/** * The global 1 halt method - Block in GMM (ring-0) and let it * try take care of the global scheduling of EMT threads. */ static DECLCALLBACK(int) vmR3HaltGlobal1Halt(PUVMCPU pUVCpu, const uint32_t fMask, uint64_t u64Now) { PUVM pUVM = pUVCpu->pUVM; PVMCPU pVCpu = pUVCpu->pVCpu; PVM pVM = pUVCpu->pVM; Assert(VMMGetCpu(pVM) == pVCpu); NOREF(u64Now); /* * Halt loop. */ //uint64_t u64NowLog, u64Start; //u64Start = u64NowLog = RTTimeNanoTS(); int rc = VINF_SUCCESS; ASMAtomicWriteBool(&pUVCpu->vm.s.fWait, true); unsigned cLoops = 0; for (;; cLoops++) { /* * Work the timers and check if we can exit. */ uint64_t const u64StartTimers = RTTimeNanoTS(); TMR3TimerQueuesDo(pVM); uint64_t const cNsElapsedTimers = RTTimeNanoTS() - u64StartTimers; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltTimers, cNsElapsedTimers); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; /* * Estimate time left to the next event. */ //u64NowLog = RTTimeNanoTS(); uint64_t u64Delta; uint64_t u64GipTime = TMTimerPollGIP(pVM, pVCpu, &u64Delta); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; /* * Block if we're not spinning and the interval isn't all that small. */ if (u64Delta >= pUVM->vm.s.Halt.Global1.cNsSpinBlockThresholdCfg) { VMMR3YieldStop(pVM); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; //RTLogPrintf("loop=%-3d u64GipTime=%'llu / %'llu now=%'llu / %'llu\n", cLoops, u64GipTime, u64Delta, u64NowLog, u64GipTime - u64NowLog); uint64_t const u64StartSchedHalt = RTTimeNanoTS(); rc = SUPR3CallVMMR0Ex(pVM->pVMR0, pVCpu->idCpu, VMMR0_DO_GVMM_SCHED_HALT, u64GipTime, NULL); uint64_t const u64EndSchedHalt = RTTimeNanoTS(); uint64_t const cNsElapsedSchedHalt = u64EndSchedHalt - u64StartSchedHalt; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlock, cNsElapsedSchedHalt); if (rc == VERR_INTERRUPTED) rc = VINF_SUCCESS; else if (RT_FAILURE(rc)) { rc = vmR3FatalWaitError(pUVCpu, "VMMR0_DO_GVMM_SCHED_HALT->%Rrc\n", rc); break; } else { int64_t const cNsOverslept = u64EndSchedHalt - u64GipTime; if (cNsOverslept > 50000) STAM_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlockOverslept, cNsOverslept); else if (cNsOverslept < -50000) STAM_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlockInsomnia, cNsElapsedSchedHalt); else STAM_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlockOnTime, cNsElapsedSchedHalt); } } /* * When spinning call upon the GVMM and do some wakups once * in a while, it's not like we're actually busy or anything. */ else if (!(cLoops & 0x1fff)) { uint64_t const u64StartSchedYield = RTTimeNanoTS(); rc = SUPR3CallVMMR0Ex(pVM->pVMR0, pVCpu->idCpu, VMMR0_DO_GVMM_SCHED_POLL, false /* don't yield */, NULL); uint64_t const cNsElapsedSchedYield = RTTimeNanoTS() - u64StartSchedYield; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltYield, cNsElapsedSchedYield); } } //RTLogPrintf("*** %u loops %'llu; lag=%RU64\n", cLoops, u64NowLog - u64Start, TMVirtualSyncGetLag(pVM)); ASMAtomicUoWriteBool(&pUVCpu->vm.s.fWait, false); return rc; }
/** * The old halt loop. */ static DECLCALLBACK(int) vmR3HaltOldDoHalt(PUVMCPU pUVCpu, const uint32_t fMask, uint64_t /* u64Now*/) { /* * Halt loop. */ PVM pVM = pUVCpu->pVM; PVMCPU pVCpu = pUVCpu->pVCpu; int rc = VINF_SUCCESS; ASMAtomicWriteBool(&pUVCpu->vm.s.fWait, true); //unsigned cLoops = 0; for (;;) { /* * Work the timers and check if we can exit. * The poll call gives us the ticks left to the next event in * addition to perhaps set an FF. */ uint64_t const u64StartTimers = RTTimeNanoTS(); TMR3TimerQueuesDo(pVM); uint64_t const cNsElapsedTimers = RTTimeNanoTS() - u64StartTimers; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltTimers, cNsElapsedTimers); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; uint64_t u64NanoTS; TMTimerPollGIP(pVM, pVCpu, &u64NanoTS); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; /* * Wait for a while. Someone will wake us up or interrupt the call if * anything needs our attention. */ if (u64NanoTS < 50000) { //RTLogPrintf("u64NanoTS=%RI64 cLoops=%d spin\n", u64NanoTS, cLoops++); /* spin */; } else { VMMR3YieldStop(pVM); //uint64_t u64Start = RTTimeNanoTS(); if (u64NanoTS < 870000) /* this is a bit speculative... works fine on linux. */ { //RTLogPrintf("u64NanoTS=%RI64 cLoops=%d yield", u64NanoTS, cLoops++); uint64_t const u64StartSchedYield = RTTimeNanoTS(); RTThreadYield(); /* this is the best we can do here */ uint64_t const cNsElapsedSchedYield = RTTimeNanoTS() - u64StartSchedYield; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltYield, cNsElapsedSchedYield); } else if (u64NanoTS < 2000000) { //RTLogPrintf("u64NanoTS=%RI64 cLoops=%d sleep 1ms", u64NanoTS, cLoops++); uint64_t const u64StartSchedHalt = RTTimeNanoTS(); rc = RTSemEventWait(pUVCpu->vm.s.EventSemWait, 1); uint64_t const cNsElapsedSchedHalt = RTTimeNanoTS() - u64StartSchedHalt; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlock, cNsElapsedSchedHalt); } else { //RTLogPrintf("u64NanoTS=%RI64 cLoops=%d sleep %dms", u64NanoTS, cLoops++, (uint32_t)RT_MIN((u64NanoTS - 500000) / 1000000, 15)); uint64_t const u64StartSchedHalt = RTTimeNanoTS(); rc = RTSemEventWait(pUVCpu->vm.s.EventSemWait, RT_MIN((u64NanoTS - 1000000) / 1000000, 15)); uint64_t const cNsElapsedSchedHalt = RTTimeNanoTS() - u64StartSchedHalt; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlock, cNsElapsedSchedHalt); } //uint64_t u64Slept = RTTimeNanoTS() - u64Start; //RTLogPrintf(" -> rc=%Rrc in %RU64 ns / %RI64 ns delta\n", rc, u64Slept, u64NanoTS - u64Slept); } if (rc == VERR_TIMEOUT) rc = VINF_SUCCESS; else if (RT_FAILURE(rc)) { rc = vmR3FatalWaitError(pUVCpu, "RTSemEventWait->%Rrc\n", rc); break; } } ASMAtomicUoWriteBool(&pUVCpu->vm.s.fWait, false); return rc; }
/** * Method 1 - Block whenever possible, and when lagging behind * switch to spinning for 10-30ms with occasional blocking until * the lag has been eliminated. */ static DECLCALLBACK(int) vmR3HaltMethod1Halt(PUVMCPU pUVCpu, const uint32_t fMask, uint64_t u64Now) { PUVM pUVM = pUVCpu->pUVM; PVMCPU pVCpu = pUVCpu->pVCpu; PVM pVM = pUVCpu->pVM; /* * To simplify things, we decide up-front whether we should switch to spinning or * not. This makes some ASSUMPTIONS about the cause of the spinning (PIT/RTC/PCNet) * and that it will generate interrupts or other events that will cause us to exit * the halt loop. */ bool fBlockOnce = false; bool fSpinning = false; uint32_t u32CatchUpPct = TMVirtualSyncGetCatchUpPct(pVM); if (u32CatchUpPct /* non-zero if catching up */) { if (pUVCpu->vm.s.Halt.Method12.u64StartSpinTS) { fSpinning = TMVirtualSyncGetLag(pVM) >= pUVM->vm.s.Halt.Method12.u32StopSpinningCfg; if (fSpinning) { uint64_t u64Lag = TMVirtualSyncGetLag(pVM); fBlockOnce = u64Now - pUVCpu->vm.s.Halt.Method12.u64LastBlockTS > RT_MAX(pUVM->vm.s.Halt.Method12.u32MinBlockIntervalCfg, RT_MIN(u64Lag / pUVM->vm.s.Halt.Method12.u32LagBlockIntervalDivisorCfg, pUVM->vm.s.Halt.Method12.u32MaxBlockIntervalCfg)); } else { //RTLogRelPrintf("Stopped spinning (%u ms)\n", (u64Now - pUVCpu->vm.s.Halt.Method12.u64StartSpinTS) / 1000000); pUVCpu->vm.s.Halt.Method12.u64StartSpinTS = 0; } } else { fSpinning = TMVirtualSyncGetLag(pVM) >= pUVM->vm.s.Halt.Method12.u32StartSpinningCfg; if (fSpinning) pUVCpu->vm.s.Halt.Method12.u64StartSpinTS = u64Now; } } else if (pUVCpu->vm.s.Halt.Method12.u64StartSpinTS) { //RTLogRelPrintf("Stopped spinning (%u ms)\n", (u64Now - pUVCpu->vm.s.Halt.Method12.u64StartSpinTS) / 1000000); pUVCpu->vm.s.Halt.Method12.u64StartSpinTS = 0; } /* * Halt loop. */ int rc = VINF_SUCCESS; ASMAtomicWriteBool(&pUVCpu->vm.s.fWait, true); unsigned cLoops = 0; for (;; cLoops++) { /* * Work the timers and check if we can exit. */ uint64_t const u64StartTimers = RTTimeNanoTS(); TMR3TimerQueuesDo(pVM); uint64_t const cNsElapsedTimers = RTTimeNanoTS() - u64StartTimers; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltTimers, cNsElapsedTimers); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; /* * Estimate time left to the next event. */ uint64_t u64NanoTS; TMTimerPollGIP(pVM, pVCpu, &u64NanoTS); if ( VM_FF_ISPENDING(pVM, VM_FF_EXTERNAL_HALTED_MASK) || VMCPU_FF_ISPENDING(pVCpu, fMask)) break; /* * Block if we're not spinning and the interval isn't all that small. */ if ( ( !fSpinning || fBlockOnce) #if 1 /* DEBUGGING STUFF - REMOVE LATER */ && u64NanoTS >= 100000) /* 0.100 ms */ #else && u64NanoTS >= 250000) /* 0.250 ms */ #endif { const uint64_t Start = pUVCpu->vm.s.Halt.Method12.u64LastBlockTS = RTTimeNanoTS(); VMMR3YieldStop(pVM); uint32_t cMilliSecs = RT_MIN(u64NanoTS / 1000000, 15); if (cMilliSecs <= pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLongAvg) cMilliSecs = 1; else cMilliSecs -= pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLongAvg; //RTLogRelPrintf("u64NanoTS=%RI64 cLoops=%3d sleep %02dms (%7RU64) ", u64NanoTS, cLoops, cMilliSecs, u64NanoTS); uint64_t const u64StartSchedHalt = RTTimeNanoTS(); rc = RTSemEventWait(pUVCpu->vm.s.EventSemWait, cMilliSecs); uint64_t const cNsElapsedSchedHalt = RTTimeNanoTS() - u64StartSchedHalt; STAM_REL_PROFILE_ADD_PERIOD(&pUVCpu->vm.s.StatHaltBlock, cNsElapsedSchedHalt); if (rc == VERR_TIMEOUT) rc = VINF_SUCCESS; else if (RT_FAILURE(rc)) { rc = vmR3FatalWaitError(pUVCpu, "RTSemEventWait->%Rrc\n", rc); break; } /* * Calc the statistics. * Update averages every 16th time, and flush parts of the history every 64th time. */ const uint64_t Elapsed = RTTimeNanoTS() - Start; pUVCpu->vm.s.Halt.Method12.cNSBlocked += Elapsed; if (Elapsed > u64NanoTS) pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLong += Elapsed - u64NanoTS; pUVCpu->vm.s.Halt.Method12.cBlocks++; if (!(pUVCpu->vm.s.Halt.Method12.cBlocks & 0xf)) { pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLongAvg = pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLong / pUVCpu->vm.s.Halt.Method12.cBlocks; if (!(pUVCpu->vm.s.Halt.Method12.cBlocks & 0x3f)) { pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLong = pUVCpu->vm.s.Halt.Method12.cNSBlockedTooLongAvg * 0x40; pUVCpu->vm.s.Halt.Method12.cBlocks = 0x40; } } //RTLogRelPrintf(" -> %7RU64 ns / %7RI64 ns delta%s\n", Elapsed, Elapsed - u64NanoTS, fBlockOnce ? " (block once)" : ""); /* * Clear the block once flag if we actually blocked. */ if ( fBlockOnce && Elapsed > 100000 /* 0.1 ms */) fBlockOnce = false; } } //if (fSpinning) RTLogRelPrintf("spun for %RU64 ns %u loops; lag=%RU64 pct=%d\n", RTTimeNanoTS() - u64Now, cLoops, TMVirtualSyncGetLag(pVM), u32CatchUpPct); ASMAtomicUoWriteBool(&pUVCpu->vm.s.fWait, false); return rc; }
/** * This is called on each EMT and will beat TM. * * @returns VINF_SUCCESS, test failure is reported via RTTEST. * @param pVM Pointer to the VM. * @param hTest The test handle. */ DECLCALLBACK(int) tstTMWorker(PVM pVM, RTTEST hTest) { VMCPUID idCpu = VMMGetCpuId(pVM); RTTestPrintfNl(hTest, RTTESTLVL_ALWAYS, "idCpu=%d STARTING\n", idCpu); /* * Create the test set. */ int rc; PTMTIMER apTimers[5]; for (size_t i = 0; i < RT_ELEMENTS(apTimers); i++) { rc = TMR3TimerCreateInternal(pVM, i & 1 ? TMCLOCK_VIRTUAL : TMCLOCK_VIRTUAL_SYNC, tstTMDummyCallback, NULL, "test timer", &apTimers[i]); RTTEST_CHECK_RET(hTest, RT_SUCCESS(rc), rc); } /* * The run loop. */ unsigned uPrevPct = 0; uint32_t const cLoops = 100000; for (uint32_t iLoop = 0; iLoop < cLoops; iLoop++) { size_t cLeft = RT_ELEMENTS(apTimers); unsigned i = iLoop % RT_ELEMENTS(apTimers); while (cLeft-- > 0) { PTMTIMER pTimer = apTimers[i]; if ( cLeft == RT_ELEMENTS(apTimers) / 2 && TMTimerIsActive(pTimer)) { rc = TMTimerStop(pTimer); RTTEST_CHECK_MSG(hTest, RT_SUCCESS(rc), (hTest, "TMTimerStop: %Rrc\n", rc)); } else { rc = TMTimerSetMicro(pTimer, 50 + cLeft); RTTEST_CHECK_MSG(hTest, RT_SUCCESS(rc), (hTest, "TMTimerSetMicro: %Rrc\n", rc)); } /* next */ i = (i + 1) % RT_ELEMENTS(apTimers); } if (i % 3) TMR3TimerQueuesDo(pVM); /* Progress report. */ unsigned uPct = (unsigned)(100.0 * iLoop / cLoops); if (uPct != uPrevPct) { uPrevPct = uPct; if (!(uPct % 10)) RTTestPrintfNl(hTest, RTTESTLVL_ALWAYS, "idCpu=%d - %3u%%\n", idCpu, uPct); } } RTTestPrintfNl(hTest, RTTESTLVL_ALWAYS, "idCpu=%d DONE\n", idCpu); return 0; }