 +
23 arch/arm/include/asm/uaccess.h | 3 +
24 arch/arm/lib/Makefile | 15 +-
25 arch/arm/lib/arm-mem.h | 159 ++++++++++++
26 arch/arm/lib/copy_from_user.S | 4 +-
27 arch/arm/lib/exports_rpi.c | 37 +++
28 arch/arm/lib/memcmp_rpi.S | 285 +++++++++++++++++++++
29 arch/arm/lib/memcpy_rpi.S | 61 +++++
30 arch/arm/lib/memcpymove.h | 506 +++++++++++++++++++++++++++++++++++++
31 arch/arm/lib/memmove_rpi.S | 61 +++++
32 arch/arm/lib/memset_rpi.S | 123 +++++++++
33 arch/arm/lib/uaccess_with_memcpy.c | 120 ++++++++-
34 arch/arm/mach-bcm/Kconfig | 7 +
35 13 files changed, 1380 insertions(+), 6 deletions(-)
36 create mode 100644 arch/arm/lib/arm-mem.h
37 create mode 100644 arch/arm/lib/exports_rpi.c
38 create mode 100644 arch/arm/lib/memcmp_rpi.S
39 create mode 100644 arch/arm/lib/memcpy_rpi.S
40 create mode 100644 arch/arm/lib/memcpymove.h
41 create mode 100644 arch/arm/lib/memmove_rpi.S
42 create mode 100644 arch/arm/lib/memset_rpi.S
43
44 --- a/arch/arm/include/asm/string.h
45 +++ b/arch/arm/include/asm/string.h
46 @@ -24,6 +24,11 @@ extern void * memchr(const void *, int,
47 #define __HAVE_ARCH_MEMSET
48 extern void * memset(void *, int, __kernel_size_t);
49
50 +#ifdef CONFIG_BCM2835_FAST_MEMCPY
51 +#define __HAVE_ARCH_MEMCMP
52 +extern int memcmp(const void *, const void *, size_t);
53 +#endif
54 +
55 extern void __memzero(void *ptr, __kernel_size_t n);
56
57 #define memset(p,v,n) \
58 --- a/arch/arm/include/asm/uaccess.h
59 +++ b/arch/arm/include/asm/uaccess.h
60 @@ -477,6 +477,9 @@ do { \
61 extern unsigned long __must_check
62 arm_copy_from_user(void *to, const void __user *from, unsigned long n);
63
64 +extern unsigned long __must_check
65 +__copy_from_user_std(void *to, const void __user *from, unsigned long n);
66 +
67 static inline unsigned long __must_check
68 __copy_from_user(void *to, const void __user *from, unsigned long n)
69 {
70 --- a/arch/arm/lib/Makefile
71 +++ b/arch/arm/lib/Makefile
72 @@ -6,9 +6,8 @@
73
74 lib-y := backtrace.o changebit.o csumipv6.o csumpartial.o \
75 csumpartialcopy.o csumpartialcopyuser.o clearbit.o \
76 - delay.o delay-loop.o findbit.o memchr.o memcpy.o \
77 - memmove.o memset.o memzero.o setbit.o \
78 - strchr.o strrchr.o \
79 + delay.o delay-loop.o findbit.o memchr.o memzero.o \
80 + setbit.o strchr.o strrchr.o \
81 testchangebit.o testclearbit.o testsetbit.o \
82 ashldi3.o ashrdi3.o lshrdi3.o muldi3.o \
83 ucmpdi2.o lib1funcs.o div64.o \
84 @@ -18,6 +17,16 @@ lib-y := backtrace.o changebit.o csumip
85 mmu-y := clear_user.o copy_page.o getuser.o putuser.o \
86 copy_from_user.o copy_to_user.o
87
88 +# Choose optimised implementations for Raspberry Pi
89 +ifeq ($(CONFIG_BCM2835_FAST_MEMCPY),y)
90 + CFLAGS_uaccess_with_memcpy.o += -DCOPY_FROM_USER_THRESHOLD=1600
91 + CFLAGS_uaccess_with_memcpy.o += -DCOPY_TO_USER_THRESHOLD=672
92 + obj-$(CONFIG_MODULES) += exports_rpi.o
93 + lib-y += memcpy_rpi.o memmove_rpi.o memset_rpi.o memcmp_rpi.o
94 +else
95 + lib-y += memcpy.o memmove.o memset.o
96 +endif
97 +
98 # using lib_ here won't override already available weak symbols
99 obj-$(CONFIG_UACCESS_WITH_MEMCPY) += uaccess_with_memcpy.o
100
101 --- /dev/null
102 +++ b/arch/arm/lib/arm-mem.h
103 @@ -0,0 +1,159 @@
104 +/*
105 +Copyright (c) 2013, Raspberry Pi Foundation
106 +Copyright (c) 2013, RISC OS Open Ltd
107 +All rights reserved.
108 +
109 +Redistribution and use in source and binary forms, with or without
110 +modification, are permitted provided that the following conditions are met:
111 + * Redistributions of source code must retain the above copyright
112 + notice, this list of conditions and the following disclaimer.
113 + * Redistributions in binary form must reproduce the above copyright
114 + notice, this list of conditions and the following disclaimer in the
115 + documentation and/or other materials provided with the distribution.
116 + * Neither the name of the copyright holder nor the
117 + names of its contributors may be used to endorse or promote products
118 + derived from this software without specific prior written permission.
119 +
120 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
121 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
122 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
123 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
124 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
125 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
126 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
127 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
128 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
129 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
130 +*/
131 +
132 +.macro myfunc fname
133 + .func fname
134 + .global fname
135 +fname:
136 +.endm
137 +
138 +.macro preload_leading_step1 backwards, ptr, base
139 +/* If the destination is already 16-byte aligned, then we need to preload
140 + * between 0 and prefetch_distance (inclusive) cache lines ahead so there
141 + * are no gaps when the inner loop starts.
142 + */
143 + .if backwards
144 + sub ptr, base, #1
145 + bic ptr, ptr, #31
146 + .else
147 + bic ptr, base, #31
148 + .endif
149 + .set OFFSET, 0
150 + .rept prefetch_distance+1
151 + pld [ptr, #OFFSET]
152 + .if backwards
153 + .set OFFSET, OFFSET-32
154 + .else
155 + .set OFFSET, OFFSET+32
156 + .endif
157 + .endr
158 +.endm
159 +
160 +.macro preload_leading_step2 backwards, ptr, base, leading_bytes, tmp
161 +/* However, if the destination is not 16-byte aligned, we may need to
162 + * preload one more cache line than that. The question we need to ask is:
163 + * are the leading bytes more than the amount by which the source
164 + * pointer will be rounded down for preloading, and if so, by how many
165 + * cache lines?
166 + */
167 + .if backwards
168 +/* Here we compare against how many bytes we are into the
169 + * cache line, counting down from the highest such address.
170 + * Effectively, we want to calculate
171 + * leading_bytes = dst&15
172 + * cacheline_offset = 31-((src-leading_bytes-1)&31)
173 + * extra_needed = leading_bytes - cacheline_offset
174 + * and test if extra_needed is <= 0, or rearranging:
175 + * leading_bytes + (src-leading_bytes-1)&31 <= 31
176 + */
177 + mov tmp, base, lsl #32-5
178 + sbc tmp, tmp, leading_bytes, lsl #32-5
179 + adds tmp, tmp, leading_bytes, lsl #32-5
180 + bcc 61f
181 + pld [ptr, #-32*(prefetch_distance+1)]
182 + .else
183 +/* Effectively, we want to calculate
184 + * leading_bytes = (-dst)&15
185 + * cacheline_offset = (src+leading_bytes)&31
186 + * extra_needed = leading_bytes - cacheline_offset
187 + * and test if extra_needed is <= 0.
188 + */
189 + mov tmp, base, lsl #32-5
190 + add tmp, tmp, leading_bytes, lsl #32-5
191 + rsbs tmp, tmp, leading_bytes, lsl #32-5
192 + bls 61f
193 + pld [ptr, #32*(prefetch_distance+1)]
194 + .endif
195 +61:
196 +.endm
197 +
198 +.macro preload_trailing backwards, base, remain, tmp
199 + /* We need either 0, 1 or 2 extra preloads */
200 + .if backwards
201 + rsb tmp, base, #0
202 + mov tmp, tmp, lsl #32-5
203 + .else
204 + mov tmp, base, lsl #32-5
205 + .endif
206 + adds tmp, tmp, remain, lsl #32-5
207 + adceqs tmp, tmp, #0
208 + /* The instruction above has two effects: ensures Z is only
209 + * set if C was clear (so Z indicates that both shifted quantities
210 + * were 0), and clears C if Z was set (so C indicates that the sum
211 + * of the shifted quantities was greater and not equal to 32) */
212 + beq 82f
213 + .if backwards
214 + sub tmp, base, #1
215 + bic tmp, tmp, #31
216 + .else
217 + bic tmp, base, #31
218 + .endif
219 + bcc 81f
220 + .if backwards
221 + pld [tmp, #-32*(prefetch_distance+1)]
222 +81:
223 + pld [tmp, #-32*prefetch_distance]
224 + .else
225 + pld [tmp, #32*(prefetch_distance+2)]
226 +81:
227 + pld [tmp, #32*(prefetch_distance+1)]
228 + .endif
229 +82:
230 +.endm
231 +
232 +.macro preload_all backwards, narrow_case, shift, base, remain, tmp0, tmp1
233 + .if backwards
234 + sub tmp0, base, #1
235 + bic tmp0, tmp0, #31
236 + pld [tmp0]
237 + sub tmp1, base, remain, lsl #shift
238 + .else
239 + bic tmp0, base, #31
240 + pld [tmp0]
241 + add tmp1, base, remain, lsl #shift
242 + sub tmp1, tmp1, #1
243 + .endif
244 + bic tmp1, tmp1, #31
245 + cmp tmp1, tmp0
246 + beq 92f
247 + .if narrow_case
248 + /* In this case, all the data fits in either 1 or 2 cache lines */
249 + pld [tmp1]
250 + .else
251 +91:
252 + .if backwards
253 + sub tmp0, tmp0, #32
254 + .else
255 + add tmp0, tmp0, #32
256 + .endif
257 + cmp tmp0, tmp1
258 + pld [tmp0]
259 + bne 91b
260 + .endif
261 +92:
262 +.endm
263 --- a/arch/arm/lib/copy_from_user.S
264 +++ b/arch/arm/lib/copy_from_user.S
265 @@ -89,11 +89,13 @@
266
267 .text
268
269 -ENTRY(arm_copy_from_user)
270 +ENTRY(__copy_from_user_std)
271 +WEAK(arm_copy_from_user)
272
273 #include "copy_template.S"
274
275 ENDPROC(arm_copy_from_user)
276 +ENDPROC(__copy_from_user_std)
277
278 .pushsection .fixup,"ax"
279 .align 0
280 --- /dev/null
281 +++ b/arch/arm/lib/exports_rpi.c
282 @@ -0,0 +1,37 @@
283 +/**
284 + * Copyright (c) 2014, Raspberry Pi (Trading) Ltd.
285 + *
286 + * Redistribution and use in source and binary forms, with or without
287 + * modification, are permitted provided that the following conditions
288 + * are met:
289 + * 1. Redistributions of source code must retain the above copyright
290 + * notice, this list of conditions, and the following disclaimer,
291 + * without modification.
292 + * 2. Redistributions in binary form must reproduce the above copyright
293 + * notice, this list of conditions and the following disclaimer in the
294 + * documentation and/or other materials provided with the distribution.
295 + * 3. The names of the above-listed copyright holders may not be used
296 + * to endorse or promote products derived from this software without
297 + * specific prior written permission.
298 + *
299 + * ALTERNATIVELY, this software may be distributed under the terms of the
300 + * GNU General Public License ("GPL") version 2, as published by the Free
301 + * Software Foundation.
302 + *
303 + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
304 + * IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
305 + * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
306 + * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
307 + * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
308 + * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
309 + * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
310 + * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
311 + * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
312 + * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
313 + * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
314 + */
315 +
316 +#include <linux/kernel.h>
317 +#include <linux/module.h>
318 +
319 +EXPORT_SYMBOL(memcmp);
320 --- /dev/null
321 +++ b/arch/arm/lib/memcmp_rpi.S
322 @@ -0,0 +1,285 @@
323 +/*
324 +Copyright (c) 2013, Raspberry Pi Foundation
325 +Copyright (c) 2013, RISC OS Open Ltd
326 +All rights reserved.
327 +
328 +Redistribution and use in source and binary forms, with or without
329 +modification, are permitted provided that the following conditions are met:
330 + * Redistributions of source code must retain the above copyright
331 + notice, this list of conditions and the following disclaimer.
332 + * Redistributions in binary form must reproduce the above copyright
333 + notice, this list of conditions and the following disclaimer in the
334 + documentation and/or other materials provided with the distribution.
335 + * Neither the name of the copyright holder nor the
336 + names of its contributors may be used to endorse or promote products
337 + derived from this software without specific prior written permission.
338 +
339 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
340 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
341 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
342 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
343 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
344 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
345 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
346 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
347 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
348 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
349 +*/
350 +
351 +#include <linux/linkage.h>
352 +#include "arm-mem.h"
353 +
354 +/* Prevent the stack from becoming executable */
355 +#if defined(__linux__) && defined(__ELF__)
356 +.section .note.GNU-stack,"",%progbits
357 +#endif
358 +
359 + .text
360 + .arch armv6
361 + .object_arch armv4
362 + .arm
363 + .altmacro
364 + .p2align 2
365 +
366 +.macro memcmp_process_head unaligned
367 + .if unaligned
368 + ldr DAT0, [S_1], #4
369 + ldr DAT1, [S_1], #4
370 + ldr DAT2, [S_1], #4
371 + ldr DAT3, [S_1], #4
372 + .else
373 + ldmia S_1!, {DAT0, DAT1, DAT2, DAT3}
374 + .endif
375 + ldmia S_2!, {DAT4, DAT5, DAT6, DAT7}
376 +.endm
377 +
378 +.macro memcmp_process_tail
379 + cmp DAT0, DAT4
380 + cmpeq DAT1, DAT5
381 + cmpeq DAT2, DAT6
382 + cmpeq DAT3, DAT7
383 + bne 200f
384 +.endm
385 +
386 +.macro memcmp_leading_31bytes
387 + movs DAT0, OFF, lsl #31
388 + ldrmib DAT0, [S_1], #1
389 + ldrcsh DAT1, [S_1], #2
390 + ldrmib DAT4, [S_2], #1
391 + ldrcsh DAT5, [S_2], #2
392 + movpl DAT0, #0
393 + movcc DAT1, #0
394 + movpl DAT4, #0
395 + movcc DAT5, #0
396 + submi N, N, #1
397 + subcs N, N, #2
398 + cmp DAT0, DAT4
399 + cmpeq DAT1, DAT5
400 + bne 200f
401 + movs DAT0, OFF, lsl #29
402 + ldrmi DAT0, [S_1], #4
403 + ldrcs DAT1, [S_1], #4
404 + ldrcs DAT2, [S_1], #4
405 + ldrmi DAT4, [S_2], #4
406 + ldmcsia S_2!, {DAT5, DAT6}
407 + movpl DAT0, #0
408 + movcc DAT1, #0
409 + movcc DAT2, #0
410 + movpl DAT4, #0
411 + movcc DAT5, #0
412 + movcc DAT6, #0
413 + submi N, N, #4
414 + subcs N, N, #8
415 + cmp DAT0, DAT4
416 + cmpeq DAT1, DAT5
417 + cmpeq DAT2, DAT6
418 + bne 200f
419 + tst OFF, #16
420 + beq 105f
421 + memcmp_process_head 1
422 + sub N, N, #16
423 + memcmp_process_tail
424 +105:
425 +.endm
426 +
427 +.macro memcmp_trailing_15bytes unaligned
428 + movs N, N, lsl #29
429 + .if unaligned
430 + ldrcs DAT0, [S_1], #4
431 + ldrcs DAT1, [S_1], #4
432 + .else
433 + ldmcsia S_1!, {DAT0, DAT1}
434 + .endif
435 + ldrmi DAT2, [S_1], #4
436 + ldmcsia S_2!, {DAT4, DAT5}
437 + ldrmi DAT6, [S_2], #4
438 + movcc DAT0, #0
439 + movcc DAT1, #0
440 + movpl DAT2, #0
441 + movcc DAT4, #0
442 + movcc DAT5, #0
443 + movpl DAT6, #0
444 + cmp DAT0, DAT4
445 + cmpeq DAT1, DAT5
446 + cmpeq DAT2, DAT6
447 + bne 200f
448 + movs N, N, lsl #2
449 + ldrcsh DAT0, [S_1], #2
450 + ldrmib DAT1, [S_1]
451 + ldrcsh DAT4, [S_2], #2
452 + ldrmib DAT5, [S_2]
453 + movcc DAT0, #0
454 + movpl DAT1, #0
455 + movcc DAT4, #0
456 + movpl DAT5, #0
457 + cmp DAT0, DAT4
458 + cmpeq DAT1, DAT5
459 + bne 200f
460 +.endm
461 +
462 +.macro memcmp_long_inner_loop unaligned
463 +110:
464 + memcmp_process_head unaligned
465 + pld [S_2, #prefetch_distance*32 + 16]
466 + memcmp_process_tail
467 + memcmp_process_head unaligned
468 + pld [S_1, OFF]
469 + memcmp_process_tail
470 + subs N, N, #32
471 + bhs 110b
472 + /* Just before the final (prefetch_distance+1) 32-byte blocks,
473 + * deal with final preloads */
474 + preload_trailing 0, S_1, N, DAT0
475 + preload_trailing 0, S_2, N, DAT0
476 + add N, N, #(prefetch_distance+2)*32 - 16
477 +120:
478 + memcmp_process_head unaligned
479 + memcmp_process_tail
480 + subs N, N, #16
481 + bhs 120b
482 + /* Trailing words and bytes */
483 + tst N, #15
484 + beq 199f
485 + memcmp_trailing_15bytes unaligned
486 +199: /* Reached end without detecting a difference */
487 + mov a1, #0
488 + setend le
489 + pop {DAT1-DAT6, pc}
490 +.endm
491 +
492 +.macro memcmp_short_inner_loop unaligned
493 + subs N, N, #16 /* simplifies inner loop termination */
494 + blo 122f
495 +120:
496 + memcmp_process_head unaligned
497 + memcmp_process_tail
498 + subs N, N, #16
499 + bhs 120b
500 +122: /* Trailing words and bytes */
501 + tst N, #15
502 + beq 199f
503 + memcmp_trailing_15bytes unaligned
504 +199: /* Reached end without detecting a difference */
505 + mov a1, #0
506 + setend le
507 + pop {DAT1-DAT6, pc}
508 +.endm
509 +
510 +/*
511 + * int memcmp(const void *s1, const void *s2, size_t n);
512 + * On entry:
513 + * a1 = pointer to buffer 1
514 + * a2 = pointer to buffer 2
515 + * a3 = number of bytes to compare (as unsigned chars)
516 + * On exit:
517 + * a1 = >0/=0/<0 if s1 >/=/< s2
518 + */
519 +
520 +.set prefetch_distance, 2
521 +
522 +ENTRY(memcmp)
523 + S_1 .req a1
524 + S_2 .req a2
525 + N .req a3
526 + DAT0 .req a4
527 + DAT1 .req v1
528 + DAT2 .req v2
529 + DAT3 .req v3
530 + DAT4 .req v4
531 + DAT5 .req v5
532 + DAT6 .req v6
533 + DAT7 .req ip
534 + OFF .req lr
535 +
536 + push {DAT1-DAT6, lr}
537 + setend be /* lowest-addressed bytes are most significant */
538 +
539 + /* To preload ahead as we go, we need at least (prefetch_distance+2) 32-byte blocks */
540 + cmp N, #(prefetch_distance+3)*32 - 1
541 + blo 170f
542 +
543 + /* Long case */
544 + /* Adjust N so that the decrement instruction can also test for
545 + * inner loop termination. We want it to stop when there are
546 + * (prefetch_distance+1) complete blocks to go. */
547 + sub N, N, #(prefetch_distance+2)*32
548 + preload_leading_step1 0, DAT0, S_1
549 + preload_leading_step1 0, DAT1, S_2
550 + tst S_2, #31
551 + beq 154f
552 + rsb OFF, S_2, #0 /* no need to AND with 15 here */
553 + preload_leading_step2 0, DAT0, S_1, OFF, DAT2
554 + preload_leading_step2 0, DAT1, S_2, OFF, DAT2
555 + memcmp_leading_31bytes
556 +154: /* Second source now cacheline (32-byte) aligned; we have at
557 + * least one prefetch to go. */
558 + /* Prefetch offset is best selected such that it lies in the
559 + * first 8 of each 32 bytes - but it's just as easy to aim for
560 + * the first one */
561 + and OFF, S_1, #31
562 + rsb OFF, OFF, #32*prefetch_distance
563 + tst S_1, #3
564 + bne 140f
565 + memcmp_long_inner_loop 0
566 +140: memcmp_long_inner_loop 1
567 +
568 +170: /* Short case */
569 + teq N, #0
570 + beq 199f
571 + preload_all 0, 0, 0, S_1, N, DAT0, DAT1
572 + preload_all 0, 0, 0, S_2, N, DAT0, DAT1
573 + tst S_2, #3
574 + beq 174f
575 +172: subs N, N, #1
576 + blo 199f
577 + ldrb DAT0, [S_1], #1
578 + ldrb DAT4, [S_2], #1
579 + cmp DAT0, DAT4
580 + bne 200f
581 + tst S_2, #3
582 + bne 172b
583 +174: /* Second source now 4-byte aligned; we have 0 or more bytes to go */
584 + tst S_1, #3
585 + bne 140f
586 + memcmp_short_inner_loop 0
587 +140: memcmp_short_inner_loop 1
588 +
589 +200: /* Difference found: determine sign. */
590 + movhi a1, #1
591 + movlo a1, #-1
592 + setend le
593 + pop {DAT1-DAT6, pc}
594 +
595 + .unreq S_1
596 + .unreq S_2
597 + .unreq N
598 + .unreq DAT0
599 + .unreq DAT1
600 + .unreq DAT2
601 + .unreq DAT3
602 + .unreq DAT4
603 + .unreq DAT5
604 + .unreq DAT6
605 + .unreq DAT7
606 + .unreq OFF
607 +ENDPROC(memcmp)
608 --- /dev/null
609 +++ b/arch/arm/lib/memcpy_rpi.S
610 @@ -0,0 +1,61 @@
611 +/*
612 +Copyright (c) 2013, Raspberry Pi Foundation
613 +Copyright (c) 2013, RISC OS Open Ltd
614 +All rights reserved.
615 +
616 +Redistribution and use in source and binary forms, with or without
617 +modification, are permitted provided that the following conditions are met:
618 + * Redistributions of source code must retain the above copyright
619 + notice, this list of conditions and the following disclaimer.
620 + * Redistributions in binary form must reproduce the above copyright
621 + notice, this list of conditions and the following disclaimer in the
622 + documentation and/or other materials provided with the distribution.
623 + * Neither the name of the copyright holder nor the
624 + names of its contributors may be used to endorse or promote products
625 + derived from this software without specific prior written permission.
626 +
627 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
628 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
629 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
630 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
631 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
632 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
633 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
634 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
635 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
636 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
637 +*/
638 +
639 +#include <linux/linkage.h>
640 +#include "arm-mem.h"
641 +#include "memcpymove.h"
642 +
643 +/* Prevent the stack from becoming executable */
644 +#if defined(__linux__) && defined(__ELF__)
645 +.section .note.GNU-stack,"",%progbits
646 +#endif
647 +
648 + .text
649 + .arch armv6
650 + .object_arch armv4
651 + .arm
652 + .altmacro
653 + .p2align 2
654 +
655 +/*
656 + * void *memcpy(void * restrict s1, const void * restrict s2, size_t n);
657 + * On entry:
658 + * a1 = pointer to destination
659 + * a2 = pointer to source
660 + * a3 = number of bytes to copy
661 + * On exit:
662 + * a1 preserved
663 + */
664 +
665 +.set prefetch_distance, 3
666 +
667 +ENTRY(mmiocpy)
668 +ENTRY(memcpy)
669 + memcpy 0
670 +ENDPROC(memcpy)
671 +ENDPROC(mmiocpy)
672 --- /dev/null
673 +++ b/arch/arm/lib/memcpymove.h
674 @@ -0,0 +1,506 @@
675 +/*
676 +Copyright (c) 2013, Raspberry Pi Foundation
677 +Copyright (c) 2013, RISC OS Open Ltd
678 +All rights reserved.
679 +
680 +Redistribution and use in source and binary forms, with or without
681 +modification, are permitted provided that the following conditions are met:
682 + * Redistributions of source code must retain the above copyright
683 + notice, this list of conditions and the following disclaimer.
684 + * Redistributions in binary form must reproduce the above copyright
685 + notice, this list of conditions and the following disclaimer in the
686 + documentation and/or other materials provided with the distribution.
687 + * Neither the name of the copyright holder nor the
688 + names of its contributors may be used to endorse or promote products
689 + derived from this software without specific prior written permission.
690 +
691 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
692 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
693 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
694 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
695 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
696 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
697 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
698 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
699 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
700 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
701 +*/
702 +
703 +.macro unaligned_words backwards, align, use_pld, words, r0, r1, r2, r3, r4, r5, r6, r7, r8
704 + .if words == 1
705 + .if backwards
706 + mov r1, r0, lsl #32-align*8
707 + ldr r0, [S, #-4]!
708 + orr r1, r1, r0, lsr #align*8
709 + str r1, [D, #-4]!
710 + .else
711 + mov r0, r1, lsr #align*8
712 + ldr r1, [S, #4]!
713 + orr r0, r0, r1, lsl #32-align*8
714 + str r0, [D], #4
715 + .endif
716 + .elseif words == 2
717 + .if backwards
718 + ldr r1, [S, #-4]!
719 + mov r2, r0, lsl #32-align*8
720 + ldr r0, [S, #-4]!
721 + orr r2, r2, r1, lsr #align*8
722 + mov r1, r1, lsl #32-align*8
723 + orr r1, r1, r0, lsr #align*8
724 + stmdb D!, {r1, r2}
725 + .else
726 + ldr r1, [S, #4]!
727 + mov r0, r2, lsr #align*8
728 + ldr r2, [S, #4]!
729 + orr r0, r0, r1, lsl #32-align*8
730 + mov r1, r1, lsr #align*8
731 + orr r1, r1, r2, lsl #32-align*8
732 + stmia D!, {r0, r1}
733 + .endif
734 + .elseif words == 4
735 + .if backwards
736 + ldmdb S!, {r2, r3}
737 + mov r4, r0, lsl #32-align*8
738 + ldmdb S!, {r0, r1}
739 + orr r4, r4, r3, lsr #align*8
740 + mov r3, r3, lsl #32-align*8
741 + orr r3, r3, r2, lsr #align*8
742 + mov r2, r2, lsl #32-align*8
743 + orr r2, r2, r1, lsr #align*8
744 + mov r1, r1, lsl #32-align*8
745 + orr r1, r1, r0, lsr #align*8
746 + stmdb D!, {r1, r2, r3, r4}
747 + .else
748 + ldmib S!, {r1, r2}
749 + mov r0, r4, lsr #align*8
750 + ldmib S!, {r3, r4}
751 + orr r0, r0, r1, lsl #32-align*8
752 + mov r1, r1, lsr #align*8
753 + orr r1, r1, r2, lsl #32-align*8
754 + mov r2, r2, lsr #align*8
755 + orr r2, r2, r3, lsl #32-align*8
756 + mov r3, r3, lsr #align*8
757 + orr r3, r3, r4, lsl #32-align*8
758 + stmia D!, {r0, r1, r2, r3}
759 + .endif
760 + .elseif words == 8
761 + .if backwards
762 + ldmdb S!, {r4, r5, r6, r7}
763 + mov r8, r0, lsl #32-align*8
764 + ldmdb S!, {r0, r1, r2, r3}
765 + .if use_pld
766 + pld [S, OFF]
767 + .endif
768 + orr r8, r8, r7, lsr #align*8
769 + mov r7, r7, lsl #32-align*8
770 + orr r7, r7, r6, lsr #align*8
771 + mov r6, r6, lsl #32-align*8
772 + orr r6, r6, r5, lsr #align*8
773 + mov r5, r5, lsl #32-align*8
774 + orr r5, r5, r4, lsr #align*8
775 + mov r4, r4, lsl #32-align*8
776 + orr r4, r4, r3, lsr #align*8
777 + mov r3, r3, lsl #32-align*8
778 + orr r3, r3, r2, lsr #align*8
779 + mov r2, r2, lsl #32-align*8
780 + orr r2, r2, r1, lsr #align*8
781 + mov r1, r1, lsl #32-align*8
782 + orr r1, r1, r0, lsr #align*8
783 + stmdb D!, {r5, r6, r7, r8}
784 + stmdb D!, {r1, r2, r3, r4}
785 + .else
786 + ldmib S!, {r1, r2, r3, r4}
787 + mov r0, r8, lsr #align*8
788 + ldmib S!, {r5, r6, r7, r8}
789 + .if use_pld
790 + pld [S, OFF]
791 + .endif
792 + orr r0, r0, r1, lsl #32-align*8
793 + mov r1, r1, lsr #align*8
794 + orr r1, r1, r2, lsl #32-align*8
795 + mov r2, r2, lsr #align*8
796 + orr r2, r2, r3, lsl #32-align*8
797 + mov r3, r3, lsr #align*8
798 + orr r3, r3, r4, lsl #32-align*8
799 + mov r4, r4, lsr #align*8
800 + orr r4, r4, r5, lsl #32-align*8
801 + mov r5, r5, lsr #align*8
802 + orr r5, r5, r6, lsl #32-align*8
803 + mov r6, r6, lsr #align*8
804 + orr r6, r6, r7, lsl #32-align*8
805 + mov r7, r7, lsr #align*8
806 + orr r7, r7, r8, lsl #32-align*8
807 + stmia D!, {r0, r1, r2, r3}
808 + stmia D!, {r4, r5, r6, r7}
809 + .endif
810 + .endif
811 +.endm
812 +
813 +.macro memcpy_leading_15bytes backwards, align
814 + movs DAT1, DAT2, lsl #31
815 + sub N, N, DAT2
816 + .if backwards
817 + ldrmib DAT0, [S, #-1]!
818 + ldrcsh DAT1, [S, #-2]!
819 + strmib DAT0, [D, #-1]!
820 + strcsh DAT1, [D, #-2]!
821 + .else
822 + ldrmib DAT0, [S], #1
823 + ldrcsh DAT1, [S], #2
824 + strmib DAT0, [D], #1
825 + strcsh DAT1, [D], #2
826 + .endif
827 + movs DAT1, DAT2, lsl #29
828 + .if backwards
829 + ldrmi DAT0, [S, #-4]!
830 + .if align == 0
831 + ldmcsdb S!, {DAT1, DAT2}
832 + .else
833 + ldrcs DAT2, [S, #-4]!
834 + ldrcs DAT1, [S, #-4]!
835 + .endif
836 + strmi DAT0, [D, #-4]!
837 + stmcsdb D!, {DAT1, DAT2}
838 + .else
839 + ldrmi DAT0, [S], #4
840 + .if align == 0
841 + ldmcsia S!, {DAT1, DAT2}
842 + .else
843 + ldrcs DAT1, [S], #4
844 + ldrcs DAT2, [S], #4
845 + .endif
846 + strmi DAT0, [D], #4
847 + stmcsia D!, {DAT1, DAT2}
848 + .endif
849 +.endm
850 +
851 +.macro memcpy_trailing_15bytes backwards, align
852 + movs N, N, lsl #29
853 + .if backwards
854 + .if align == 0
855 + ldmcsdb S!, {DAT0, DAT1}
856 + .else
857 + ldrcs DAT1, [S, #-4]!
858 + ldrcs DAT0, [S, #-4]!
859 + .endif
860 + ldrmi DAT2, [S, #-4]!
861 + stmcsdb D!, {DAT0, DAT1}
862 + strmi DAT2, [D, #-4]!
863 + .else
864 + .if align == 0
865 + ldmcsia S!, {DAT0, DAT1}
866 + .else
867 + ldrcs DAT0, [S], #4
868 + ldrcs DAT1, [S], #4
869 + .endif
870 + ldrmi DAT2, [S], #4
871 + stmcsia D!, {DAT0, DAT1}
872 + strmi DAT2, [D], #4
873 + .endif
874 + movs N, N, lsl #2
875 + .if backwards
876 + ldrcsh DAT0, [S, #-2]!
877 + ldrmib DAT1, [S, #-1]
878 + strcsh DAT0, [D, #-2]!
879 + strmib DAT1, [D, #-1]
880 + .else
881 + ldrcsh DAT0, [S], #2
882 + ldrmib DAT1, [S]
883 + strcsh DAT0, [D], #2
884 + strmib DAT1, [D]
885 + .endif
886 +.endm
887 +
888 +.macro memcpy_long_inner_loop backwards, align
889 + .if align != 0
890 + .if backwards
891 + ldr DAT0, [S, #-align]!
892 + .else
893 + ldr LAST, [S, #-align]!
894 + .endif
895 + .endif
896 +110:
897 + .if align == 0
898 + .if backwards
899 + ldmdb S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
900 + pld [S, OFF]
901 + stmdb D!, {DAT4, DAT5, DAT6, LAST}
902 + stmdb D!, {DAT0, DAT1, DAT2, DAT3}
903 + .else
904 + ldmia S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
905 + pld [S, OFF]
906 + stmia D!, {DAT0, DAT1, DAT2, DAT3}
907 + stmia D!, {DAT4, DAT5, DAT6, LAST}
908 + .endif
909 + .else
910 + unaligned_words backwards, align, 1, 8, DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, DAT7, LAST
911 + .endif
912 + subs N, N, #32
913 + bhs 110b
914 + /* Just before the final (prefetch_distance+1) 32-byte blocks, deal with final preloads */
915 + preload_trailing backwards, S, N, OFF
916 + add N, N, #(prefetch_distance+2)*32 - 32
917 +120:
918 + .if align == 0
919 + .if backwards
920 + ldmdb S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
921 + stmdb D!, {DAT4, DAT5, DAT6, LAST}
922 + stmdb D!, {DAT0, DAT1, DAT2, DAT3}
923 + .else
924 + ldmia S!, {DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, LAST}
925 + stmia D!, {DAT0, DAT1, DAT2, DAT3}
926 + stmia D!, {DAT4, DAT5, DAT6, LAST}
927 + .endif
928 + .else
929 + unaligned_words backwards, align, 0, 8, DAT0, DAT1, DAT2, DAT3, DAT4, DAT5, DAT6, DAT7, LAST
930 + .endif
931 + subs N, N, #32
932 + bhs 120b
933 + tst N, #16
934 + .if align == 0
935 + .if backwards
936 + ldmnedb S!, {DAT0, DAT1, DAT2, LAST}
937 + stmnedb D!, {DAT0, DAT1, DAT2, LAST}
938 + .else
939 + ldmneia S!, {DAT0, DAT1, DAT2, LAST}
940 + stmneia D!, {DAT0, DAT1, DAT2, LAST}
941 + .endif
942 + .else
943 + beq 130f
944 + unaligned_words backwards, align, 0, 4, DAT0, DAT1, DAT2, DAT3, LAST
945 +130:
946 + .endif
947 + /* Trailing words and bytes */
948 + tst N, #15
949 + beq 199f
950 + .if align != 0
951 + add S, S, #align
952 + .endif
953 + memcpy_trailing_15bytes backwards, align
954 +199:
955 + pop {DAT3, DAT4, DAT5, DAT6, DAT7}
956 + pop {D, DAT1, DAT2, pc}
957 +.endm
958 +
959 +.macro memcpy_medium_inner_loop backwards, align
960 +120:
961 + .if backwards
962 + .if align == 0
963 + ldmdb S!, {DAT0, DAT1, DAT2, LAST}
964 + .else
965 + ldr LAST, [S, #-4]!
966 + ldr DAT2, [S, #-4]!
967 + ldr DAT1, [S, #-4]!
968 + ldr DAT0, [S, #-4]!
969 + .endif
970 + stmdb D!, {DAT0, DAT1, DAT2, LAST}
971 + .else
972 + .if align == 0
973 + ldmia S!, {DAT0, DAT1, DAT2, LAST}
974 + .else
975 + ldr DAT0, [S], #4
976 + ldr DAT1, [S], #4
977 + ldr DAT2, [S], #4
978 + ldr LAST, [S], #4
979 + .endif
980 + stmia D!, {DAT0, DAT1, DAT2, LAST}
981 + .endif
982 + subs N, N, #16
983 + bhs 120b
984 + /* Trailing words and bytes */
985 + tst N, #15
986 + beq 199f
987 + memcpy_trailing_15bytes backwards, align
988 +199:
989 + pop {D, DAT1, DAT2, pc}
990 +.endm
991 +
992 +.macro memcpy_short_inner_loop backwards, align
993 + tst N, #16
994 + .if backwards
995 + .if align == 0
996 + ldmnedb S!, {DAT0, DAT1, DAT2, LAST}
997 + .else
998 + ldrne LAST, [S, #-4]!
999 + ldrne DAT2, [S, #-4]!
1000 + ldrne DAT1, [S, #-4]!
1001 + ldrne DAT0, [S, #-4]!
1002 + .endif
1003 + stmnedb D!, {DAT0, DAT1, DAT2, LAST}
1004 + .else
1005 + .if align == 0
1006 + ldmneia S!, {DAT0, DAT1, DAT2, LAST}
1007 + .else
1008 + ldrne DAT0, [S], #4
1009 + ldrne DAT1, [S], #4
1010 + ldrne DAT2, [S], #4
1011 + ldrne LAST, [S], #4
1012 + .endif
1013 + stmneia D!, {DAT0, DAT1, DAT2, LAST}
1014 + .endif
1015 + memcpy_trailing_15bytes backwards, align
1016 +199:
1017 + pop {D, DAT1, DAT2, pc}
1018 +.endm
1019 +
1020 +.macro memcpy backwards
1021 + D .req a1
1022 + S .req a2
1023 + N .req a3
1024 + DAT0 .req a4
1025 + DAT1 .req v1
1026 + DAT2 .req v2
1027 + DAT3 .req v3
1028 + DAT4 .req v4
1029 + DAT5 .req v5
1030 + DAT6 .req v6
1031 + DAT7 .req sl
1032 + LAST .req ip
1033 + OFF .req lr
1034 +
1035 + .cfi_startproc
1036 +
1037 + push {D, DAT1, DAT2, lr}
1038 +
1039 + .cfi_def_cfa_offset 16
1040 + .cfi_rel_offset D, 0
1041 + .cfi_undefined S
1042 + .cfi_undefined N
1043 + .cfi_undefined DAT0
1044 + .cfi_rel_offset DAT1, 4
1045 + .cfi_rel_offset DAT2, 8
1046 + .cfi_undefined LAST
1047 + .cfi_rel_offset lr, 12
1048 +
1049 + .if backwards
1050 + add D, D, N
1051 + add S, S, N
1052 + .endif
1053 +
1054 + /* See if we're guaranteed to have at least one 16-byte aligned 16-byte write */
1055 + cmp N, #31
1056 + blo 170f
1057 + /* To preload ahead as we go, we need at least (prefetch_distance+2) 32-byte blocks */
1058 + cmp N, #(prefetch_distance+3)*32 - 1
1059 + blo 160f
1060 +
1061 + /* Long case */
1062 + push {DAT3, DAT4, DAT5, DAT6, DAT7}
1063 +
1064 + .cfi_def_cfa_offset 36
1065 + .cfi_rel_offset D, 20
1066 + .cfi_rel_offset DAT1, 24
1067 + .cfi_rel_offset DAT2, 28
1068 + .cfi_rel_offset DAT3, 0
1069 + .cfi_rel_offset DAT4, 4
1070 + .cfi_rel_offset DAT5, 8
1071 + .cfi_rel_offset DAT6, 12
1072 + .cfi_rel_offset DAT7, 16
1073 + .cfi_rel_offset lr, 32
1074 +
1075 + /* Adjust N so that the decrement instruction can also test for
1076 + * inner loop termination. We want it to stop when there are
1077 + * (prefetch_distance+1) complete blocks to go. */
1078 + sub N, N, #(prefetch_distance+2)*32
1079 + preload_leading_step1 backwards, DAT0, S
1080 + .if backwards
1081 + /* Bug in GAS: it accepts, but mis-assembles the instruction
1082 + * ands DAT2, D, #60, 2
1083 + * which sets DAT2 to the number of leading bytes until destination is aligned and also clears C (sets borrow)
1084 + */
1085 + .word 0xE210513C
1086 + beq 154f
1087 + .else
1088 + ands DAT2, D, #15
1089 + beq 154f
1090 + rsb DAT2, DAT2, #16 /* number of leading bytes until destination aligned */
1091 + .endif
1092 + preload_leading_step2 backwards, DAT0, S, DAT2, OFF
1093 + memcpy_leading_15bytes backwards, 1
1094 +154: /* Destination now 16-byte aligned; we have at least one prefetch as well as at least one 16-byte output block */
1095 + /* 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 */
1096 + .if backwards
1097 + rsb OFF, S, #3
1098 + and OFF, OFF, #28
1099 + sub OFF, OFF, #32*(prefetch_distance+1)
1100 + .else
1101 + and OFF, S, #28
1102 + rsb OFF, OFF, #32*prefetch_distance
1103 + .endif
1104 + movs DAT0, S, lsl #31
1105 + bhi 157f
1106 + bcs 156f
1107 + bmi 155f
1108 + memcpy_long_inner_loop backwards, 0
1109 +155: memcpy_long_inner_loop backwards, 1
1110 +156: memcpy_long_inner_loop backwards, 2
1111 +157: memcpy_long_inner_loop backwards, 3
1112 +
1113 + .cfi_def_cfa_offset 16
1114 + .cfi_rel_offset D, 0
1115 + .cfi_rel_offset DAT1, 4
1116 + .cfi_rel_offset DAT2, 8
1117 + .cfi_same_value DAT3
1118 + .cfi_same_value DAT4
1119 + .cfi_same_value DAT5
1120 + .cfi_same_value DAT6
1121 + .cfi_same_value DAT7
1122 + .cfi_rel_offset lr, 12
1123 +
1124 +160: /* Medium case */
1125 + preload_all backwards, 0, 0, S, N, DAT2, OFF
1126 + sub N, N, #16 /* simplifies inner loop termination */
1127 + .if backwards
1128 + ands DAT2, D, #15
1129 + beq 164f
1130 + .else
1131 + ands DAT2, D, #15
1132 + beq 164f
1133 + rsb DAT2, DAT2, #16
1134 + .endif
1135 + memcpy_leading_15bytes backwards, align
1136 +164: /* Destination now 16-byte aligned; we have at least one 16-byte output block */
1137 + tst S, #3
1138 + bne 140f
1139 + memcpy_medium_inner_loop backwards, 0
1140 +140: memcpy_medium_inner_loop backwards, 1
1141 +
1142 +170: /* Short case, less than 31 bytes, so no guarantee of at least one 16-byte block */
1143 + teq N, #0
1144 + beq 199f
1145 + preload_all backwards, 1, 0, S, N, DAT2, LAST
1146 + tst D, #3
1147 + beq 174f
1148 +172: subs N, N, #1
1149 + blo 199f
1150 + .if backwards
1151 + ldrb DAT0, [S, #-1]!
1152 + strb DAT0, [D, #-1]!
1153 + .else
1154 + ldrb DAT0, [S], #1
1155 + strb DAT0, [D], #1
1156 + .endif
1157 + tst D, #3
1158 + bne 172b
1159 +174: /* Destination now 4-byte aligned; we have 0 or more output bytes to go */
1160 + tst S, #3
1161 + bne 140f
1162 + memcpy_short_inner_loop backwards, 0
1163 +140: memcpy_short_inner_loop backwards, 1
1164 +
1165 + .cfi_endproc
1166 +
1167 + .unreq D
1168 + .unreq S
1169 + .unreq N
1170 + .unreq DAT0
1171 + .unreq DAT1
1172 + .unreq DAT2
1173 + .unreq DAT3
1174 + .unreq DAT4
1175 + .unreq DAT5
1176 + .unreq DAT6
1177 + .unreq DAT7
1178 + .unreq LAST
1179 + .unreq OFF
1180 +.endm
1181 --- /dev/null
1182 +++ b/arch/arm/lib/memmove_rpi.S
1183 @@ -0,0 +1,61 @@
1184 +/*
1185 +Copyright (c) 2013, Raspberry Pi Foundation
1186 +Copyright (c) 2013, RISC OS Open Ltd
1187 +All rights reserved.
1188 +
1189 +Redistribution and use in source and binary forms, with or without
1190 +modification, are permitted provided that the following conditions are met:
1191 + * Redistributions of source code must retain the above copyright
1192 + notice, this list of conditions and the following disclaimer.
1193 + * Redistributions in binary form must reproduce the above copyright
1194 + notice, this list of conditions and the following disclaimer in the
1195 + documentation and/or other materials provided with the distribution.
1196 + * Neither the name of the copyright holder nor the
1197 + names of its contributors may be used to endorse or promote products
1198 + derived from this software without specific prior written permission.
1199 +
1200 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
1201 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
1202 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
1203 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
1204 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
1205 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
1206 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
1207 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
1208 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
1209 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1210 +*/
1211 +
1212 +#include <linux/linkage.h>
1213 +#include "arm-mem.h"
1214 +#include "memcpymove.h"
1215 +
1216 +/* Prevent the stack from becoming executable */
1217 +#if defined(__linux__) && defined(__ELF__)
1218 +.section .note.GNU-stack,"",%progbits
1219 +#endif
1220 +
1221 + .text
1222 + .arch armv6
1223 + .object_arch armv4
1224 + .arm
1225 + .altmacro
1226 + .p2align 2
1227 +
1228 +/*
1229 + * void *memmove(void *s1, const void *s2, size_t n);
1230 + * On entry:
1231 + * a1 = pointer to destination
1232 + * a2 = pointer to source
1233 + * a3 = number of bytes to copy
1234 + * On exit:
1235 + * a1 preserved
1236 + */
1237 +
1238 +.set prefetch_distance, 3
1239 +
1240 +ENTRY(memmove)
1241 + cmp a2, a1
1242 + bpl memcpy /* pl works even over -1 - 0 and 0x7fffffff - 0x80000000 boundaries */
1243 + memcpy 1
1244 +ENDPROC(memmove)
1245 --- /dev/null
1246 +++ b/arch/arm/lib/memset_rpi.S
1247 @@ -0,0 +1,123 @@
1248 +/*
1249 +Copyright (c) 2013, Raspberry Pi Foundation
1250 +Copyright (c) 2013, RISC OS Open Ltd
1251 +All rights reserved.
1252 +
1253 +Redistribution and use in source and binary forms, with or without
1254 +modification, are permitted provided that the following conditions are met:
1255 + * Redistributions of source code must retain the above copyright
1256 + notice, this list of conditions and the following disclaimer.
1257 + * Redistributions in binary form must reproduce the above copyright
1258 + notice, this list of conditions and the following disclaimer in the
1259 + documentation and/or other materials provided with the distribution.
1260 + * Neither the name of the copyright holder nor the
1261 + names of its contributors may be used to endorse or promote products
1262 + derived from this software without specific prior written permission.
1263 +
1264 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
1265 +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
1266 +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
1267 +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
1268 +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
1269 +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
1270 +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
1271 +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
1272 +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
1273 +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1274 +*/
1275 +
1276 +#include <linux/linkage.h>
1277 +#include "arm-mem.h"
1278 +
1279 +/* Prevent the stack from becoming executable */
1280 +#if defined(__linux__) && defined(__ELF__)
1281 +.section .note.GNU-stack,"",%progbits
1282 +#endif
1283 +
1284 + .text
1285 + .arch armv6
1286 + .object_arch armv4
1287 + .arm
1288 + .altmacro
1289 + .p2align 2
1290 +
1291 +/*
1292 + * void *memset(void *s, int c, size_t n);
1293 + * On entry:
1294 + * a1 = pointer to buffer to fill
1295 + * a2 = byte pattern to fill with (caller-narrowed)
1296 + * a3 = number of bytes to fill
1297 + * On exit:
1298 + * a1 preserved
1299 + */
1300 +ENTRY(mmioset)
1301 +ENTRY(memset)
1302 + S .req a1
1303 + DAT0 .req a2
1304 + N .req a3
1305 + DAT1 .req a4
1306 + DAT2 .req ip
1307 + DAT3 .req lr
1308 +
1309 + orr DAT0, DAT0, lsl #8
1310 + push {S, lr}
1311 + orr DAT0, DAT0, lsl #16
1312 + mov DAT1, DAT0
1313 +
1314 + /* See if we're guaranteed to have at least one 16-byte aligned 16-byte write */
1315 + cmp N, #31
1316 + blo 170f
1317 +
1318 +161: sub N, N, #16 /* simplifies inner loop termination */
1319 + /* Leading words and bytes */
1320 + tst S, #15
1321 + beq 164f
1322 + rsb DAT3, S, #0 /* bits 0-3 = number of leading bytes until aligned */
1323 + movs DAT2, DAT3, lsl #31
1324 + submi N, N, #1
1325 + strmib DAT0, [S], #1
1326 + subcs N, N, #2
1327 + strcsh DAT0, [S], #2
1328 + movs DAT2, DAT3, lsl #29
1329 + submi N, N, #4
1330 + strmi DAT0, [S], #4
1331 + subcs N, N, #8
1332 + stmcsia S!, {DAT0, DAT1}
1333 +164: /* Delayed set up of DAT2 and DAT3 so we could use them as scratch registers above */
1334 + mov DAT2, DAT0
1335 + mov DAT3, DAT0
1336 + /* Now the inner loop of 16-byte stores */
1337 +165: stmia S!, {DAT0, DAT1, DAT2, DAT3}
1338 + subs N, N, #16
1339 + bhs 165b
1340 +166: /* Trailing words and bytes */
1341 + movs N, N, lsl #29
1342 + stmcsia S!, {DAT0, DAT1}
1343 + strmi DAT0, [S], #4
1344 + movs N, N, lsl #2
1345 + strcsh DAT0, [S], #2
1346 + strmib DAT0, [S]
1347 +199: pop {S, pc}
1348 +
1349 +170: /* Short case */
1350 + mov DAT2, DAT0
1351 + mov DAT3, DAT0
1352 + tst S, #3
1353 + beq 174f
1354 +172: subs N, N, #1
1355 + blo 199b
1356 + strb DAT0, [S], #1
1357 + tst S, #3
1358 + bne 172b
1359 +174: tst N, #16
1360 + stmneia S!, {DAT0, DAT1, DAT2, DAT3}
1361 + b 166b
1362 +
1363 + .unreq S
1364 + .unreq DAT0
1365 + .unreq N
1366 + .unreq DAT1
1367 + .unreq DAT2
1368 + .unreq DAT3
1369 +ENDPROC(memset)
1370 +ENDPROC(mmioset)
1371 --- a/arch/arm/lib/uaccess_with_memcpy.c
1372 +++ b/arch/arm/lib/uaccess_with_memcpy.c
1373 @@ -22,6 +22,14 @@
1374 #include <asm/current.h>
1375 #include <asm/page.h>
1376
1377 +#ifndef COPY_FROM_USER_THRESHOLD
1378 +#define COPY_FROM_USER_THRESHOLD 64
1379 +#endif
1380 +
1381 +#ifndef COPY_TO_USER_THRESHOLD
1382 +#define COPY_TO_USER_THRESHOLD 64
1383 +#endif
1384 +
1385 static int
1386 pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
1387 {
1388 @@ -84,7 +92,44 @@ pin_page_for_write(const void __user *_a
1389 return 1;
1390 }
1391
1392 -static unsigned long noinline
1393 +static int
1394 +pin_page_for_read(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp)
1395 +{
1396 + unsigned long addr = (unsigned long)_addr;
1397 + pgd_t *pgd;
1398 + pmd_t *pmd;
1399 + pte_t *pte;
1400 + pud_t *pud;
1401 + spinlock_t *ptl;
1402 +
1403 + pgd = pgd_offset(current->mm, addr);
1404 + if (unlikely(pgd_none(*pgd) || pgd_bad(*pgd)))
1405 + {
1406 + return 0;
1407 + }
1408 + pud = pud_offset(pgd, addr);
1409 + if (unlikely(pud_none(*pud) || pud_bad(*pud)))
1410 + {
1411 + return 0;
1412 + }
1413 +
1414 + pmd = pmd_offset(pud, addr);
1415 + if (unlikely(pmd_none(*pmd) || pmd_bad(*pmd)))
1416 + return 0;
1417 +
1418 + pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl);
1419 + if (unlikely(!pte_present(*pte) || !pte_young(*pte))) {
1420 + pte_unmap_unlock(pte, ptl);
1421 + return 0;
1422 + }
1423 +
1424 + *ptep = pte;
1425 + *ptlp = ptl;
1426 +
1427 + return 1;
1428 +}
1429 +
1430 +unsigned long noinline
1431 __copy_to_user_memcpy(void __user *to, const void *from, unsigned long n)
1432 {
1433 unsigned long ua_flags;
1434 @@ -137,6 +182,57 @@ out:
1435 return n;
1436 }
1437
1438 +unsigned long noinline
1439 +__copy_from_user_memcpy(void *to, const void __user *from, unsigned long n)
1440 +{
1441 + unsigned long ua_flags;
1442 + int atomic;
1443 +
1444 + if (unlikely(segment_eq(get_fs(), KERNEL_DS))) {
1445 + memcpy(to, (const void *)from, n);
1446 + return 0;
1447 + }
1448 +
1449 + /* the mmap semaphore is taken only if not in an atomic context */
1450 + atomic = in_atomic();
1451 +
1452 + if (!atomic)
1453 + down_read(&current->mm->mmap_sem);
1454 + while (n) {
1455 + pte_t *pte;
1456 + spinlock_t *ptl;
1457 + int tocopy;
1458 +
1459 + while (!pin_page_for_read(from, &pte, &ptl)) {
1460 + char temp;
1461 + if (!atomic)
1462 + up_read(&current->mm->mmap_sem);
1463 + if (__get_user(temp, (char __user *)from))
1464 + goto out;
1465 + if (!atomic)
1466 + down_read(&current->mm->mmap_sem);
1467 + }
1468 +
1469 + tocopy = (~(unsigned long)from & ~PAGE_MASK) + 1;
1470 + if (tocopy > n)
1471 + tocopy = n;
1472 +
1473 + ua_flags = uaccess_save_and_enable();
1474 + memcpy(to, (const void *)from, tocopy);
1475 + uaccess_restore(ua_flags);
1476 + to += tocopy;
1477 + from += tocopy;
1478 + n -= tocopy;
1479 +
1480 + pte_unmap_unlock(pte, ptl);
1481 + }
1482 + if (!atomic)
1483 + up_read(&current->mm->mmap_sem);
1484 +
1485 +out:
1486 + return n;
1487 +}
1488 +
1489 unsigned long
1490 arm_copy_to_user(void __user *to, const void *from, unsigned long n)
1491 {
1492 @@ -147,7 +243,7 @@ arm_copy_to_user(void __user *to, const
1493 * With frame pointer disabled, tail call optimization kicks in
1494 * as well making this test almost invisible.
1495 */
1496 - if (n < 64) {
1497 + if (n < COPY_TO_USER_THRESHOLD) {
1498 unsigned long ua_flags = uaccess_save_and_enable();
1499 n = __copy_to_user_std(to, from, n);
1500 uaccess_restore(ua_flags);
1501 @@ -156,6 +252,26 @@ arm_copy_to_user(void __user *to, const
1502 }
1503 return n;
1504 }
1505 +
1506 +unsigned long __must_check
1507 +arm_copy_from_user(void *to, const void __user *from, unsigned long n)
1508 +{
1509 + /*
1510 + * This test is stubbed out of the main function above to keep
1511 + * the overhead for small copies low by avoiding a large
1512 + * register dump on the stack just to reload them right away.
1513 + * With frame pointer disabled, tail call optimization kicks in
1514 + * as well making this test almost invisible.
1515 + */
1516 + if (n < COPY_TO_USER_THRESHOLD) {
1517 + unsigned long ua_flags = uaccess_save_and_enable();
1518 + n = __copy_from_user_std(to, from, n);
1519 + uaccess_restore(ua_flags);
1520 + } else {
1521 + n = __copy_from_user_memcpy(to, from, n);
1522 + }
1523 + return n;
1524 +}
1525
1526 static unsigned long noinline
1527 __clear_user_memset(void __user *addr, unsigned long n)
1528 --- a/arch/arm/mach-bcm/Kconfig
1529 +++ b/arch/arm/mach-bcm/Kconfig
1530 @@ -174,6 +174,13 @@ config ARCH_BCM_53573
1531 The base chip is BCM53573 and there are some packaging modifications
1532 like BCM47189 and BCM47452.
1533
1534 +config BCM2835_FAST_MEMCPY
1535 + bool "Enable optimized __copy_to_user and __copy_from_user"
1536 + depends on ARCH_BCM2835 && ARCH_MULTI_V6
1537 + default y
1538 + help
1539 + Optimized versions of __copy_to_user and __copy_from_user for Pi1.
1540 +
1541 config ARCH_BCM_63XX
1542 bool "Broadcom BCM63xx DSL SoC"
1543 depends on ARCH_MULTI_V7
Felix Fietkaus staging tree