blob: 009f98216b7eb316c12847e42f50ac77d4f4b8a3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/* Written 2003 by Andi Kleen, based on a kernel by Evandro Menezes */
Dave Jones038b0a62006-10-04 03:38:54 -04002
Jan Beulich8d379da2006-09-26 10:52:32 +02003#include <linux/linkage.h>
Borislav Petkov090a3f62015-01-12 18:19:40 +01004#include <asm/cpufeature.h>
Andy Lutomirski59e97e42011-07-13 09:24:10 -04005#include <asm/alternative-asm.h>
Jan Beulich8d379da2006-09-26 10:52:32 +02006
Borislav Petkov090a3f62015-01-12 18:19:40 +01007/*
8 * Some CPUs run faster using the string copy instructions (sane microcode).
9 * It is also a lot simpler. Use this when possible. But, don't use streaming
10 * copy unless the CPU indicates X86_FEATURE_REP_GOOD. Could vary the
11 * prefetch distance based on SMP/UP.
12 */
Jan Beulich8d379da2006-09-26 10:52:32 +020013 ALIGN
Borislav Petkov090a3f62015-01-12 18:19:40 +010014ENTRY(copy_page)
Borislav Petkov090a3f62015-01-12 18:19:40 +010015 ALTERNATIVE "jmp copy_page_regs", "", X86_FEATURE_REP_GOOD
Ma Ling269833b2012-10-18 03:52:45 +080016 movl $4096/8, %ecx
17 rep movsq
Jan Beulich8d379da2006-09-26 10:52:32 +020018 ret
Borislav Petkov090a3f62015-01-12 18:19:40 +010019ENDPROC(copy_page)
Jan Beulich8d379da2006-09-26 10:52:32 +020020
Borislav Petkov090a3f62015-01-12 18:19:40 +010021ENTRY(copy_page_regs)
Ma Ling269833b2012-10-18 03:52:45 +080022 subq $2*8, %rsp
Ma Ling269833b2012-10-18 03:52:45 +080023 movq %rbx, (%rsp)
Ma Ling269833b2012-10-18 03:52:45 +080024 movq %r12, 1*8(%rsp)
Andi Kleen7bcd3f32006-02-03 21:51:02 +010025
Ma Ling269833b2012-10-18 03:52:45 +080026 movl $(4096/64)-5, %ecx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010027 .p2align 4
28.Loop64:
Ma Ling269833b2012-10-18 03:52:45 +080029 dec %rcx
30 movq 0x8*0(%rsi), %rax
31 movq 0x8*1(%rsi), %rbx
32 movq 0x8*2(%rsi), %rdx
33 movq 0x8*3(%rsi), %r8
34 movq 0x8*4(%rsi), %r9
35 movq 0x8*5(%rsi), %r10
36 movq 0x8*6(%rsi), %r11
37 movq 0x8*7(%rsi), %r12
Andi Kleen7bcd3f32006-02-03 21:51:02 +010038
39 prefetcht0 5*64(%rsi)
40
Ma Ling269833b2012-10-18 03:52:45 +080041 movq %rax, 0x8*0(%rdi)
42 movq %rbx, 0x8*1(%rdi)
43 movq %rdx, 0x8*2(%rdi)
44 movq %r8, 0x8*3(%rdi)
45 movq %r9, 0x8*4(%rdi)
46 movq %r10, 0x8*5(%rdi)
47 movq %r11, 0x8*6(%rdi)
48 movq %r12, 0x8*7(%rdi)
Andi Kleen7bcd3f32006-02-03 21:51:02 +010049
Ma Ling269833b2012-10-18 03:52:45 +080050 leaq 64 (%rsi), %rsi
51 leaq 64 (%rdi), %rdi
Andi Kleen7bcd3f32006-02-03 21:51:02 +010052
Ma Ling269833b2012-10-18 03:52:45 +080053 jnz .Loop64
Andi Kleen7bcd3f32006-02-03 21:51:02 +010054
Ma Ling269833b2012-10-18 03:52:45 +080055 movl $5, %ecx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010056 .p2align 4
57.Loop2:
Ma Ling269833b2012-10-18 03:52:45 +080058 decl %ecx
Andi Kleen7bcd3f32006-02-03 21:51:02 +010059
Ma Ling269833b2012-10-18 03:52:45 +080060 movq 0x8*0(%rsi), %rax
61 movq 0x8*1(%rsi), %rbx
62 movq 0x8*2(%rsi), %rdx
63 movq 0x8*3(%rsi), %r8
64 movq 0x8*4(%rsi), %r9
65 movq 0x8*5(%rsi), %r10
66 movq 0x8*6(%rsi), %r11
67 movq 0x8*7(%rsi), %r12
Andi Kleen7bcd3f32006-02-03 21:51:02 +010068
Ma Ling269833b2012-10-18 03:52:45 +080069 movq %rax, 0x8*0(%rdi)
70 movq %rbx, 0x8*1(%rdi)
71 movq %rdx, 0x8*2(%rdi)
72 movq %r8, 0x8*3(%rdi)
73 movq %r9, 0x8*4(%rdi)
74 movq %r10, 0x8*5(%rdi)
75 movq %r11, 0x8*6(%rdi)
76 movq %r12, 0x8*7(%rdi)
Andi Kleen7bcd3f32006-02-03 21:51:02 +010077
Ma Ling269833b2012-10-18 03:52:45 +080078 leaq 64(%rdi), %rdi
79 leaq 64(%rsi), %rsi
Andi Kleen7bcd3f32006-02-03 21:51:02 +010080 jnz .Loop2
81
Ma Ling269833b2012-10-18 03:52:45 +080082 movq (%rsp), %rbx
Ma Ling269833b2012-10-18 03:52:45 +080083 movq 1*8(%rsp), %r12
Ma Ling269833b2012-10-18 03:52:45 +080084 addq $2*8, %rsp
Andi Kleen7bcd3f32006-02-03 21:51:02 +010085 ret
Borislav Petkov090a3f62015-01-12 18:19:40 +010086ENDPROC(copy_page_regs)