diff options
Diffstat (limited to 'arch/x86/lib/clear_page_64.S')
| -rw-r--r-- | arch/x86/lib/clear_page_64.S | 183 |
1 files changed, 67 insertions, 116 deletions
diff --git a/arch/x86/lib/clear_page_64.S b/arch/x86/lib/clear_page_64.S index ecbfb4dd3b01..f74a3e704a1c 100644 --- a/arch/x86/lib/clear_page_64.S +++ b/arch/x86/lib/clear_page_64.S @@ -57,134 +57,85 @@ EXPORT_SYMBOL_GPL(clear_page_erms) * Input: * rdi destination * rcx count + * rax is zero * * Output: * rcx: uncleared bytes or 0 if successful. */ -SYM_FUNC_START(clear_user_original) - /* - * Copy only the lower 32 bits of size as that is enough to handle the rest bytes, - * i.e., no need for a 'q' suffix and thus a REX prefix. - */ - mov %ecx,%eax - shr $3,%rcx - jz .Lrest_bytes +SYM_FUNC_START(rep_stos_alternative) + cmpq $64,%rcx + jae .Lunrolled - # do the qwords first - .p2align 4 -.Lqwords: - movq $0,(%rdi) - lea 8(%rdi),%rdi - dec %rcx - jnz .Lqwords + cmp $8,%ecx + jae .Lword -.Lrest_bytes: - and $7, %eax - jz .Lexit + testl %ecx,%ecx + je .Lexit - # now do the rest bytes -.Lbytes: - movb $0,(%rdi) +.Lclear_user_tail: +0: movb %al,(%rdi) inc %rdi - dec %eax - jnz .Lbytes - + dec %rcx + jnz .Lclear_user_tail .Lexit: - /* - * %rax still needs to be cleared in the exception case because this function is called - * from inline asm and the compiler expects %rax to be zero when exiting the inline asm, - * in case it might reuse it somewhere. - */ - xor %eax,%eax - RET - -.Lqwords_exception: - # convert remaining qwords back into bytes to return to caller - shl $3, %rcx - and $7, %eax - add %rax,%rcx - jmp .Lexit - -.Lbytes_exception: - mov %eax,%ecx - jmp .Lexit - - _ASM_EXTABLE_UA(.Lqwords, .Lqwords_exception) - _ASM_EXTABLE_UA(.Lbytes, .Lbytes_exception) -SYM_FUNC_END(clear_user_original) -EXPORT_SYMBOL(clear_user_original) - -/* - * Alternative clear user-space when CPU feature X86_FEATURE_REP_GOOD is - * present. - * Input: - * rdi destination - * rcx count - * - * Output: - * rcx: uncleared bytes or 0 if successful. - */ -SYM_FUNC_START(clear_user_rep_good) - # call the original thing for less than a cacheline - cmp $64, %rcx - jb clear_user_original - -.Lprep: - # copy lower 32-bits for rest bytes - mov %ecx, %edx - shr $3, %rcx - jz .Lrep_good_rest_bytes - -.Lrep_good_qwords: - rep stosq - -.Lrep_good_rest_bytes: - and $7, %edx - jz .Lrep_good_exit - -.Lrep_good_bytes: - mov %edx, %ecx - rep stosb - -.Lrep_good_exit: - # see .Lexit comment above - xor %eax, %eax RET -.Lrep_good_qwords_exception: - # convert remaining qwords back into bytes to return to caller - shl $3, %rcx - and $7, %edx - add %rdx, %rcx - jmp .Lrep_good_exit + _ASM_EXTABLE_UA( 0b, .Lexit) - _ASM_EXTABLE_UA(.Lrep_good_qwords, .Lrep_good_qwords_exception) - _ASM_EXTABLE_UA(.Lrep_good_bytes, .Lrep_good_exit) -SYM_FUNC_END(clear_user_rep_good) -EXPORT_SYMBOL(clear_user_rep_good) +.Lword: +1: movq %rax,(%rdi) + addq $8,%rdi + sub $8,%ecx + je .Lexit + cmp $8,%ecx + jae .Lword + jmp .Lclear_user_tail -/* - * Alternative clear user-space when CPU feature X86_FEATURE_ERMS is present. - * Input: - * rdi destination - * rcx count - * - * Output: - * rcx: uncleared bytes or 0 if successful. - * - */ -SYM_FUNC_START(clear_user_erms) - # call the original thing for less than a cacheline - cmp $64, %rcx - jb clear_user_original - -.Lerms_bytes: - rep stosb - -.Lerms_exit: - xorl %eax,%eax + .p2align 4 +.Lunrolled: +10: movq %rax,(%rdi) +11: movq %rax,8(%rdi) +12: movq %rax,16(%rdi) +13: movq %rax,24(%rdi) +14: movq %rax,32(%rdi) +15: movq %rax,40(%rdi) +16: movq %rax,48(%rdi) +17: movq %rax,56(%rdi) + addq $64,%rdi + subq $64,%rcx + cmpq $64,%rcx + jae .Lunrolled + cmpl $8,%ecx + jae .Lword + testl %ecx,%ecx + jne .Lclear_user_tail RET - _ASM_EXTABLE_UA(.Lerms_bytes, .Lerms_exit) -SYM_FUNC_END(clear_user_erms) -EXPORT_SYMBOL(clear_user_erms) + /* + * If we take an exception on any of the + * word stores, we know that %rcx isn't zero, + * so we can just go to the tail clearing to + * get the exact count. + * + * The unrolled case might end up clearing + * some bytes twice. Don't care. + * + * We could use the value in %rdi to avoid + * a second fault on the exact count case, + * but do we really care? No. + * + * Finally, we could try to align %rdi at the + * top of the unrolling. But unaligned stores + * just aren't that common or expensive. + */ + _ASM_EXTABLE_UA( 1b, .Lclear_user_tail) + _ASM_EXTABLE_UA(10b, .Lclear_user_tail) + _ASM_EXTABLE_UA(11b, .Lclear_user_tail) + _ASM_EXTABLE_UA(12b, .Lclear_user_tail) + _ASM_EXTABLE_UA(13b, .Lclear_user_tail) + _ASM_EXTABLE_UA(14b, .Lclear_user_tail) + _ASM_EXTABLE_UA(15b, .Lclear_user_tail) + _ASM_EXTABLE_UA(16b, .Lclear_user_tail) + _ASM_EXTABLE_UA(17b, .Lclear_user_tail) +SYM_FUNC_END(rep_stos_alternative) +EXPORT_SYMBOL(rep_stos_alternative) |