From f8fe689a81f906d1b91bb3220acde2a4ecb14c5b Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Mon, 6 May 2024 05:01:46 +0200 Subject: Adding upstream version 6.0.4-dfsg. Signed-off-by: Daniel Baumann --- src/VBox/Runtime/common/time/timesupA.mac | 885 ++++++++++++++++++++++++++++++ 1 file changed, 885 insertions(+) create mode 100644 src/VBox/Runtime/common/time/timesupA.mac (limited to 'src/VBox/Runtime/common/time/timesupA.mac') diff --git a/src/VBox/Runtime/common/time/timesupA.mac b/src/VBox/Runtime/common/time/timesupA.mac new file mode 100644 index 00000000..275be08f --- /dev/null +++ b/src/VBox/Runtime/common/time/timesupA.mac @@ -0,0 +1,885 @@ +; $Id: timesupA.mac $ +;; @file +; IPRT - Time using SUPLib, the Assembly Code Template. +; + +; +; Copyright (C) 2006-2019 Oracle Corporation +; +; This file is part of VirtualBox Open Source Edition (OSE), as +; available from http://www.virtualbox.org. This file is free software; +; you can redistribute it and/or modify it under the terms of the GNU +; General Public License (GPL) as published by the Free Software +; Foundation, in version 2 as it comes in the "COPYING" file of the +; VirtualBox OSE distribution. VirtualBox OSE is distributed in the +; hope that it will be useful, but WITHOUT ANY WARRANTY of any kind. +; +; The contents of this file may alternatively be used under the terms +; of the Common Development and Distribution License Version 1.0 +; (CDDL) only, as it comes in the "COPYING.CDDL" file of the +; VirtualBox OSE distribution, in which case the provisions of the +; CDDL are applicable instead of those of the GPL. +; +; You may elect to license modified versions of this file under the +; terms and conditions of either the GPL or the CDDL or both. +; + +%ifdef RT_ARCH_X86 +;; +; The x86 assembly implementation of the assembly routines. +; +; @returns Nanosecond timestamp. +; @param pData Pointer to the nanosecond timestamp data. +; +BEGINPROC rtTimeNanoTSInternalAsm + ; + ; Variable definitions. + ; +%define pData [ebp + 08h] +%define u64RetNanoTS_Hi [ebp - 04h] +%define u64RetNanoTS [ebp - 08h] +%define u32UpdateIntervalNS [ebp - 0ch] +%define u32UpdateIntervalTSC [ebp - 10h] +%define u64TSC_Hi [ebp - 14h] +%define u64TSC [ebp - 18h] +%define u64CurNanoTS_Hi [ebp - 1ch] +%define u64CurNanoTS [ebp - 20h] +%define u64PrevNanoTS_Hi [ebp - 24h] +%define u64PrevNanoTS [ebp - 28h] +%define u32TransactionId [ebp - 2ch] +%define u32ApicIdPlus [ebp - 30h] +%define TmpVar [ebp - 34h] +%define SavedEBX [ebp - 38h] +%define SavedEDI [ebp - 3ch] +%define SavedESI [ebp - 40h] + + ; + ; Prolog. + ; + push ebp + mov ebp, esp + sub esp, 40h + mov SavedEBX, ebx + mov SavedEDI, edi + mov SavedESI, esi + + + ;; + ;; Read the GIP data and the previous value. + ;; +.ReadGip: + + + ; + ; Load pGip. + ; +%ifdef IMPORTED_SUPLIB + %ifdef IN_RING0 + mov esi, IMP(g_SUPGlobalInfoPage) + %else + mov esi, IMP(g_pSUPGlobalInfoPage) + mov esi, [esi] + %endif +%else + mov esi, [NAME(g_pSUPGlobalInfoPage)] +%endif + or esi, esi + jz .Rediscover + cmp dword [esi + SUPGLOBALINFOPAGE.u32Magic], SUPGLOBALINFOPAGE_MAGIC + jne .Rediscover + + ; + ; Calc pGipCPU, setting u32ApicIdPlus if necessary. + ; +%ifdef NEED_APIC_ID + ; u8ApicId = ASMGetApicId(); + mov eax, 1 + cpuid ; expensive + %ifdef NEED_TRANSACTION_ID + mov u32ApicIdPlus, ebx + %endif + ; pGipCpu/pGipCpuDelta = &pGip->aCPU[pGip->aiCpuFromApicId[u8ApicId]]; + shr ebx, 24 + movzx ebx, word [esi + ebx * 2 + SUPGLOBALINFOPAGE.aiCpuFromApicId] + mov eax, SUPGIPCPU_size + mul ebx + lea edi, [esi + eax + SUPGLOBALINFOPAGE.aCPUs] ; edi == &pGip->aCPU[u8ApicId]; +%endif + +%ifdef NEED_TRANSACTION_ID + ; + ; Serialized loading of u32TransactionId. + ; + %ifdef ASYNC_GIP + mov ebx, [edi + SUPGIPCPU.u32TransactionId] + %else + mov ebx, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u32TransactionId] + %endif + mov u32TransactionId, ebx + %ifdef USE_LFENCE + lfence + %else + lock xor dword TmpVar, 0 + %endif +%endif + + ; + ; Load the data and TSC with delta applied. + ; + mov eax, [esi + SUPGLOBALINFOPAGE.u32UpdateIntervalNS] + mov u32UpdateIntervalNS, eax +%ifdef ASYNC_GIP ; esi is now free. + mov edx, [edi + SUPGIPCPU.u32UpdateIntervalTSC] +%else + mov edx, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u32UpdateIntervalTSC] +%endif + mov u32UpdateIntervalTSC, edx + + rdtsc +%ifdef WITH_TSC_DELTA + cmp dword [edi + SUPGIPCPU.i64TSCDelta], 0xffffffff + je .TscDeltaPossiblyInvalid +.TscDeltaValid: + sub eax, dword [edi + SUPGIPCPU.i64TSCDelta] + sbb edx, dword [edi + SUPGIPCPU.i64TSCDelta + 4] +.TscDeltaNotValid: ; edi is now free. +%endif + +%ifdef ASYNC_GIP + mov ecx, [edi + SUPGIPCPU.u64NanoTS] + mov esi, [edi + SUPGIPCPU.u64NanoTS + 4] +%else + mov ecx, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u64NanoTS] + mov ebx, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u64NanoTS + 4] +%endif + mov u64CurNanoTS, ecx + mov u64CurNanoTS_Hi, ebx +%ifdef ASYNC_GIP + mov ecx, [edi + SUPGIPCPU.u64TSC] + mov ebx, [edi + SUPGIPCPU.u64TSC + 4] +%else + mov ecx, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u64TSC] + mov ebx, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u64TSC + 4] +%endif + mov u64TSC, ecx + mov u64TSC_Hi, ebx + + ; u64PrevNanoTS = ASMAtomicReadU64(pu64Prev); + ; This serializes load/save. And with the dependency on the + ; RDTSC result, we try to make sure it has completed as well. +%ifdef ASYNC_GIP + mov esi, pData + mov esi, [esi + RTTIMENANOTSDATA.pu64Prev] +%else + mov edi, pData + mov edi, [esi + RTTIMENANOTSDATA.pu64Prev] +%endif + mov ebx, eax + mov ecx, edx +%ifdef ASYNC_GIP + lock cmpxchg8b [esi] +%else + lock cmpxchg8b [edi] +%endif + mov u64PrevNanoTS, eax + mov u64PrevNanoTS_Hi, edx + +%undef SAVED_u64RetNanoTS +%ifdef NEED_TRANSACTION_ID + ; + ; Check that the GIP and CPU didn't change. + ; We've already serialized all the loads and stores at this point. + ; + %ifdef NEED_APIC_ID + mov u64RetNanoTS, ebx + mov u64RetNanoTS_Hi, ecx + %define SAVED_u64RetNanoTS + mov eax, 1 + cpuid + cmp u32ApicIdPlus, ebx + jne .ReadGip + %endif + %ifdef ASYNC_GIP + mov esi, [edi + SUPGIPCPU.u32TransactionId] + %else + mov esi, [esi + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u32TransactionId] + %endif + cmp esi, u32TransactionId + jne .ReadGip + test esi, 1 + jnz .ReadGip +%endif ; NEED_TRANSACTION_ID +%ifdef SAVED_u64RetNanoTS + mov ebx, u64RetNanoTS + mov ecx, u64RetNanoTS_Hi +%endif + + ;; + ;; Calc the timestamp. + ;; + ; u64RetNanoTS -= u64TSC; + sub ebx, u64TSC + sbb ecx, u64TSC_Hi + + ; if (u64RetNanoTS > u32UpdateIntervalTSC) -> jump + or ecx, ecx + jnz .OverFlow + cmp ebx, u32UpdateIntervalTSC + ja .OverFlow + mov eax, ebx +.ContinueCalcs: ; eax <= u32UpdateIntervalTSC + mul dword u32UpdateIntervalNS + div dword u32UpdateIntervalTSC + xor edx, edx + + ; u64RetNanoTS += u64CurNanoTS; + add eax, u64CurNanoTS + adc edx, u64CurNanoTS_Hi + + ;; + ;; Compare it with the previous one. + ;; + ; if (RT_LIKELY( u64RetNanoTS > u64PrevNanoTS + ; && u64RetNanoTS < u64PrevNanoTS + UINT64_C(86000000000000) /* 24h */)) + ;; @todo optimize this compare (/me too tired). + mov ecx, u64PrevNanoTS_Hi + mov ebx, u64PrevNanoTS + cmp edx, ecx + ja .Compare2 + jb .DeltaPrevTooBig + cmp eax, ebx + jbe .DeltaPrevTooBig + +.Compare2: + add ebx, 0x6F736000 + adc ecx, 0x00004E37 + cmp edx, ecx + jb .CompareDone + ja .DeltaPrevTooBig + cmp eax, ebx + jae .DeltaPrevTooBig +.CompareDone: + + + ;; + ;; Update the previous value with the u64RetNanoTS value. + ;; +.Update: + ; if (RT_LIKELY(ASMAtomicCmpXchgU64(&pData->u64Prev, u64RetNanoTS, u64PrevNanoTS))) + mov ebx, eax + mov ecx, edx + mov esi, pData + mov esi, [esi + RTTIMENANOTSDATA.pu64Prev] + mov eax, u64PrevNanoTS + mov edx, u64PrevNanoTS_Hi + lock cmpxchg8b [esi] + jnz .UpdateFailed + +.Updated: + mov eax, ebx + mov edx, ecx + +.Done: + mov esi, SavedESI + mov edi, SavedEDI + mov ebx, SavedEBX + leave + ret + + + ;; + ;; We've expired the interval, cap it. If we're here for the 2nd + ;; time without any GIP update in-between, the checks against + ;; pData->u64Prev below will force 1ns stepping. + ;; +.OverFlow: + ; u64Delta = u32UpdateIntervalTSC; + mov esi, pData + inc dword [esi + RTTIMENANOTSDATA.cExpired] + mov eax, u32UpdateIntervalTSC + jmp .ContinueCalcs + + + ;; + ;; u64DeltaPrev >= 24h + ;; + ;; eax:edx = u64RetNanoTS (to be adjusted) + ;; +.DeltaPrevTooBig: + ; uint64_t u64DeltaPrev = u64RetNanoTS - u64PrevNanoTS; + mov ebx, eax + sub ebx, u64PrevNanoTS + mov ecx, edx + sbb ecx, u64PrevNanoTS_Hi ; ebx:ecx = u64DeltaPrev + + ; else if ( (int64_t)u64DeltaPrev <= 0 + ; && (int64_t)u64DeltaPrev + u32UpdateIntervalNS * 2 >= 0) + ; { + ; /* Occasional - u64RetNanoTS is in the recent 'past' relative the previous call. */ + ; pData->c1nsSteps++; + ; u64RetNanoTS = u64PrevNanoTS + 1; + ; } + mov esi, u32UpdateIntervalNS + cmp ecx, 0 + jl .PrevNotZero2ndTest + jg .DeltaPrevNotInRecentPast + cmp ebx, 0 + ja .DeltaPrevNotInRecentPast + +.PrevNotZero2ndTest: + add esi, esi ; ASSUMES: u32UpdateIntervalNS * 2 <= 32-bit. + xor edi, edi + add esi, ebx + adc edi, ecx + test edi, edi + js .DeltaPrevNotInRecentPast + +.DeltaPrevInRecentPast: + mov esi, pData + inc dword [esi + RTTIMENANOTSDATA.c1nsSteps] + mov eax, u64PrevNanoTS + mov edx, u64PrevNanoTS_Hi + add eax, 1 + adc edx, 0 + jmp .Update + +.DeltaPrevNotInRecentPast: + ; else if (!u64PrevNanoTS) /* We're resuming (see TMVirtualResume). */ + ; /* do nothing */; + cmp dword u64PrevNanoTS, 0 + jne .DeltaPrevNotZero + cmp dword u64PrevNanoTS_Hi, 0 + jne .DeltaPrevNotZero + jmp .Update + +.DeltaPrevNotZero: + ; else + ; { + ; /* Something has gone bust, if negative offset it's real bad. */ + ; rtTimeNanoTSInternalBitch(pVM, + ; } + + ; call C function that does the bitching. + mov u64RetNanoTS, eax + mov u64RetNanoTS_Hi, edx + + mov edi, u64PrevNanoTS_Hi + mov esi, u64PrevNanoTS + push edi + push esi ; 4 - u64PrevNanoTS + push ecx + push ebx ; 3 - u64DeltaPrev + push edx + push eax ; 2 - u64RetNanoTS + mov eax, pData + push eax ; 1 - pData + call dword [eax + RTTIMENANOTSDATA.pfnBad] + add esp, 4*7 + + mov eax, u64RetNanoTS + mov edx, u64RetNanoTS_Hi + jmp .Update + + + ;; + ;; Attempt updating the previous value, provided we're still ahead of it. + ;; + ;; There is no point in recalculating u64NanoTS because we got preempted or if + ;; we raced somebody while the GIP was updated, since these are events + ;; that might occur at any point in the return path as well. + ;; + ;; eax:edx = *pData->u64Prev + ;; ebx:ecx = u64RetNanoTS + ;; + ALIGNCODE(16) +.UpdateFailed: + mov edi, pData + lock inc dword [edi + RTTIMENANOTSDATA.cUpdateRaces] + ; for (i = 0; i < 10; i++) + mov edi, 10 +.UpdateLoop: + ; if (u64PrevNanoTS >= u64NanoTS) + ; break; + cmp edx, ecx + jg .Updated + jne .UpdateLoopLess + cmp eax, ebx + jae .Updated +.UpdateLoopLess: + ; retry + lock cmpxchg8b [esi] + jz .Updated + dec edi + jnz .UpdateLoop + jmp .Updated + + + ;; + ;; The GIP is seemingly invalid, redo the discovery. + ;; +.Rediscover: + mov eax, pData + push eax + call [eax + RTTIMENANOTSDATA.pfnRediscover] + add esp, 4h + jmp .Done + + +%ifdef WITH_TSC_DELTA + ;; + ;; Unlikely branch for when we think the TSC delta might be invalid. + ;; +.TscDeltaPossiblyInvalid: + cmp dword [edi + SUPGIPCPU.i64TSCDelta + 4], 0x7fffffff + jne .TscDeltaValid + jmp .TscDeltaNotValid +%endif + + ; + ; Cleanup variables + ; +%undef pData +%undef u64Delta_Hi +%undef u64Delta +%undef u32UpdateIntervalNS +%undef u32UpdateIntervalTSC +%undef u64TSC_Hi +%undef u64TSC +%undef u64NanoTS_Hi +%undef u64NanoTS +%undef u64PrevNanoTS_Hi +%undef u64PrevNanoTS +%undef u32TransactionId +%undef u8ApicId + +%else ; AMD64 + +;; +; The AMD64 assembly implementation of the assembly routines. +; +; @returns Nanosecond timestamp. +; @param pData gcc:rdi msc:rcx Pointer to the nanosecond timestamp data. +; +BEGINPROC rtTimeNanoTSInternalAsm + ; + ; Define variables and stack frame. + ; +%define SavedRBX [rbp - 08h] +%define SavedR12 [rbp - 10h] +%define SavedR13 [rbp - 18h] +%define SavedRDI [rbp - 20h] +%define SavedRSI [rbp - 28h] +%define TmpVar [rbp - 30h] +%define TmpVar2 [rbp - 38h] +%ifdef NEED_TRANSACTION_ID + %ifdef NEED_APIC_ID + %define SavedR14 [rbp - 40h] + %define SavedR15 [rbp - 48h] + %endif +%endif + +%define pData rdi + +%ifdef ASYNC_GIP + %define u64TSC rsi + %define pGip rsi + %ifdef NEED_APIC_ID + %define pGipCPU r8 + %endif +%else + %define u64TSC r8 + %define pGip rsi + %ifdef NEED_APIC_ID + %define pGipCPU r8 + %endif +%endif +%define u32TransactionId r9d +%define u64CurNanoTS r10 +%define u64PrevNanoTS r11 ; not parameter register +%define u32UpdateIntervalTSC r12d +%define u32UpdateIntervalTSC_64 r12 +%define u32UpdateIntervalNS r13d +%define u32UpdateIntervalNS_64 r13 +%undef u64SavedRetNanoTS +%undef u32ApicIdPlus +%ifdef NEED_TRANSACTION_ID + %ifdef NEED_APIC_ID + %define u64SavedRetNanoTS r14 + %define u32ApicIdPlus r15d + %endif +%endif + + ; + ; The prolog. + ; + push rbp + mov rbp, rsp +%ifdef ASM_CALL64_MSC + sub rsp, 50h+20h +%else + sub rsp, 50h +%endif + mov SavedRBX, rbx + mov SavedR12, r12 + mov SavedR13, r13 +%ifdef ASM_CALL64_MSC + mov SavedRDI, rdi + mov SavedRSI, rsi + mov pData, rcx +%else + ;mov pData, rdi - already in rdi. +%endif +%ifdef SavedR14 + mov SavedR14, r14 +%endif +%ifdef SavedR15 + mov SavedR15, r15 +%endif + + + ;; + ;; Data fetch loop. + ;; We take great pain ensuring that data consistency here. + ;; +.ReadGip: + + ; + ; Load pGip - finding the GIP is fun... + ; +%ifdef RT_OS_WINDOWS + %ifdef IMPORTED_SUPLIB + %ifdef IN_RING0 + mov rax, qword IMP(g_SUPGlobalInfoPage) + mov pGip, rax + %else + mov pGip, [IMP(g_pSUPGlobalInfoPage) wrt rip] + mov pGip, [pGip] + %endif + %else + mov pGip, [NAME(g_pSUPGlobalInfoPage) wrt rip] + %endif +%else + %ifdef IN_RING0 + mov rax, qword NAME(g_SUPGlobalInfoPage) + mov pGip, rax + %else + mov pGip, [rel NAME(g_pSUPGlobalInfoPage) wrt ..gotpcrel] + mov pGip, [pGip] + %endif +%endif + or pGip, pGip + jz .Rediscover + cmp dword [pGip + SUPGLOBALINFOPAGE.u32Magic], SUPGLOBALINFOPAGE_MAGIC + jne .Rediscover + + ; + ; pGipCPU, setting u32ApicIdPlus if necessary. + ; +%ifdef NEED_APIC_ID + ; u8ApicId = ASMGetApicId(); + mov eax, 1 + cpuid ; expensive + %ifdef NEED_TRANSACTION_ID + mov u32ApicIdPlus, ebx + %endif + ; pGipCPU = &pGip->aCPU[pGip->aiCpuFromApicId[u8ApicId]]; + shr ebx, 24 + movzx eax, word [pGip + rbx * 2 + SUPGLOBALINFOPAGE.aiCpuFromApicId] + imul eax, SUPGIPCPU_size + lea pGipCPU, [pGip + rax + SUPGLOBALINFOPAGE.aCPUs] +%endif + +%ifdef NEED_TRANSACTION_ID + ; + ; Serialized loading of u32TransactionId. + ; + %ifdef ASYNC_GIP + mov u32TransactionId, [pGipCPU + SUPGIPCPU.u32TransactionId] + %else + mov u32TransactionId, [pGip + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u32TransactionId] + %endif + %ifdef USE_LFENCE + lfence + %else + lock xor dword TmpVar, 0 + %endif +%endif + + ; + ; Load the data and TSC. + ; + mov u32UpdateIntervalNS, [pGip + SUPGLOBALINFOPAGE.u32UpdateIntervalNS] +%ifdef ASYNC_GIP + mov u32UpdateIntervalTSC, [pGipCPU + SUPGIPCPU.u32UpdateIntervalTSC] +%else + mov u32UpdateIntervalTSC, [pGip + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u32UpdateIntervalTSC] +%endif + + rdtsc + mov u64PrevNanoTS, [pData + RTTIMENANOTSDATA.pu64Prev] + mov u64PrevNanoTS, [u64PrevNanoTS] + shl rdx, 32 + or rax, rdx ; rax is u64RetNanoTS. +%ifdef WITH_TSC_DELTA + mov rdx, [pGipCPU + SUPGIPCPU.i64TSCDelta] + mov u64CurNanoTS, 0x7fffffffffffffff ; INT64_MAX - temporarily borrowing u64CurNanoTS + cmp rdx, u64CurNanoTS + je .TscDeltaNotValid + sub rax, rdx +.TscDeltaNotValid: +%endif +%ifdef u64SavedRetNanoTS ; doing this here may save a tick or so? + mov u64SavedRetNanoTS, rax +%endif + +%ifdef ASYNC_GIP + mov u64CurNanoTS, [pGipCPU + SUPGIPCPU.u64NanoTS] + mov u64TSC, [pGipCPU + SUPGIPCPU.u64TSC] ; transhes pGIP! +%else + mov u64CurNanoTS, [pGip + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u64NanoTS] + mov u64TSC, [pGip + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u64TSC] ; trashes pGipCPU! +%endif + + +%ifdef NEED_TRANSACTION_ID + ; + ; Check that the GIP and CPU didn't change. + ; + ; It is crucial that the rdtsc instruction has completed before + ; we check the transaction id. The LOCK prefixed instruction with + ; dependency on the RDTSC result should do the trick, I think. + ; CPUID is serializing, so the async path is safe by default. + ; + %ifdef NEED_APIC_ID + mov eax, 1 + cpuid + cmp u32ApicIdPlus, ebx + jne .ReadGip + %else + lock xor qword TmpVar, rax + %endif + %ifdef ASYNC_GIP + cmp u32TransactionId, [pGipCPU + SUPGIPCPU.u32TransactionId] + %else + cmp u32TransactionId, [pGip + SUPGLOBALINFOPAGE.aCPUs + SUPGIPCPU.u32TransactionId] + %endif + jne .ReadGip + test u32TransactionId, 1 + jnz .ReadGip + %ifdef u64SavedRetNanoTS + mov rax, u64SavedRetNanoTS ; rax is u64RetNanoTS. + %endif +%endif ; NEED_TRANSACTION_ID + + + ;; + ;; Calc the timestamp. + ;; + ; u64RetNanoTS -= u64TSC; + sub rax, u64TSC + xor edx, edx + + ; if (u64RetNanoTS > u32UpdateIntervalTSC) -> jump + cmp rax, u32UpdateIntervalTSC_64 + ja .OverFlow +.ContinueCalcs: ; edx = 0; eax <= u32UpdateIntervalTSC + mul u32UpdateIntervalNS + div u32UpdateIntervalTSC + + ; u64RetNanoTS += u64CurNanoTS; + add rax, u64CurNanoTS + + + ;; + ;; Compare it with the previous one. + ;; + ; if (RT_LIKELY( u64RetNanoTS > u64PrevNanoTS + ; && u64RetNanoTS < u64PrevNanoTS + UINT64_C(86000000000000) /* 24h */)) + ; /* Frequent - less than 24h since last call. */; + cmp rax, u64PrevNanoTS + jbe .DeltaPrevTooBig + mov ecx, 5 + shl rcx, 44 ; close enough + add rcx, u64PrevNanoTS + cmp rax, rcx + jae .DeltaPrevTooBig + + + ;; + ;; Update the previous value. + ;; +.Update: + ; if (RT_LIKELY(ASMAtomicCmpXchgU64(&pData->u64Prev, u64RetNanoTS, u64PrevNanoTS))) + mov rbx, [pData + RTTIMENANOTSDATA.pu64Prev] + mov rcx, rax + mov rax, u64PrevNanoTS + lock cmpxchg [rbx], rcx + jnz .UpdateFailed + +.Updated: + mov rax, rcx + +.Done: + mov rbx, SavedRBX + mov r12, SavedR12 + mov r13, SavedR13 +%ifdef SavedR14 + mov r14, SavedR14 +%endif +%ifdef SavedR15 + mov r15, SavedR15 +%endif +%ifdef ASM_CALL64_MSC + mov rdi, SavedRDI + mov rsi, SavedRSI +%endif + leave + ret + + + ;; + ;; We've expired the interval, cap it. If we're here for the 2nd + ;; time without any GIP update in-between, the checks against + ;; pData->u64Prev below will force 1ns stepping. + ;; +ALIGNCODE(16) +.OverFlow: + ; u64RetNanoTS = u32UpdateIntervalTSC; + inc dword [pData + RTTIMENANOTSDATA.cExpired] + mov eax, u32UpdateIntervalTSC + jmp .ContinueCalcs + + + ;; + ;; u64DeltaPrev >= 24h + ;; + ;; rax = u64RetNanoTS (to be adjusted) + ;; +ALIGNCODE(16) +.DeltaPrevTooBig: + ; uint64_t u64DeltaPrev = u64RetNanoTS - u64PrevNanoTS; + mov rbx, rax + sub rbx, u64PrevNanoTS + + ; else if ( (int64_t)u64DeltaPrev <= 0 + ; && (int64_t)u64DeltaPrev + u32UpdateIntervalNS * 2 >= 0) + ; { + ; /* Occasional - u64NanoTS is in the recent 'past' relative the previous call. */ + ; pData->c1nsSteps++; + ; u64RetNanoTS = u64PrevNanoTS + 1; + ; } + test rbx, rbx + jg .DeltaPrevNotInRecentPast + + lea rdx, [u32UpdateIntervalNS_64 + u32UpdateIntervalNS_64] + add rdx, rbx + js .DeltaPrevNotInRecentPast + + ; body + inc dword [pData + RTTIMENANOTSDATA.c1nsSteps] + lea rax, [u64PrevNanoTS + 1] + jmp .Update + + ; else if (!u64PrevNanoTS) /* We're resuming (see TMVirtualResume) / first call. */ + ; /* do nothing */; +.DeltaPrevNotInRecentPast: + or u64PrevNanoTS, u64PrevNanoTS + jz .Update + + ; else + ; { + ; /* Something has gone bust, if negative offset it's real bad. */ + ; rtTimeNanoTSInternalBitch(pVM, + ; } + + ; call C function that does the bitching. + mov TmpVar, rax + mov TmpVar2, pData + +%ifdef ASM_CALL64_MSC + mov rcx, pData ; param 1 - pData + mov rdx, rax ; param 2 - u64RetNanoTS + mov r8, rbx ; param 3 - u64DeltaPrev + mov r9, u64PrevNanoTS ; param 4 - u64PrevNanoTS +%else + ;mov rdi, pData - already in rdi; param 1 - pData + mov rsi, rax ; param 2 - u64RetNanoTS + mov rdx, rbx ; param 3 - u64DeltaPrev + mov rcx, u64PrevNanoTS ; param 4 - u64PrevNanoTS +%endif + call qword [pData + RTTIMENANOTSDATA.pfnBad] + + mov rax, TmpVar + mov pData, TmpVar2 + jmp .Update + + + ;; + ;; Attempt updating the previous value, provided we're still ahead of it. + ;; + ;; There is no point in recalculating u64NanoTS because we got preempted or if + ;; we raced somebody while the GIP was updated, since these are events + ;; that might occur at any point in the return path as well. + ;; + ;; rax = *pData->u64Prev; + ;; rcx = u64RetNanoTS + ;; +ALIGNCODE(16) +.UpdateFailed: + lock inc dword [pData + RTTIMENANOTSDATA.cUpdateRaces] + ; for (i = 0; i < 10; i++) + mov edx, 10 +.UpdateLoop: + ; if (u64PrevNanoTS >= u64RetNanoTS) + ; break; + cmp rax, rcx + jge .Updated +.UpdateLoopLess: + ; retry + lock cmpxchg [rbx], rcx + jz .Updated + dec edx + jnz .UpdateLoop + jmp .Updated + + + ;; + ;; The GIP is seemingly invalid, redo the discovery. + ;; +.Rediscover: +%ifdef ASM_CALL64_MSC + mov rcx, pData +%else + ; mov rdi, pData - already in rdi +%endif + call [pData + RTTIMENANOTSDATA.pfnRediscover] + jmp .Done + + + ; + ; Cleanup variables + ; +%undef SavedRBX +%undef SavedR12 +%undef SavedR13 +%undef SavedR14 +%undef SavedR15 +%undef SavedRDI +%undef SavedRSI +%undef pData +%undef TmpVar +%undef u64TSC +%undef pGip +%undef pGipCPU +%undef u32TransactionId +%undef u64CurNanoTS +%undef u64PrevNanoTS +%undef u32UpdateIntervalTSC +%undef u32UpdateIntervalTSC_64 +%undef u32UpdateIntervalNS +%undef u64SavedRetNanoTS +%undef u32ApicIdPlus + +%endif ; AMD64 +ENDPROC rtTimeNanoTSInternalAsm + -- cgit v1.2.3