Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* Written 2003 by Andi Kleen, based on a kernel by Evandro Menezes */ |
Dave Jones | 038b0a6 | 2006-10-04 03:38:54 -0400 | [diff] [blame] | 2 | |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 3 | #include <linux/linkage.h> |
| 4 | #include <asm/dwarf2.h> |
Andy Lutomirski | 59e97e4 | 2011-07-13 09:24:10 -0400 | [diff] [blame] | 5 | #include <asm/alternative-asm.h> |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 6 | |
| 7 | ALIGN |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 8 | copy_page_rep: |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 9 | CFI_STARTPROC |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 10 | movl $4096/8, %ecx |
| 11 | rep movsq |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 12 | ret |
| 13 | CFI_ENDPROC |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 14 | ENDPROC(copy_page_rep) |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 15 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 16 | /* |
| 17 | * Don't use streaming copy unless the CPU indicates X86_FEATURE_REP_GOOD. |
| 18 | * Could vary the prefetch distance based on SMP/UP. |
| 19 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 20 | |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 21 | ENTRY(copy_page) |
| 22 | CFI_STARTPROC |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 23 | subq $2*8, %rsp |
Jan Beulich | 4269329 | 2012-01-05 16:12:25 +0000 | [diff] [blame] | 24 | CFI_ADJUST_CFA_OFFSET 2*8 |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 25 | movq %rbx, (%rsp) |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 26 | CFI_REL_OFFSET rbx, 0 |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 27 | movq %r12, 1*8(%rsp) |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 28 | CFI_REL_OFFSET r12, 1*8 |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 29 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 30 | movl $(4096/64)-5, %ecx |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 31 | .p2align 4 |
| 32 | .Loop64: |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 33 | dec %rcx |
| 34 | movq 0x8*0(%rsi), %rax |
| 35 | movq 0x8*1(%rsi), %rbx |
| 36 | movq 0x8*2(%rsi), %rdx |
| 37 | movq 0x8*3(%rsi), %r8 |
| 38 | movq 0x8*4(%rsi), %r9 |
| 39 | movq 0x8*5(%rsi), %r10 |
| 40 | movq 0x8*6(%rsi), %r11 |
| 41 | movq 0x8*7(%rsi), %r12 |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 42 | |
| 43 | prefetcht0 5*64(%rsi) |
| 44 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 45 | movq %rax, 0x8*0(%rdi) |
| 46 | movq %rbx, 0x8*1(%rdi) |
| 47 | movq %rdx, 0x8*2(%rdi) |
| 48 | movq %r8, 0x8*3(%rdi) |
| 49 | movq %r9, 0x8*4(%rdi) |
| 50 | movq %r10, 0x8*5(%rdi) |
| 51 | movq %r11, 0x8*6(%rdi) |
| 52 | movq %r12, 0x8*7(%rdi) |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 53 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 54 | leaq 64 (%rsi), %rsi |
| 55 | leaq 64 (%rdi), %rdi |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 56 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 57 | jnz .Loop64 |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 58 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 59 | movl $5, %ecx |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 60 | .p2align 4 |
| 61 | .Loop2: |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 62 | decl %ecx |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 63 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 64 | movq 0x8*0(%rsi), %rax |
| 65 | movq 0x8*1(%rsi), %rbx |
| 66 | movq 0x8*2(%rsi), %rdx |
| 67 | movq 0x8*3(%rsi), %r8 |
| 68 | movq 0x8*4(%rsi), %r9 |
| 69 | movq 0x8*5(%rsi), %r10 |
| 70 | movq 0x8*6(%rsi), %r11 |
| 71 | movq 0x8*7(%rsi), %r12 |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 72 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 73 | movq %rax, 0x8*0(%rdi) |
| 74 | movq %rbx, 0x8*1(%rdi) |
| 75 | movq %rdx, 0x8*2(%rdi) |
| 76 | movq %r8, 0x8*3(%rdi) |
| 77 | movq %r9, 0x8*4(%rdi) |
| 78 | movq %r10, 0x8*5(%rdi) |
| 79 | movq %r11, 0x8*6(%rdi) |
| 80 | movq %r12, 0x8*7(%rdi) |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 81 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 82 | leaq 64(%rdi), %rdi |
| 83 | leaq 64(%rsi), %rsi |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 84 | jnz .Loop2 |
| 85 | |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 86 | movq (%rsp), %rbx |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 87 | CFI_RESTORE rbx |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 88 | movq 1*8(%rsp), %r12 |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 89 | CFI_RESTORE r12 |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 90 | addq $2*8, %rsp |
Jan Beulich | 4269329 | 2012-01-05 16:12:25 +0000 | [diff] [blame] | 91 | CFI_ADJUST_CFA_OFFSET -2*8 |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 92 | ret |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 93 | .Lcopy_page_end: |
| 94 | CFI_ENDPROC |
| 95 | ENDPROC(copy_page) |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 96 | |
| 97 | /* Some CPUs run faster using the string copy instructions. |
| 98 | It is also a lot simpler. Use this when possible */ |
| 99 | |
| 100 | #include <asm/cpufeature.h> |
| 101 | |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 102 | .section .altinstr_replacement,"ax" |
| 103 | 1: .byte 0xeb /* jmp <disp8> */ |
Ma Ling | 269833b | 2012-10-18 03:52:45 +0800 | [diff] [blame] | 104 | .byte (copy_page_rep - copy_page) - (2f - 1b) /* offset */ |
Jan Beulich | 8d379da | 2006-09-26 10:52:32 +0200 | [diff] [blame] | 105 | 2: |
| 106 | .previous |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 107 | .section .altinstructions,"a" |
Andy Lutomirski | 59e97e4 | 2011-07-13 09:24:10 -0400 | [diff] [blame] | 108 | altinstruction_entry copy_page, 1b, X86_FEATURE_REP_GOOD, \ |
| 109 | .Lcopy_page_end-copy_page, 2b-1b |
Andi Kleen | 7bcd3f3 | 2006-02-03 21:51:02 +0100 | [diff] [blame] | 110 | .previous |