SUPDrvGip.cpp@ 78363

Last change on this file since 78363 was 77727, checked in by vboxsync, 6 years ago
IPRT,HostDrivers: Fixed some warnings.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 180.9 KB

Line
1	/* $Id: SUPDrvGip.cpp 77727 2019-03-15 14:14:18Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code for GIP.
4	*/
5
6	/*
7	* Copyright (C) 2006-2019 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*********************************************************************************************************************************
29	* Header Files *
30	*********************************************************************************************************************************/
31	#define LOG_GROUP LOG_GROUP_SUP_DRV
32	#define SUPDRV_AGNOSTIC
33	#include "SUPDrvInternal.h"
34	#ifndef PAGE_SHIFT
35	# include <iprt/param.h>
36	#endif
37	#include <iprt/asm.h>
38	#include <iprt/asm-amd64-x86.h>
39	#include <iprt/asm-math.h>
40	#include <iprt/cpuset.h>
41	#include <iprt/handletable.h>
42	#include <iprt/mem.h>
43	#include <iprt/mp.h>
44	#include <iprt/power.h>
45	#include <iprt/process.h>
46	#include <iprt/semaphore.h>
47	#include <iprt/spinlock.h>
48	#include <iprt/thread.h>
49	#include <iprt/uuid.h>
50	#include <iprt/net.h>
51	#include <iprt/crc.h>
52	#include <iprt/string.h>
53	#include <iprt/timer.h>
54	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_FREEBSD)
55	# include <iprt/rand.h>
56	# include <iprt/path.h>
57	#endif
58	#include <iprt/uint128.h>
59	#include <iprt/x86.h>
60
61	#include <VBox/param.h>
62	#include <VBox/log.h>
63	#include <VBox/err.h>
64
65	#if defined(RT_OS_SOLARIS) \|\| defined(RT_OS_DARWIN)
66	# include "dtrace/SUPDrv.h"
67	#else
68	/* ... */
69	#endif
70
71
72	/*********************************************************************************************************************************
73	* Defined Constants And Macros *
74	*********************************************************************************************************************************/
75	/** The frequency by which we recalculate the u32UpdateHz and
76	* u32UpdateIntervalNS GIP members. The value must be a power of 2.
77	*
78	* Warning: Bumping this too high might overflow u32UpdateIntervalNS.
79	*/
80	#define GIP_UPDATEHZ_RECALC_FREQ 0x800
81
82	/** A reserved TSC value used for synchronization as well as measurement of
83	* TSC deltas. */
84	#define GIP_TSC_DELTA_RSVD UINT64_MAX
85	/** The number of TSC delta measurement loops in total (includes primer and
86	* read-time loops). */
87	#define GIP_TSC_DELTA_LOOPS 96
88	/** The number of cache primer loops. */
89	#define GIP_TSC_DELTA_PRIMER_LOOPS 4
90	/** The number of loops until we keep computing the minumum read time. */
91	#define GIP_TSC_DELTA_READ_TIME_LOOPS 24
92
93	/** The TSC frequency refinement period in seconds.
94	* The timer fires after 200ms, then every second, this value just says when
95	* to stop it after that. */
96	#define GIP_TSC_REFINE_PERIOD_IN_SECS 12
97	/** The TSC-delta threshold for the SUPGIPUSETSCDELTA_PRACTICALLY_ZERO rating */
98	#define GIP_TSC_DELTA_THRESHOLD_PRACTICALLY_ZERO 32
99	/** The TSC-delta threshold for the SUPGIPUSETSCDELTA_ROUGHLY_ZERO rating */
100	#define GIP_TSC_DELTA_THRESHOLD_ROUGHLY_ZERO 448
101	/** The TSC delta value for the initial GIP master - 0 in regular builds.
102	* To test the delta code this can be set to a non-zero value. */
103	#if 0
104	# define GIP_TSC_DELTA_INITIAL_MASTER_VALUE INT64_C(170139095182512) /* 0x00009abd9854acb0 */
105	#else
106	# define GIP_TSC_DELTA_INITIAL_MASTER_VALUE INT64_C(0)
107	#endif
108
109	AssertCompile(GIP_TSC_DELTA_PRIMER_LOOPS < GIP_TSC_DELTA_READ_TIME_LOOPS);
110	AssertCompile(GIP_TSC_DELTA_PRIMER_LOOPS + GIP_TSC_DELTA_READ_TIME_LOOPS < GIP_TSC_DELTA_LOOPS);
111
112	/** @def VBOX_SVN_REV
113	* The makefile should define this if it can. */
114	#ifndef VBOX_SVN_REV
115	# define VBOX_SVN_REV 0
116	#endif
117
118	#if 0 /* Don't start the GIP timers. Useful when debugging the IPRT timer code. */
119	# define DO_NOT_START_GIP
120	#endif
121
122
123	/*********************************************************************************************************************************
124	* Internal Functions *
125	*********************************************************************************************************************************/
126	static DECLCALLBACK(void) supdrvGipSyncAndInvariantTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
127	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
128	static int supdrvGipSetFlags(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, uint32_t fOrMask, uint32_t fAndMask);
129	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS, uint64_t uCpuHz);
130	static void supdrvTscResetSamples(PSUPDRVDEVEXT pDevExt, bool fClearDeltas);
131	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
132	static int supdrvTscDeltaThreadInit(PSUPDRVDEVEXT pDevExt);
133	static void supdrvTscDeltaTerm(PSUPDRVDEVEXT pDevExt);
134	static void supdrvTscDeltaThreadStartMeasurement(PSUPDRVDEVEXT pDevExt, bool fForceAll);
135	#else
136	static int supdrvTscMeasureInitialDeltas(PSUPDRVDEVEXT pDevExt);
137	static int supdrvTscMeasureDeltaOne(PSUPDRVDEVEXT pDevExt, uint32_t idxWorker);
138	#endif
139
140
141	/*********************************************************************************************************************************
142	* Global Variables *
143	*********************************************************************************************************************************/
144	DECLEXPORT(PSUPGLOBALINFOPAGE) g_pSUPGlobalInfoPage = NULL;
145
146
147
148	/*
149	*
150	* Misc Common GIP Code
151	* Misc Common GIP Code
152	* Misc Common GIP Code
153	*
154	*
155	*/
156
157
158	/**
159	* Finds the GIP CPU index corresponding to @a idCpu.
160	*
161	* @returns GIP CPU array index, UINT32_MAX if not found.
162	* @param pGip The GIP.
163	* @param idCpu The CPU ID.
164	*/
165	static uint32_t supdrvGipFindCpuIndexForCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
166	{
167	uint32_t i;
168	for (i = 0; i < pGip->cCpus; i++)
169	if (pGip->aCPUs[i].idCpu == idCpu)
170	return i;
171	return UINT32_MAX;
172	}
173
174
175
176	/*
177	*
178	* GIP Mapping and Unmapping Related Code.
179	* GIP Mapping and Unmapping Related Code.
180	* GIP Mapping and Unmapping Related Code.
181	*
182	*
183	*/
184
185
186	/**
187	* (Re-)initializes the per-cpu structure prior to starting or resuming the GIP
188	* updating.
189	*
190	* @param pGipCpu The per CPU structure for this CPU.
191	* @param u64NanoTS The current time.
192	*/
193	static void supdrvGipReInitCpu(PSUPGIPCPU pGipCpu, uint64_t u64NanoTS)
194	{
195	/*
196	* Here we don't really care about applying the TSC delta. The re-initialization of this
197	* value is not relevant especially while (re)starting the GIP as the first few ones will
198	* be ignored anyway, see supdrvGipDoUpdateCpu().
199	*/
200	pGipCpu->u64TSC = ASMReadTSC() - pGipCpu->u32UpdateIntervalTSC;
201	pGipCpu->u64NanoTS = u64NanoTS;
202	}
203
204
205	/**
206	* Set the current TSC and NanoTS value for the CPU.
207	*
208	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
209	* @param pvUser1 Pointer to the ring-0 GIP mapping.
210	* @param pvUser2 Pointer to the variable holding the current time.
211	*/
212	static DECLCALLBACK(void) supdrvGipReInitCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
213	{
214	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser1;
215	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
216
217	if (RT_LIKELY(iCpu < pGip->cCpus && pGip->aCPUs[iCpu].idCpu == idCpu))
218	supdrvGipReInitCpu(&pGip->aCPUs[iCpu], (uint64_t )pvUser2);
219
220	NOREF(pvUser2);
221	NOREF(idCpu);
222	}
223
224
225	/**
226	* State structure for supdrvGipDetectGetGipCpuCallback.
227	*/
228	typedef struct SUPDRVGIPDETECTGETCPU
229	{
230	/** Bitmap of APIC IDs that has been seen (initialized to zero).
231	* Used to detect duplicate APIC IDs (paranoia). */
232	uint8_t volatile bmApicId[256 / 8];
233	/** Mask of supported GIP CPU getter methods (SUPGIPGETCPU_XXX) (all bits set
234	* initially). The callback clears the methods not detected. */
235	uint32_t volatile fSupported;
236	/** The first callback detecting any kind of range issues (initialized to
237	* NIL_RTCPUID). */
238	RTCPUID volatile idCpuProblem;
239	} SUPDRVGIPDETECTGETCPU;
240	/** Pointer to state structure for supdrvGipDetectGetGipCpuCallback. */
241	typedef SUPDRVGIPDETECTGETCPU *PSUPDRVGIPDETECTGETCPU;
242
243
244	/**
245	* Checks for alternative ways of getting the CPU ID.
246	*
247	* This also checks the APIC ID, CPU ID and CPU set index values against the
248	* GIP tables.
249	*
250	* @param idCpu The CPU ID. Unused - we have to use the APIC ID.
251	* @param pvUser1 Pointer to the state structure.
252	* @param pvUser2 Pointer to the GIP.
253	*/
254	static DECLCALLBACK(void) supdrvGipDetectGetGipCpuCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
255	{
256	PSUPDRVGIPDETECTGETCPU pState = (PSUPDRVGIPDETECTGETCPU)pvUser1;
257	PSUPGLOBALINFOPAGE pGip = (PSUPGLOBALINFOPAGE)pvUser2;
258	uint32_t fSupported = 0;
259	uint16_t idApic;
260	int iCpuSet;
261	NOREF(pGip);
262
263	AssertMsg(idCpu == RTMpCpuId(), ("idCpu=%#x RTMpCpuId()=%#x\n", idCpu, RTMpCpuId())); /* paranoia^3 */
264
265	/*
266	* Check that the CPU ID and CPU set index are interchangable.
267	*/
268	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
269	if ((RTCPUID)iCpuSet == idCpu)
270	{
271	AssertCompile(RT_IS_POWER_OF_TWO(RTCPUSET_MAX_CPUS));
272	if ( iCpuSet >= 0
273	&& iCpuSet < RTCPUSET_MAX_CPUS
274	&& RT_IS_POWER_OF_TWO(RTCPUSET_MAX_CPUS))
275	{
276	PSUPGIPCPU pGipCpu = SUPGetGipCpuBySetIndex(pGip, iCpuSet);
277
278	/*
279	* Check whether the IDTR.LIMIT contains a CPU number.
280	*/
281	#ifdef RT_ARCH_X86
282	uint16_t const cbIdt = sizeof(X86DESC64SYSTEM) * 256;
283	#else
284	uint16_t const cbIdt = sizeof(X86DESCGATE) * 256;
285	#endif
286	RTIDTR Idtr;
287	ASMGetIDTR(&Idtr);
288	if (Idtr.cbIdt >= cbIdt)
289	{
290	uint32_t uTmp = Idtr.cbIdt - cbIdt;
291	uTmp &= RTCPUSET_MAX_CPUS - 1;
292	if (uTmp == idCpu)
293	{
294	RTIDTR Idtr2;
295	ASMGetIDTR(&Idtr2);
296	if (Idtr2.cbIdt == Idtr.cbIdt)
297	fSupported \|= SUPGIPGETCPU_IDTR_LIMIT_MASK_MAX_SET_CPUS;
298	}
299	}
300
301	/*
302	* Check whether RDTSCP is an option.
303	*/
304	if (ASMHasCpuId())
305	{
306	if ( ASMIsValidExtRange(ASMCpuId_EAX(UINT32_C(0x80000000)))
307	&& (ASMCpuId_EDX(UINT32_C(0x80000001)) & X86_CPUID_EXT_FEATURE_EDX_RDTSCP) )
308	{
309	uint32_t uAux;
310	ASMReadTscWithAux(&uAux);
311	if ((uAux & (RTCPUSET_MAX_CPUS - 1)) == idCpu)
312	{
313	ASMNopPause();
314	ASMReadTscWithAux(&uAux);
315	if ((uAux & (RTCPUSET_MAX_CPUS - 1)) == idCpu)
316	fSupported \|= SUPGIPGETCPU_RDTSCP_MASK_MAX_SET_CPUS;
317	}
318
319	if (pGipCpu)
320	{
321	uint32_t const uGroupedAux = (uint8_t)pGipCpu->iCpuGroupMember \| ((uint32_t)pGipCpu->iCpuGroup << 8);
322	if ( (uAux & UINT16_MAX) == uGroupedAux
323	&& pGipCpu->iCpuGroupMember <= UINT8_MAX)
324	{
325	ASMNopPause();
326	ASMReadTscWithAux(&uAux);
327	if ((uAux & UINT16_MAX) == uGroupedAux)
328	fSupported \|= SUPGIPGETCPU_RDTSCP_GROUP_IN_CH_NUMBER_IN_CL;
329	}
330	}
331	}
332	}
333	}
334	}
335
336	/*
337	* Check that the APIC ID is unique.
338	*/
339	idApic = ASMGetApicId();
340	if (RT_LIKELY( idApic < RT_ELEMENTS(pGip->aiCpuFromApicId)
341	&& !ASMAtomicBitTestAndSet(pState->bmApicId, idApic)))
342	fSupported \|= SUPGIPGETCPU_APIC_ID;
343	else
344	{
345	AssertCompile(sizeof(pState->bmApicId) * 8 == RT_ELEMENTS(pGip->aiCpuFromApicId));
346	ASMAtomicCmpXchgU32(&pState->idCpuProblem, idCpu, NIL_RTCPUID);
347	LogRel(("supdrvGipDetectGetGipCpuCallback: idCpu=%#x iCpuSet=%d idApic=%#x - duplicate APIC ID.\n",
348	idCpu, iCpuSet, idApic));
349	}
350
351	/*
352	* Check that the iCpuSet is within the expected range.
353	*/
354	if (RT_UNLIKELY( iCpuSet < 0
355	\|\| (unsigned)iCpuSet >= RTCPUSET_MAX_CPUS
356	\|\| (unsigned)iCpuSet >= RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)))
357	{
358	ASMAtomicCmpXchgU32(&pState->idCpuProblem, idCpu, NIL_RTCPUID);
359	LogRel(("supdrvGipDetectGetGipCpuCallback: idCpu=%#x iCpuSet=%d idApic=%#x - CPU set index is out of range.\n",
360	idCpu, iCpuSet, idApic));
361	}
362	else
363	{
364	RTCPUID idCpu2 = RTMpCpuIdFromSetIndex(iCpuSet);
365	if (RT_UNLIKELY(idCpu2 != idCpu))
366	{
367	ASMAtomicCmpXchgU32(&pState->idCpuProblem, idCpu, NIL_RTCPUID);
368	LogRel(("supdrvGipDetectGetGipCpuCallback: idCpu=%#x iCpuSet=%d idApic=%#x - CPU id/index roundtrip problem: %#x\n",
369	idCpu, iCpuSet, idApic, idCpu2));
370	}
371	}
372
373	/*
374	* Update the supported feature mask before we return.
375	*/
376	ASMAtomicAndU32(&pState->fSupported, fSupported);
377
378	NOREF(pvUser2);
379	}
380
381
382	/**
383	* Increase the timer freqency on hosts where this is possible (NT).
384	*
385	* The idea is that more interrupts is better for us... Also, it's better than
386	* we increase the timer frequence, because we might end up getting inaccurate
387	* callbacks if someone else does it.
388	*
389	* @param pDevExt Sets u32SystemTimerGranularityGrant if increased.
390	*/
391	static void supdrvGipRequestHigherTimerFrequencyFromSystem(PSUPDRVDEVEXT pDevExt)
392	{
393	if (pDevExt->u32SystemTimerGranularityGrant == 0)
394	{
395	uint32_t u32SystemResolution;
396	if ( RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 976563 /* 1024 HZ */, &u32SystemResolution))
397	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1000000 /* 1000 HZ */, &u32SystemResolution))
398	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 1953125 /* 512 HZ */, &u32SystemResolution))
399	\|\| RT_SUCCESS_NP(RTTimerRequestSystemGranularity( 2000000 /* 500 HZ */, &u32SystemResolution))
400	)
401	{
402	#if 0 /* def VBOX_STRICT - this is somehow triggers bogus assertions on windows 10 */
403	uint32_t u32After = RTTimerGetSystemGranularity();
404	AssertMsg(u32After <= u32SystemResolution, ("u32After=%u u32SystemResolution=%u\n", u32After, u32SystemResolution));
405	#endif
406	pDevExt->u32SystemTimerGranularityGrant = u32SystemResolution;
407	}
408	}
409	}
410
411
412	/**
413	* Undoes supdrvGipRequestHigherTimerFrequencyFromSystem.
414	*
415	* @param pDevExt Clears u32SystemTimerGranularityGrant.
416	*/
417	static void supdrvGipReleaseHigherTimerFrequencyFromSystem(PSUPDRVDEVEXT pDevExt)
418	{
419	if (pDevExt->u32SystemTimerGranularityGrant)
420	{
421	int rc2 = RTTimerReleaseSystemGranularity(pDevExt->u32SystemTimerGranularityGrant);
422	AssertRC(rc2);
423	pDevExt->u32SystemTimerGranularityGrant = 0;
424	}
425	}
426
427
428	/**
429	* Maps the GIP into userspace and/or get the physical address of the GIP.
430	*
431	* @returns IPRT status code.
432	* @param pSession Session to which the GIP mapping should belong.
433	* @param ppGipR3 Where to store the address of the ring-3 mapping. (optional)
434	* @param pHCPhysGip Where to store the physical address. (optional)
435	*
436	* @remark There is no reference counting on the mapping, so one call to this function
437	* count globally as one reference. One call to SUPR0GipUnmap() is will unmap GIP
438	* and remove the session as a GIP user.
439	*/
440	SUPR0DECL(int) SUPR0GipMap(PSUPDRVSESSION pSession, PRTR3PTR ppGipR3, PRTHCPHYS pHCPhysGip)
441	{
442	int rc;
443	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
444	RTR3PTR pGipR3 = NIL_RTR3PTR;
445	RTHCPHYS HCPhys = NIL_RTHCPHYS;
446	LogFlow(("SUPR0GipMap: pSession=%p ppGipR3=%p pHCPhysGip=%p\n", pSession, ppGipR3, pHCPhysGip));
447
448	/*
449	* Validate
450	*/
451	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
452	AssertPtrNullReturn(ppGipR3, VERR_INVALID_POINTER);
453	AssertPtrNullReturn(pHCPhysGip, VERR_INVALID_POINTER);
454
455	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
456	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
457	#else
458	RTSemFastMutexRequest(pDevExt->mtxGip);
459	#endif
460	if (pDevExt->pGip)
461	{
462	/*
463	* Map it?
464	*/
465	rc = VINF_SUCCESS;
466	if (ppGipR3)
467	{
468	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
469	rc = RTR0MemObjMapUser(&pSession->GipMapObjR3, pDevExt->GipMemObj, (RTR3PTR)-1, 0,
470	RTMEM_PROT_READ, NIL_RTR0PROCESS);
471	if (RT_SUCCESS(rc))
472	pGipR3 = RTR0MemObjAddressR3(pSession->GipMapObjR3);
473	}
474
475	/*
476	* Get physical address.
477	*/
478	if (pHCPhysGip && RT_SUCCESS(rc))
479	HCPhys = pDevExt->HCPhysGip;
480
481	/*
482	* Reference globally.
483	*/
484	if (!pSession->fGipReferenced && RT_SUCCESS(rc))
485	{
486	pSession->fGipReferenced = 1;
487	pDevExt->cGipUsers++;
488	if (pDevExt->cGipUsers == 1)
489	{
490	PSUPGLOBALINFOPAGE pGipR0 = pDevExt->pGip;
491	uint64_t u64NanoTS;
492
493	/*
494	* GIP starts/resumes updating again. On windows we bump the
495	* host timer frequency to make sure we don't get stuck in guest
496	* mode and to get better timer (and possibly clock) accuracy.
497	*/
498	LogFlow(("SUPR0GipMap: Resumes GIP updating\n"));
499
500	supdrvGipRequestHigherTimerFrequencyFromSystem(pDevExt);
501
502	/*
503	* document me
504	*/
505	if (pGipR0->aCPUs[0].u32TransactionId != 2 /* not the first time */)
506	{
507	unsigned i;
508	for (i = 0; i < pGipR0->cCpus; i++)
509	ASMAtomicUoWriteU32(&pGipR0->aCPUs[i].u32TransactionId,
510	(pGipR0->aCPUs[i].u32TransactionId + GIP_UPDATEHZ_RECALC_FREQ * 2)
511	& ~(GIP_UPDATEHZ_RECALC_FREQ * 2 - 1));
512	ASMAtomicWriteU64(&pGipR0->u64NanoTSLastUpdateHz, 0);
513	}
514
515	/*
516	* document me
517	*/
518	u64NanoTS = RTTimeSystemNanoTS() - pGipR0->u32UpdateIntervalNS;
519	if ( pGipR0->u32Mode == SUPGIPMODE_INVARIANT_TSC
520	\|\| pGipR0->u32Mode == SUPGIPMODE_SYNC_TSC
521	\|\| RTMpGetOnlineCount() == 1)
522	supdrvGipReInitCpu(&pGipR0->aCPUs[0], u64NanoTS);
523	else
524	RTMpOnAll(supdrvGipReInitCpuCallback, pGipR0, &u64NanoTS);
525
526	/*
527	* Detect alternative ways to figure the CPU ID in ring-3 and
528	* raw-mode context. Check the sanity of the APIC IDs, CPU IDs,
529	* and CPU set indexes while we're at it.
530	*/
531	if (RT_SUCCESS(rc))
532	{
533	SUPDRVGIPDETECTGETCPU DetectState;
534	RT_BZERO((void *)&DetectState.bmApicId, sizeof(DetectState.bmApicId));
535	DetectState.fSupported = UINT32_MAX;
536	DetectState.idCpuProblem = NIL_RTCPUID;
537	rc = RTMpOnAll(supdrvGipDetectGetGipCpuCallback, &DetectState, pGipR0);
538	if (DetectState.idCpuProblem == NIL_RTCPUID)
539	{
540	if ( DetectState.fSupported != UINT32_MAX
541	&& DetectState.fSupported != 0)
542	{
543	if (pGipR0->fGetGipCpu != DetectState.fSupported)
544	{
545	pGipR0->fGetGipCpu = DetectState.fSupported;
546	LogRel(("SUPR0GipMap: fGetGipCpu=%#x\n", DetectState.fSupported));
547	}
548	}
549	else
550	{
551	LogRel(("SUPR0GipMap: No supported ways of getting the APIC ID or CPU number in ring-3! (%#x)\n",
552	DetectState.fSupported));
553	rc = VERR_UNSUPPORTED_CPU;
554	}
555	}
556	else
557	{
558	LogRel(("SUPR0GipMap: APIC ID, CPU ID or CPU set index problem detected on CPU #%u (%#x)!\n",
559	DetectState.idCpuProblem, DetectState.idCpuProblem));
560	rc = VERR_INVALID_CPU_ID;
561	}
562	}
563
564	/*
565	* Start the GIP timer if all is well..
566	*/
567	if (RT_SUCCESS(rc))
568	{
569	#ifndef DO_NOT_START_GIP
570	rc = RTTimerStart(pDevExt->pGipTimer, 0 /* fire ASAP */); AssertRC(rc);
571	#endif
572	rc = VINF_SUCCESS;
573	}
574
575	/*
576	* Bail out on error.
577	*/
578	if (RT_FAILURE(rc))
579	{
580	LogRel(("SUPR0GipMap: failed rc=%Rrc\n", rc));
581	pDevExt->cGipUsers = 0;
582	pSession->fGipReferenced = 0;
583	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
584	{
585	int rc2 = RTR0MemObjFree(pSession->GipMapObjR3, false); AssertRC(rc2);
586	if (RT_SUCCESS(rc2))
587	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
588	}
589	HCPhys = NIL_RTHCPHYS;
590	pGipR3 = NIL_RTR3PTR;
591	}
592	}
593	}
594	}
595	else
596	{
597	rc = VERR_GENERAL_FAILURE;
598	Log(("SUPR0GipMap: GIP is not available!\n"));
599	}
600	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
601	RTSemMutexRelease(pDevExt->mtxGip);
602	#else
603	RTSemFastMutexRelease(pDevExt->mtxGip);
604	#endif
605
606	/*
607	* Write returns.
608	*/
609	if (pHCPhysGip)
610	*pHCPhysGip = HCPhys;
611	if (ppGipR3)
612	*ppGipR3 = pGipR3;
613
614	#ifdef DEBUG_DARWIN_GIP
615	OSDBGPRINT(("SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
616	#else
617	LogFlow(( "SUPR0GipMap: returns %d pHCPhysGip=%lx pGipR3=%p\n", rc, (unsigned long)HCPhys, (void )pGipR3));
618	#endif
619	return rc;
620	}
621
622
623	/**
624	* Unmaps any user mapping of the GIP and terminates all GIP access
625	* from this session.
626	*
627	* @returns IPRT status code.
628	* @param pSession Session to which the GIP mapping should belong.
629	*/
630	SUPR0DECL(int) SUPR0GipUnmap(PSUPDRVSESSION pSession)
631	{
632	int rc = VINF_SUCCESS;
633	PSUPDRVDEVEXT pDevExt = pSession->pDevExt;
634	#ifdef DEBUG_DARWIN_GIP
635	OSDBGPRINT(("SUPR0GipUnmap: pSession=%p pGip=%p GipMapObjR3=%p\n",
636	pSession,
637	pSession->GipMapObjR3 != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pSession->GipMapObjR3) : NULL,
638	pSession->GipMapObjR3));
639	#else
640	LogFlow(("SUPR0GipUnmap: pSession=%p\n", pSession));
641	#endif
642	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
643
644	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
645	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
646	#else
647	RTSemFastMutexRequest(pDevExt->mtxGip);
648	#endif
649
650	/*
651	* GIP test-mode session?
652	*/
653	if ( pSession->fGipTestMode
654	&& pDevExt->pGip)
655	{
656	supdrvGipSetFlags(pDevExt, pSession, 0, ~SUPGIP_FLAGS_TESTING_ENABLE);
657	Assert(!pSession->fGipTestMode);
658	}
659
660	/*
661	* Unmap anything?
662	*/
663	if (pSession->GipMapObjR3 != NIL_RTR0MEMOBJ)
664	{
665	rc = RTR0MemObjFree(pSession->GipMapObjR3, false);
666	AssertRC(rc);
667	if (RT_SUCCESS(rc))
668	pSession->GipMapObjR3 = NIL_RTR0MEMOBJ;
669	}
670
671	/*
672	* Dereference global GIP.
673	*/
674	if (pSession->fGipReferenced && !rc)
675	{
676	pSession->fGipReferenced = 0;
677	if ( pDevExt->cGipUsers > 0
678	&& !--pDevExt->cGipUsers)
679	{
680	LogFlow(("SUPR0GipUnmap: Suspends GIP updating\n"));
681	#ifndef DO_NOT_START_GIP
682	rc = RTTimerStop(pDevExt->pGipTimer); AssertRC(rc); rc = VINF_SUCCESS;
683	#endif
684	supdrvGipReleaseHigherTimerFrequencyFromSystem(pDevExt);
685	}
686	}
687
688	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
689	RTSemMutexRelease(pDevExt->mtxGip);
690	#else
691	RTSemFastMutexRelease(pDevExt->mtxGip);
692	#endif
693
694	return rc;
695	}
696
697
698	/**
699	* Gets the GIP pointer.
700	*
701	* @returns Pointer to the GIP or NULL.
702	*/
703	SUPDECL(PSUPGLOBALINFOPAGE) SUPGetGIP(void)
704	{
705	return g_pSUPGlobalInfoPage;
706	}
707
708
709
710
711
712	/*
713	*
714	*
715	* GIP Initialization, Termination and CPU Offline / Online Related Code.
716	* GIP Initialization, Termination and CPU Offline / Online Related Code.
717	* GIP Initialization, Termination and CPU Offline / Online Related Code.
718	*
719	*
720	*/
721
722	/**
723	* Used by supdrvGipInitRefineInvariantTscFreqTimer and supdrvGipInitMeasureTscFreq
724	* to update the TSC frequency related GIP variables.
725	*
726	* @param pGip The GIP.
727	* @param nsElapsed The number of nanoseconds elapsed.
728	* @param cElapsedTscTicks The corresponding number of TSC ticks.
729	* @param iTick The tick number for debugging.
730	*/
731	static void supdrvGipInitSetCpuFreq(PSUPGLOBALINFOPAGE pGip, uint64_t nsElapsed, uint64_t cElapsedTscTicks, uint32_t iTick)
732	{
733	/*
734	* Calculate the frequency.
735	*/
736	uint64_t uCpuHz;
737	if ( cElapsedTscTicks < UINT64_MAX / RT_NS_1SEC
738	&& nsElapsed < UINT32_MAX)
739	uCpuHz = ASMMultU64ByU32DivByU32(cElapsedTscTicks, RT_NS_1SEC, (uint32_t)nsElapsed);
740	else
741	{
742	RTUINT128U CpuHz, Tmp, Divisor;
743	CpuHz.s.Lo = CpuHz.s.Hi = 0;
744	RTUInt128MulU64ByU64(&Tmp, cElapsedTscTicks, RT_NS_1SEC_64);
745	RTUInt128Div(&CpuHz, &Tmp, RTUInt128AssignU64(&Divisor, nsElapsed));
746	uCpuHz = CpuHz.s.Lo;
747	}
748
749	/*
750	* Update the GIP.
751	*/
752	ASMAtomicWriteU64(&pGip->u64CpuHz, uCpuHz);
753	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
754	{
755	ASMAtomicWriteU64(&pGip->aCPUs[0].u64CpuHz, uCpuHz);
756
757	/* For inspecting the frequency calcs using tstGIP-2, debugger or similar. */
758	if (iTick + 1 < pGip->cCpus)
759	ASMAtomicWriteU64(&pGip->aCPUs[iTick + 1].u64CpuHz, uCpuHz);
760	}
761	}
762
763
764	/**
765	* Timer callback function for TSC frequency refinement in invariant GIP mode.
766	*
767	* This is started during driver init and fires once
768	* GIP_TSC_REFINE_PERIOD_IN_SECS seconds later.
769	*
770	* @param pTimer The timer.
771	* @param pvUser Opaque pointer to the device instance data.
772	* @param iTick The timer tick.
773	*/
774	static DECLCALLBACK(void) supdrvGipInitRefineInvariantTscFreqTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
775	{
776	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
777	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
778	RTCPUID idCpu;
779	uint64_t cNsElapsed;
780	uint64_t cTscTicksElapsed;
781	uint64_t nsNow;
782	uint64_t uTsc;
783	RTCCUINTREG fEFlags;
784
785	/* Paranoia. */
786	AssertReturnVoid(pGip);
787	AssertReturnVoid(pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC);
788
789	/*
790	* If we got a power event, stop the refinement process.
791	*/
792	if (pDevExt->fInvTscRefinePowerEvent)
793	{
794	int rc = RTTimerStop(pTimer); AssertRC(rc);
795	return;
796	}
797
798	/*
799	* Read the TSC and time, noting which CPU we are on.
800	*
801	* Don't bother spinning until RTTimeSystemNanoTS changes, since on
802	* systems where it matters we're in a context where we cannot waste that
803	* much time (DPC watchdog, called from clock interrupt).
804	*/
805	fEFlags = ASMIntDisableFlags();
806	uTsc = ASMReadTSC();
807	nsNow = RTTimeSystemNanoTS();
808	idCpu = RTMpCpuId();
809	ASMSetFlags(fEFlags);
810
811	cNsElapsed = nsNow - pDevExt->nsStartInvarTscRefine;
812	cTscTicksElapsed = uTsc - pDevExt->uTscStartInvarTscRefine;
813
814	/*
815	* If the above measurement was taken on a different CPU than the one we
816	* started the process on, cTscTicksElapsed will need to be adjusted with
817	* the TSC deltas of both the CPUs.
818	*
819	* We ASSUME that the delta calculation process takes less time than the
820	* TSC frequency refinement timer. If it doesn't, we'll complain and
821	* drop the frequency refinement.
822	*
823	* Note! We cannot entirely trust enmUseTscDelta here because it's
824	* downgraded after each delta calculation.
825	*/
826	if ( idCpu != pDevExt->idCpuInvarTscRefine
827	&& pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
828	{
829	uint32_t iStartCpuSet = RTMpCpuIdToSetIndex(pDevExt->idCpuInvarTscRefine);
830	uint32_t iStopCpuSet = RTMpCpuIdToSetIndex(idCpu);
831	uint16_t iStartGipCpu = iStartCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
832	? pGip->aiCpuFromCpuSetIdx[iStartCpuSet] : UINT16_MAX;
833	uint16_t iStopGipCpu = iStopCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
834	? pGip->aiCpuFromCpuSetIdx[iStopCpuSet] : UINT16_MAX;
835	int64_t iStartTscDelta = iStartGipCpu < pGip->cCpus ? pGip->aCPUs[iStartGipCpu].i64TSCDelta : INT64_MAX;
836	int64_t iStopTscDelta = iStopGipCpu < pGip->cCpus ? pGip->aCPUs[iStopGipCpu].i64TSCDelta : INT64_MAX;
837	if (RT_LIKELY(iStartTscDelta != INT64_MAX && iStopTscDelta != INT64_MAX))
838	{
839	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_PRACTICALLY_ZERO)
840	{
841	/* cTscTicksElapsed = (uTsc - iStopTscDelta) - (pDevExt->uTscStartInvarTscRefine - iStartTscDelta); */
842	cTscTicksElapsed += iStartTscDelta - iStopTscDelta;
843	}
844	}
845	/*
846	* Allow 5 times the refinement period to elapse before we give up on the TSC delta
847	* calculations.
848	*/
849	else if (cNsElapsed > GIP_TSC_REFINE_PERIOD_IN_SECS * 5 * RT_NS_1SEC_64)
850	{
851	SUPR0Printf("vboxdrv: Failed to refine invariant TSC frequency because deltas are unavailable after %u (%u) seconds\n",
852	(uint32_t)(cNsElapsed / RT_NS_1SEC), GIP_TSC_REFINE_PERIOD_IN_SECS);
853	SUPR0Printf("vboxdrv: start: %u, %u, %#llx stop: %u, %u, %#llx\n",
854	iStartCpuSet, iStartGipCpu, iStartTscDelta, iStopCpuSet, iStopGipCpu, iStopTscDelta);
855	int rc = RTTimerStop(pTimer); AssertRC(rc);
856	return;
857	}
858	}
859
860	/*
861	* Calculate and update the CPU frequency variables in GIP.
862	*
863	* If there is a GIP user already and we've already refined the frequency
864	* a couple of times, don't update it as we want a stable frequency value
865	* for all VMs.
866	*/
867	if ( pDevExt->cGipUsers == 0
868	\|\| cNsElapsed < RT_NS_1SEC * 2)
869	{
870	supdrvGipInitSetCpuFreq(pGip, cNsElapsed, cTscTicksElapsed, (uint32_t)iTick);
871
872	/*
873	* Stop the timer once we've reached the defined refinement period.
874	*/
875	if (cNsElapsed > GIP_TSC_REFINE_PERIOD_IN_SECS * RT_NS_1SEC_64)
876	{
877	int rc = RTTimerStop(pTimer);
878	AssertRC(rc);
879	}
880	}
881	else
882	{
883	int rc = RTTimerStop(pTimer);
884	AssertRC(rc);
885	}
886	}
887
888
889	/**
890	* @callback_method_impl{FNRTPOWERNOTIFICATION}
891	*/
892	static DECLCALLBACK(void) supdrvGipPowerNotificationCallback(RTPOWEREVENT enmEvent, void *pvUser)
893	{
894	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
895	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
896
897	/*
898	* If the TSC frequency refinement timer is running, we need to cancel it so it
899	* doesn't screw up the frequency after a long suspend.
900	*
901	* Recalculate all TSC-deltas on host resume as it may have changed, seen
902	* on Windows 7 running on the Dell Optiplex Intel Core i5-3570.
903	*/
904	if (enmEvent == RTPOWEREVENT_RESUME)
905	{
906	ASMAtomicWriteBool(&pDevExt->fInvTscRefinePowerEvent, true);
907	if ( RT_LIKELY(pGip)
908	&& pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED
909	&& !supdrvOSAreCpusOfflinedOnSuspend())
910	{
911	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
912	supdrvTscDeltaThreadStartMeasurement(pDevExt, true /* fForceAll */);
913	#else
914	RTCpuSetCopy(&pDevExt->TscDeltaCpuSet, &pGip->OnlineCpuSet);
915	supdrvTscMeasureInitialDeltas(pDevExt);
916	#endif
917	}
918	}
919	else if (enmEvent == RTPOWEREVENT_SUSPEND)
920	ASMAtomicWriteBool(&pDevExt->fInvTscRefinePowerEvent, true);
921	}
922
923
924	/**
925	* Start the TSC-frequency refinment timer for the invariant TSC GIP mode.
926	*
927	* We cannot use this in the synchronous and asynchronous tsc GIP modes because
928	* the CPU may change the TSC frequence between now and when the timer fires
929	* (supdrvInitAsyncRefineTscTimer).
930	*
931	* @param pDevExt Pointer to the device instance data.
932	*/
933	static void supdrvGipInitStartTimerForRefiningInvariantTscFreq(PSUPDRVDEVEXT pDevExt)
934	{
935	uint64_t u64NanoTS;
936	RTCCUINTREG fEFlags;
937	int rc;
938
939	/*
940	* Register a power management callback.
941	*/
942	pDevExt->fInvTscRefinePowerEvent = false;
943	rc = RTPowerNotificationRegister(supdrvGipPowerNotificationCallback, pDevExt);
944	AssertRC(rc); /* ignore */
945
946	/*
947	* Record the TSC and NanoTS as the starting anchor point for refinement
948	* of the TSC. We try get as close to a clock tick as possible on systems
949	* which does not provide high resolution time.
950	*/
951	u64NanoTS = RTTimeSystemNanoTS();
952	while (RTTimeSystemNanoTS() == u64NanoTS)
953	ASMNopPause();
954
955	fEFlags = ASMIntDisableFlags();
956	pDevExt->uTscStartInvarTscRefine = ASMReadTSC();
957	pDevExt->nsStartInvarTscRefine = RTTimeSystemNanoTS();
958	pDevExt->idCpuInvarTscRefine = RTMpCpuId();
959	ASMSetFlags(fEFlags);
960
961	/*
962	* Create a timer that runs on the same CPU so we won't have a depencency
963	* on the TSC-delta and can run in parallel to it. On systems that does not
964	* implement CPU specific timers we'll apply deltas in the timer callback,
965	* just like we do for CPUs going offline.
966	*
967	* The longer the refinement interval the better the accuracy, at least in
968	* theory. If it's too long though, ring-3 may already be starting its
969	* first VMs before we're done. On most systems we will be loading the
970	* support driver during boot and VMs won't be started for a while yet,
971	* it is really only a problem during development (especially with
972	* on-demand driver starting on windows).
973	*
974	* To avoid wasting time doing a long supdrvGipInitMeasureTscFreq() call
975	* to calculate the frequency during driver loading, the timer is set
976	* to fire after 200 ms the first time. It will then reschedule itself
977	* to fire every second until GIP_TSC_REFINE_PERIOD_IN_SECS has been
978	* reached or it notices that there is a user land client with GIP
979	* mapped (we want a stable frequency for all VMs).
980	*/
981	rc = RTTimerCreateEx(&pDevExt->pInvarTscRefineTimer, RT_NS_1SEC,
982	RTTIMER_FLAGS_CPU(RTMpCpuIdToSetIndex(pDevExt->idCpuInvarTscRefine)),
983	supdrvGipInitRefineInvariantTscFreqTimer, pDevExt);
984	if (RT_SUCCESS(rc))
985	{
986	rc = RTTimerStart(pDevExt->pInvarTscRefineTimer, 2*RT_NS_100MS);
987	if (RT_SUCCESS(rc))
988	return;
989	RTTimerDestroy(pDevExt->pInvarTscRefineTimer);
990	}
991
992	if (rc == VERR_CPU_OFFLINE \|\| rc == VERR_NOT_SUPPORTED)
993	{
994	rc = RTTimerCreateEx(&pDevExt->pInvarTscRefineTimer, RT_NS_1SEC, RTTIMER_FLAGS_CPU_ANY,
995	supdrvGipInitRefineInvariantTscFreqTimer, pDevExt);
996	if (RT_SUCCESS(rc))
997	{
998	rc = RTTimerStart(pDevExt->pInvarTscRefineTimer, 2*RT_NS_100MS);
999	if (RT_SUCCESS(rc))
1000	return;
1001	RTTimerDestroy(pDevExt->pInvarTscRefineTimer);
1002	}
1003	}
1004
1005	pDevExt->pInvarTscRefineTimer = NULL;
1006	OSDBGPRINT(("vboxdrv: Failed to create or start TSC frequency refinement timer: rc=%Rrc\n", rc));
1007	}
1008
1009
1010	/**
1011	* @callback_method_impl{PFNRTMPWORKER,
1012	* RTMpOnSpecific callback for reading TSC and time on the CPU we started
1013	* the measurements on.}
1014	*/
1015	static DECLCALLBACK(void) supdrvGipInitReadTscAndNanoTsOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
1016	{
1017	RTCCUINTREG fEFlags = ASMIntDisableFlags();
1018	uint64_t puTscStop = (uint64_t )pvUser1;
1019	uint64_t pnsStop = (uint64_t )pvUser2;
1020	RT_NOREF1(idCpu);
1021
1022	*puTscStop = ASMReadTSC();
1023	*pnsStop = RTTimeSystemNanoTS();
1024
1025	ASMSetFlags(fEFlags);
1026	}
1027
1028
1029	/**
1030	* Measures the TSC frequency of the system.
1031	*
1032	* The TSC frequency can vary on systems which are not reported as invariant.
1033	* On such systems the object of this function is to find out what the nominal,
1034	* maximum TSC frequency under 'normal' CPU operation.
1035	*
1036	* @returns VBox status code.
1037	* @param pGip Pointer to the GIP.
1038	* @param fRough Set if we're doing the rough calculation that the
1039	* TSC measuring code needs, where accuracy isn't all
1040	* that important (too high is better than too low).
1041	* When clear we try for best accuracy that we can
1042	* achieve in reasonably short time.
1043	*/
1044	static int supdrvGipInitMeasureTscFreq(PSUPGLOBALINFOPAGE pGip, bool fRough)
1045	{
1046	uint32_t nsTimerIncr = RTTimerGetSystemGranularity();
1047	int cTriesLeft = fRough ? 4 : 2;
1048	while (cTriesLeft-- > 0)
1049	{
1050	RTCCUINTREG fEFlags;
1051	uint64_t nsStart;
1052	uint64_t nsStop;
1053	uint64_t uTscStart;
1054	uint64_t uTscStop;
1055	RTCPUID idCpuStart;
1056	RTCPUID idCpuStop;
1057
1058	/*
1059	* Synchronize with the host OS clock tick on systems without high
1060	* resolution time API (older Windows version for example).
1061	*/
1062	nsStart = RTTimeSystemNanoTS();
1063	while (RTTimeSystemNanoTS() == nsStart)
1064	ASMNopPause();
1065
1066	/*
1067	* Read the TSC and current time, noting which CPU we're on.
1068	*/
1069	fEFlags = ASMIntDisableFlags();
1070	uTscStart = ASMReadTSC();
1071	nsStart = RTTimeSystemNanoTS();
1072	idCpuStart = RTMpCpuId();
1073	ASMSetFlags(fEFlags);
1074
1075	/*
1076	* Delay for a while.
1077	*/
1078	if (pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC)
1079	{
1080	/*
1081	* Sleep-wait since the TSC frequency is constant, it eases host load.
1082	* Shorter interval produces more variance in the frequency (esp. Windows).
1083	*/
1084	uint64_t msElapsed = 0;
1085	uint64_t msDelay = ( ((fRough ? 16 : 200) * RT_NS_1MS + nsTimerIncr - 1) / nsTimerIncr * nsTimerIncr - RT_NS_100US )
1086	/ RT_NS_1MS;
1087	do
1088	{
1089	RTThreadSleep((RTMSINTERVAL)(msDelay - msElapsed));
1090	nsStop = RTTimeSystemNanoTS();
1091	msElapsed = (nsStop - nsStart) / RT_NS_1MS;
1092	} while (msElapsed < msDelay);
1093
1094	while (RTTimeSystemNanoTS() == nsStop)
1095	ASMNopPause();
1096	}
1097	else
1098	{
1099	/*
1100	* Busy-wait keeping the frequency up.
1101	*/
1102	do
1103	{
1104	ASMNopPause();
1105	nsStop = RTTimeSystemNanoTS();
1106	} while (nsStop - nsStart < RT_NS_100MS);
1107	}
1108
1109	/*
1110	* Read the TSC and time again.
1111	*/
1112	fEFlags = ASMIntDisableFlags();
1113	uTscStop = ASMReadTSC();
1114	nsStop = RTTimeSystemNanoTS();
1115	idCpuStop = RTMpCpuId();
1116	ASMSetFlags(fEFlags);
1117
1118	/*
1119	* If the CPU changes, things get a bit complicated and what we
1120	* can get away with depends on the GIP mode / TSC reliability.
1121	*/
1122	if (idCpuStop != idCpuStart)
1123	{
1124	bool fDoXCall = false;
1125
1126	/*
1127	* Synchronous TSC mode: we're probably fine as it's unlikely
1128	* that we were rescheduled because of TSC throttling or power
1129	* management reasons, so just go ahead.
1130	*/
1131	if (pGip->u32Mode == SUPGIPMODE_SYNC_TSC)
1132	{
1133	/* Probably ok, maybe we should retry once?. */
1134	Assert(pGip->enmUseTscDelta == SUPGIPUSETSCDELTA_NOT_APPLICABLE);
1135	}
1136	/*
1137	* If we're just doing the rough measurement, do the cross call and
1138	* get on with things (we don't have deltas!).
1139	*/
1140	else if (fRough)
1141	fDoXCall = true;
1142	/*
1143	* Invariant TSC mode: It doesn't matter if we have delta available
1144	* for both CPUs. That is not something we can assume at this point.
1145	*
1146	* Note! We cannot necessarily trust enmUseTscDelta here because it's
1147	* downgraded after each delta calculation and the delta
1148	* calculations may not be complete yet.
1149	*/
1150	else if (pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC)
1151	{
1152	/** @todo This section of code is never reached atm, consider dropping it later on... */
1153	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1154	{
1155	uint32_t iStartCpuSet = RTMpCpuIdToSetIndex(idCpuStart);
1156	uint32_t iStopCpuSet = RTMpCpuIdToSetIndex(idCpuStop);
1157	uint16_t iStartGipCpu = iStartCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
1158	? pGip->aiCpuFromCpuSetIdx[iStartCpuSet] : UINT16_MAX;
1159	uint16_t iStopGipCpu = iStopCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
1160	? pGip->aiCpuFromCpuSetIdx[iStopCpuSet] : UINT16_MAX;
1161	int64_t iStartTscDelta = iStartGipCpu < pGip->cCpus ? pGip->aCPUs[iStartGipCpu].i64TSCDelta : INT64_MAX;
1162	int64_t iStopTscDelta = iStopGipCpu < pGip->cCpus ? pGip->aCPUs[iStopGipCpu].i64TSCDelta : INT64_MAX;
1163	if (RT_LIKELY(iStartTscDelta != INT64_MAX && iStopTscDelta != INT64_MAX))
1164	{
1165	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_PRACTICALLY_ZERO)
1166	{
1167	uTscStart -= iStartTscDelta;
1168	uTscStop -= iStopTscDelta;
1169	}
1170	}
1171	/*
1172	* Invalid CPU indexes are not caused by online/offline races, so
1173	* we have to trigger driver load failure if that happens as GIP
1174	* and IPRT assumptions are busted on this system.
1175	*/
1176	else if (iStopGipCpu >= pGip->cCpus \|\| iStartGipCpu >= pGip->cCpus)
1177	{
1178	SUPR0Printf("vboxdrv: Unexpected CPU index in supdrvGipInitMeasureTscFreq.\n");
1179	SUPR0Printf("vboxdrv: start: %u, %u, %#llx stop: %u, %u, %#llx\n",
1180	iStartCpuSet, iStartGipCpu, iStartTscDelta, iStopCpuSet, iStopGipCpu, iStopTscDelta);
1181	return VERR_INVALID_CPU_INDEX;
1182	}
1183	/*
1184	* No valid deltas. We retry, if we're on our last retry
1185	* we do the cross call instead just to get a result. The
1186	* frequency will be refined in a few seconds anyway.
1187	*/
1188	else if (cTriesLeft > 0)
1189	continue;
1190	else
1191	fDoXCall = true;
1192	}
1193	}
1194	/*
1195	* Asynchronous TSC mode: This is bad, as the reason we usually
1196	* use this mode is to deal with variable TSC frequencies and
1197	* deltas. So, we need to get the TSC from the same CPU as
1198	* started it, we also need to keep that CPU busy. So, retry
1199	* and fall back to the cross call on the last attempt.
1200	*/
1201	else
1202	{
1203	Assert(pGip->u32Mode == SUPGIPMODE_ASYNC_TSC);
1204	if (cTriesLeft > 0)
1205	continue;
1206	fDoXCall = true;
1207	}
1208
1209	if (fDoXCall)
1210	{
1211	/*
1212	* Try read the TSC and timestamp on the start CPU.
1213	*/
1214	int rc = RTMpOnSpecific(idCpuStart, supdrvGipInitReadTscAndNanoTsOnCpu, &uTscStop, &nsStop);
1215	if (RT_FAILURE(rc) && (!fRough \|\| cTriesLeft > 0))
1216	continue;
1217	}
1218	}
1219
1220	/*
1221	* Calculate the TSC frequency and update it (shared with the refinement timer).
1222	*/
1223	supdrvGipInitSetCpuFreq(pGip, nsStop - nsStart, uTscStop - uTscStart, 0);
1224	return VINF_SUCCESS;
1225	}
1226
1227	Assert(!fRough);
1228	return VERR_SUPDRV_TSC_FREQ_MEASUREMENT_FAILED;
1229	}
1230
1231
1232	/**
1233	* Finds our (@a idCpu) entry, or allocates a new one if not found.
1234	*
1235	* @returns Index of the CPU in the cache set.
1236	* @param pGip The GIP.
1237	* @param idCpu The CPU ID.
1238	*/
1239	static uint32_t supdrvGipFindOrAllocCpuIndexForCpuId(PSUPGLOBALINFOPAGE pGip, RTCPUID idCpu)
1240	{
1241	uint32_t i, cTries;
1242
1243	/*
1244	* ASSUMES that CPU IDs are constant.
1245	*/
1246	for (i = 0; i < pGip->cCpus; i++)
1247	if (pGip->aCPUs[i].idCpu == idCpu)
1248	return i;
1249
1250	cTries = 0;
1251	do
1252	{
1253	for (i = 0; i < pGip->cCpus; i++)
1254	{
1255	bool fRc;
1256	ASMAtomicCmpXchgSize(&pGip->aCPUs[i].idCpu, idCpu, NIL_RTCPUID, fRc);
1257	if (fRc)
1258	return i;
1259	}
1260	} while (cTries++ < 32);
1261	AssertReleaseFailed();
1262	return i - 1;
1263	}
1264
1265
1266	/**
1267	* The calling CPU should be accounted as online, update GIP accordingly.
1268	*
1269	* This is used by supdrvGipCreate() as well as supdrvGipMpEvent().
1270	*
1271	* @param pDevExt The device extension.
1272	* @param idCpu The CPU ID.
1273	*/
1274	static void supdrvGipMpEventOnlineOrInitOnCpu(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
1275	{
1276	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
1277	int iCpuSet = 0;
1278	uint16_t idApic = UINT16_MAX;
1279	uint32_t i = 0;
1280	uint64_t u64NanoTS = 0;
1281
1282	AssertPtrReturnVoid(pGip);
1283	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
1284	AssertRelease(idCpu == RTMpCpuId());
1285	Assert(pGip->cPossibleCpus == RTMpGetCount());
1286
1287	/*
1288	* Do this behind a spinlock with interrupts disabled as this can fire
1289	* on all CPUs simultaneously, see @bugref{6110}.
1290	*/
1291	RTSpinlockAcquire(pDevExt->hGipSpinlock);
1292
1293	/*
1294	* Update the globals.
1295	*/
1296	ASMAtomicWriteU16(&pGip->cPresentCpus, RTMpGetPresentCount());
1297	ASMAtomicWriteU16(&pGip->cOnlineCpus, RTMpGetOnlineCount());
1298	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
1299	if (iCpuSet >= 0)
1300	{
1301	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
1302	RTCpuSetAddByIndex(&pGip->OnlineCpuSet, iCpuSet);
1303	RTCpuSetAddByIndex(&pGip->PresentCpuSet, iCpuSet);
1304	}
1305
1306	/*
1307	* Update the entry.
1308	*/
1309	u64NanoTS = RTTimeSystemNanoTS() - pGip->u32UpdateIntervalNS;
1310	i = supdrvGipFindOrAllocCpuIndexForCpuId(pGip, idCpu);
1311
1312	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS, pGip->u64CpuHz);
1313
1314	idApic = ASMGetApicId();
1315	ASMAtomicWriteU16(&pGip->aCPUs[i].idApic, idApic);
1316	ASMAtomicWriteS16(&pGip->aCPUs[i].iCpuSet, (int16_t)iCpuSet);
1317	ASMAtomicWriteSize(&pGip->aCPUs[i].idCpu, idCpu);
1318
1319	pGip->aCPUs[i].iCpuGroup = 0;
1320	pGip->aCPUs[i].iCpuGroupMember = iCpuSet;
1321	#ifdef RT_OS_WINDOWS
1322	supdrvOSGipInitGroupBitsForCpu(pDevExt, pGip, &pGip->aCPUs[i]);
1323	#endif
1324
1325	/*
1326	* Update the APIC ID and CPU set index mappings.
1327	*/
1328	ASMAtomicWriteU16(&pGip->aiCpuFromApicId[idApic], i);
1329	ASMAtomicWriteU16(&pGip->aiCpuFromCpuSetIdx[iCpuSet], i);
1330
1331	/* Add this CPU to this set of CPUs we need to calculate the TSC-delta for. */
1332	RTCpuSetAddByIndex(&pDevExt->TscDeltaCpuSet, RTMpCpuIdToSetIndex(idCpu));
1333
1334	/* Update the Mp online/offline counter. */
1335	ASMAtomicIncU32(&pDevExt->cMpOnOffEvents);
1336
1337	/* Commit it. */
1338	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_ONLINE);
1339
1340	RTSpinlockRelease(pDevExt->hGipSpinlock);
1341	}
1342
1343
1344	/**
1345	* RTMpOnSpecific callback wrapper for supdrvGipMpEventOnlineOrInitOnCpu().
1346	*
1347	* @param idCpu The CPU ID we are running on.
1348	* @param pvUser1 Opaque pointer to the device instance data.
1349	* @param pvUser2 Not used.
1350	*/
1351	static DECLCALLBACK(void) supdrvGipMpEventOnlineCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
1352	{
1353	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser1;
1354	NOREF(pvUser2);
1355	supdrvGipMpEventOnlineOrInitOnCpu(pDevExt, idCpu);
1356	}
1357
1358
1359	/**
1360	* The CPU should be accounted as offline, update the GIP accordingly.
1361	*
1362	* This is used by supdrvGipMpEvent.
1363	*
1364	* @param pDevExt The device extension.
1365	* @param idCpu The CPU ID.
1366	*/
1367	static void supdrvGipMpEventOffline(PSUPDRVDEVEXT pDevExt, RTCPUID idCpu)
1368	{
1369	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
1370	int iCpuSet;
1371	unsigned i;
1372
1373	AssertPtrReturnVoid(pGip);
1374	RTSpinlockAcquire(pDevExt->hGipSpinlock);
1375
1376	iCpuSet = RTMpCpuIdToSetIndex(idCpu);
1377	AssertReturnVoid(iCpuSet >= 0);
1378
1379	i = pGip->aiCpuFromCpuSetIdx[iCpuSet];
1380	AssertReturnVoid(i < pGip->cCpus);
1381	AssertReturnVoid(pGip->aCPUs[i].idCpu == idCpu);
1382
1383	Assert(RTCpuSetIsMemberByIndex(&pGip->PossibleCpuSet, iCpuSet));
1384	RTCpuSetDelByIndex(&pGip->OnlineCpuSet, iCpuSet);
1385
1386	/* Update the Mp online/offline counter. */
1387	ASMAtomicIncU32(&pDevExt->cMpOnOffEvents);
1388
1389	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1390	{
1391	/* Reset the TSC delta, we will recalculate it lazily. */
1392	ASMAtomicWriteS64(&pGip->aCPUs[i].i64TSCDelta, INT64_MAX);
1393	/* Remove this CPU from the set of CPUs that we have obtained the TSC deltas. */
1394	RTCpuSetDelByIndex(&pDevExt->TscDeltaObtainedCpuSet, iCpuSet);
1395	}
1396
1397	/* Commit it. */
1398	ASMAtomicWriteSize(&pGip->aCPUs[i].enmState, SUPGIPCPUSTATE_OFFLINE);
1399
1400	RTSpinlockRelease(pDevExt->hGipSpinlock);
1401	}
1402
1403
1404	/**
1405	* Multiprocessor event notification callback.
1406	*
1407	* This is used to make sure that the GIP master gets passed on to
1408	* another CPU. It also updates the associated CPU data.
1409	*
1410	* @param enmEvent The event.
1411	* @param idCpu The cpu it applies to.
1412	* @param pvUser Pointer to the device extension.
1413	*/
1414	static DECLCALLBACK(void) supdrvGipMpEvent(RTMPEVENT enmEvent, RTCPUID idCpu, void *pvUser)
1415	{
1416	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
1417	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
1418
1419	if (pGip)
1420	{
1421	RTTHREADPREEMPTSTATE PreemptState = RTTHREADPREEMPTSTATE_INITIALIZER;
1422	switch (enmEvent)
1423	{
1424	case RTMPEVENT_ONLINE:
1425	{
1426	RTThreadPreemptDisable(&PreemptState);
1427	if (idCpu == RTMpCpuId())
1428	{
1429	supdrvGipMpEventOnlineOrInitOnCpu(pDevExt, idCpu);
1430	RTThreadPreemptRestore(&PreemptState);
1431	}
1432	else
1433	{
1434	RTThreadPreemptRestore(&PreemptState);
1435	RTMpOnSpecific(idCpu, supdrvGipMpEventOnlineCallback, pDevExt, NULL /* pvUser2 */);
1436	}
1437
1438	/*
1439	* Recompute TSC-delta for the newly online'd CPU.
1440	*/
1441	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1442	{
1443	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
1444	supdrvTscDeltaThreadStartMeasurement(pDevExt, false /* fForceAll */);
1445	#else
1446	uint32_t iCpu = supdrvGipFindOrAllocCpuIndexForCpuId(pGip, idCpu);
1447	supdrvTscMeasureDeltaOne(pDevExt, iCpu);
1448	#endif
1449	}
1450	break;
1451	}
1452
1453	case RTMPEVENT_OFFLINE:
1454	supdrvGipMpEventOffline(pDevExt, idCpu);
1455	break;
1456	}
1457	}
1458
1459	/*
1460	* Make sure there is a master GIP.
1461	*/
1462	if (enmEvent == RTMPEVENT_OFFLINE)
1463	{
1464	RTCPUID idGipMaster = ASMAtomicReadU32(&pDevExt->idGipMaster);
1465	if (idGipMaster == idCpu)
1466	{
1467	/*
1468	* The GIP master is going offline, find a new one.
1469	*/
1470	bool fIgnored;
1471	unsigned i;
1472	RTCPUID idNewGipMaster = NIL_RTCPUID;
1473	RTCPUSET OnlineCpus;
1474	RTMpGetOnlineSet(&OnlineCpus);
1475
1476	for (i = 0; i < RTCPUSET_MAX_CPUS; i++)
1477	if (RTCpuSetIsMemberByIndex(&OnlineCpus, i))
1478	{
1479	RTCPUID idCurCpu = RTMpCpuIdFromSetIndex(i);
1480	if (idCurCpu != idGipMaster)
1481	{
1482	idNewGipMaster = idCurCpu;
1483	break;
1484	}
1485	}
1486
1487	Log(("supdrvGipMpEvent: Gip master %#lx -> %#lx\n", (long)idGipMaster, (long)idNewGipMaster));
1488	ASMAtomicCmpXchgSize(&pDevExt->idGipMaster, idNewGipMaster, idGipMaster, fIgnored);
1489	NOREF(fIgnored);
1490	}
1491	}
1492	}
1493
1494
1495	/**
1496	* On CPU initialization callback for RTMpOnAll.
1497	*
1498	* @param idCpu The CPU ID.
1499	* @param pvUser1 The device extension.
1500	* @param pvUser2 The GIP.
1501	*/
1502	static DECLCALLBACK(void) supdrvGipInitOnCpu(RTCPUID idCpu, void pvUser1, void pvUser2)
1503	{
1504	/* This is good enough, even though it will update some of the globals a
1505	bit to much. */
1506	supdrvGipMpEventOnlineOrInitOnCpu((PSUPDRVDEVEXT)pvUser1, idCpu);
1507	NOREF(pvUser2);
1508	}
1509
1510
1511	/**
1512	* Callback used by supdrvDetermineAsyncTSC to read the TSC on a CPU.
1513	*
1514	* @param idCpu Ignored.
1515	* @param pvUser1 Where to put the TSC.
1516	* @param pvUser2 Ignored.
1517	*/
1518	static DECLCALLBACK(void) supdrvGipInitDetermineAsyncTscWorker(RTCPUID idCpu, void pvUser1, void pvUser2)
1519	{
1520	Assert(RTMpCpuIdToSetIndex(idCpu) == (intptr_t)pvUser2);
1521	ASMAtomicWriteU64((uint64_t volatile *)pvUser1, ASMReadTSC());
1522	RT_NOREF2(idCpu, pvUser2);
1523	}
1524
1525
1526	/**
1527	* Determine if Async GIP mode is required because of TSC drift.
1528	*
1529	* When using the default/normal timer code it is essential that the time stamp counter
1530	* (TSC) runs never backwards, that is, a read operation to the counter should return
1531	* a bigger value than any previous read operation. This is guaranteed by the latest
1532	* AMD CPUs and by newer Intel CPUs which never enter the C2 state (P4). In any other
1533	* case we have to choose the asynchronous timer mode.
1534	*
1535	* @param poffMin Pointer to the determined difference between different
1536	* cores (optional, can be NULL).
1537	* @return false if the time stamp counters appear to be synchronized, true otherwise.
1538	*/
1539	static bool supdrvGipInitDetermineAsyncTsc(uint64_t *poffMin)
1540	{
1541	/*
1542	* Just iterate all the cpus 8 times and make sure that the TSC is
1543	* ever increasing. We don't bother taking TSC rollover into account.
1544	*/
1545	int iEndCpu = RTMpGetArraySize();
1546	int iCpu;
1547	int cLoops = 8;
1548	bool fAsync = false;
1549	int rc = VINF_SUCCESS;
1550	uint64_t offMax = 0;
1551	uint64_t offMin = ~(uint64_t)0;
1552	uint64_t PrevTsc = ASMReadTSC();
1553
1554	while (cLoops-- > 0)
1555	{
1556	for (iCpu = 0; iCpu < iEndCpu; iCpu++)
1557	{
1558	uint64_t CurTsc;
1559	rc = RTMpOnSpecific(RTMpCpuIdFromSetIndex(iCpu), supdrvGipInitDetermineAsyncTscWorker,
1560	&CurTsc, (void *)(uintptr_t)iCpu);
1561	if (RT_SUCCESS(rc))
1562	{
1563	if (CurTsc <= PrevTsc)
1564	{
1565	fAsync = true;
1566	offMin = offMax = PrevTsc - CurTsc;
1567	Log(("supdrvGipInitDetermineAsyncTsc: iCpu=%d cLoops=%d CurTsc=%llx PrevTsc=%llx\n",
1568	iCpu, cLoops, CurTsc, PrevTsc));
1569	break;
1570	}
1571
1572	/* Gather statistics (except the first time). */
1573	if (iCpu != 0 \|\| cLoops != 7)
1574	{
1575	uint64_t off = CurTsc - PrevTsc;
1576	if (off < offMin)
1577	offMin = off;
1578	if (off > offMax)
1579	offMax = off;
1580	Log2(("%d/%d: off=%llx\n", cLoops, iCpu, off));
1581	}
1582
1583	/* Next */
1584	PrevTsc = CurTsc;
1585	}
1586	else if (rc == VERR_NOT_SUPPORTED)
1587	break;
1588	else
1589	AssertMsg(rc == VERR_CPU_NOT_FOUND \|\| rc == VERR_CPU_OFFLINE, ("%d\n", rc));
1590	}
1591
1592	/* broke out of the loop. */
1593	if (iCpu < iEndCpu)
1594	break;
1595	}
1596
1597	if (poffMin)
1598	poffMin = offMin; / Almost RTMpOnSpecific profiling. */
1599	Log(("supdrvGipInitDetermineAsyncTsc: returns %d; iEndCpu=%d rc=%d offMin=%llx offMax=%llx\n",
1600	fAsync, iEndCpu, rc, offMin, offMax));
1601	#if !defined(RT_OS_SOLARIS) && !defined(RT_OS_OS2) && !defined(RT_OS_WINDOWS)
1602	OSDBGPRINT(("vboxdrv: fAsync=%d offMin=%#lx offMax=%#lx\n", fAsync, (long)offMin, (long)offMax));
1603	#endif
1604	return fAsync;
1605	}
1606
1607
1608	/**
1609	* supdrvGipInit() worker that determines the GIP TSC mode.
1610	*
1611	* @returns The most suitable TSC mode.
1612	* @param pDevExt Pointer to the device instance data.
1613	*/
1614	static SUPGIPMODE supdrvGipInitDetermineTscMode(PSUPDRVDEVEXT pDevExt)
1615	{
1616	uint64_t u64DiffCoresIgnored;
1617	uint32_t uEAX, uEBX, uECX, uEDX;
1618
1619	/*
1620	* Establish whether the CPU advertises TSC as invariant, we need that in
1621	* a couple of places below.
1622	*/
1623	bool fInvariantTsc = false;
1624	if (ASMHasCpuId())
1625	{
1626	uEAX = ASMCpuId_EAX(0x80000000);
1627	if (ASMIsValidExtRange(uEAX) && uEAX >= 0x80000007)
1628	{
1629	uEDX = ASMCpuId_EDX(0x80000007);
1630	if (uEDX & X86_CPUID_AMD_ADVPOWER_EDX_TSCINVAR)
1631	fInvariantTsc = true;
1632	}
1633	}
1634
1635	/*
1636	* On single CPU systems, we don't need to consider ASYNC mode.
1637	*/
1638	if (RTMpGetCount() <= 1)
1639	return fInvariantTsc ? SUPGIPMODE_INVARIANT_TSC : SUPGIPMODE_SYNC_TSC;
1640
1641	/*
1642	* Allow the user and/or OS specific bits to force async mode.
1643	*/
1644	if (supdrvOSGetForcedAsyncTscMode(pDevExt))
1645	return SUPGIPMODE_ASYNC_TSC;
1646
1647	/*
1648	* Use invariant mode if the CPU says TSC is invariant.
1649	*/
1650	if (fInvariantTsc)
1651	return SUPGIPMODE_INVARIANT_TSC;
1652
1653	/*
1654	* TSC is not invariant and we're on SMP, this presents two problems:
1655	*
1656	* (1) There might be a skew between the CPU, so that cpu0
1657	* returns a TSC that is slightly different from cpu1.
1658	* This screw may be due to (2), bad TSC initialization
1659	* or slightly different TSC rates.
1660	*
1661	* (2) Power management (and other things) may cause the TSC
1662	* to run at a non-constant speed, and cause the speed
1663	* to be different on the cpus. This will result in (1).
1664	*
1665	* If any of the above is detected, we will have to use ASYNC mode.
1666	*/
1667	/* (1). Try check for current differences between the cpus. */
1668	if (supdrvGipInitDetermineAsyncTsc(&u64DiffCoresIgnored))
1669	return SUPGIPMODE_ASYNC_TSC;
1670
1671	/* (2) If it's an AMD CPU with power management, we won't trust its TSC. */
1672	ASMCpuId(0, &uEAX, &uEBX, &uECX, &uEDX);
1673	if ( ASMIsValidStdRange(uEAX)
1674	&& ASMIsAmdCpuEx(uEBX, uECX, uEDX))
1675	{
1676	/* Check for APM support. */
1677	uEAX = ASMCpuId_EAX(0x80000000);
1678	if (ASMIsValidExtRange(uEAX) && uEAX >= 0x80000007)
1679	{
1680	uEDX = ASMCpuId_EDX(0x80000007);
1681	if (uEDX & 0x3e) /* STC\|TM\|THERMTRIP\|VID\|FID. Ignore TS. */
1682	return SUPGIPMODE_ASYNC_TSC;
1683	}
1684	}
1685
1686	return SUPGIPMODE_SYNC_TSC;
1687	}
1688
1689
1690	/**
1691	* Initializes per-CPU GIP information.
1692	*
1693	* @param pGip Pointer to the GIP.
1694	* @param pCpu Pointer to which GIP CPU to initialize.
1695	* @param u64NanoTS The current nanosecond timestamp.
1696	* @param uCpuHz The CPU frequency to set, 0 if the caller doesn't know.
1697	*/
1698	static void supdrvGipInitCpu(PSUPGLOBALINFOPAGE pGip, PSUPGIPCPU pCpu, uint64_t u64NanoTS, uint64_t uCpuHz)
1699	{
1700	pCpu->u32TransactionId = 2;
1701	pCpu->u64NanoTS = u64NanoTS;
1702	pCpu->u64TSC = ASMReadTSC();
1703	pCpu->u64TSCSample = GIP_TSC_DELTA_RSVD;
1704	pCpu->i64TSCDelta = pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED ? INT64_MAX : 0;
1705
1706	ASMAtomicWriteSize(&pCpu->enmState, SUPGIPCPUSTATE_INVALID);
1707	ASMAtomicWriteU32(&pCpu->idCpu, NIL_RTCPUID);
1708	ASMAtomicWriteS16(&pCpu->iCpuSet, -1);
1709	ASMAtomicWriteU16(&pCpu->iCpuGroup, 0);
1710	ASMAtomicWriteU16(&pCpu->iCpuGroupMember, UINT16_MAX);
1711	ASMAtomicWriteU16(&pCpu->idApic, UINT16_MAX);
1712	ASMAtomicWriteU32(&pCpu->iReservedForNumaNode, 0);
1713
1714	/*
1715	* The first time we're called, we don't have a CPU frequency handy,
1716	* so pretend it's a 4 GHz CPU. On CPUs that are online, we'll get
1717	* called again and at that point we have a more plausible CPU frequency
1718	* value handy. The frequency history will also be adjusted again on
1719	* the 2nd timer callout (maybe we can skip that now?).
1720	*/
1721	if (!uCpuHz)
1722	{
1723	pCpu->u64CpuHz = _4G - 1;
1724	pCpu->u32UpdateIntervalTSC = (uint32_t)((_4G - 1) / pGip->u32UpdateHz);
1725	}
1726	else
1727	{
1728	pCpu->u64CpuHz = uCpuHz;
1729	pCpu->u32UpdateIntervalTSC = (uint32_t)(uCpuHz / pGip->u32UpdateHz);
1730	}
1731	pCpu->au32TSCHistory[0]
1732	= pCpu->au32TSCHistory[1]
1733	= pCpu->au32TSCHistory[2]
1734	= pCpu->au32TSCHistory[3]
1735	= pCpu->au32TSCHistory[4]
1736	= pCpu->au32TSCHistory[5]
1737	= pCpu->au32TSCHistory[6]
1738	= pCpu->au32TSCHistory[7]
1739	= pCpu->u32UpdateIntervalTSC;
1740	}
1741
1742
1743	/**
1744	* Initializes the GIP data.
1745	*
1746	* @returns VBox status code.
1747	* @param pDevExt Pointer to the device instance data.
1748	* @param pGip Pointer to the read-write kernel mapping of the GIP.
1749	* @param HCPhys The physical address of the GIP.
1750	* @param u64NanoTS The current nanosecond timestamp.
1751	* @param uUpdateHz The update frequency.
1752	* @param uUpdateIntervalNS The update interval in nanoseconds.
1753	* @param cCpus The CPU count.
1754	* @param cbGipCpuGroups The supdrvOSGipGetGroupTableSize return value we
1755	* used when allocating the GIP structure.
1756	*/
1757	static int supdrvGipInit(PSUPDRVDEVEXT pDevExt, PSUPGLOBALINFOPAGE pGip, RTHCPHYS HCPhys,
1758	uint64_t u64NanoTS, unsigned uUpdateHz, unsigned uUpdateIntervalNS,
1759	unsigned cCpus, size_t cbGipCpuGroups)
1760	{
1761	size_t const cbGip = RT_ALIGN_Z(RT_UOFFSETOF_DYN(SUPGLOBALINFOPAGE, aCPUs[cCpus]) + cbGipCpuGroups, PAGE_SIZE);
1762	unsigned i;
1763	#ifdef DEBUG_DARWIN_GIP
1764	OSDBGPRINT(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
1765	#else
1766	LogFlow(("supdrvGipInit: pGip=%p HCPhys=%lx u64NanoTS=%llu uUpdateHz=%d cCpus=%u\n", pGip, (long)HCPhys, u64NanoTS, uUpdateHz, cCpus));
1767	#endif
1768
1769	/*
1770	* Initialize the structure.
1771	*/
1772	memset(pGip, 0, cbGip);
1773
1774	pGip->u32Magic = SUPGLOBALINFOPAGE_MAGIC;
1775	pGip->u32Version = SUPGLOBALINFOPAGE_VERSION;
1776	pGip->u32Mode = supdrvGipInitDetermineTscMode(pDevExt);
1777	if ( pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC
1778	/\|\| pGip->u32Mode == SUPGIPMODE_SYNC_TSC /)
1779	pGip->enmUseTscDelta = supdrvOSAreTscDeltasInSync() /* Allow OS override (windows). */
1780	? SUPGIPUSETSCDELTA_ZERO_CLAIMED : SUPGIPUSETSCDELTA_PRACTICALLY_ZERO /* downgrade later */;
1781	else
1782	pGip->enmUseTscDelta = SUPGIPUSETSCDELTA_NOT_APPLICABLE;
1783	pGip->cCpus = (uint16_t)cCpus;
1784	pGip->cPages = (uint16_t)(cbGip / PAGE_SIZE);
1785	pGip->u32UpdateHz = uUpdateHz;
1786	pGip->u32UpdateIntervalNS = uUpdateIntervalNS;
1787	pGip->fGetGipCpu = SUPGIPGETCPU_APIC_ID;
1788	RTCpuSetEmpty(&pGip->OnlineCpuSet);
1789	RTCpuSetEmpty(&pGip->PresentCpuSet);
1790	RTMpGetSet(&pGip->PossibleCpuSet);
1791	pGip->cOnlineCpus = RTMpGetOnlineCount();
1792	pGip->cPresentCpus = RTMpGetPresentCount();
1793	pGip->cPossibleCpus = RTMpGetCount();
1794	pGip->cPossibleCpuGroups = 1;
1795	pGip->idCpuMax = RTMpGetMaxCpuId();
1796	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromApicId); i++)
1797	pGip->aiCpuFromApicId[i] = UINT16_MAX;
1798	for (i = 0; i < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx); i++)
1799	pGip->aiCpuFromCpuSetIdx[i] = UINT16_MAX;
1800	for (i = 0; i < RT_ELEMENTS(pGip->aoffCpuGroup); i++)
1801	pGip->aoffCpuGroup[i] = UINT16_MAX;
1802	for (i = 0; i < cCpus; i++)
1803	supdrvGipInitCpu(pGip, &pGip->aCPUs[i], u64NanoTS, 0 /uCpuHz/);
1804	#ifdef RT_OS_WINDOWS
1805	int rc = supdrvOSInitGipGroupTable(pDevExt, pGip, cbGipCpuGroups);
1806	AssertRCReturn(rc, rc);
1807	#endif
1808
1809	/*
1810	* Link it to the device extension.
1811	*/
1812	pDevExt->pGip = pGip;
1813	pDevExt->HCPhysGip = HCPhys;
1814	pDevExt->cGipUsers = 0;
1815
1816	return VINF_SUCCESS;
1817	}
1818
1819
1820	/**
1821	* Creates the GIP.
1822	*
1823	* @returns VBox status code.
1824	* @param pDevExt Instance data. GIP stuff may be updated.
1825	*/
1826	int VBOXCALL supdrvGipCreate(PSUPDRVDEVEXT pDevExt)
1827	{
1828	PSUPGLOBALINFOPAGE pGip;
1829	size_t cbGip;
1830	size_t cbGipCpuGroups;
1831	RTHCPHYS HCPhysGip;
1832	uint32_t u32SystemResolution;
1833	uint32_t u32Interval;
1834	uint32_t u32MinInterval;
1835	uint32_t uMod;
1836	unsigned cCpus;
1837	int rc;
1838
1839	LogFlow(("supdrvGipCreate:\n"));
1840
1841	/*
1842	* Assert order.
1843	*/
1844	Assert(pDevExt->u32SystemTimerGranularityGrant == 0);
1845	Assert(pDevExt->GipMemObj == NIL_RTR0MEMOBJ);
1846	Assert(!pDevExt->pGipTimer);
1847	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
1848	Assert(pDevExt->mtxGip != NIL_RTSEMMUTEX);
1849	Assert(pDevExt->mtxTscDelta != NIL_RTSEMMUTEX);
1850	#else
1851	Assert(pDevExt->mtxGip != NIL_RTSEMFASTMUTEX);
1852	Assert(pDevExt->mtxTscDelta != NIL_RTSEMFASTMUTEX);
1853	#endif
1854
1855	/*
1856	* Check the CPU count.
1857	*/
1858	cCpus = RTMpGetArraySize();
1859	if ( cCpus > RTCPUSET_MAX_CPUS
1860	#if RTCPUSET_MAX_CPUS != 256
1861	\|\| cCpus > 256 /* ApicId is used for the mappings */
1862	#endif
1863	)
1864	{
1865	SUPR0Printf("VBoxDrv: Too many CPUs (%u) for the GIP (max %u)\n", cCpus, RT_MIN(RTCPUSET_MAX_CPUS, 256));
1866	return VERR_TOO_MANY_CPUS;
1867	}
1868
1869	/*
1870	* Allocate a contiguous set of pages with a default kernel mapping.
1871	*/
1872	#ifdef RT_OS_WINDOWS
1873	cbGipCpuGroups = supdrvOSGipGetGroupTableSize(pDevExt);
1874	#else
1875	cbGipCpuGroups = 0;
1876	#endif
1877	cbGip = RT_UOFFSETOF_DYN(SUPGLOBALINFOPAGE, aCPUs[cCpus]) + cbGipCpuGroups;
1878	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, cbGip, false /fExecutable/);
1879	if (RT_FAILURE(rc))
1880	{
1881	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
1882	return rc;
1883	}
1884	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
1885	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
1886
1887	/*
1888	* Find a reasonable update interval and initialize the structure.
1889	*/
1890	supdrvGipRequestHigherTimerFrequencyFromSystem(pDevExt);
1891	/** @todo figure out why using a 100Ms interval upsets timekeeping in VMs.
1892	* See @bugref{6710}. */
1893	u32MinInterval = RT_NS_10MS;
1894	u32SystemResolution = RTTimerGetSystemGranularity();
1895	u32Interval = u32MinInterval;
1896	uMod = u32MinInterval % u32SystemResolution;
1897	if (uMod)
1898	u32Interval += u32SystemResolution - uMod;
1899
1900	rc = supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), RT_NS_1SEC / u32Interval /=Hz/, u32Interval,
1901	cCpus, cbGipCpuGroups);
1902
1903	/*
1904	* Important sanity check... (Sets rc)
1905	*/
1906	if (RT_UNLIKELY( pGip->enmUseTscDelta == SUPGIPUSETSCDELTA_ZERO_CLAIMED
1907	&& pGip->u32Mode == SUPGIPMODE_ASYNC_TSC
1908	&& !supdrvOSGetForcedAsyncTscMode(pDevExt)))
1909	{
1910	OSDBGPRINT(("supdrvGipCreate: Host-OS/user claims the TSC-deltas are zero but we detected async. TSC! Bad.\n"));
1911	rc = VERR_INTERNAL_ERROR_2;
1912	}
1913
1914	/* It doesn't make sense to do TSC-delta detection on systems we detect as async. */
1915	AssertStmt( pGip->u32Mode != SUPGIPMODE_ASYNC_TSC
1916	\|\| pGip->enmUseTscDelta <= SUPGIPUSETSCDELTA_ZERO_CLAIMED,
1917	rc = VERR_INTERNAL_ERROR_3);
1918
1919	/*
1920	* Do the TSC frequency measurements.
1921	*
1922	* If we're in invariant TSC mode, just to a quick preliminary measurement
1923	* that the TSC-delta measurement code can use to yield cross calls.
1924	*
1925	* If we're in any of the other two modes, neither which require MP init,
1926	* notifications or deltas for the job, do the full measurement now so
1927	* that supdrvGipInitOnCpu() can populate the TSC interval and history
1928	* array with more reasonable values.
1929	*/
1930	if (RT_SUCCESS(rc))
1931	{
1932	if (pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC)
1933	{
1934	rc = supdrvGipInitMeasureTscFreq(pGip, true /fRough/); /* cannot fail */
1935	supdrvGipInitStartTimerForRefiningInvariantTscFreq(pDevExt);
1936	}
1937	else
1938	rc = supdrvGipInitMeasureTscFreq(pGip, false /fRough/);
1939	if (RT_SUCCESS(rc))
1940	{
1941	/*
1942	* Start TSC-delta measurement thread before we start getting MP
1943	* events that will try kick it into action (includes the
1944	* RTMpOnAll/supdrvGipInitOnCpu call below).
1945	*/
1946	RTCpuSetEmpty(&pDevExt->TscDeltaCpuSet);
1947	RTCpuSetEmpty(&pDevExt->TscDeltaObtainedCpuSet);
1948	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
1949	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1950	rc = supdrvTscDeltaThreadInit(pDevExt);
1951	#endif
1952	if (RT_SUCCESS(rc))
1953	{
1954	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
1955	if (RT_SUCCESS(rc))
1956	{
1957	/*
1958	* Do GIP initialization on all online CPUs. Wake up the
1959	* TSC-delta thread afterwards.
1960	*/
1961	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
1962	if (RT_SUCCESS(rc))
1963	{
1964	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
1965	supdrvTscDeltaThreadStartMeasurement(pDevExt, true /* fForceAll */);
1966	#else
1967	uint16_t iCpu;
1968	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1969	{
1970	/*
1971	* Measure the TSC deltas now that we have MP notifications.
1972	*/
1973	int cTries = 5;
1974	do
1975	{
1976	rc = supdrvTscMeasureInitialDeltas(pDevExt);
1977	if ( rc != VERR_TRY_AGAIN
1978	&& rc != VERR_CPU_OFFLINE)
1979	break;
1980	} while (--cTries > 0);
1981	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
1982	Log(("supdrvTscDeltaInit: cpu[%u] delta %lld\n", iCpu, pGip->aCPUs[iCpu].i64TSCDelta));
1983	}
1984	else
1985	{
1986	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
1987	AssertMsg(!pGip->aCPUs[iCpu].i64TSCDelta, ("iCpu=%u %lld mode=%d\n", iCpu, pGip->aCPUs[iCpu].i64TSCDelta, pGip->u32Mode));
1988	}
1989	if (RT_SUCCESS(rc))
1990	#endif
1991	{
1992	/*
1993	* Create the timer.
1994	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
1995	*/
1996	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
1997	{
1998	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL,
1999	supdrvGipAsyncTimer, pDevExt);
2000	if (rc == VERR_NOT_SUPPORTED)
2001	{
2002	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
2003	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
2004	}
2005	}
2006	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
2007	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0 /* fFlags */,
2008	supdrvGipSyncAndInvariantTimer, pDevExt);
2009	if (RT_SUCCESS(rc))
2010	{
2011	/*
2012	* We're good.
2013	*/
2014	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
2015	supdrvGipReleaseHigherTimerFrequencyFromSystem(pDevExt);
2016
2017	g_pSUPGlobalInfoPage = pGip;
2018	return VINF_SUCCESS;
2019	}
2020
2021	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
2022	Assert(!pDevExt->pGipTimer);
2023	}
2024	}
2025	else
2026	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed. rc=%Rrc\n", rc));
2027	}
2028	else
2029	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
2030	}
2031	else
2032	OSDBGPRINT(("supdrvGipCreate: supdrvTscDeltaInit failed. rc=%Rrc\n", rc));
2033	}
2034	else
2035	OSDBGPRINT(("supdrvGipCreate: supdrvTscMeasureInitialDeltas failed. rc=%Rrc\n", rc));
2036	}
2037
2038	/* Releases timer frequency increase too. */
2039	supdrvGipDestroy(pDevExt);
2040	return rc;
2041	}
2042
2043
2044	/**
2045	* Invalidates the GIP data upon termination.
2046	*
2047	* @param pGip Pointer to the read-write kernel mapping of the GIP.
2048	*/
2049	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
2050	{
2051	unsigned i;
2052	pGip->u32Magic = 0;
2053	for (i = 0; i < pGip->cCpus; i++)
2054	{
2055	pGip->aCPUs[i].u64NanoTS = 0;
2056	pGip->aCPUs[i].u64TSC = 0;
2057	pGip->aCPUs[i].iTSCHistoryHead = 0;
2058	pGip->aCPUs[i].u64TSCSample = 0;
2059	pGip->aCPUs[i].i64TSCDelta = INT64_MAX;
2060	}
2061	}
2062
2063
2064	/**
2065	* Terminates the GIP.
2066	*
2067	* @param pDevExt Instance data. GIP stuff may be updated.
2068	*/
2069	void VBOXCALL supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
2070	{
2071	int rc;
2072	#ifdef DEBUG_DARWIN_GIP
2073	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
2074	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
2075	pDevExt->pGipTimer, pDevExt->GipMemObj));
2076	#endif
2077
2078	/*
2079	* Stop receiving MP notifications before tearing anything else down.
2080	*/
2081	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
2082
2083	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
2084	/*
2085	* Terminate the TSC-delta measurement thread and resources.
2086	*/
2087	supdrvTscDeltaTerm(pDevExt);
2088	#endif
2089
2090	/*
2091	* Destroy the TSC-refinement timer.
2092	*/
2093	if (pDevExt->pInvarTscRefineTimer)
2094	{
2095	RTTimerDestroy(pDevExt->pInvarTscRefineTimer);
2096	pDevExt->pInvarTscRefineTimer = NULL;
2097	}
2098
2099	/*
2100	* Invalid the GIP data.
2101	*/
2102	if (pDevExt->pGip)
2103	{
2104	supdrvGipTerm(pDevExt->pGip);
2105	pDevExt->pGip = NULL;
2106	}
2107	g_pSUPGlobalInfoPage = NULL;
2108
2109	/*
2110	* Destroy the timer and free the GIP memory object.
2111	*/
2112	if (pDevExt->pGipTimer)
2113	{
2114	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
2115	pDevExt->pGipTimer = NULL;
2116	}
2117
2118	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
2119	{
2120	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
2121	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
2122	}
2123
2124	/*
2125	* Finally, make sure we've release the system timer resolution request
2126	* if one actually succeeded and is still pending.
2127	*/
2128	supdrvGipReleaseHigherTimerFrequencyFromSystem(pDevExt);
2129	}
2130
2131
2132
2133
2134	/*
2135	*
2136	*
2137	* GIP Update Timer Related Code
2138	* GIP Update Timer Related Code
2139	* GIP Update Timer Related Code
2140	*
2141	*
2142	*/
2143
2144
2145	/**
2146	* Worker routine for supdrvGipUpdate() and supdrvGipUpdatePerCpu() that
2147	* updates all the per cpu data except the transaction id.
2148	*
2149	* @param pDevExt The device extension.
2150	* @param pGipCpu Pointer to the per cpu data.
2151	* @param u64NanoTS The current time stamp.
2152	* @param u64TSC The current TSC.
2153	* @param iTick The current timer tick.
2154	*
2155	* @remarks Can be called with interrupts disabled!
2156	*/
2157	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
2158	{
2159	uint64_t u64TSCDelta;
2160	bool fUpdateCpuHz;
2161	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2162	AssertPtrReturnVoid(pGip);
2163
2164	/* Delta between this and the previous update. */
2165	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
2166
2167	/*
2168	* Update the NanoTS.
2169	*/
2170	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
2171
2172	/*
2173	* Calc TSC delta.
2174	*/
2175	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
2176	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
2177
2178	/*
2179	* Determine if we need to update the CPU (TSC) frequency calculation.
2180	*
2181	* We don't need to keep recalculating the frequency when it's invariant,
2182	* unless the special tstGIP-2 testing mode is enabled.
2183	*/
2184	fUpdateCpuHz = pGip->u32Mode != SUPGIPMODE_INVARIANT_TSC;
2185	if (!(pGip->fFlags & SUPGIP_FLAGS_TESTING))
2186	{ /* likely*/ }
2187	else
2188	{
2189	uint32_t fGipFlags = pGip->fFlags;
2190	if (fGipFlags & (SUPGIP_FLAGS_TESTING_ENABLE \| SUPGIP_FLAGS_TESTING_START))
2191	{
2192	if (fGipFlags & SUPGIP_FLAGS_TESTING_START)
2193	{
2194	/* Cache the TSC frequency before forcing updates due to test mode. */
2195	if (!fUpdateCpuHz)
2196	pDevExt->uGipTestModeInvariantCpuHz = pGip->aCPUs[0].u64CpuHz;
2197	ASMAtomicAndU32(&pGip->fFlags, ~SUPGIP_FLAGS_TESTING_START);
2198	}
2199	fUpdateCpuHz = true;
2200	}
2201	else if (fGipFlags & SUPGIP_FLAGS_TESTING_STOP)
2202	{
2203	/* Restore the cached TSC frequency if any. */
2204	if (!fUpdateCpuHz)
2205	{
2206	Assert(pDevExt->uGipTestModeInvariantCpuHz);
2207	ASMAtomicWriteU64(&pGip->aCPUs[0].u64CpuHz, pDevExt->uGipTestModeInvariantCpuHz);
2208	}
2209	ASMAtomicAndU32(&pGip->fFlags, ~(SUPGIP_FLAGS_TESTING_STOP \| SUPGIP_FLAGS_TESTING));
2210	}
2211	}
2212
2213	/*
2214	* Calculate the CPU (TSC) frequency if necessary.
2215	*/
2216	if (fUpdateCpuHz)
2217	{
2218	uint64_t u64CpuHz;
2219	uint32_t u32UpdateIntervalTSC;
2220	uint32_t u32UpdateIntervalTSCSlack;
2221	uint32_t u32TransactionId;
2222	unsigned iTSCHistoryHead;
2223
2224	if (u64TSCDelta >> 32)
2225	{
2226	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
2227	pGipCpu->cErrors++;
2228	}
2229
2230	/*
2231	* On the 2nd and 3rd callout, reset the history with the current TSC
2232	* interval since the values entered by supdrvGipInit are totally off.
2233	* The interval on the 1st callout completely unreliable, the 2nd is a bit
2234	* better, while the 3rd should be most reliable.
2235	*/
2236	/** @todo Could we drop this now that we initializes the history
2237	* with nominal TSC frequency values? */
2238	u32TransactionId = pGipCpu->u32TransactionId;
2239	if (RT_UNLIKELY( ( u32TransactionId == 5
2240	\|\| u32TransactionId == 7)
2241	&& ( iTick == 2
2242	\|\| iTick == 3) ))
2243	{
2244	unsigned i;
2245	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
2246	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
2247	}
2248
2249	/*
2250	* Validate the NanoTS deltas between timer fires with an arbitrary threshold of 0.5%.
2251	* Wait until we have at least one full history since the above history reset. The
2252	* assumption is that the majority of the previous history values will be tolerable.
2253	* See @bugref{6710#c67}.
2254	*/
2255	/** @todo Could we drop the fudging there now that we initializes the history
2256	* with nominal TSC frequency values? */
2257	if ( u32TransactionId > 23 /* 7 + (8 * 2) */
2258	&& pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
2259	{
2260	uint32_t uNanoTsThreshold = pGip->u32UpdateIntervalNS / 200;
2261	if ( pGipCpu->u32PrevUpdateIntervalNS > pGip->u32UpdateIntervalNS + uNanoTsThreshold
2262	\|\| pGipCpu->u32PrevUpdateIntervalNS < pGip->u32UpdateIntervalNS - uNanoTsThreshold)
2263	{
2264	uint32_t u32;
2265	u32 = pGipCpu->au32TSCHistory[0];
2266	u32 += pGipCpu->au32TSCHistory[1];
2267	u32 += pGipCpu->au32TSCHistory[2];
2268	u32 += pGipCpu->au32TSCHistory[3];
2269	u32 >>= 2;
2270	u64TSCDelta = pGipCpu->au32TSCHistory[4];
2271	u64TSCDelta += pGipCpu->au32TSCHistory[5];
2272	u64TSCDelta += pGipCpu->au32TSCHistory[6];
2273	u64TSCDelta += pGipCpu->au32TSCHistory[7];
2274	u64TSCDelta >>= 2;
2275	u64TSCDelta += u32;
2276	u64TSCDelta >>= 1;
2277	}
2278	}
2279
2280	/*
2281	* TSC History.
2282	*/
2283	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
2284	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
2285	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
2286	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
2287
2288	/*
2289	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
2290	*
2291	* On Windows, we have an occasional (but recurring) sour value that messed up
2292	* the history but taking only 1 interval reduces the precision overall.
2293	*/
2294	if ( pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC
2295	\|\| pGip->u32UpdateHz >= 1000)
2296	{
2297	uint32_t u32;
2298	u32 = pGipCpu->au32TSCHistory[0];
2299	u32 += pGipCpu->au32TSCHistory[1];
2300	u32 += pGipCpu->au32TSCHistory[2];
2301	u32 += pGipCpu->au32TSCHistory[3];
2302	u32 >>= 2;
2303	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
2304	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
2305	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
2306	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
2307	u32UpdateIntervalTSC >>= 2;
2308	u32UpdateIntervalTSC += u32;
2309	u32UpdateIntervalTSC >>= 1;
2310
2311	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11. */
2312	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
2313	}
2314	else if (pGip->u32UpdateHz >= 90)
2315	{
2316	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
2317	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
2318	u32UpdateIntervalTSC >>= 1;
2319
2320	/* value chosen on a 2GHz thinkpad running windows */
2321	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
2322	}
2323	else
2324	{
2325	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
2326
2327	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
2328	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
2329	}
2330	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
2331
2332	/*
2333	* CpuHz.
2334	*/
2335	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, RT_NS_1SEC);
2336	u64CpuHz /= pGip->u32UpdateIntervalNS;
2337	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
2338	}
2339	}
2340
2341
2342	/**
2343	* Updates the GIP.
2344	*
2345	* @param pDevExt The device extension.
2346	* @param u64NanoTS The current nanosecond timestamp.
2347	* @param u64TSC The current TSC timestamp.
2348	* @param idCpu The CPU ID.
2349	* @param iTick The current timer tick.
2350	*
2351	* @remarks Can be called with interrupts disabled!
2352	*/
2353	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
2354	{
2355	/*
2356	* Determine the relevant CPU data.
2357	*/
2358	PSUPGIPCPU pGipCpu;
2359	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2360	AssertPtrReturnVoid(pGip);
2361
2362	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
2363	pGipCpu = &pGip->aCPUs[0];
2364	else
2365	{
2366	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
2367	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
2368	return;
2369	pGipCpu = &pGip->aCPUs[iCpu];
2370	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
2371	return;
2372	}
2373
2374	/*
2375	* Start update transaction.
2376	*/
2377	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
2378	{
2379	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
2380	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
2381	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2382	pGipCpu->cErrors++;
2383	return;
2384	}
2385
2386	/*
2387	* Recalc the update frequency every 0x800th time.
2388	*/
2389	if ( pGip->u32Mode != SUPGIPMODE_INVARIANT_TSC /* cuz we're not recalculating the frequency on invariant hosts. */
2390	&& !(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
2391	{
2392	if (pGip->u64NanoTSLastUpdateHz)
2393	{
2394	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
2395	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
2396	uint32_t u32UpdateHz = (uint32_t)((RT_NS_1SEC_64 * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
2397	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
2398	{
2399	/** @todo r=ramshankar: Changing u32UpdateHz might screw up TSC frequency
2400	* calculation on non-invariant hosts if it changes the history decision
2401	* taken in supdrvGipDoUpdateCpu(). */
2402	uint64_t u64Interval = u64Delta / GIP_UPDATEHZ_RECALC_FREQ;
2403	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
2404	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, (uint32_t)u64Interval);
2405	}
2406	#endif
2407	}
2408	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS \| 1);
2409	}
2410
2411	/*
2412	* Update the data.
2413	*/
2414	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
2415
2416	/*
2417	* Complete transaction.
2418	*/
2419	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2420	}
2421
2422
2423	/**
2424	* Updates the per cpu GIP data for the calling cpu.
2425	*
2426	* @param pDevExt The device extension.
2427	* @param u64NanoTS The current nanosecond timestamp.
2428	* @param u64TSC The current TSC timesaver.
2429	* @param idCpu The CPU ID.
2430	* @param idApic The APIC id for the CPU index.
2431	* @param iTick The current timer tick.
2432	*
2433	* @remarks Can be called with interrupts disabled!
2434	*/
2435	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
2436	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
2437	{
2438	uint32_t iCpu;
2439	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2440
2441	/*
2442	* Avoid a potential race when a CPU online notification doesn't fire on
2443	* the onlined CPU but the tick creeps in before the event notification is
2444	* run.
2445	*/
2446	if (RT_LIKELY(iTick != 1))
2447	{ /* likely*/ }
2448	else
2449	{
2450	iCpu = supdrvGipFindOrAllocCpuIndexForCpuId(pGip, idCpu);
2451	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
2452	supdrvGipMpEventOnlineOrInitOnCpu(pDevExt, idCpu);
2453	}
2454
2455	iCpu = pGip->aiCpuFromApicId[idApic];
2456	if (RT_LIKELY(iCpu < pGip->cCpus))
2457	{
2458	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
2459	if (pGipCpu->idCpu == idCpu)
2460	{
2461	/*
2462	* Start update transaction.
2463	*/
2464	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
2465	{
2466	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
2467	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2468	pGipCpu->cErrors++;
2469	return;
2470	}
2471
2472	/*
2473	* Update the data.
2474	*/
2475	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
2476
2477	/*
2478	* Complete transaction.
2479	*/
2480	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2481	}
2482	}
2483	}
2484
2485
2486	/**
2487	* Timer callback function for the sync and invariant GIP modes.
2488	*
2489	* @param pTimer The timer.
2490	* @param pvUser Opaque pointer to the device extension.
2491	* @param iTick The timer tick.
2492	*/
2493	static DECLCALLBACK(void) supdrvGipSyncAndInvariantTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
2494	{
2495	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
2496	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2497	RTCCUINTREG fEFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
2498	uint64_t u64TSC = ASMReadTSC();
2499	uint64_t u64NanoTS = RTTimeSystemNanoTS();
2500	RT_NOREF1(pTimer);
2501
2502	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_PRACTICALLY_ZERO)
2503	{
2504	/*
2505	* The calculations in supdrvGipUpdate() is somewhat timing sensitive,
2506	* missing timer ticks is not an option for GIP because the GIP users
2507	* will end up incrementing the time in 1ns per time getter call until
2508	* there is a complete timer update. So, if the delta has yet to be
2509	* calculated, we just pretend it is zero for now (the GIP users
2510	* probably won't have it for a wee while either and will do the same).
2511	*
2512	* We could maybe on some platforms try cross calling a CPU with a
2513	* working delta here, but it's not worth the hassle since the
2514	* likelihood of this happening is really low. On Windows, Linux, and
2515	* Solaris timers fire on the CPU they were registered/started on.
2516	* Darwin timers doesn't necessarily (they are high priority threads).
2517	*/
2518	uint32_t iCpuSet = RTMpCpuIdToSetIndex(RTMpCpuId());
2519	uint16_t iGipCpu = RT_LIKELY(iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx))
2520	? pGip->aiCpuFromCpuSetIdx[iCpuSet] : UINT16_MAX;
2521	Assert(!ASMIntAreEnabled());
2522	if (RT_LIKELY(iGipCpu < pGip->cCpus))
2523	{
2524	int64_t iTscDelta = pGip->aCPUs[iGipCpu].i64TSCDelta;
2525	if (iTscDelta != INT64_MAX)
2526	u64TSC -= iTscDelta;
2527	}
2528	}
2529
2530	supdrvGipUpdate(pDevExt, u64NanoTS, u64TSC, NIL_RTCPUID, iTick);
2531
2532	ASMSetFlags(fEFlags);
2533	}
2534
2535
2536	/**
2537	* Timer callback function for async GIP mode.
2538	* @param pTimer The timer.
2539	* @param pvUser Opaque pointer to the device extension.
2540	* @param iTick The timer tick.
2541	*/
2542	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
2543	{
2544	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
2545	RTCCUINTREG fEFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
2546	RTCPUID idCpu = RTMpCpuId();
2547	uint64_t u64TSC = ASMReadTSC();
2548	uint64_t NanoTS = RTTimeSystemNanoTS();
2549	RT_NOREF1(pTimer);
2550
2551	/** @todo reset the transaction number and whatnot when iTick == 1. */
2552	if (pDevExt->idGipMaster == idCpu)
2553	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
2554	else
2555	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
2556
2557	ASMSetFlags(fEFlags);
2558	}
2559
2560
2561
2562
2563	/*
2564	*
2565	*
2566	* TSC Delta Measurements And Related Code
2567	* TSC Delta Measurements And Related Code
2568	* TSC Delta Measurements And Related Code
2569	*
2570	*
2571	*/
2572
2573
2574	/*
2575	* Select TSC delta measurement algorithm.
2576	*/
2577	#if 0
2578	# define GIP_TSC_DELTA_METHOD_1
2579	#else
2580	# define GIP_TSC_DELTA_METHOD_2
2581	#endif
2582
2583	/** For padding variables to keep them away from other cache lines. Better too
2584	* large than too small!
2585	* @remarks Current AMD64 and x86 CPUs seems to use 64 bytes. There are claims
2586	* that NetBurst had 128 byte cache lines while the 486 thru Pentium
2587	* III had 32 bytes cache lines. */
2588	#define GIP_TSC_DELTA_CACHE_LINE_SIZE 128
2589
2590
2591	/**
2592	* TSC delta measurement algorithm \#2 result entry.
2593	*/
2594	typedef struct SUPDRVTSCDELTAMETHOD2ENTRY
2595	{
2596	uint32_t iSeqMine;
2597	uint32_t iSeqOther;
2598	uint64_t uTsc;
2599	} SUPDRVTSCDELTAMETHOD2ENTRY;
2600
2601	/**
2602	* TSC delta measurement algorithm \#2 Data.
2603	*/
2604	typedef struct SUPDRVTSCDELTAMETHOD2
2605	{
2606	/** Padding to make sure the iCurSeqNo is in its own cache line. */
2607	uint64_t au64CacheLinePaddingBefore[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2608	/** The current sequence number of this worker. */
2609	uint32_t volatile iCurSeqNo;
2610	/** Padding to make sure the iCurSeqNo is in its own cache line. */
2611	uint32_t au64CacheLinePaddingAfter[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint32_t) - 1];
2612	/** Result table. */
2613	SUPDRVTSCDELTAMETHOD2ENTRY aResults[64];
2614	} SUPDRVTSCDELTAMETHOD2;
2615	/** Pointer to the data for TSC delta measurement algorithm \#2 .*/
2616	typedef SUPDRVTSCDELTAMETHOD2 *PSUPDRVTSCDELTAMETHOD2;
2617
2618
2619	/**
2620	* The TSC delta synchronization struct, version 2.
2621	*
2622	* The synchronization variable is completely isolated in its own cache line
2623	* (provided our max cache line size estimate is correct).
2624	*/
2625	typedef struct SUPTSCDELTASYNC2
2626	{
2627	/** Padding to make sure the uVar1 is in its own cache line. */
2628	uint64_t au64CacheLinePaddingBefore[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2629
2630	/** The synchronization variable, holds values GIP_TSC_DELTA_SYNC_. /
2631	volatile uint32_t uSyncVar;
2632	/** Sequence synchronizing variable used for post 'GO' synchronization. */
2633	volatile uint32_t uSyncSeq;
2634
2635	/** Padding to make sure the uVar1 is in its own cache line. */
2636	uint64_t au64CacheLinePaddingAfter[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t) - 2];
2637
2638	/** Start RDTSC value. Put here mainly to save stack space. */
2639	uint64_t uTscStart;
2640	/** Copy of SUPDRVGIPTSCDELTARGS::cMaxTscTicks. */
2641	uint64_t cMaxTscTicks;
2642	} SUPTSCDELTASYNC2;
2643	AssertCompileSize(SUPTSCDELTASYNC2, GIP_TSC_DELTA_CACHE_LINE_SIZE * 2 + sizeof(uint64_t));
2644	typedef SUPTSCDELTASYNC2 *PSUPTSCDELTASYNC2;
2645
2646	/** Prestart wait. */
2647	#define GIP_TSC_DELTA_SYNC2_PRESTART_WAIT UINT32_C(0x0ffe)
2648	/** Prestart aborted. */
2649	#define GIP_TSC_DELTA_SYNC2_PRESTART_ABORT UINT32_C(0x0fff)
2650	/** Ready (on your mark). */
2651	#define GIP_TSC_DELTA_SYNC2_READY UINT32_C(0x1000)
2652	/** Steady (get set). */
2653	#define GIP_TSC_DELTA_SYNC2_STEADY UINT32_C(0x1001)
2654	/** Go! */
2655	#define GIP_TSC_DELTA_SYNC2_GO UINT32_C(0x1002)
2656	/** Used by the verification test. */
2657	#define GIP_TSC_DELTA_SYNC2_GO_GO UINT32_C(0x1003)
2658
2659	/** We reached the time limit. */
2660	#define GIP_TSC_DELTA_SYNC2_TIMEOUT UINT32_C(0x1ffe)
2661	/** The other party won't touch the sync struct ever again. */
2662	#define GIP_TSC_DELTA_SYNC2_FINAL UINT32_C(0x1fff)
2663
2664
2665	/**
2666	* Argument package/state passed by supdrvTscMeasureDeltaOne() to the RTMpOn
2667	* callback worker.
2668	* @todo add
2669	*/
2670	typedef struct SUPDRVGIPTSCDELTARGS
2671	{
2672	/** The device extension. */
2673	PSUPDRVDEVEXT pDevExt;
2674	/** Pointer to the GIP CPU array entry for the worker. */
2675	PSUPGIPCPU pWorker;
2676	/** Pointer to the GIP CPU array entry for the master. */
2677	PSUPGIPCPU pMaster;
2678	/** The maximum number of ticks to spend in supdrvTscMeasureDeltaCallback.
2679	* (This is what we need a rough TSC frequency for.) */
2680	uint64_t cMaxTscTicks;
2681	/** Used to abort synchronization setup. */
2682	bool volatile fAbortSetup;
2683
2684	/** Padding to make sure the master variables live in its own cache lines. */
2685	uint64_t au64CacheLinePaddingBefore[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2686
2687	/** @name Master
2688	* @{ */
2689	/** The time the master spent in the MP worker. */
2690	uint64_t cElapsedMasterTscTicks;
2691	/** The iTry value when stopped at. */
2692	uint32_t iTry;
2693	/** Set if the run timed out. */
2694	bool volatile fTimedOut;
2695	/** Pointer to the master's synchronization struct (on stack). */
2696	PSUPTSCDELTASYNC2 volatile pSyncMaster;
2697	/** Master data union. */
2698	union
2699	{
2700	/** Data (master) for delta verification. */
2701	struct
2702	{
2703	/** Verification test TSC values for the master. */
2704	uint64_t volatile auTscs[32];
2705	} Verify;
2706	/** Data (master) for measurement method \#2. */
2707	struct
2708	{
2709	/** Data and sequence number. */
2710	SUPDRVTSCDELTAMETHOD2 Data;
2711	/** The lag setting for the next run. */
2712	bool fLag;
2713	/** Number of hits. */
2714	uint32_t cHits;
2715	} M2;
2716	} uMaster;
2717	/** The verifier verdict, VINF_SUCCESS if ok, VERR_OUT_OF_RANGE if not,
2718	* VERR_TRY_AGAIN on timeout. */
2719	int32_t rcVerify;
2720	#ifdef TSCDELTA_VERIFY_WITH_STATS
2721	/** The maximum difference between TSC read during delta verification. */
2722	int64_t cMaxVerifyTscTicks;
2723	/** The minimum difference between two TSC reads during verification. */
2724	int64_t cMinVerifyTscTicks;
2725	/** The bad TSC diff, worker relative to master (= worker - master).
2726	* Negative value means the worker is behind the master. */
2727	int64_t iVerifyBadTscDiff;
2728	#endif
2729	/** @} */
2730
2731	/** Padding to make sure the worker variables live is in its own cache line. */
2732	uint64_t au64CacheLinePaddingBetween[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2733
2734	/** @name Proletarian
2735	* @{ */
2736	/** Pointer to the worker's synchronization struct (on stack). */
2737	PSUPTSCDELTASYNC2 volatile pSyncWorker;
2738	/** The time the worker spent in the MP worker. */
2739	uint64_t cElapsedWorkerTscTicks;
2740	/** Worker data union. */
2741	union
2742	{
2743	/** Data (worker) for delta verification. */
2744	struct
2745	{
2746	/** Verification test TSC values for the worker. */
2747	uint64_t volatile auTscs[32];
2748	} Verify;
2749	/** Data (worker) for measurement method \#2. */
2750	struct
2751	{
2752	/** Data and sequence number. */
2753	SUPDRVTSCDELTAMETHOD2 Data;
2754	/** The lag setting for the next run (set by master). */
2755	bool fLag;
2756	} M2;
2757	} uWorker;
2758	/** @} */
2759
2760	/** Padding to make sure the above is in its own cache line. */
2761	uint64_t au64CacheLinePaddingAfter[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2762	} SUPDRVGIPTSCDELTARGS;
2763	typedef SUPDRVGIPTSCDELTARGS *PSUPDRVGIPTSCDELTARGS;
2764
2765
2766	/** @name Macros that implements the basic synchronization steps common to
2767	* the algorithms.
2768	*
2769	* Must be used from loop as the timeouts are implemented via 'break' statements
2770	* at the moment.
2771	*
2772	* @{
2773	*/
2774	#if defined(DEBUG_bird) /* \|\| defined(VBOX_STRICT) */
2775	# define TSCDELTA_DBG_VARS() uint32_t iDbgCounter
2776	# define TSCDELTA_DBG_START_LOOP() do { iDbgCounter = 0; } while (0)
2777	# define TSCDELTA_DBG_CHECK_LOOP() \
2778	do { iDbgCounter++; if ((iDbgCounter & UINT32_C(0x01ffffff)) == 0) RT_BREAKPOINT(); } while (0)
2779	#else
2780	# define TSCDELTA_DBG_VARS() ((void)0)
2781	# define TSCDELTA_DBG_START_LOOP() ((void)0)
2782	# define TSCDELTA_DBG_CHECK_LOOP() ((void)0)
2783	#endif
2784	#if 0
2785	# define TSCDELTA_DBG_SYNC_MSG(a_Args) SUPR0Printf a_Args
2786	#else
2787	# define TSCDELTA_DBG_SYNC_MSG(a_Args) ((void)0)
2788	#endif
2789	#if 0
2790	# define TSCDELTA_DBG_SYNC_MSG2(a_Args) SUPR0Printf a_Args
2791	#else
2792	# define TSCDELTA_DBG_SYNC_MSG2(a_Args) ((void)0)
2793	#endif
2794	#if 0
2795	# define TSCDELTA_DBG_SYNC_MSG9(a_Args) SUPR0Printf a_Args
2796	#else
2797	# define TSCDELTA_DBG_SYNC_MSG9(a_Args) ((void)0)
2798	#endif
2799
2800
2801	static bool supdrvTscDeltaSync2_Before(PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
2802	bool fIsMaster, PRTCCUINTREG pfEFlags, PSUPDRVGIPTSCDELTARGS pArgs)
2803	{
2804	uint32_t iMySeq = fIsMaster ? 0 : 256;
2805	uint32_t const iMaxSeq = iMySeq + 16; /* For the last loop, darn linux/freebsd C-ishness. */
2806	uint32_t u32Tmp;
2807	uint32_t iSync2Loops = 0;
2808	RTCCUINTREG fEFlags;
2809	TSCDELTA_DBG_VARS();
2810
2811	pfEFlags = X86_EFL_IF \| X86_EFL_1; / should shut up most nagging compilers. */
2812
2813	/*
2814	* The master tells the worker to get on it's mark.
2815	*/
2816	if (fIsMaster)
2817	{
2818	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_STEADY, GIP_TSC_DELTA_SYNC2_READY)))
2819	{ /* likely*/ }
2820	else
2821	{
2822	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #1 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2823	return false;
2824	}
2825	}
2826
2827	/*
2828	* Wait for the on your mark signal (ack in the master case). We process timeouts here.
2829	*/
2830	ASMAtomicWriteU32(&(pMySync)->uSyncSeq, 0);
2831	for (;;)
2832	{
2833	fEFlags = ASMIntDisableFlags();
2834	u32Tmp = ASMAtomicReadU32(&pMySync->uSyncVar);
2835	if (u32Tmp == GIP_TSC_DELTA_SYNC2_STEADY)
2836	break;
2837	ASMSetFlags(fEFlags);
2838	ASMNopPause();
2839
2840	/* Abort? */
2841	if (u32Tmp != GIP_TSC_DELTA_SYNC2_READY)
2842	{
2843	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #2 u32Tmp=%#x\n", fIsMaster ? "master" : "worker", u32Tmp));
2844	return false;
2845	}
2846
2847	/* Check for timeouts every so often (not every loop in case RDTSC is
2848	trapping or something). Must check the first time around. */
2849	#if 0 /* For debugging the timeout paths. */
2850	static uint32_t volatile xxx;
2851	#endif
2852	if ( ( (iSync2Loops & 0x3ff) == 0
2853	&& ASMReadTSC() - pMySync->uTscStart > pMySync->cMaxTscTicks)
2854	#if 0 /* This is crazy, I know, but enable this code and the results are markedly better when enabled on the 1.4GHz AMD (debug). */
2855	\|\| (!fIsMaster && (++xxx & 0xf) == 0)
2856	#endif
2857	)
2858	{
2859	/* Try switch our own state into timeout mode so the master cannot tell us to 'GO',
2860	ignore the timeout if we've got the go ahead already (simpler). */
2861	if (ASMAtomicCmpXchgU32(&pMySync->uSyncVar, GIP_TSC_DELTA_SYNC2_TIMEOUT, GIP_TSC_DELTA_SYNC2_READY))
2862	{
2863	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: timeout\n", fIsMaster ? "master" : "worker"));
2864	ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_TIMEOUT, GIP_TSC_DELTA_SYNC2_STEADY);
2865	ASMAtomicWriteBool(&pArgs->fTimedOut, true);
2866	return false;
2867	}
2868	}
2869	iSync2Loops++;
2870	}
2871
2872	/*
2873	* Interrupts are now disabled and will remain disabled until we do
2874	* TSCDELTA_MASTER_SYNC_AFTER / TSCDELTA_OTHER_SYNC_AFTER.
2875	*/
2876	*pfEFlags = fEFlags;
2877
2878	/*
2879	* The worker tells the master that it is on its mark and that the master
2880	* need to get into position as well.
2881	*/
2882	if (!fIsMaster)
2883	{
2884	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_STEADY, GIP_TSC_DELTA_SYNC2_READY)))
2885	{ /* likely */ }
2886	else
2887	{
2888	ASMSetFlags(fEFlags);
2889	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #3 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2890	return false;
2891	}
2892	}
2893
2894	/*
2895	* The master sends the 'go' to the worker and wait for ACK.
2896	*/
2897	if (fIsMaster)
2898	{
2899	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO, GIP_TSC_DELTA_SYNC2_STEADY)))
2900	{ /* likely */ }
2901	else
2902	{
2903	ASMSetFlags(fEFlags);
2904	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #4 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2905	return false;
2906	}
2907	}
2908
2909	/*
2910	* Wait for the 'go' signal (ack in the master case).
2911	*/
2912	TSCDELTA_DBG_START_LOOP();
2913	for (;;)
2914	{
2915	u32Tmp = ASMAtomicReadU32(&pMySync->uSyncVar);
2916	if (u32Tmp == GIP_TSC_DELTA_SYNC2_GO)
2917	break;
2918	if (RT_LIKELY(u32Tmp == GIP_TSC_DELTA_SYNC2_STEADY))
2919	{ /* likely */ }
2920	else
2921	{
2922	ASMSetFlags(fEFlags);
2923	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #5 u32Tmp=%#x\n", fIsMaster ? "master" : "worker", u32Tmp));
2924	return false;
2925	}
2926
2927	TSCDELTA_DBG_CHECK_LOOP();
2928	ASMNopPause();
2929	}
2930
2931	/*
2932	* The worker acks the 'go' (shouldn't fail).
2933	*/
2934	if (!fIsMaster)
2935	{
2936	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO, GIP_TSC_DELTA_SYNC2_STEADY)))
2937	{ /* likely */ }
2938	else
2939	{
2940	ASMSetFlags(fEFlags);
2941	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #6 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2942	return false;
2943	}
2944	}
2945
2946	/*
2947	* Try enter mostly lockstep execution with it.
2948	*/
2949	for (;;)
2950	{
2951	uint32_t iOtherSeq1, iOtherSeq2;
2952	ASMCompilerBarrier();
2953	ASMSerializeInstruction();
2954
2955	ASMAtomicWriteU32(&pMySync->uSyncSeq, iMySeq);
2956	ASMNopPause();
2957	iOtherSeq1 = ASMAtomicXchgU32(&pOtherSync->uSyncSeq, iMySeq);
2958	ASMNopPause();
2959	iOtherSeq2 = ASMAtomicReadU32(&pMySync->uSyncSeq);
2960
2961	ASMCompilerBarrier();
2962	if (iOtherSeq1 == iOtherSeq2)
2963	return true;
2964
2965	/* Did the other guy give up? Should we give up? */
2966	if ( iOtherSeq1 == UINT32_MAX
2967	\|\| iOtherSeq2 == UINT32_MAX)
2968	return true;
2969	if (++iMySeq >= iMaxSeq)
2970	{
2971	ASMAtomicWriteU32(&pMySync->uSyncSeq, UINT32_MAX);
2972	return true;
2973	}
2974	ASMNopPause();
2975	}
2976	}
2977
2978	#define TSCDELTA_MASTER_SYNC_BEFORE(a_pMySync, a_pOtherSync, a_pfEFlags, a_pArgs) \
2979	if (RT_LIKELY(supdrvTscDeltaSync2_Before(a_pMySync, a_pOtherSync, true /fIsMaster/, a_pfEFlags, a_pArgs))) \
2980	{ /likely/ } \
2981	else if (true) \
2982	{ \
2983	TSCDELTA_DBG_SYNC_MSG9(("sync/before/master: #89\n")); \
2984	break; \
2985	} else do {} while (0)
2986	#define TSCDELTA_OTHER_SYNC_BEFORE(a_pMySync, a_pOtherSync, a_pfEFlags, a_pArgs) \
2987	if (RT_LIKELY(supdrvTscDeltaSync2_Before(a_pMySync, a_pOtherSync, false /fIsMaster/, a_pfEFlags, a_pArgs))) \
2988	{ /likely/ } \
2989	else if (true) \
2990	{ \
2991	TSCDELTA_DBG_SYNC_MSG9(("sync/before/other: #89\n")); \
2992	break; \
2993	} else do {} while (0)
2994
2995
2996	static bool supdrvTscDeltaSync2_After(PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
2997	bool fIsMaster, RTCCUINTREG fEFlags)
2998	{
2999	TSCDELTA_DBG_VARS();
3000	RT_NOREF1(pOtherSync);
3001
3002	/*
3003	* Wait for the 'ready' signal. In the master's case, this means the
3004	* worker has completed its data collection, while in the worker's case it
3005	* means the master is done processing the data and it's time for the next
3006	* loop iteration (or whatever).
3007	*/
3008	ASMSetFlags(fEFlags);
3009	TSCDELTA_DBG_START_LOOP();
3010	for (;;)
3011	{
3012	uint32_t u32Tmp = ASMAtomicReadU32(&pMySync->uSyncVar);
3013	if ( u32Tmp == GIP_TSC_DELTA_SYNC2_READY
3014	\|\| (u32Tmp == GIP_TSC_DELTA_SYNC2_STEADY && !fIsMaster) /* kicked twice => race */ )
3015	return true;
3016	ASMNopPause();
3017	if (RT_LIKELY(u32Tmp == GIP_TSC_DELTA_SYNC2_GO))
3018	{ /* likely */}
3019	else
3020	{
3021	TSCDELTA_DBG_SYNC_MSG(("sync/after/other: #1 u32Tmp=%#x\n", u32Tmp));
3022	return false; /* shouldn't ever happen! */
3023	}
3024	TSCDELTA_DBG_CHECK_LOOP();
3025	ASMNopPause();
3026	}
3027	}
3028
3029	#define TSCDELTA_MASTER_SYNC_AFTER(a_pMySync, a_pOtherSync, a_fEFlags) \
3030	if (RT_LIKELY(supdrvTscDeltaSync2_After(a_pMySync, a_pOtherSync, true /fIsMaster/, a_fEFlags))) \
3031	{ /* likely */ } \
3032	else if (true) \
3033	{ \
3034	TSCDELTA_DBG_SYNC_MSG9(("sync/after/master: #97\n")); \
3035	break; \
3036	} else do {} while (0)
3037
3038	#define TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(a_pMySync, a_pOtherSync) \
3039	/* \
3040	* Tell the worker that we're done processing the data and ready for the next round. \
3041	*/ \
3042	if (RT_LIKELY(ASMAtomicCmpXchgU32(&(a_pOtherSync)->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_GO))) \
3043	{ /* likely */ } \
3044	else if (true)\
3045	{ \
3046	TSCDELTA_DBG_SYNC_MSG(("sync/after/master: #99 uSyncVar=%#x\n", (a_pOtherSync)->uSyncVar)); \
3047	break; \
3048	} else do {} while (0)
3049
3050	#define TSCDELTA_OTHER_SYNC_AFTER(a_pMySync, a_pOtherSync, a_fEFlags) \
3051	if (true) { \
3052	/* \
3053	* Tell the master that we're done collecting data and wait for the next round to start. \
3054	*/ \
3055	if (RT_LIKELY(ASMAtomicCmpXchgU32(&(a_pOtherSync)->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_GO))) \
3056	{ /* likely */ } \
3057	else \
3058	{ \
3059	ASMSetFlags(a_fEFlags); \
3060	TSCDELTA_DBG_SYNC_MSG(("sync/after/other: #0 uSyncVar=%#x\n", (a_pOtherSync)->uSyncVar)); \
3061	break; \
3062	} \
3063	if (RT_LIKELY(supdrvTscDeltaSync2_After(a_pMySync, a_pOtherSync, false /fIsMaster/, a_fEFlags))) \
3064	{ /* likely */ } \
3065	else \
3066	{ \
3067	TSCDELTA_DBG_SYNC_MSG9(("sync/after/other: #98\n")); \
3068	break; \
3069	} \
3070	} else do {} while (0)
3071	/** @} */
3072
3073
3074	#ifdef GIP_TSC_DELTA_METHOD_1
3075	/**
3076	* TSC delta measurement algorithm \#1 (GIP_TSC_DELTA_METHOD_1).
3077	*
3078	*
3079	* We ignore the first few runs of the loop in order to prime the
3080	* cache. Also, we need to be careful about using 'pause' instruction
3081	* in critical busy-wait loops in this code - it can cause undesired
3082	* behaviour with hyperthreading.
3083	*
3084	* We try to minimize the measurement error by computing the minimum
3085	* read time of the compare statement in the worker by taking TSC
3086	* measurements across it.
3087	*
3088	* It must be noted that the computed minimum read time is mostly to
3089	* eliminate huge deltas when the worker is too early and doesn't by
3090	* itself help produce more accurate deltas. We allow two times the
3091	* computed minimum as an arbitrary acceptable threshold. Therefore,
3092	* it is still possible to get negative deltas where there are none
3093	* when the worker is earlier. As long as these occasional negative
3094	* deltas are lower than the time it takes to exit guest-context and
3095	* the OS to reschedule EMT on a different CPU, we won't expose a TSC
3096	* that jumped backwards. It is due to the existence of the negative
3097	* deltas that we don't recompute the delta with the master and
3098	* worker interchanged to eliminate the remaining measurement error.
3099	*
3100	*
3101	* @param pArgs The argument/state data.
3102	* @param pMySync My synchronization structure.
3103	* @param pOtherSync My partner's synchronization structure.
3104	* @param fIsMaster Set if master, clear if worker.
3105	* @param iTry The attempt number.
3106	*/
3107	static void supdrvTscDeltaMethod1Loop(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
3108	bool fIsMaster, uint32_t iTry)
3109	{
3110	PSUPGIPCPU pGipCpuWorker = pArgs->pWorker;
3111	PSUPGIPCPU pGipCpuMaster = pArgs->pMaster;
3112	uint64_t uMinCmpReadTime = UINT64_MAX;
3113	unsigned iLoop;
3114	NOREF(iTry);
3115
3116	for (iLoop = 0; iLoop < GIP_TSC_DELTA_LOOPS; iLoop++)
3117	{
3118	RTCCUINTREG fEFlags;
3119	if (fIsMaster)
3120	{
3121	/*
3122	* The master.
3123	*/
3124	AssertMsg(pGipCpuMaster->u64TSCSample == GIP_TSC_DELTA_RSVD,
3125	("%#llx idMaster=%#x idWorker=%#x (idGipMaster=%#x)\n",
3126	pGipCpuMaster->u64TSCSample, pGipCpuMaster->idCpu, pGipCpuWorker->idCpu, pArgs->pDevExt->idGipMaster));
3127	TSCDELTA_MASTER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3128
3129	do
3130	{
3131	ASMSerializeInstruction();
3132	ASMAtomicWriteU64(&pGipCpuMaster->u64TSCSample, ASMReadTSC());
3133	} while (pGipCpuMaster->u64TSCSample == GIP_TSC_DELTA_RSVD);
3134
3135	TSCDELTA_MASTER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3136
3137	/* Process the data. */
3138	if (iLoop > GIP_TSC_DELTA_PRIMER_LOOPS + GIP_TSC_DELTA_READ_TIME_LOOPS)
3139	{
3140	if (pGipCpuWorker->u64TSCSample != GIP_TSC_DELTA_RSVD)
3141	{
3142	int64_t iDelta = pGipCpuWorker->u64TSCSample
3143	- (pGipCpuMaster->u64TSCSample - pGipCpuMaster->i64TSCDelta);
3144	if ( iDelta >= GIP_TSC_DELTA_INITIAL_MASTER_VALUE
3145	? iDelta < pGipCpuWorker->i64TSCDelta
3146	: iDelta > pGipCpuWorker->i64TSCDelta \|\| pGipCpuWorker->i64TSCDelta == INT64_MAX)
3147	pGipCpuWorker->i64TSCDelta = iDelta;
3148	}
3149	}
3150
3151	/* Reset our TSC sample and tell the worker to move on. */
3152	ASMAtomicWriteU64(&pGipCpuMaster->u64TSCSample, GIP_TSC_DELTA_RSVD);
3153	TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(pMySync, pOtherSync);
3154	}
3155	else
3156	{
3157	/*
3158	* The worker.
3159	*/
3160	uint64_t uTscWorker;
3161	uint64_t uTscWorkerFlushed;
3162	uint64_t uCmpReadTime;
3163
3164	ASMAtomicReadU64(&pGipCpuMaster->u64TSCSample); /* Warm the cache line. */
3165	TSCDELTA_OTHER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3166
3167	/*
3168	* Keep reading the TSC until we notice that the master has read his. Reading
3169	* the TSC -after- the master has updated the memory is way too late. We thus
3170	* compensate by trying to measure how long it took for the worker to notice
3171	* the memory flushed from the master.
3172	*/
3173	do
3174	{
3175	ASMSerializeInstruction();
3176	uTscWorker = ASMReadTSC();
3177	} while (pGipCpuMaster->u64TSCSample == GIP_TSC_DELTA_RSVD);
3178	ASMSerializeInstruction();
3179	uTscWorkerFlushed = ASMReadTSC();
3180
3181	uCmpReadTime = uTscWorkerFlushed - uTscWorker;
3182	if (iLoop > GIP_TSC_DELTA_PRIMER_LOOPS + GIP_TSC_DELTA_READ_TIME_LOOPS)
3183	{
3184	/* This is totally arbitrary a.k.a I don't like it but I have no better ideas for now. */
3185	if (uCmpReadTime < (uMinCmpReadTime << 1))
3186	{
3187	ASMAtomicWriteU64(&pGipCpuWorker->u64TSCSample, uTscWorker);
3188	if (uCmpReadTime < uMinCmpReadTime)
3189	uMinCmpReadTime = uCmpReadTime;
3190	}
3191	else
3192	ASMAtomicWriteU64(&pGipCpuWorker->u64TSCSample, GIP_TSC_DELTA_RSVD);
3193	}
3194	else if (iLoop > GIP_TSC_DELTA_PRIMER_LOOPS)
3195	{
3196	if (uCmpReadTime < uMinCmpReadTime)
3197	uMinCmpReadTime = uCmpReadTime;
3198	}
3199
3200	TSCDELTA_OTHER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3201	}
3202	}
3203
3204	TSCDELTA_DBG_SYNC_MSG9(("sync/method1loop/%s: #92 iLoop=%u MyState=%#x\n", fIsMaster ? "master" : "worker", iLoop,
3205	pMySync->uSyncVar));
3206
3207	/*
3208	* We must reset the worker TSC sample value in case it gets picked as a
3209	* GIP master later on (it's trashed above, naturally).
3210	*/
3211	if (!fIsMaster)
3212	ASMAtomicWriteU64(&pGipCpuWorker->u64TSCSample, GIP_TSC_DELTA_RSVD);
3213	}
3214	#endif /* GIP_TSC_DELTA_METHOD_1 */
3215
3216
3217	#ifdef GIP_TSC_DELTA_METHOD_2
3218	/*
3219	* TSC delta measurement algorithm \#2 configuration and code - Experimental!!
3220	*/
3221
3222	# define GIP_TSC_DELTA_M2_LOOPS (7 + GIP_TSC_DELTA_M2_PRIMER_LOOPS)
3223	# define GIP_TSC_DELTA_M2_PRIMER_LOOPS 0
3224
3225
3226	static void supdrvTscDeltaMethod2ProcessDataOnMaster(PSUPDRVGIPTSCDELTARGS pArgs)
3227	{
3228	int64_t iMasterTscDelta = pArgs->pMaster->i64TSCDelta;
3229	int64_t iBestDelta = pArgs->pWorker->i64TSCDelta;
3230	uint32_t idxResult;
3231	uint32_t cHits = 0;
3232
3233	/*
3234	* Look for matching entries in the master and worker tables.
3235	*/
3236	for (idxResult = 0; idxResult < RT_ELEMENTS(pArgs->uMaster.M2.Data.aResults); idxResult++)
3237	{
3238	uint32_t idxOther = pArgs->uMaster.M2.Data.aResults[idxResult].iSeqOther;
3239	if (idxOther & 1)
3240	{
3241	idxOther >>= 1;
3242	if (idxOther < RT_ELEMENTS(pArgs->uWorker.M2.Data.aResults))
3243	{
3244	if (pArgs->uWorker.M2.Data.aResults[idxOther].iSeqOther == pArgs->uMaster.M2.Data.aResults[idxResult].iSeqMine)
3245	{
3246	int64_t iDelta;
3247	iDelta = pArgs->uWorker.M2.Data.aResults[idxOther].uTsc
3248	- (pArgs->uMaster.M2.Data.aResults[idxResult].uTsc - iMasterTscDelta);
3249	if ( iDelta >= GIP_TSC_DELTA_INITIAL_MASTER_VALUE
3250	? iDelta < iBestDelta
3251	: iDelta > iBestDelta \|\| iBestDelta == INT64_MAX)
3252	iBestDelta = iDelta;
3253	cHits++;
3254	}
3255	}
3256	}
3257	}
3258
3259	/*
3260	* Save the results.
3261	*/
3262	if (cHits > 2)
3263	pArgs->pWorker->i64TSCDelta = iBestDelta;
3264	pArgs->uMaster.M2.cHits += cHits;
3265	}
3266
3267
3268	/**
3269	* The core function of the 2nd TSC delta measurement algorithm.
3270	*
3271	* The idea here is that we have the two CPUs execute the exact same code
3272	* collecting a largish set of TSC samples. The code has one data dependency on
3273	* the other CPU which intention it is to synchronize the execution as well as
3274	* help cross references the two sets of TSC samples (the sequence numbers).
3275	*
3276	* The @a fLag parameter is used to modify the execution a tiny bit on one or
3277	* both of the CPUs. When @a fLag differs between the CPUs, it is thought that
3278	* it will help with making the CPUs enter lock step execution occasionally.
3279	*
3280	*/
3281	static void supdrvTscDeltaMethod2CollectData(PSUPDRVTSCDELTAMETHOD2 pMyData, uint32_t volatile *piOtherSeqNo, bool fLag)
3282	{
3283	SUPDRVTSCDELTAMETHOD2ENTRY *pEntry = &pMyData->aResults[0];
3284	uint32_t cLeft = RT_ELEMENTS(pMyData->aResults);
3285
3286	ASMAtomicWriteU32(&pMyData->iCurSeqNo, 0);
3287	ASMSerializeInstruction();
3288	while (cLeft-- > 0)
3289	{
3290	uint64_t uTsc;
3291	uint32_t iSeqMine = ASMAtomicIncU32(&pMyData->iCurSeqNo);
3292	uint32_t iSeqOther = ASMAtomicReadU32(piOtherSeqNo);
3293	ASMCompilerBarrier();
3294	ASMSerializeInstruction(); /* Way better result than with ASMMemoryFenceSSE2() in this position! */
3295	uTsc = ASMReadTSC();
3296	ASMAtomicIncU32(&pMyData->iCurSeqNo);
3297	ASMCompilerBarrier();
3298	ASMSerializeInstruction();
3299	pEntry->iSeqMine = iSeqMine;
3300	pEntry->iSeqOther = iSeqOther;
3301	pEntry->uTsc = uTsc;
3302	pEntry++;
3303	ASMSerializeInstruction();
3304	if (fLag)
3305	ASMNopPause();
3306	}
3307	}
3308
3309
3310	/**
3311	* TSC delta measurement algorithm \#2 (GIP_TSC_DELTA_METHOD_2).
3312	*
3313	* See supdrvTscDeltaMethod2CollectData for algorithm details.
3314	*
3315	* @param pArgs The argument/state data.
3316	* @param pMySync My synchronization structure.
3317	* @param pOtherSync My partner's synchronization structure.
3318	* @param fIsMaster Set if master, clear if worker.
3319	* @param iTry The attempt number.
3320	*/
3321	static void supdrvTscDeltaMethod2Loop(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
3322	bool fIsMaster, uint32_t iTry)
3323	{
3324	unsigned iLoop;
3325	RT_NOREF1(iTry);
3326
3327	for (iLoop = 0; iLoop < GIP_TSC_DELTA_M2_LOOPS; iLoop++)
3328	{
3329	RTCCUINTREG fEFlags;
3330	if (fIsMaster)
3331	{
3332	/*
3333	* Adjust the loop lag fudge.
3334	*/
3335	# if GIP_TSC_DELTA_M2_PRIMER_LOOPS > 0
3336	if (iLoop < GIP_TSC_DELTA_M2_PRIMER_LOOPS)
3337	{
3338	/* Lag during the priming to be nice to everyone.. */
3339	pArgs->uMaster.M2.fLag = true;
3340	pArgs->uWorker.M2.fLag = true;
3341	}
3342	else
3343	# endif
3344	if (iLoop < (GIP_TSC_DELTA_M2_LOOPS - GIP_TSC_DELTA_M2_PRIMER_LOOPS) / 4)
3345	{
3346	/* 25 % of the body without lagging. */
3347	pArgs->uMaster.M2.fLag = false;
3348	pArgs->uWorker.M2.fLag = false;
3349	}
3350	else if (iLoop < (GIP_TSC_DELTA_M2_LOOPS - GIP_TSC_DELTA_M2_PRIMER_LOOPS) / 4 * 2)
3351	{
3352	/* 25 % of the body with both lagging. */
3353	pArgs->uMaster.M2.fLag = true;
3354	pArgs->uWorker.M2.fLag = true;
3355	}
3356	else
3357	{
3358	/* 50% of the body with alternating lag. */
3359	pArgs->uMaster.M2.fLag = (iLoop & 1) == 0;
3360	pArgs->uWorker.M2.fLag= (iLoop & 1) == 1;
3361	}
3362
3363	/*
3364	* Sync up with the worker and collect data.
3365	*/
3366	TSCDELTA_MASTER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3367	supdrvTscDeltaMethod2CollectData(&pArgs->uMaster.M2.Data, &pArgs->uWorker.M2.Data.iCurSeqNo, pArgs->uMaster.M2.fLag);
3368	TSCDELTA_MASTER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3369
3370	/*
3371	* Process the data.
3372	*/
3373	# if GIP_TSC_DELTA_M2_PRIMER_LOOPS > 0
3374	if (iLoop >= GIP_TSC_DELTA_M2_PRIMER_LOOPS)
3375	# endif
3376	supdrvTscDeltaMethod2ProcessDataOnMaster(pArgs);
3377
3378	TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(pMySync, pOtherSync);
3379	}
3380	else
3381	{
3382	/*
3383	* The worker.
3384	*/
3385	TSCDELTA_OTHER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3386	supdrvTscDeltaMethod2CollectData(&pArgs->uWorker.M2.Data, &pArgs->uMaster.M2.Data.iCurSeqNo, pArgs->uWorker.M2.fLag);
3387	TSCDELTA_OTHER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3388	}
3389	}
3390	}
3391
3392	#endif /* GIP_TSC_DELTA_METHOD_2 */
3393
3394
3395
3396	static int supdrvTscDeltaVerify(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync,
3397	PSUPTSCDELTASYNC2 pOtherSync, bool fIsMaster, int64_t iWorkerTscDelta)
3398	{
3399	/PSUPGIPCPU pGipCpuWorker = pArgs->pWorker; - unused /
3400	PSUPGIPCPU pGipCpuMaster = pArgs->pMaster;
3401	uint32_t i;
3402	TSCDELTA_DBG_VARS();
3403
3404	for (;;)
3405	{
3406	RTCCUINTREG fEFlags;
3407	AssertCompile((RT_ELEMENTS(pArgs->uMaster.Verify.auTscs) & 1) == 0);
3408	AssertCompile(RT_ELEMENTS(pArgs->uMaster.Verify.auTscs) == RT_ELEMENTS(pArgs->uWorker.Verify.auTscs));
3409
3410	if (fIsMaster)
3411	{
3412	uint64_t uTscWorker;
3413	TSCDELTA_MASTER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3414
3415	/*
3416	* Collect TSC, master goes first.
3417	*/
3418	for (i = 0; i < RT_ELEMENTS(pArgs->uMaster.Verify.auTscs); i += 2)
3419	{
3420	/* Read, kick & wait #1. */
3421	uint64_t uTsc = ASMReadTSC();
3422	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO_GO);
3423	ASMSerializeInstruction();
3424	pArgs->uMaster.Verify.auTscs[i] = uTsc;
3425	TSCDELTA_DBG_START_LOOP();
3426	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO)
3427	{
3428	TSCDELTA_DBG_CHECK_LOOP();
3429	ASMNopPause();
3430	}
3431
3432	/* Read, kick & wait #2. */
3433	uTsc = ASMReadTSC();
3434	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO);
3435	ASMSerializeInstruction();
3436	pArgs->uMaster.Verify.auTscs[i + 1] = uTsc;
3437	TSCDELTA_DBG_START_LOOP();
3438	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO_GO)
3439	{
3440	TSCDELTA_DBG_CHECK_LOOP();
3441	ASMNopPause();
3442	}
3443	}
3444
3445	TSCDELTA_MASTER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3446
3447	/*
3448	* Process the data.
3449	*/
3450	#ifdef TSCDELTA_VERIFY_WITH_STATS
3451	pArgs->cMaxVerifyTscTicks = INT64_MIN;
3452	pArgs->cMinVerifyTscTicks = INT64_MAX;
3453	pArgs->iVerifyBadTscDiff = 0;
3454	#endif
3455	ASMAtomicWriteS32(&pArgs->rcVerify, VINF_SUCCESS);
3456	uTscWorker = 0;
3457	for (i = 0; i < RT_ELEMENTS(pArgs->uMaster.Verify.auTscs); i++)
3458	{
3459	/* Master vs previous worker entry. */
3460	uint64_t uTscMaster = pArgs->uMaster.Verify.auTscs[i] - pGipCpuMaster->i64TSCDelta;
3461	int64_t iDiff;
3462	if (i > 0)
3463	{
3464	iDiff = uTscMaster - uTscWorker;
3465	#ifdef TSCDELTA_VERIFY_WITH_STATS
3466	if (iDiff > pArgs->cMaxVerifyTscTicks)
3467	pArgs->cMaxVerifyTscTicks = iDiff;
3468	if (iDiff < pArgs->cMinVerifyTscTicks)
3469	pArgs->cMinVerifyTscTicks = iDiff;
3470	#endif
3471	if (iDiff < 0)
3472	{
3473	#ifdef TSCDELTA_VERIFY_WITH_STATS
3474	pArgs->iVerifyBadTscDiff = -iDiff;
3475	#endif
3476	ASMAtomicWriteS32(&pArgs->rcVerify, VERR_OUT_OF_RANGE);
3477	break;
3478	}
3479	}
3480
3481	/* Worker vs master. */
3482	uTscWorker = pArgs->uWorker.Verify.auTscs[i] - iWorkerTscDelta;
3483	iDiff = uTscWorker - uTscMaster;
3484	#ifdef TSCDELTA_VERIFY_WITH_STATS
3485	if (iDiff > pArgs->cMaxVerifyTscTicks)
3486	pArgs->cMaxVerifyTscTicks = iDiff;
3487	if (iDiff < pArgs->cMinVerifyTscTicks)
3488	pArgs->cMinVerifyTscTicks = iDiff;
3489	#endif
3490	if (iDiff < 0)
3491	{
3492	#ifdef TSCDELTA_VERIFY_WITH_STATS
3493	pArgs->iVerifyBadTscDiff = iDiff;
3494	#endif
3495	ASMAtomicWriteS32(&pArgs->rcVerify, VERR_OUT_OF_RANGE);
3496	break;
3497	}
3498	}
3499
3500	/* Done. */
3501	TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(pMySync, pOtherSync);
3502	}
3503	else
3504	{
3505	/*
3506	* The worker, master leads.
3507	*/
3508	TSCDELTA_OTHER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3509
3510	for (i = 0; i < RT_ELEMENTS(pArgs->uWorker.Verify.auTscs); i += 2)
3511	{
3512	uint64_t uTsc;
3513
3514	/* Wait, Read and Kick #1. */
3515	TSCDELTA_DBG_START_LOOP();
3516	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO)
3517	{
3518	TSCDELTA_DBG_CHECK_LOOP();
3519	ASMNopPause();
3520	}
3521	uTsc = ASMReadTSC();
3522	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO_GO);
3523	ASMSerializeInstruction();
3524	pArgs->uWorker.Verify.auTscs[i] = uTsc;
3525
3526	/* Wait, Read and Kick #2. */
3527	TSCDELTA_DBG_START_LOOP();
3528	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO_GO)
3529	{
3530	TSCDELTA_DBG_CHECK_LOOP();
3531	ASMNopPause();
3532	}
3533	uTsc = ASMReadTSC();
3534	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO);
3535	ASMSerializeInstruction();
3536	pArgs->uWorker.Verify.auTscs[i + 1] = uTsc;
3537	}
3538
3539	TSCDELTA_OTHER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3540	}
3541	return pArgs->rcVerify;
3542	}
3543
3544	/*
3545	* Timed out, please retry.
3546	*/
3547	ASMAtomicWriteS32(&pArgs->rcVerify, VERR_TRY_AGAIN);
3548	return VERR_TIMEOUT;
3549	}
3550
3551
3552
3553	/**
3554	* Handles the special abort procedure during synchronization setup in
3555	* supdrvTscMeasureDeltaCallbackUnwrapped().
3556	*
3557	* @returns 0 (dummy, ignored)
3558	* @param pArgs Pointer to argument/state data.
3559	* @param pMySync Pointer to my sync structure.
3560	* @param fIsMaster Set if we're the master, clear if worker.
3561	* @param fTimeout Set if it's a timeout.
3562	*/
3563	DECL_NO_INLINE(static, int)
3564	supdrvTscMeasureDeltaCallbackAbortSyncSetup(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync, bool fIsMaster, bool fTimeout)
3565	{
3566	PSUPTSCDELTASYNC2 volatile *ppMySync = fIsMaster ? &pArgs->pSyncMaster : &pArgs->pSyncWorker;
3567	PSUPTSCDELTASYNC2 volatile *ppOtherSync = fIsMaster ? &pArgs->pSyncWorker : &pArgs->pSyncMaster;
3568	TSCDELTA_DBG_VARS();
3569	RT_NOREF1(pMySync);
3570
3571	/*
3572	* Clear our sync pointer and make sure the abort flag is set.
3573	*/
3574	ASMAtomicWriteNullPtr(ppMySync);
3575	ASMAtomicWriteBool(&pArgs->fAbortSetup, true);
3576	if (fTimeout)
3577	ASMAtomicWriteBool(&pArgs->fTimedOut, true);
3578
3579	/*
3580	* Make sure the other party is out of there and won't be touching our
3581	* sync state again (would cause stack corruption).
3582	*/
3583	TSCDELTA_DBG_START_LOOP();
3584	while (ASMAtomicReadPtrT(ppOtherSync, PSUPTSCDELTASYNC2) != NULL)
3585	{
3586	ASMNopPause();
3587	ASMNopPause();
3588	ASMNopPause();
3589	TSCDELTA_DBG_CHECK_LOOP();
3590	}
3591
3592	return 0;
3593	}
3594
3595
3596	/**
3597	* This is used by supdrvTscMeasureInitialDeltas() to read the TSC on two CPUs
3598	* and compute the delta between them.
3599	*
3600	* To reduce code size a good when timeout handling was added, a dummy return
3601	* value had to be added (saves 1-3 lines per timeout case), thus this
3602	* 'Unwrapped' function and the dummy 0 return value.
3603	*
3604	* @returns 0 (dummy, ignored)
3605	* @param idCpu The CPU we are current scheduled on.
3606	* @param pArgs Pointer to a parameter package.
3607	*
3608	* @remarks Measuring TSC deltas between the CPUs is tricky because we need to
3609	* read the TSC at exactly the same time on both the master and the
3610	* worker CPUs. Due to DMA, bus arbitration, cache locality,
3611	* contention, SMI, pipelining etc. there is no guaranteed way of
3612	* doing this on x86 CPUs.
3613	*/
3614	static int supdrvTscMeasureDeltaCallbackUnwrapped(RTCPUID idCpu, PSUPDRVGIPTSCDELTARGS pArgs)
3615	{
3616	PSUPDRVDEVEXT pDevExt = pArgs->pDevExt;
3617	PSUPGIPCPU pGipCpuWorker = pArgs->pWorker;
3618	PSUPGIPCPU pGipCpuMaster = pArgs->pMaster;
3619	bool const fIsMaster = idCpu == pGipCpuMaster->idCpu;
3620	uint32_t iTry;
3621	PSUPTSCDELTASYNC2 volatile *ppMySync = fIsMaster ? &pArgs->pSyncMaster : &pArgs->pSyncWorker;
3622	PSUPTSCDELTASYNC2 volatile *ppOtherSync = fIsMaster ? &pArgs->pSyncWorker : &pArgs->pSyncMaster;
3623	SUPTSCDELTASYNC2 MySync;
3624	PSUPTSCDELTASYNC2 pOtherSync;
3625	int rc;
3626	TSCDELTA_DBG_VARS();
3627
3628	/* A bit of paranoia first. */
3629	if (!pGipCpuMaster \|\| !pGipCpuWorker)
3630	return 0;
3631
3632	/*
3633	* If the CPU isn't part of the measurement, return immediately.
3634	*/
3635	if ( !fIsMaster
3636	&& idCpu != pGipCpuWorker->idCpu)
3637	return 0;
3638
3639	/*
3640	* Set up my synchronization stuff and wait for the other party to show up.
3641	*
3642	* We don't wait forever since the other party may be off fishing (offline,
3643	* spinning with ints disables, whatever), we must play nice to the rest of
3644	* the system as this context generally isn't one in which we will get
3645	* preempted and we may hold up a number of lower priority interrupts.
3646	*/
3647	ASMAtomicWriteU32(&MySync.uSyncVar, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT);
3648	ASMAtomicWritePtr(ppMySync, &MySync);
3649	MySync.uTscStart = ASMReadTSC();
3650	MySync.cMaxTscTicks = pArgs->cMaxTscTicks;
3651
3652	/* Look for the partner, might not be here yet... Special abort considerations. */
3653	iTry = 0;
3654	TSCDELTA_DBG_START_LOOP();
3655	while ((pOtherSync = ASMAtomicReadPtrT(ppOtherSync, PSUPTSCDELTASYNC2)) == NULL)
3656	{
3657	ASMNopPause();
3658	if ( ASMAtomicReadBool(&pArgs->fAbortSetup)
3659	\|\| !RTMpIsCpuOnline(fIsMaster ? pGipCpuWorker->idCpu : pGipCpuMaster->idCpu) )
3660	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3661	if ( (iTry++ & 0xff) == 0
3662	&& ASMReadTSC() - MySync.uTscStart > pArgs->cMaxTscTicks)
3663	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, true /fTimeout/);
3664	TSCDELTA_DBG_CHECK_LOOP();
3665	ASMNopPause();
3666	}
3667
3668	/* I found my partner, waiting to be found... Special abort considerations. */
3669	if (fIsMaster)
3670	if (!ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT)) /* parnaoia */
3671	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3672
3673	iTry = 0;
3674	TSCDELTA_DBG_START_LOOP();
3675	while (ASMAtomicReadU32(&MySync.uSyncVar) == GIP_TSC_DELTA_SYNC2_PRESTART_WAIT)
3676	{
3677	ASMNopPause();
3678	if (ASMAtomicReadBool(&pArgs->fAbortSetup))
3679	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3680	if ( (iTry++ & 0xff) == 0
3681	&& ASMReadTSC() - MySync.uTscStart > pArgs->cMaxTscTicks)
3682	{
3683	if ( fIsMaster
3684	&& !ASMAtomicCmpXchgU32(&MySync.uSyncVar, GIP_TSC_DELTA_SYNC2_PRESTART_ABORT, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT))
3685	break; /* race #1: slave has moved on, handle timeout in loop instead. */
3686	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, true /fTimeout/);
3687	}
3688	TSCDELTA_DBG_CHECK_LOOP();
3689	}
3690
3691	if (!fIsMaster)
3692	if (!ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT)) /* race #1 */
3693	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3694
3695	/** @todo Add a resumable state to pArgs so we don't waste time if we time
3696	* out or something. Timeouts are legit, any of the two CPUs may get
3697	* interrupted. */
3698
3699	/*
3700	* Start by seeing if we have a zero delta between the two CPUs.
3701	* This should normally be the case.
3702	*/
3703	rc = supdrvTscDeltaVerify(pArgs, &MySync, pOtherSync, fIsMaster, GIP_TSC_DELTA_INITIAL_MASTER_VALUE);
3704	if (RT_SUCCESS(rc))
3705	{
3706	if (fIsMaster)
3707	{
3708	ASMAtomicWriteS64(&pGipCpuWorker->i64TSCDelta, GIP_TSC_DELTA_INITIAL_MASTER_VALUE);
3709	RTCpuSetDelByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet);
3710	RTCpuSetAddByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpuWorker->iCpuSet);
3711	}
3712	}
3713	/*
3714	* If the verification didn't time out, do regular delta measurements.
3715	* We retry this until we get a reasonable value.
3716	*/
3717	else if (rc != VERR_TIMEOUT)
3718	{
3719	Assert(pGipCpuWorker->i64TSCDelta == INT64_MAX);
3720	for (iTry = 0; iTry < 12; iTry++)
3721	{
3722	/*
3723	* Check the state before we start.
3724	*/
3725	uint32_t u32Tmp = ASMAtomicReadU32(&MySync.uSyncVar);
3726	if ( u32Tmp != GIP_TSC_DELTA_SYNC2_READY
3727	&& (fIsMaster \|\| u32Tmp != GIP_TSC_DELTA_SYNC2_STEADY) /* worker may be late prepping for the next round */ )
3728	{
3729	TSCDELTA_DBG_SYNC_MSG(("sync/loop/%s: #0 iTry=%u MyState=%#x\n", fIsMaster ? "master" : "worker", iTry, u32Tmp));
3730	break;
3731	}
3732
3733	/*
3734	* Do the measurements.
3735	*/
3736	#ifdef GIP_TSC_DELTA_METHOD_1
3737	supdrvTscDeltaMethod1Loop(pArgs, &MySync, pOtherSync, fIsMaster, iTry);
3738	#elif defined(GIP_TSC_DELTA_METHOD_2)
3739	supdrvTscDeltaMethod2Loop(pArgs, &MySync, pOtherSync, fIsMaster, iTry);
3740	#else
3741	# error "huh??"
3742	#endif
3743
3744	/*
3745	* Check the state.
3746	*/
3747	u32Tmp = ASMAtomicReadU32(&MySync.uSyncVar);
3748	if ( u32Tmp != GIP_TSC_DELTA_SYNC2_READY
3749	&& (fIsMaster \|\| u32Tmp != GIP_TSC_DELTA_SYNC2_STEADY) /* worker may be late prepping for the next round */ )
3750	{
3751	if (fIsMaster)
3752	TSCDELTA_DBG_SYNC_MSG(("sync/loop/master: #1 iTry=%u MyState=%#x\n", iTry, u32Tmp));
3753	else
3754	TSCDELTA_DBG_SYNC_MSG2(("sync/loop/worker: #1 iTry=%u MyState=%#x\n", iTry, u32Tmp));
3755	break;
3756	}
3757
3758	/*
3759	* Success? If so, stop trying. Master decides.
3760	*/
3761	if (fIsMaster)
3762	{
3763	if (pGipCpuWorker->i64TSCDelta != INT64_MAX)
3764	{
3765	RTCpuSetDelByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet);
3766	RTCpuSetAddByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpuWorker->iCpuSet);
3767	TSCDELTA_DBG_SYNC_MSG2(("sync/loop/master: #9 iTry=%u MyState=%#x\n", iTry, MySync.uSyncVar));
3768	break;
3769	}
3770	}
3771	}
3772	if (fIsMaster)
3773	pArgs->iTry = iTry;
3774	}
3775
3776	/*
3777	* End the synchronization dance. We tell the other that we're done,
3778	* then wait for the same kind of reply.
3779	*/
3780	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_FINAL);
3781	ASMAtomicWriteNullPtr(ppMySync);
3782	iTry = 0;
3783	TSCDELTA_DBG_START_LOOP();
3784	while (ASMAtomicReadU32(&MySync.uSyncVar) != GIP_TSC_DELTA_SYNC2_FINAL)
3785	{
3786	iTry++;
3787	if ( iTry == 0
3788	&& !RTMpIsCpuOnline(fIsMaster ? pGipCpuWorker->idCpu : pGipCpuMaster->idCpu))
3789	break; /* this really shouldn't happen. */
3790	TSCDELTA_DBG_CHECK_LOOP();
3791	ASMNopPause();
3792	}
3793
3794	/*
3795	* Collect some runtime stats.
3796	*/
3797	if (fIsMaster)
3798	pArgs->cElapsedMasterTscTicks = ASMReadTSC() - MySync.uTscStart;
3799	else
3800	pArgs->cElapsedWorkerTscTicks = ASMReadTSC() - MySync.uTscStart;
3801	return 0;
3802	}
3803
3804	/**
3805	* Callback used by supdrvTscMeasureInitialDeltas() to read the TSC on two CPUs
3806	* and compute the delta between them.
3807	*
3808	* @param idCpu The CPU we are current scheduled on.
3809	* @param pvUser1 Pointer to a parameter package (SUPDRVGIPTSCDELTARGS).
3810	* @param pvUser2 Unused.
3811	*/
3812	static DECLCALLBACK(void) supdrvTscMeasureDeltaCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3813	{
3814	supdrvTscMeasureDeltaCallbackUnwrapped(idCpu, (PSUPDRVGIPTSCDELTARGS)pvUser1);
3815	RT_NOREF1(pvUser2);
3816	}
3817
3818
3819	/**
3820	* Measures the TSC delta between the master GIP CPU and one specified worker
3821	* CPU.
3822	*
3823	* @returns VBox status code.
3824	* @retval VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED on pure measurement
3825	* failure.
3826	* @param pDevExt Pointer to the device instance data.
3827	* @param idxWorker The index of the worker CPU from the GIP's array of
3828	* CPUs.
3829	*
3830	* @remarks This must be called with preemption enabled!
3831	*/
3832	static int supdrvTscMeasureDeltaOne(PSUPDRVDEVEXT pDevExt, uint32_t idxWorker)
3833	{
3834	int rc;
3835	int rc2;
3836	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
3837	RTCPUID idMaster = pDevExt->idGipMaster;
3838	PSUPGIPCPU pGipCpuWorker = &pGip->aCPUs[idxWorker];
3839	PSUPGIPCPU pGipCpuMaster;
3840	uint32_t iGipCpuMaster;
3841	uint32_t u32Tmp;
3842
3843	/* Validate input a bit. */
3844	AssertReturn(pGip, VERR_INVALID_PARAMETER);
3845	Assert(pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED);
3846	Assert(RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3847
3848	/*
3849	* Don't attempt measuring the delta for the GIP master.
3850	*/
3851	if (pGipCpuWorker->idCpu == idMaster)
3852	{
3853	if (pGipCpuWorker->i64TSCDelta == INT64_MAX) /* This shouldn't happen, but just in case. */
3854	ASMAtomicWriteS64(&pGipCpuWorker->i64TSCDelta, GIP_TSC_DELTA_INITIAL_MASTER_VALUE);
3855	return VINF_SUCCESS;
3856	}
3857
3858	/*
3859	* One measurement at a time, at least for now. We might be using
3860	* broadcast IPIs so, so be nice to the rest of the system.
3861	*/
3862	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3863	rc = RTSemMutexRequest(pDevExt->mtxTscDelta, RT_INDEFINITE_WAIT);
3864	#else
3865	rc = RTSemFastMutexRequest(pDevExt->mtxTscDelta);
3866	#endif
3867	if (RT_FAILURE(rc))
3868	return rc;
3869
3870	/*
3871	* If the CPU has hyper-threading and the APIC IDs of the master and worker are adjacent,
3872	* try pick a different master. (This fudge only works with multi core systems.)
3873	* ASSUMES related threads have adjacent APIC IDs. ASSUMES two threads per core.
3874	*
3875	* We skip this on AMDs for now as their HTT is different from Intel's and
3876	* it doesn't seem to have any favorable effect on the results.
3877	*
3878	* If the master is offline, we need a new master too, so share the code.
3879	*/
3880	iGipCpuMaster = supdrvGipFindCpuIndexForCpuId(pGip, idMaster);
3881	AssertReturn(iGipCpuMaster < pGip->cCpus, VERR_INVALID_CPU_ID);
3882	pGipCpuMaster = &pGip->aCPUs[iGipCpuMaster];
3883	if ( ( (pGipCpuMaster->idApic & ~1) == (pGipCpuWorker->idApic & ~1)
3884	&& pGip->cOnlineCpus > 2
3885	&& ASMHasCpuId()
3886	&& ASMIsValidStdRange(ASMCpuId_EAX(0))
3887	&& (ASMCpuId_EDX(1) & X86_CPUID_FEATURE_EDX_HTT)
3888	&& ( !ASMIsAmdCpu()
3889	\|\| ASMGetCpuFamily(u32Tmp = ASMCpuId_EAX(1)) > 0x15
3890	\|\| ( ASMGetCpuFamily(u32Tmp) == 0x15 /* Piledriver+, not bulldozer (FX-4150 didn't like it). */
3891	&& ASMGetCpuModelAMD(u32Tmp) >= 0x02) ) )
3892	\|\| !RTMpIsCpuOnline(idMaster) )
3893	{
3894	uint32_t i;
3895	for (i = 0; i < pGip->cCpus; i++)
3896	if ( i != iGipCpuMaster
3897	&& i != idxWorker
3898	&& pGip->aCPUs[i].enmState == SUPGIPCPUSTATE_ONLINE
3899	&& pGip->aCPUs[i].i64TSCDelta != INT64_MAX
3900	&& pGip->aCPUs[i].idCpu != NIL_RTCPUID
3901	&& pGip->aCPUs[i].idCpu != idMaster /* paranoia starts here... */
3902	&& pGip->aCPUs[i].idCpu != pGipCpuWorker->idCpu
3903	&& pGip->aCPUs[i].idApic != pGipCpuWorker->idApic
3904	&& pGip->aCPUs[i].idApic != pGipCpuMaster->idApic
3905	&& RTMpIsCpuOnline(pGip->aCPUs[i].idCpu))
3906	{
3907	iGipCpuMaster = i;
3908	pGipCpuMaster = &pGip->aCPUs[i];
3909	idMaster = pGipCpuMaster->idCpu;
3910	break;
3911	}
3912	}
3913
3914	if (RTCpuSetIsMemberByIndex(&pGip->OnlineCpuSet, pGipCpuWorker->iCpuSet))
3915	{
3916	/*
3917	* Initialize data package for the RTMpOnPair callback.
3918	*/
3919	PSUPDRVGIPTSCDELTARGS pArgs = (PSUPDRVGIPTSCDELTARGS)RTMemAllocZ(sizeof(*pArgs));
3920	if (pArgs)
3921	{
3922	pArgs->pWorker = pGipCpuWorker;
3923	pArgs->pMaster = pGipCpuMaster;
3924	pArgs->pDevExt = pDevExt;
3925	pArgs->pSyncMaster = NULL;
3926	pArgs->pSyncWorker = NULL;
3927	pArgs->cMaxTscTicks = ASMAtomicReadU64(&pGip->u64CpuHz) / 512; /* 1953 us */
3928
3929	/*
3930	* Do the RTMpOnPair call. We reset i64TSCDelta first so we
3931	* and supdrvTscMeasureDeltaCallback can use it as a success check.
3932	*/
3933	/** @todo Store the i64TSCDelta result in pArgs first? Perhaps deals with
3934	* that when doing the restart loop reorg. */
3935	ASMAtomicWriteS64(&pGipCpuWorker->i64TSCDelta, INT64_MAX);
3936	rc = RTMpOnPair(pGipCpuMaster->idCpu, pGipCpuWorker->idCpu, RTMPON_F_CONCURRENT_EXEC,
3937	supdrvTscMeasureDeltaCallback, pArgs, NULL);
3938	if (RT_SUCCESS(rc))
3939	{
3940	#if 0
3941	SUPR0Printf("mponpair ticks: %9llu %9llu max: %9llu iTry: %u%s\n", pArgs->cElapsedMasterTscTicks,
3942	pArgs->cElapsedWorkerTscTicks, pArgs->cMaxTscTicks, pArgs->iTry,
3943	pArgs->fTimedOut ? " timed out" :"");
3944	#endif
3945	#if 0
3946	SUPR0Printf("rcVerify=%d iVerifyBadTscDiff=%lld cMinVerifyTscTicks=%lld cMaxVerifyTscTicks=%lld\n",
3947	pArgs->rcVerify, pArgs->iVerifyBadTscDiff, pArgs->cMinVerifyTscTicks, pArgs->cMaxVerifyTscTicks);
3948	#endif
3949	if (RT_LIKELY(pGipCpuWorker->i64TSCDelta != INT64_MAX))
3950	{
3951	/*
3952	* Work the TSC delta applicability rating. It starts
3953	* optimistic in supdrvGipInit, we downgrade it here.
3954	*/
3955	SUPGIPUSETSCDELTA enmRating;
3956	if ( pGipCpuWorker->i64TSCDelta > GIP_TSC_DELTA_THRESHOLD_ROUGHLY_ZERO
3957	\|\| pGipCpuWorker->i64TSCDelta < -GIP_TSC_DELTA_THRESHOLD_ROUGHLY_ZERO)
3958	enmRating = SUPGIPUSETSCDELTA_NOT_ZERO;
3959	else if ( pGipCpuWorker->i64TSCDelta > GIP_TSC_DELTA_THRESHOLD_PRACTICALLY_ZERO
3960	\|\| pGipCpuWorker->i64TSCDelta < -GIP_TSC_DELTA_THRESHOLD_PRACTICALLY_ZERO)
3961	enmRating = SUPGIPUSETSCDELTA_ROUGHLY_ZERO;
3962	else
3963	enmRating = SUPGIPUSETSCDELTA_PRACTICALLY_ZERO;
3964	if (pGip->enmUseTscDelta < enmRating)
3965	{
3966	AssertCompile(sizeof(pGip->enmUseTscDelta) == sizeof(uint32_t));
3967	ASMAtomicWriteU32((uint32_t volatile *)&pGip->enmUseTscDelta, enmRating);
3968	}
3969	}
3970	else
3971	rc = VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED;
3972	}
3973	/** @todo return try-again if we get an offline CPU error. */
3974
3975	RTMemFree(pArgs);
3976	}
3977	else
3978	rc = VERR_NO_MEMORY;
3979	}
3980	else
3981	rc = VERR_CPU_OFFLINE;
3982
3983	/*
3984	* We're done now.
3985	*/
3986	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3987	rc2 = RTSemMutexRelease(pDevExt->mtxTscDelta); AssertRC(rc2);
3988	#else
3989	rc2 = RTSemFastMutexRelease(pDevExt->mtxTscDelta); AssertRC(rc2);
3990	#endif
3991	return rc;
3992	}
3993
3994
3995	/**
3996	* Resets the TSC-delta related TSC samples and optionally the deltas
3997	* themselves.
3998	*
3999	* @param pDevExt Pointer to the device instance data.
4000	* @param fResetTscDeltas Whether the TSC-deltas are also to be reset.
4001	*
4002	* @remarks This might be called while holding a spinlock!
4003	*/
4004	static void supdrvTscResetSamples(PSUPDRVDEVEXT pDevExt, bool fResetTscDeltas)
4005	{
4006	unsigned iCpu;
4007	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4008	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
4009	{
4010	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
4011	ASMAtomicWriteU64(&pGipCpu->u64TSCSample, GIP_TSC_DELTA_RSVD);
4012	if (fResetTscDeltas)
4013	{
4014	RTCpuSetDelByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpu->iCpuSet);
4015	ASMAtomicWriteS64(&pGipCpu->i64TSCDelta, INT64_MAX);
4016	}
4017	}
4018	}
4019
4020
4021	/**
4022	* Picks an online CPU as the master TSC for TSC-delta computations.
4023	*
4024	* @returns VBox status code.
4025	* @param pDevExt Pointer to the device instance data.
4026	* @param pidxMaster Where to store the CPU array index of the chosen
4027	* master. Optional, can be NULL.
4028	*/
4029	static int supdrvTscPickMaster(PSUPDRVDEVEXT pDevExt, uint32_t *pidxMaster)
4030	{
4031	/*
4032	* Pick the first CPU online as the master TSC and make it the new GIP master based
4033	* on the APIC ID.
4034	*
4035	* Technically we can simply use "idGipMaster" but doing this gives us master as CPU 0
4036	* in most cases making it nicer/easier for comparisons. It is safe to update the GIP
4037	* master as this point since the sync/async timer isn't created yet.
4038	*/
4039	unsigned iCpu;
4040	uint32_t idxMaster = UINT32_MAX;
4041	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4042	for (iCpu = 0; iCpu < RT_ELEMENTS(pGip->aiCpuFromApicId); iCpu++)
4043	{
4044	uint16_t idxCpu = pGip->aiCpuFromApicId[iCpu];
4045	if (idxCpu != UINT16_MAX)
4046	{
4047	PSUPGIPCPU pGipCpu = &pGip->aCPUs[idxCpu];
4048	if (RTCpuSetIsMemberByIndex(&pGip->OnlineCpuSet, pGipCpu->iCpuSet))
4049	{
4050	idxMaster = idxCpu;
4051	pGipCpu->i64TSCDelta = GIP_TSC_DELTA_INITIAL_MASTER_VALUE;
4052	ASMAtomicWriteSize(&pDevExt->idGipMaster, pGipCpu->idCpu);
4053	if (pidxMaster)
4054	*pidxMaster = idxMaster;
4055	return VINF_SUCCESS;
4056	}
4057	}
4058	}
4059	return VERR_CPU_OFFLINE;
4060	}
4061
4062
4063	/**
4064	* Performs the initial measurements of the TSC deltas between CPUs.
4065	*
4066	* This is called by supdrvGipCreate(), supdrvGipPowerNotificationCallback() or
4067	* triggered by it if threaded.
4068	*
4069	* @returns VBox status code.
4070	* @param pDevExt Pointer to the device instance data.
4071	*
4072	* @remarks Must be called only after supdrvGipInitOnCpu() as this function uses
4073	* idCpu, GIP's online CPU set which are populated in
4074	* supdrvGipInitOnCpu().
4075	*/
4076	static int supdrvTscMeasureInitialDeltas(PSUPDRVDEVEXT pDevExt)
4077	{
4078	PSUPGIPCPU pGipCpuMaster;
4079	unsigned iCpu;
4080	unsigned iOddEven;
4081	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4082	uint32_t idxMaster = UINT32_MAX;
4083	uint32_t cMpOnOffEvents = ASMAtomicReadU32(&pDevExt->cMpOnOffEvents);
4084
4085	Assert(pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED);
4086	supdrvTscResetSamples(pDevExt, true /* fClearDeltas */);
4087	int rc = supdrvTscPickMaster(pDevExt, &idxMaster);
4088	if (RT_FAILURE(rc))
4089	{
4090	SUPR0Printf("Failed to pick a CPU master for TSC-delta measurements rc=%Rrc\n", rc);
4091	return rc;
4092	}
4093	AssertReturn(idxMaster < pGip->cCpus, VERR_INVALID_CPU_INDEX);
4094	pGipCpuMaster = &pGip->aCPUs[idxMaster];
4095	Assert(pDevExt->idGipMaster == pGipCpuMaster->idCpu);
4096
4097	/*
4098	* If there is only a single CPU online we have nothing to do.
4099	*/
4100	if (pGip->cOnlineCpus <= 1)
4101	{
4102	AssertReturn(pGip->cOnlineCpus > 0, VERR_INTERNAL_ERROR_5);
4103	return VINF_SUCCESS;
4104	}
4105
4106	/*
4107	* Loop thru the GIP CPU array and get deltas for each CPU (except the
4108	* master). We do the CPUs with the even numbered APIC IDs first so that
4109	* we've got alternative master CPUs to pick from on hyper-threaded systems.
4110	*/
4111	for (iOddEven = 0; iOddEven < 2; iOddEven++)
4112	{
4113	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
4114	{
4115	PSUPGIPCPU pGipCpuWorker = &pGip->aCPUs[iCpu];
4116	if ( iCpu != idxMaster
4117	&& (iOddEven > 0 \|\| (pGipCpuWorker->idApic & 1) == 0)
4118	&& RTCpuSetIsMemberByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet))
4119	{
4120	rc = supdrvTscMeasureDeltaOne(pDevExt, iCpu);
4121	if (RT_FAILURE(rc))
4122	{
4123	SUPR0Printf("supdrvTscMeasureDeltaOne failed. rc=%d CPU[%u].idCpu=%u Master[%u].idCpu=%u\n", rc, iCpu,
4124	pGipCpuWorker->idCpu, idxMaster, pDevExt->idGipMaster, pGipCpuMaster->idCpu);
4125	break;
4126	}
4127
4128	if (ASMAtomicReadU32(&pDevExt->cMpOnOffEvents) != cMpOnOffEvents)
4129	{
4130	SUPR0Printf("One or more CPUs transitioned between online & offline states. I'm confused, retry...\n");
4131	rc = VERR_TRY_AGAIN;
4132	break;
4133	}
4134	}
4135	}
4136	}
4137
4138	return rc;
4139	}
4140
4141
4142	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
4143
4144	/**
4145	* Switches the TSC-delta measurement thread into the butchered state.
4146	*
4147	* @returns VBox status code.
4148	* @param pDevExt Pointer to the device instance data.
4149	* @param fSpinlockHeld Whether the TSC-delta spinlock is held or not.
4150	* @param pszFailed An error message to log.
4151	* @param rcFailed The error code to exit the thread with.
4152	*/
4153	static int supdrvTscDeltaThreadButchered(PSUPDRVDEVEXT pDevExt, bool fSpinlockHeld, const char *pszFailed, int rcFailed)
4154	{
4155	if (!fSpinlockHeld)
4156	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4157
4158	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Butchered;
4159	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4160	OSDBGPRINT(("supdrvTscDeltaThreadButchered: %s. rc=%Rrc\n", pszFailed, rcFailed));
4161	return rcFailed;
4162	}
4163
4164
4165	/**
4166	* The TSC-delta measurement thread.
4167	*
4168	* @returns VBox status code.
4169	* @param hThread The thread handle.
4170	* @param pvUser Opaque pointer to the device instance data.
4171	*/
4172	static DECLCALLBACK(int) supdrvTscDeltaThread(RTTHREAD hThread, void *pvUser)
4173	{
4174	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4175	int rc = VERR_INTERNAL_ERROR_2;
4176	for (;;)
4177	{
4178	/*
4179	* Switch on the current state.
4180	*/
4181	SUPDRVTSCDELTATHREADSTATE enmState;
4182	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4183	enmState = pDevExt->enmTscDeltaThreadState;
4184	switch (enmState)
4185	{
4186	case kTscDeltaThreadState_Creating:
4187	{
4188	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Listening;
4189	rc = RTSemEventSignal(pDevExt->hTscDeltaEvent);
4190	if (RT_FAILURE(rc))
4191	return supdrvTscDeltaThreadButchered(pDevExt, true /* fSpinlockHeld */, "RTSemEventSignal", rc);
4192	RT_FALL_THRU();
4193	}
4194
4195	case kTscDeltaThreadState_Listening:
4196	{
4197	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4198
4199	/*
4200	* Linux counts uninterruptible sleeps as load, hence we shall do a
4201	* regular, interruptible sleep here and ignore wake ups due to signals.
4202	* See task_contributes_to_load() in include/linux/sched.h in the Linux sources.
4203	*/
4204	rc = RTThreadUserWaitNoResume(hThread, pDevExt->cMsTscDeltaTimeout);
4205	if ( RT_FAILURE(rc)
4206	&& rc != VERR_TIMEOUT
4207	&& rc != VERR_INTERRUPTED)
4208	return supdrvTscDeltaThreadButchered(pDevExt, false /* fSpinlockHeld */, "RTThreadUserWait", rc);
4209	RTThreadUserReset(hThread);
4210	break;
4211	}
4212
4213	case kTscDeltaThreadState_WaitAndMeasure:
4214	{
4215	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Measuring;
4216	rc = RTSemEventSignal(pDevExt->hTscDeltaEvent); /* (Safe on windows as long as spinlock isn't IRQ safe.) */
4217	if (RT_FAILURE(rc))
4218	return supdrvTscDeltaThreadButchered(pDevExt, true /* fSpinlockHeld */, "RTSemEventSignal", rc);
4219	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4220	RTThreadSleep(1);
4221	RT_FALL_THRU();
4222	}
4223
4224	case kTscDeltaThreadState_Measuring:
4225	{
4226	if (pDevExt->fTscThreadRecomputeAllDeltas)
4227	{
4228	int cTries = 8;
4229	int cMsWaitPerTry = 10;
4230	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4231	Assert(pGip);
4232	do
4233	{
4234	RTCpuSetCopy(&pDevExt->TscDeltaCpuSet, &pGip->OnlineCpuSet);
4235	rc = supdrvTscMeasureInitialDeltas(pDevExt);
4236	if ( RT_SUCCESS(rc)
4237	\|\| ( RT_FAILURE(rc)
4238	&& rc != VERR_TRY_AGAIN
4239	&& rc != VERR_CPU_OFFLINE))
4240	{
4241	break;
4242	}
4243	RTThreadSleep(cMsWaitPerTry);
4244	} while (cTries-- > 0);
4245	pDevExt->fTscThreadRecomputeAllDeltas = false;
4246	}
4247	else
4248	{
4249	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4250	unsigned iCpu;
4251
4252	/* Measure TSC-deltas only for the CPUs that are in the set. */
4253	rc = VINF_SUCCESS;
4254	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
4255	{
4256	PSUPGIPCPU pGipCpuWorker = &pGip->aCPUs[iCpu];
4257	if (RTCpuSetIsMemberByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet))
4258	{
4259	if (pGipCpuWorker->i64TSCDelta == INT64_MAX)
4260	{
4261	int rc2 = supdrvTscMeasureDeltaOne(pDevExt, iCpu);
4262	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
4263	rc = rc2;
4264	}
4265	else
4266	{
4267	/*
4268	* The thread/someone must've called SUPR0TscDeltaMeasureBySetIndex(),
4269	* mark the delta as fine to get the timer thread off our back.
4270	*/
4271	RTCpuSetDelByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet);
4272	RTCpuSetAddByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpuWorker->iCpuSet);
4273	}
4274	}
4275	}
4276	}
4277	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4278	if (pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Measuring)
4279	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Listening;
4280	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4281	Assert(rc != VERR_NOT_AVAILABLE); /* VERR_NOT_AVAILABLE is used as init value, see supdrvTscDeltaThreadInit(). */
4282	ASMAtomicWriteS32(&pDevExt->rcTscDelta, rc);
4283	break;
4284	}
4285
4286	case kTscDeltaThreadState_Terminating:
4287	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Destroyed;
4288	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4289	return VINF_SUCCESS;
4290
4291	case kTscDeltaThreadState_Butchered:
4292	default:
4293	return supdrvTscDeltaThreadButchered(pDevExt, true /* fSpinlockHeld */, "Invalid state", VERR_INVALID_STATE);
4294	}
4295	}
4296	/* not reached */
4297	}
4298
4299
4300	/**
4301	* Waits for the TSC-delta measurement thread to respond to a state change.
4302	*
4303	* @returns VINF_SUCCESS on success, VERR_TIMEOUT if it doesn't respond in time,
4304	* other error code on internal error.
4305	*
4306	* @param pDevExt The device instance data.
4307	* @param enmCurState The current state.
4308	* @param enmNewState The new state we're waiting for it to enter.
4309	*/
4310	static int supdrvTscDeltaThreadWait(PSUPDRVDEVEXT pDevExt, SUPDRVTSCDELTATHREADSTATE enmCurState,
4311	SUPDRVTSCDELTATHREADSTATE enmNewState)
4312	{
4313	SUPDRVTSCDELTATHREADSTATE enmActualState;
4314	int rc;
4315
4316	/*
4317	* Wait a short while for the expected state transition.
4318	*/
4319	RTSemEventWait(pDevExt->hTscDeltaEvent, RT_MS_1SEC);
4320	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4321	enmActualState = pDevExt->enmTscDeltaThreadState;
4322	if (enmActualState == enmNewState)
4323	{
4324	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4325	rc = VINF_SUCCESS;
4326	}
4327	else if (enmActualState == enmCurState)
4328	{
4329	/*
4330	* Wait longer if the state has not yet transitioned to the one we want.
4331	*/
4332	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4333	rc = RTSemEventWait(pDevExt->hTscDeltaEvent, 50 * RT_MS_1SEC);
4334	if ( RT_SUCCESS(rc)
4335	\|\| rc == VERR_TIMEOUT)
4336	{
4337	/*
4338	* Check the state whether we've succeeded.
4339	*/
4340	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4341	enmActualState = pDevExt->enmTscDeltaThreadState;
4342	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4343	if (enmActualState == enmNewState)
4344	rc = VINF_SUCCESS;
4345	else if (enmActualState == enmCurState)
4346	{
4347	rc = VERR_TIMEOUT;
4348	OSDBGPRINT(("supdrvTscDeltaThreadWait: timed out state transition. enmActualState=%d enmNewState=%d\n",
4349	enmActualState, enmNewState));
4350	}
4351	else
4352	{
4353	rc = VERR_INTERNAL_ERROR;
4354	OSDBGPRINT(("supdrvTscDeltaThreadWait: invalid state transition from %d to %d, expected %d\n", enmCurState,
4355	enmActualState, enmNewState));
4356	}
4357	}
4358	else
4359	OSDBGPRINT(("supdrvTscDeltaThreadWait: RTSemEventWait failed. rc=%Rrc\n", rc));
4360	}
4361	else
4362	{
4363	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4364	OSDBGPRINT(("supdrvTscDeltaThreadWait: invalid state %d when transitioning from %d to %d\n",
4365	enmActualState, enmCurState, enmNewState));
4366	rc = VERR_INTERNAL_ERROR;
4367	}
4368
4369	return rc;
4370	}
4371
4372
4373	/**
4374	* Signals the TSC-delta thread to start measuring TSC-deltas.
4375	*
4376	* @param pDevExt Pointer to the device instance data.
4377	* @param fForceAll Force re-calculating TSC-deltas on all CPUs.
4378	*/
4379	static void supdrvTscDeltaThreadStartMeasurement(PSUPDRVDEVEXT pDevExt, bool fForceAll)
4380	{
4381	if (pDevExt->hTscDeltaThread != NIL_RTTHREAD)
4382	{
4383	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4384	if ( pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Listening
4385	\|\| pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Measuring)
4386	{
4387	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_WaitAndMeasure;
4388	if (fForceAll)
4389	pDevExt->fTscThreadRecomputeAllDeltas = true;
4390	}
4391	else if ( pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_WaitAndMeasure
4392	&& fForceAll)
4393	pDevExt->fTscThreadRecomputeAllDeltas = true;
4394	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4395	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4396	}
4397	}
4398
4399
4400	/**
4401	* Terminates the actual thread running supdrvTscDeltaThread().
4402	*
4403	* This is an internal worker function for supdrvTscDeltaThreadInit() and
4404	* supdrvTscDeltaTerm().
4405	*
4406	* @param pDevExt Pointer to the device instance data.
4407	*/
4408	static void supdrvTscDeltaThreadTerminate(PSUPDRVDEVEXT pDevExt)
4409	{
4410	int rc;
4411	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4412	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Terminating;
4413	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4414	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4415	rc = RTThreadWait(pDevExt->hTscDeltaThread, 50 * RT_MS_1SEC, NULL /* prc */);
4416	if (RT_FAILURE(rc))
4417	{
4418	/* Signal a few more times before giving up. */
4419	int cTriesLeft = 5;
4420	while (--cTriesLeft > 0)
4421	{
4422	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4423	rc = RTThreadWait(pDevExt->hTscDeltaThread, 2 * RT_MS_1SEC, NULL /* prc */);
4424	if (rc != VERR_TIMEOUT)
4425	break;
4426	}
4427	}
4428	}
4429
4430
4431	/**
4432	* Initializes and spawns the TSC-delta measurement thread.
4433	*
4434	* A thread is required for servicing re-measurement requests from events like
4435	* CPUs coming online, suspend/resume etc. as it cannot be done synchronously
4436	* under all contexts on all OSs.
4437	*
4438	* @returns VBox status code.
4439	* @param pDevExt Pointer to the device instance data.
4440	*
4441	* @remarks Must only be called -after- initializing GIP and setting up MP
4442	* notifications!
4443	*/
4444	static int supdrvTscDeltaThreadInit(PSUPDRVDEVEXT pDevExt)
4445	{
4446	int rc;
4447	Assert(pDevExt->pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED);
4448	rc = RTSpinlockCreate(&pDevExt->hTscDeltaSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "VBoxTscSpnLck");
4449	if (RT_SUCCESS(rc))
4450	{
4451	rc = RTSemEventCreate(&pDevExt->hTscDeltaEvent);
4452	if (RT_SUCCESS(rc))
4453	{
4454	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Creating;
4455	pDevExt->cMsTscDeltaTimeout = 60000;
4456	rc = RTThreadCreate(&pDevExt->hTscDeltaThread, supdrvTscDeltaThread, pDevExt, 0 /* cbStack */,
4457	RTTHREADTYPE_DEFAULT, RTTHREADFLAGS_WAITABLE, "VBoxTscThread");
4458	if (RT_SUCCESS(rc))
4459	{
4460	rc = supdrvTscDeltaThreadWait(pDevExt, kTscDeltaThreadState_Creating, kTscDeltaThreadState_Listening);
4461	if (RT_SUCCESS(rc))
4462	{
4463	ASMAtomicWriteS32(&pDevExt->rcTscDelta, VERR_NOT_AVAILABLE);
4464	return rc;
4465	}
4466
4467	OSDBGPRINT(("supdrvTscDeltaInit: supdrvTscDeltaThreadWait failed. rc=%Rrc\n", rc));
4468	supdrvTscDeltaThreadTerminate(pDevExt);
4469	}
4470	else
4471	OSDBGPRINT(("supdrvTscDeltaInit: RTThreadCreate failed. rc=%Rrc\n", rc));
4472	RTSemEventDestroy(pDevExt->hTscDeltaEvent);
4473	pDevExt->hTscDeltaEvent = NIL_RTSEMEVENT;
4474	}
4475	else
4476	OSDBGPRINT(("supdrvTscDeltaInit: RTSemEventCreate failed. rc=%Rrc\n", rc));
4477	RTSpinlockDestroy(pDevExt->hTscDeltaSpinlock);
4478	pDevExt->hTscDeltaSpinlock = NIL_RTSPINLOCK;
4479	}
4480	else
4481	OSDBGPRINT(("supdrvTscDeltaInit: RTSpinlockCreate failed. rc=%Rrc\n", rc));
4482
4483	return rc;
4484	}
4485
4486
4487	/**
4488	* Terminates the TSC-delta measurement thread and cleanup.
4489	*
4490	* @param pDevExt Pointer to the device instance data.
4491	*/
4492	static void supdrvTscDeltaTerm(PSUPDRVDEVEXT pDevExt)
4493	{
4494	if ( pDevExt->hTscDeltaSpinlock != NIL_RTSPINLOCK
4495	&& pDevExt->hTscDeltaEvent != NIL_RTSEMEVENT)
4496	{
4497	supdrvTscDeltaThreadTerminate(pDevExt);
4498	}
4499
4500	if (pDevExt->hTscDeltaSpinlock != NIL_RTSPINLOCK)
4501	{
4502	RTSpinlockDestroy(pDevExt->hTscDeltaSpinlock);
4503	pDevExt->hTscDeltaSpinlock = NIL_RTSPINLOCK;
4504	}
4505
4506	if (pDevExt->hTscDeltaEvent != NIL_RTSEMEVENT)
4507	{
4508	RTSemEventDestroy(pDevExt->hTscDeltaEvent);
4509	pDevExt->hTscDeltaEvent = NIL_RTSEMEVENT;
4510	}
4511
4512	ASMAtomicWriteS32(&pDevExt->rcTscDelta, VERR_NOT_AVAILABLE);
4513	}
4514
4515	#endif /* SUPDRV_USE_TSC_DELTA_THREAD */
4516
4517	/**
4518	* Measure the TSC delta for the CPU given by its CPU set index.
4519	*
4520	* @returns VBox status code.
4521	* @retval VERR_INTERRUPTED if interrupted while waiting.
4522	* @retval VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED if we were unable to get a
4523	* measurement.
4524	* @retval VERR_CPU_OFFLINE if the specified CPU is offline.
4525	*
4526	* @param pSession The caller's session. GIP must've been mapped.
4527	* @param iCpuSet The CPU set index of the CPU to measure.
4528	* @param fFlags Flags, SUP_TSCDELTA_MEASURE_F_XXX.
4529	* @param cMsWaitRetry Number of milliseconds to wait between each retry.
4530	* @param cMsWaitThread Number of milliseconds to wait for the thread to get
4531	* ready.
4532	* @param cTries Number of times to try, pass 0 for the default.
4533	*/
4534	SUPR0DECL(int) SUPR0TscDeltaMeasureBySetIndex(PSUPDRVSESSION pSession, uint32_t iCpuSet, uint32_t fFlags,
4535	RTMSINTERVAL cMsWaitRetry, RTMSINTERVAL cMsWaitThread, uint32_t cTries)
4536	{
4537	PSUPDRVDEVEXT pDevExt;
4538	PSUPGLOBALINFOPAGE pGip;
4539	uint16_t iGipCpu;
4540	int rc;
4541	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
4542	uint64_t msTsStartWait;
4543	uint32_t iWaitLoop;
4544	#endif
4545
4546	/*
4547	* Validate and adjust the input.
4548	*/
4549	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4550	if (!pSession->fGipReferenced)
4551	return VERR_WRONG_ORDER;
4552
4553	pDevExt = pSession->pDevExt;
4554	AssertReturn(SUP_IS_DEVEXT_VALID(pDevExt), VERR_INVALID_PARAMETER);
4555
4556	pGip = pDevExt->pGip;
4557	AssertPtrReturn(pGip, VERR_INTERNAL_ERROR_2);
4558
4559	AssertReturn(iCpuSet < RTCPUSET_MAX_CPUS, VERR_INVALID_CPU_INDEX);
4560	AssertReturn(iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx), VERR_INVALID_CPU_INDEX);
4561	iGipCpu = pGip->aiCpuFromCpuSetIdx[iCpuSet];
4562	AssertReturn(iGipCpu < pGip->cCpus, VERR_INVALID_CPU_INDEX);
4563
4564	if (fFlags & ~SUP_TSCDELTA_MEASURE_F_VALID_MASK)
4565	return VERR_INVALID_FLAGS;
4566
4567	/*
4568	* The request is a noop if the TSC delta isn't being used.
4569	*/
4570	if (pGip->enmUseTscDelta <= SUPGIPUSETSCDELTA_ZERO_CLAIMED)
4571	return VINF_SUCCESS;
4572
4573	if (cTries == 0)
4574	cTries = 12;
4575	else if (cTries > 256)
4576	cTries = 256;
4577
4578	if (cMsWaitRetry == 0)
4579	cMsWaitRetry = 2;
4580	else if (cMsWaitRetry > 1000)
4581	cMsWaitRetry = 1000;
4582
4583	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
4584	/*
4585	* Has the TSC already been measured and we're not forced to redo it?
4586	*/
4587	if ( pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX
4588	&& !(fFlags & SUP_TSCDELTA_MEASURE_F_FORCE))
4589	return VINF_SUCCESS;
4590
4591	/*
4592	* Asynchronous request? Forward it to the thread, no waiting.
4593	*/
4594	if (fFlags & SUP_TSCDELTA_MEASURE_F_ASYNC)
4595	{
4596	/** @todo Async. doesn't implement options like retries, waiting. We'll need
4597	* to pass those options to the thread somehow and implement it in the
4598	* thread. Check if anyone uses/needs fAsync before implementing this. */
4599	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4600	RTCpuSetAddByIndex(&pDevExt->TscDeltaCpuSet, iCpuSet);
4601	if ( pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Listening
4602	\|\| pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Measuring)
4603	{
4604	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_WaitAndMeasure;
4605	rc = VINF_SUCCESS;
4606	}
4607	else if (pDevExt->enmTscDeltaThreadState != kTscDeltaThreadState_WaitAndMeasure)
4608	rc = VERR_THREAD_IS_DEAD;
4609	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4610	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4611	return VINF_SUCCESS;
4612	}
4613
4614	/*
4615	* If a TSC-delta measurement request is already being serviced by the thread,
4616	* wait 'cTries' times if a retry-timeout is provided, otherwise bail as busy.
4617	*/
4618	msTsStartWait = RTTimeSystemMilliTS();
4619	for (iWaitLoop = 0;; iWaitLoop++)
4620	{
4621	uint64_t cMsElapsed;
4622	SUPDRVTSCDELTATHREADSTATE enmState;
4623	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4624	enmState = pDevExt->enmTscDeltaThreadState;
4625	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4626
4627	if (enmState == kTscDeltaThreadState_Measuring)
4628	{ /* Must wait, the thread is busy. */ }
4629	else if (enmState == kTscDeltaThreadState_WaitAndMeasure)
4630	{ /* Must wait, this state only says what will happen next. */ }
4631	else if (enmState == kTscDeltaThreadState_Terminating)
4632	{ /* Must wait, this state only says what should happen next. */ }
4633	else
4634	break; /* All other states, the thread is either idly listening or dead. */
4635
4636	/* Wait or fail. */
4637	if (cMsWaitThread == 0)
4638	return VERR_SUPDRV_TSC_DELTA_MEASUREMENT_BUSY;
4639	cMsElapsed = RTTimeSystemMilliTS() - msTsStartWait;
4640	if (cMsElapsed >= cMsWaitThread)
4641	return VERR_SUPDRV_TSC_DELTA_MEASUREMENT_BUSY;
4642
4643	rc = RTThreadSleep(RT_MIN((RTMSINTERVAL)(cMsWaitThread - cMsElapsed), RT_MIN(iWaitLoop + 1, 10)));
4644	if (rc == VERR_INTERRUPTED)
4645	return rc;
4646	}
4647	#endif /* SUPDRV_USE_TSC_DELTA_THREAD */
4648
4649	/*
4650	* Try measure the TSC delta the given number of times.
4651	*/
4652	for (;;)
4653	{
4654	/* Unless we're forced to measure the delta, check whether it's done already. */
4655	if ( !(fFlags & SUP_TSCDELTA_MEASURE_F_FORCE)
4656	&& pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX)
4657	{
4658	rc = VINF_SUCCESS;
4659	break;
4660	}
4661
4662	/* Measure it. */
4663	rc = supdrvTscMeasureDeltaOne(pDevExt, iGipCpu);
4664	if (rc != VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED)
4665	{
4666	Assert(pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX \|\| RT_FAILURE_NP(rc));
4667	break;
4668	}
4669
4670	/* Retry? */
4671	if (cTries <= 1)
4672	break;
4673	cTries--;
4674
4675	/* Always delay between retries (be nice to the rest of the system
4676	and avoid the BSOD hounds). */
4677	rc = RTThreadSleep(cMsWaitRetry);
4678	if (rc == VERR_INTERRUPTED)
4679	break;
4680	}
4681
4682	return rc;
4683	}
4684
4685
4686	/**
4687	* Service a TSC-delta measurement request.
4688	*
4689	* @returns VBox status code.
4690	* @param pDevExt Pointer to the device instance data.
4691	* @param pSession The support driver session.
4692	* @param pReq Pointer to the TSC-delta measurement request.
4693	*/
4694	int VBOXCALL supdrvIOCtl_TscDeltaMeasure(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPTSCDELTAMEASURE pReq)
4695	{
4696	uint32_t cTries;
4697	uint32_t iCpuSet;
4698	uint32_t fFlags;
4699	RTMSINTERVAL cMsWaitRetry;
4700	RT_NOREF1(pDevExt);
4701
4702	/*
4703	* Validate and adjust/resolve the input so they can be passed onto SUPR0TscDeltaMeasureBySetIndex.
4704	*/
4705	AssertPtr(pDevExt); AssertPtr(pSession); AssertPtr(pReq); /* paranoia^2 */
4706
4707	if (pReq->u.In.idCpu == NIL_RTCPUID)
4708	return VERR_INVALID_CPU_ID;
4709	iCpuSet = RTMpCpuIdToSetIndex(pReq->u.In.idCpu);
4710	if (iCpuSet >= RTCPUSET_MAX_CPUS)
4711	return VERR_INVALID_CPU_ID;
4712
4713	cTries = pReq->u.In.cRetries == 0 ? 0 : (uint32_t)pReq->u.In.cRetries + 1;
4714
4715	cMsWaitRetry = RT_MAX(pReq->u.In.cMsWaitRetry, 5);
4716
4717	fFlags = 0;
4718	if (pReq->u.In.fAsync)
4719	fFlags \|= SUP_TSCDELTA_MEASURE_F_ASYNC;
4720	if (pReq->u.In.fForce)
4721	fFlags \|= SUP_TSCDELTA_MEASURE_F_FORCE;
4722
4723	return SUPR0TscDeltaMeasureBySetIndex(pSession, iCpuSet, fFlags, cMsWaitRetry,
4724	cTries == 0 ? 5 * RT_MS_1SEC : cMsWaitRetry * cTries /cMsWaitThread/,
4725	cTries);
4726	}
4727
4728
4729	/**
4730	* Reads TSC with delta applied.
4731	*
4732	* Will try to resolve delta value INT64_MAX before applying it. This is the
4733	* main purpose of this function, to handle the case where the delta needs to be
4734	* determined.
4735	*
4736	* @returns VBox status code.
4737	* @param pDevExt Pointer to the device instance data.
4738	* @param pSession The support driver session.
4739	* @param pReq Pointer to the TSC-read request.
4740	*/
4741	int VBOXCALL supdrvIOCtl_TscRead(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPTSCREAD pReq)
4742	{
4743	PSUPGLOBALINFOPAGE pGip;
4744	int rc;
4745
4746	/*
4747	* Validate. We require the client to have mapped GIP (no asserting on
4748	* ring-3 preconditions).
4749	*/
4750	AssertPtr(pDevExt); AssertPtr(pReq); AssertPtr(pSession); /* paranoia^2 */
4751	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
4752	return VERR_WRONG_ORDER;
4753	pGip = pDevExt->pGip;
4754	AssertReturn(pGip, VERR_INTERNAL_ERROR_2);
4755
4756	/*
4757	* We're usually here because we need to apply delta, but we shouldn't be
4758	* upset if the GIP is some different mode.
4759	*/
4760	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
4761	{
4762	uint32_t cTries = 0;
4763	for (;;)
4764	{
4765	/*
4766	* Start by gathering the data, using CLI for disabling preemption
4767	* while we do that.
4768	*/
4769	RTCCUINTREG fEFlags = ASMIntDisableFlags();
4770	int iCpuSet = RTMpCpuIdToSetIndex(RTMpCpuId());
4771	int iGipCpu = 0; /* gcc maybe used uninitialized */
4772	if (RT_LIKELY( (unsigned)iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
4773	&& (iGipCpu = pGip->aiCpuFromCpuSetIdx[iCpuSet]) < pGip->cCpus ))
4774	{
4775	int64_t i64Delta = pGip->aCPUs[iGipCpu].i64TSCDelta;
4776	pReq->u.Out.idApic = pGip->aCPUs[iGipCpu].idApic;
4777	pReq->u.Out.u64AdjustedTsc = ASMReadTSC();
4778	ASMSetFlags(fEFlags);
4779
4780	/*
4781	* If we're lucky we've got a delta, but no predictions here
4782	* as this I/O control is normally only used when the TSC delta
4783	* is set to INT64_MAX.
4784	*/
4785	if (i64Delta != INT64_MAX)
4786	{
4787	pReq->u.Out.u64AdjustedTsc -= i64Delta;
4788	rc = VINF_SUCCESS;
4789	break;
4790	}
4791
4792	/* Give up after a few times. */
4793	if (cTries >= 4)
4794	{
4795	rc = VWRN_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED;
4796	break;
4797	}
4798
4799	/* Need to measure the delta an try again. */
4800	rc = supdrvTscMeasureDeltaOne(pDevExt, iGipCpu);
4801	Assert(pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX \|\| RT_FAILURE_NP(rc));
4802	/** @todo should probably delay on failure... dpc watchdogs */
4803	}
4804	else
4805	{
4806	/* This really shouldn't happen. */
4807	AssertMsgFailed(("idCpu=%#x iCpuSet=%#x (%d)\n", RTMpCpuId(), iCpuSet, iCpuSet));
4808	pReq->u.Out.idApic = ASMGetApicId();
4809	pReq->u.Out.u64AdjustedTsc = ASMReadTSC();
4810	ASMSetFlags(fEFlags);
4811	rc = VERR_INTERNAL_ERROR_5; /** @todo change to warning. */
4812	break;
4813	}
4814	}
4815	}
4816	else
4817	{
4818	/*
4819	* No delta to apply. Easy. Deal with preemption the lazy way.
4820	*/
4821	RTCCUINTREG fEFlags = ASMIntDisableFlags();
4822	int iCpuSet = RTMpCpuIdToSetIndex(RTMpCpuId());
4823	int iGipCpu = 0; /* gcc may be used uninitialized */
4824	if (RT_LIKELY( (unsigned)iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
4825	&& (iGipCpu = pGip->aiCpuFromCpuSetIdx[iCpuSet]) < pGip->cCpus ))
4826	pReq->u.Out.idApic = pGip->aCPUs[iGipCpu].idApic;
4827	else
4828	pReq->u.Out.idApic = ASMGetApicId();
4829	pReq->u.Out.u64AdjustedTsc = ASMReadTSC();
4830	ASMSetFlags(fEFlags);
4831	rc = VINF_SUCCESS;
4832	}
4833
4834	return rc;
4835	}
4836
4837
4838	/**
4839	* Worker for supdrvIOCtl_GipSetFlags.
4840	*
4841	* @returns VBox status code.
4842	* @retval VERR_WRONG_ORDER if an enable-once-per-session flag is set again for
4843	* a session.
4844	*
4845	* @param pDevExt Pointer to the device instance data.
4846	* @param pSession The support driver session.
4847	* @param fOrMask The OR mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4848	* @param fAndMask The AND mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4849	*
4850	* @remarks Caller must own the GIP mutex.
4851	*
4852	* @remarks This function doesn't validate any of the flags.
4853	*/
4854	static int supdrvGipSetFlags(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, uint32_t fOrMask, uint32_t fAndMask)
4855	{
4856	uint32_t cRefs;
4857	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4858	AssertMsg((fOrMask & fAndMask) == fOrMask, ("%#x & %#x\n", fOrMask, fAndMask)); /* ASSUMED by code below */
4859
4860	/*
4861	* Compute GIP test-mode flags.
4862	*/
4863	if (fOrMask & SUPGIP_FLAGS_TESTING_ENABLE)
4864	{
4865	if (!pSession->fGipTestMode)
4866	{
4867	Assert(pDevExt->cGipTestModeRefs < _64K);
4868	pSession->fGipTestMode = true;
4869	cRefs = ++pDevExt->cGipTestModeRefs;
4870	if (cRefs == 1)
4871	{
4872	fOrMask \|= SUPGIP_FLAGS_TESTING \| SUPGIP_FLAGS_TESTING_START;
4873	fAndMask &= ~SUPGIP_FLAGS_TESTING_STOP;
4874	}
4875	}
4876	else
4877	{
4878	LogRelMax(10, ("supdrvGipSetFlags: SUPGIP_FLAGS_TESTING_ENABLE already set for this session\n"));
4879	return VERR_WRONG_ORDER;
4880	}
4881	}
4882	else if ( !(fAndMask & SUPGIP_FLAGS_TESTING_ENABLE)
4883	&& pSession->fGipTestMode)
4884	{
4885	Assert(pDevExt->cGipTestModeRefs > 0);
4886	Assert(pDevExt->cGipTestModeRefs < _64K);
4887	pSession->fGipTestMode = false;
4888	cRefs = --pDevExt->cGipTestModeRefs;
4889	if (!cRefs)
4890	fOrMask \|= SUPGIP_FLAGS_TESTING_STOP;
4891	else
4892	fAndMask \|= SUPGIP_FLAGS_TESTING_ENABLE;
4893	}
4894
4895	/*
4896	* Commit the flags. This should be done as atomically as possible
4897	* since the flag consumers won't be holding the GIP mutex.
4898	*/
4899	ASMAtomicOrU32(&pGip->fFlags, fOrMask);
4900	ASMAtomicAndU32(&pGip->fFlags, fAndMask);
4901
4902	return VINF_SUCCESS;
4903	}
4904
4905
4906	/**
4907	* Sets GIP test mode parameters.
4908	*
4909	* @returns VBox status code.
4910	* @param pDevExt Pointer to the device instance data.
4911	* @param pSession The support driver session.
4912	* @param fOrMask The OR mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4913	* @param fAndMask The AND mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4914	*/
4915	int VBOXCALL supdrvIOCtl_GipSetFlags(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, uint32_t fOrMask, uint32_t fAndMask)
4916	{
4917	PSUPGLOBALINFOPAGE pGip;
4918	int rc;
4919
4920	/*
4921	* Validate. We require the client to have mapped GIP (no asserting on
4922	* ring-3 preconditions).
4923	*/
4924	AssertPtr(pDevExt); AssertPtr(pSession); /* paranoia^2 */
4925	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
4926	return VERR_WRONG_ORDER;
4927	pGip = pDevExt->pGip;
4928	AssertReturn(pGip, VERR_INTERNAL_ERROR_3);
4929
4930	if (fOrMask & ~SUPGIP_FLAGS_VALID_MASK)
4931	return VERR_INVALID_PARAMETER;
4932	if ((fAndMask & ~SUPGIP_FLAGS_VALID_MASK) != ~SUPGIP_FLAGS_VALID_MASK)
4933	return VERR_INVALID_PARAMETER;
4934
4935	/*
4936	* Don't confuse supdrvGipSetFlags or anyone else by both setting
4937	* and clearing the same flags. AND takes precedence.
4938	*/
4939	fOrMask &= fAndMask;
4940
4941	/*
4942	* Take the loader lock to avoid having to think about races between two
4943	* clients changing the flags at the same time (state is not simple).
4944	*/
4945	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4946	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
4947	#else
4948	RTSemFastMutexRequest(pDevExt->mtxGip);
4949	#endif
4950
4951	rc = supdrvGipSetFlags(pDevExt, pSession, fOrMask, fAndMask);
4952
4953	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4954	RTSemMutexRelease(pDevExt->mtxGip);
4955	#else
4956	RTSemFastMutexRelease(pDevExt->mtxGip);
4957	#endif
4958	return rc;
4959	}
4960

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrvGip.cpp@ 78363

Download in other formats: