VirtualBox

About
Screenshots
Downloads
Documentation
End-user docs
Technical docs
Contribute
Community

Browse Source

Context Navigation

source: vbox/trunk/src/VBox/HostDrivers/Support/SUPDrvGip.cpp@ 64281

Last change on this file since 64281 was 64281, checked in by vboxsync, 8 years ago
IPRT,SUP: Major vboxdrv and GIP version change; more flexible processor group handling on Windows.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 180.8 KB

Line
1	/* $Id: SUPDrvGip.cpp 64281 2016-10-15 16:46:29Z vboxsync $ */
2	/** @file
3	* VBoxDrv - The VirtualBox Support Driver - Common code for GIP.
4	*/
5
6	/*
7	* Copyright (C) 2006-2016 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	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_OFFSETOF(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	cbGipCpuGroups = supdrvOSGipGetGroupTableSize(pDevExt);
1873	cbGip = RT_UOFFSETOF(SUPGLOBALINFOPAGE, aCPUs[cCpus]) + cbGipCpuGroups;
1874	rc = RTR0MemObjAllocCont(&pDevExt->GipMemObj, cbGip, false /fExecutable/);
1875	if (RT_FAILURE(rc))
1876	{
1877	OSDBGPRINT(("supdrvGipCreate: failed to allocate the GIP page. rc=%d\n", rc));
1878	return rc;
1879	}
1880	pGip = (PSUPGLOBALINFOPAGE)RTR0MemObjAddress(pDevExt->GipMemObj); AssertPtr(pGip);
1881	HCPhysGip = RTR0MemObjGetPagePhysAddr(pDevExt->GipMemObj, 0); Assert(HCPhysGip != NIL_RTHCPHYS);
1882
1883	/*
1884	* Find a reasonable update interval and initialize the structure.
1885	*/
1886	supdrvGipRequestHigherTimerFrequencyFromSystem(pDevExt);
1887	/** @todo figure out why using a 100Ms interval upsets timekeeping in VMs.
1888	* See @bugref{6710}. */
1889	u32MinInterval = RT_NS_10MS;
1890	u32SystemResolution = RTTimerGetSystemGranularity();
1891	u32Interval = u32MinInterval;
1892	uMod = u32MinInterval % u32SystemResolution;
1893	if (uMod)
1894	u32Interval += u32SystemResolution - uMod;
1895
1896	rc = supdrvGipInit(pDevExt, pGip, HCPhysGip, RTTimeSystemNanoTS(), RT_NS_1SEC / u32Interval /=Hz/, u32Interval,
1897	cCpus, cbGipCpuGroups);
1898
1899	/*
1900	* Important sanity check... (Sets rc)
1901	*/
1902	if (RT_UNLIKELY( pGip->enmUseTscDelta == SUPGIPUSETSCDELTA_ZERO_CLAIMED
1903	&& pGip->u32Mode == SUPGIPMODE_ASYNC_TSC
1904	&& !supdrvOSGetForcedAsyncTscMode(pDevExt)))
1905	{
1906	OSDBGPRINT(("supdrvGipCreate: Host-OS/user claims the TSC-deltas are zero but we detected async. TSC! Bad.\n"));
1907	rc = VERR_INTERNAL_ERROR_2;
1908	}
1909
1910	/* It doesn't make sense to do TSC-delta detection on systems we detect as async. */
1911	AssertStmt( pGip->u32Mode != SUPGIPMODE_ASYNC_TSC
1912	\|\| pGip->enmUseTscDelta <= SUPGIPUSETSCDELTA_ZERO_CLAIMED,
1913	rc = VERR_INTERNAL_ERROR_3);
1914
1915	/*
1916	* Do the TSC frequency measurements.
1917	*
1918	* If we're in invariant TSC mode, just to a quick preliminary measurement
1919	* that the TSC-delta measurement code can use to yield cross calls.
1920	*
1921	* If we're in any of the other two modes, neither which require MP init,
1922	* notifications or deltas for the job, do the full measurement now so
1923	* that supdrvGipInitOnCpu() can populate the TSC interval and history
1924	* array with more reasonable values.
1925	*/
1926	if (RT_SUCCESS(rc))
1927	{
1928	if (pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC)
1929	{
1930	rc = supdrvGipInitMeasureTscFreq(pGip, true /fRough/); /* cannot fail */
1931	supdrvGipInitStartTimerForRefiningInvariantTscFreq(pDevExt);
1932	}
1933	else
1934	rc = supdrvGipInitMeasureTscFreq(pGip, false /fRough/);
1935	if (RT_SUCCESS(rc))
1936	{
1937	/*
1938	* Start TSC-delta measurement thread before we start getting MP
1939	* events that will try kick it into action (includes the
1940	* RTMpOnAll/supdrvGipInitOnCpu call below).
1941	*/
1942	RTCpuSetEmpty(&pDevExt->TscDeltaCpuSet);
1943	RTCpuSetEmpty(&pDevExt->TscDeltaObtainedCpuSet);
1944	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
1945	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1946	rc = supdrvTscDeltaThreadInit(pDevExt);
1947	#endif
1948	if (RT_SUCCESS(rc))
1949	{
1950	rc = RTMpNotificationRegister(supdrvGipMpEvent, pDevExt);
1951	if (RT_SUCCESS(rc))
1952	{
1953	/*
1954	* Do GIP initialization on all online CPUs. Wake up the
1955	* TSC-delta thread afterwards.
1956	*/
1957	rc = RTMpOnAll(supdrvGipInitOnCpu, pDevExt, pGip);
1958	if (RT_SUCCESS(rc))
1959	{
1960	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
1961	supdrvTscDeltaThreadStartMeasurement(pDevExt, true /* fForceAll */);
1962	#else
1963	uint16_t iCpu;
1964	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
1965	{
1966	/*
1967	* Measure the TSC deltas now that we have MP notifications.
1968	*/
1969	int cTries = 5;
1970	do
1971	{
1972	rc = supdrvTscMeasureInitialDeltas(pDevExt);
1973	if ( rc != VERR_TRY_AGAIN
1974	&& rc != VERR_CPU_OFFLINE)
1975	break;
1976	} while (--cTries > 0);
1977	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
1978	Log(("supdrvTscDeltaInit: cpu[%u] delta %lld\n", iCpu, pGip->aCPUs[iCpu].i64TSCDelta));
1979	}
1980	else
1981	{
1982	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
1983	AssertMsg(!pGip->aCPUs[iCpu].i64TSCDelta, ("iCpu=%u %lld mode=%d\n", iCpu, pGip->aCPUs[iCpu].i64TSCDelta, pGip->u32Mode));
1984	}
1985	if (RT_SUCCESS(rc))
1986	#endif
1987	{
1988	/*
1989	* Create the timer.
1990	* If CPU_ALL isn't supported we'll have to fall back to synchronous mode.
1991	*/
1992	if (pGip->u32Mode == SUPGIPMODE_ASYNC_TSC)
1993	{
1994	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, RTTIMER_FLAGS_CPU_ALL,
1995	supdrvGipAsyncTimer, pDevExt);
1996	if (rc == VERR_NOT_SUPPORTED)
1997	{
1998	OSDBGPRINT(("supdrvGipCreate: omni timer not supported, falling back to synchronous mode\n"));
1999	pGip->u32Mode = SUPGIPMODE_SYNC_TSC;
2000	}
2001	}
2002	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
2003	rc = RTTimerCreateEx(&pDevExt->pGipTimer, u32Interval, 0 /* fFlags */,
2004	supdrvGipSyncAndInvariantTimer, pDevExt);
2005	if (RT_SUCCESS(rc))
2006	{
2007	/*
2008	* We're good.
2009	*/
2010	Log(("supdrvGipCreate: %u ns interval.\n", u32Interval));
2011	supdrvGipReleaseHigherTimerFrequencyFromSystem(pDevExt);
2012
2013	g_pSUPGlobalInfoPage = pGip;
2014	return VINF_SUCCESS;
2015	}
2016
2017	OSDBGPRINT(("supdrvGipCreate: failed create GIP timer at %u ns interval. rc=%Rrc\n", u32Interval, rc));
2018	Assert(!pDevExt->pGipTimer);
2019	}
2020	}
2021	else
2022	OSDBGPRINT(("supdrvGipCreate: RTMpOnAll failed. rc=%Rrc\n", rc));
2023	}
2024	else
2025	OSDBGPRINT(("supdrvGipCreate: failed to register MP event notfication. rc=%Rrc\n", rc));
2026	}
2027	else
2028	OSDBGPRINT(("supdrvGipCreate: supdrvTscDeltaInit failed. rc=%Rrc\n", rc));
2029	}
2030	else
2031	OSDBGPRINT(("supdrvGipCreate: supdrvTscMeasureInitialDeltas failed. rc=%Rrc\n", rc));
2032	}
2033
2034	/* Releases timer frequency increase too. */
2035	supdrvGipDestroy(pDevExt);
2036	return rc;
2037	}
2038
2039
2040	/**
2041	* Invalidates the GIP data upon termination.
2042	*
2043	* @param pGip Pointer to the read-write kernel mapping of the GIP.
2044	*/
2045	static void supdrvGipTerm(PSUPGLOBALINFOPAGE pGip)
2046	{
2047	unsigned i;
2048	pGip->u32Magic = 0;
2049	for (i = 0; i < pGip->cCpus; i++)
2050	{
2051	pGip->aCPUs[i].u64NanoTS = 0;
2052	pGip->aCPUs[i].u64TSC = 0;
2053	pGip->aCPUs[i].iTSCHistoryHead = 0;
2054	pGip->aCPUs[i].u64TSCSample = 0;
2055	pGip->aCPUs[i].i64TSCDelta = INT64_MAX;
2056	}
2057	}
2058
2059
2060	/**
2061	* Terminates the GIP.
2062	*
2063	* @param pDevExt Instance data. GIP stuff may be updated.
2064	*/
2065	void VBOXCALL supdrvGipDestroy(PSUPDRVDEVEXT pDevExt)
2066	{
2067	int rc;
2068	#ifdef DEBUG_DARWIN_GIP
2069	OSDBGPRINT(("supdrvGipDestroy: pDevExt=%p pGip=%p pGipTimer=%p GipMemObj=%p\n", pDevExt,
2070	pDevExt->GipMemObj != NIL_RTR0MEMOBJ ? RTR0MemObjAddress(pDevExt->GipMemObj) : NULL,
2071	pDevExt->pGipTimer, pDevExt->GipMemObj));
2072	#endif
2073
2074	/*
2075	* Stop receiving MP notifications before tearing anything else down.
2076	*/
2077	RTMpNotificationDeregister(supdrvGipMpEvent, pDevExt);
2078
2079	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
2080	/*
2081	* Terminate the TSC-delta measurement thread and resources.
2082	*/
2083	supdrvTscDeltaTerm(pDevExt);
2084	#endif
2085
2086	/*
2087	* Destroy the TSC-refinement timer.
2088	*/
2089	if (pDevExt->pInvarTscRefineTimer)
2090	{
2091	RTTimerDestroy(pDevExt->pInvarTscRefineTimer);
2092	pDevExt->pInvarTscRefineTimer = NULL;
2093	}
2094
2095	/*
2096	* Invalid the GIP data.
2097	*/
2098	if (pDevExt->pGip)
2099	{
2100	supdrvGipTerm(pDevExt->pGip);
2101	pDevExt->pGip = NULL;
2102	}
2103	g_pSUPGlobalInfoPage = NULL;
2104
2105	/*
2106	* Destroy the timer and free the GIP memory object.
2107	*/
2108	if (pDevExt->pGipTimer)
2109	{
2110	rc = RTTimerDestroy(pDevExt->pGipTimer); AssertRC(rc);
2111	pDevExt->pGipTimer = NULL;
2112	}
2113
2114	if (pDevExt->GipMemObj != NIL_RTR0MEMOBJ)
2115	{
2116	rc = RTR0MemObjFree(pDevExt->GipMemObj, true /* free mappings */); AssertRC(rc);
2117	pDevExt->GipMemObj = NIL_RTR0MEMOBJ;
2118	}
2119
2120	/*
2121	* Finally, make sure we've release the system timer resolution request
2122	* if one actually succeeded and is still pending.
2123	*/
2124	supdrvGipReleaseHigherTimerFrequencyFromSystem(pDevExt);
2125	}
2126
2127
2128
2129
2130	/*
2131	*
2132	*
2133	* GIP Update Timer Related Code
2134	* GIP Update Timer Related Code
2135	* GIP Update Timer Related Code
2136	*
2137	*
2138	*/
2139
2140
2141	/**
2142	* Worker routine for supdrvGipUpdate() and supdrvGipUpdatePerCpu() that
2143	* updates all the per cpu data except the transaction id.
2144	*
2145	* @param pDevExt The device extension.
2146	* @param pGipCpu Pointer to the per cpu data.
2147	* @param u64NanoTS The current time stamp.
2148	* @param u64TSC The current TSC.
2149	* @param iTick The current timer tick.
2150	*
2151	* @remarks Can be called with interrupts disabled!
2152	*/
2153	static void supdrvGipDoUpdateCpu(PSUPDRVDEVEXT pDevExt, PSUPGIPCPU pGipCpu, uint64_t u64NanoTS, uint64_t u64TSC, uint64_t iTick)
2154	{
2155	uint64_t u64TSCDelta;
2156	bool fUpdateCpuHz;
2157	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2158	AssertPtrReturnVoid(pGip);
2159
2160	/* Delta between this and the previous update. */
2161	ASMAtomicUoWriteU32(&pGipCpu->u32PrevUpdateIntervalNS, (uint32_t)(u64NanoTS - pGipCpu->u64NanoTS));
2162
2163	/*
2164	* Update the NanoTS.
2165	*/
2166	ASMAtomicWriteU64(&pGipCpu->u64NanoTS, u64NanoTS);
2167
2168	/*
2169	* Calc TSC delta.
2170	*/
2171	u64TSCDelta = u64TSC - pGipCpu->u64TSC;
2172	ASMAtomicWriteU64(&pGipCpu->u64TSC, u64TSC);
2173
2174	/*
2175	* Determine if we need to update the CPU (TSC) frequency calculation.
2176	*
2177	* We don't need to keep recalculating the frequency when it's invariant,
2178	* unless the special tstGIP-2 testing mode is enabled.
2179	*/
2180	fUpdateCpuHz = pGip->u32Mode != SUPGIPMODE_INVARIANT_TSC;
2181	if (!(pGip->fFlags & SUPGIP_FLAGS_TESTING))
2182	{ /* likely*/ }
2183	else
2184	{
2185	uint32_t fGipFlags = pGip->fFlags;
2186	if (fGipFlags & (SUPGIP_FLAGS_TESTING_ENABLE \| SUPGIP_FLAGS_TESTING_START))
2187	{
2188	if (fGipFlags & SUPGIP_FLAGS_TESTING_START)
2189	{
2190	/* Cache the TSC frequency before forcing updates due to test mode. */
2191	if (!fUpdateCpuHz)
2192	pDevExt->uGipTestModeInvariantCpuHz = pGip->aCPUs[0].u64CpuHz;
2193	ASMAtomicAndU32(&pGip->fFlags, ~SUPGIP_FLAGS_TESTING_START);
2194	}
2195	fUpdateCpuHz = true;
2196	}
2197	else if (fGipFlags & SUPGIP_FLAGS_TESTING_STOP)
2198	{
2199	/* Restore the cached TSC frequency if any. */
2200	if (!fUpdateCpuHz)
2201	{
2202	Assert(pDevExt->uGipTestModeInvariantCpuHz);
2203	ASMAtomicWriteU64(&pGip->aCPUs[0].u64CpuHz, pDevExt->uGipTestModeInvariantCpuHz);
2204	}
2205	ASMAtomicAndU32(&pGip->fFlags, ~(SUPGIP_FLAGS_TESTING_STOP \| SUPGIP_FLAGS_TESTING));
2206	}
2207	}
2208
2209	/*
2210	* Calculate the CPU (TSC) frequency if necessary.
2211	*/
2212	if (fUpdateCpuHz)
2213	{
2214	uint64_t u64CpuHz;
2215	uint32_t u32UpdateIntervalTSC;
2216	uint32_t u32UpdateIntervalTSCSlack;
2217	uint32_t u32TransactionId;
2218	unsigned iTSCHistoryHead;
2219
2220	if (u64TSCDelta >> 32)
2221	{
2222	u64TSCDelta = pGipCpu->u32UpdateIntervalTSC;
2223	pGipCpu->cErrors++;
2224	}
2225
2226	/*
2227	* On the 2nd and 3rd callout, reset the history with the current TSC
2228	* interval since the values entered by supdrvGipInit are totally off.
2229	* The interval on the 1st callout completely unreliable, the 2nd is a bit
2230	* better, while the 3rd should be most reliable.
2231	*/
2232	/** @todo Could we drop this now that we initializes the history
2233	* with nominal TSC frequency values? */
2234	u32TransactionId = pGipCpu->u32TransactionId;
2235	if (RT_UNLIKELY( ( u32TransactionId == 5
2236	\|\| u32TransactionId == 7)
2237	&& ( iTick == 2
2238	\|\| iTick == 3) ))
2239	{
2240	unsigned i;
2241	for (i = 0; i < RT_ELEMENTS(pGipCpu->au32TSCHistory); i++)
2242	ASMAtomicUoWriteU32(&pGipCpu->au32TSCHistory[i], (uint32_t)u64TSCDelta);
2243	}
2244
2245	/*
2246	* Validate the NanoTS deltas between timer fires with an arbitrary threshold of 0.5%.
2247	* Wait until we have at least one full history since the above history reset. The
2248	* assumption is that the majority of the previous history values will be tolerable.
2249	* See @bugref{6710#c67}.
2250	*/
2251	/** @todo Could we drop the fudging there now that we initializes the history
2252	* with nominal TSC frequency values? */
2253	if ( u32TransactionId > 23 /* 7 + (8 * 2) */
2254	&& pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
2255	{
2256	uint32_t uNanoTsThreshold = pGip->u32UpdateIntervalNS / 200;
2257	if ( pGipCpu->u32PrevUpdateIntervalNS > pGip->u32UpdateIntervalNS + uNanoTsThreshold
2258	\|\| pGipCpu->u32PrevUpdateIntervalNS < pGip->u32UpdateIntervalNS - uNanoTsThreshold)
2259	{
2260	uint32_t u32;
2261	u32 = pGipCpu->au32TSCHistory[0];
2262	u32 += pGipCpu->au32TSCHistory[1];
2263	u32 += pGipCpu->au32TSCHistory[2];
2264	u32 += pGipCpu->au32TSCHistory[3];
2265	u32 >>= 2;
2266	u64TSCDelta = pGipCpu->au32TSCHistory[4];
2267	u64TSCDelta += pGipCpu->au32TSCHistory[5];
2268	u64TSCDelta += pGipCpu->au32TSCHistory[6];
2269	u64TSCDelta += pGipCpu->au32TSCHistory[7];
2270	u64TSCDelta >>= 2;
2271	u64TSCDelta += u32;
2272	u64TSCDelta >>= 1;
2273	}
2274	}
2275
2276	/*
2277	* TSC History.
2278	*/
2279	Assert(RT_ELEMENTS(pGipCpu->au32TSCHistory) == 8);
2280	iTSCHistoryHead = (pGipCpu->iTSCHistoryHead + 1) & 7;
2281	ASMAtomicWriteU32(&pGipCpu->iTSCHistoryHead, iTSCHistoryHead);
2282	ASMAtomicWriteU32(&pGipCpu->au32TSCHistory[iTSCHistoryHead], (uint32_t)u64TSCDelta);
2283
2284	/*
2285	* UpdateIntervalTSC = average of last 8,2,1 intervals depending on update HZ.
2286	*
2287	* On Windows, we have an occasional (but recurring) sour value that messed up
2288	* the history but taking only 1 interval reduces the precision overall.
2289	*/
2290	if ( pGip->u32Mode == SUPGIPMODE_INVARIANT_TSC
2291	\|\| pGip->u32UpdateHz >= 1000)
2292	{
2293	uint32_t u32;
2294	u32 = pGipCpu->au32TSCHistory[0];
2295	u32 += pGipCpu->au32TSCHistory[1];
2296	u32 += pGipCpu->au32TSCHistory[2];
2297	u32 += pGipCpu->au32TSCHistory[3];
2298	u32 >>= 2;
2299	u32UpdateIntervalTSC = pGipCpu->au32TSCHistory[4];
2300	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[5];
2301	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[6];
2302	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[7];
2303	u32UpdateIntervalTSC >>= 2;
2304	u32UpdateIntervalTSC += u32;
2305	u32UpdateIntervalTSC >>= 1;
2306
2307	/* Value chosen for a 2GHz Athlon64 running linux 2.6.10/11. */
2308	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 14;
2309	}
2310	else if (pGip->u32UpdateHz >= 90)
2311	{
2312	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
2313	u32UpdateIntervalTSC += pGipCpu->au32TSCHistory[(iTSCHistoryHead - 1) & 7];
2314	u32UpdateIntervalTSC >>= 1;
2315
2316	/* value chosen on a 2GHz thinkpad running windows */
2317	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 7;
2318	}
2319	else
2320	{
2321	u32UpdateIntervalTSC = (uint32_t)u64TSCDelta;
2322
2323	/* This value hasn't be checked yet.. waiting for OS/2 and 33Hz timers.. :-) */
2324	u32UpdateIntervalTSCSlack = u32UpdateIntervalTSC >> 6;
2325	}
2326	ASMAtomicWriteU32(&pGipCpu->u32UpdateIntervalTSC, u32UpdateIntervalTSC + u32UpdateIntervalTSCSlack);
2327
2328	/*
2329	* CpuHz.
2330	*/
2331	u64CpuHz = ASMMult2xU32RetU64(u32UpdateIntervalTSC, RT_NS_1SEC);
2332	u64CpuHz /= pGip->u32UpdateIntervalNS;
2333	ASMAtomicWriteU64(&pGipCpu->u64CpuHz, u64CpuHz);
2334	}
2335	}
2336
2337
2338	/**
2339	* Updates the GIP.
2340	*
2341	* @param pDevExt The device extension.
2342	* @param u64NanoTS The current nanosecond timestamp.
2343	* @param u64TSC The current TSC timestamp.
2344	* @param idCpu The CPU ID.
2345	* @param iTick The current timer tick.
2346	*
2347	* @remarks Can be called with interrupts disabled!
2348	*/
2349	static void supdrvGipUpdate(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC, RTCPUID idCpu, uint64_t iTick)
2350	{
2351	/*
2352	* Determine the relevant CPU data.
2353	*/
2354	PSUPGIPCPU pGipCpu;
2355	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2356	AssertPtrReturnVoid(pGip);
2357
2358	if (pGip->u32Mode != SUPGIPMODE_ASYNC_TSC)
2359	pGipCpu = &pGip->aCPUs[0];
2360	else
2361	{
2362	unsigned iCpu = pGip->aiCpuFromApicId[ASMGetApicId()];
2363	if (RT_UNLIKELY(iCpu >= pGip->cCpus))
2364	return;
2365	pGipCpu = &pGip->aCPUs[iCpu];
2366	if (RT_UNLIKELY(pGipCpu->idCpu != idCpu))
2367	return;
2368	}
2369
2370	/*
2371	* Start update transaction.
2372	*/
2373	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
2374	{
2375	/* this can happen on win32 if we're taking to long and there are more CPUs around. shouldn't happen though. */
2376	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
2377	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2378	pGipCpu->cErrors++;
2379	return;
2380	}
2381
2382	/*
2383	* Recalc the update frequency every 0x800th time.
2384	*/
2385	if ( pGip->u32Mode != SUPGIPMODE_INVARIANT_TSC /* cuz we're not recalculating the frequency on invariant hosts. */
2386	&& !(pGipCpu->u32TransactionId & (GIP_UPDATEHZ_RECALC_FREQ * 2 - 2)))
2387	{
2388	if (pGip->u64NanoTSLastUpdateHz)
2389	{
2390	#ifdef RT_ARCH_AMD64 /** @todo fix 64-bit div here to work on x86 linux. */
2391	uint64_t u64Delta = u64NanoTS - pGip->u64NanoTSLastUpdateHz;
2392	uint32_t u32UpdateHz = (uint32_t)((RT_NS_1SEC_64 * GIP_UPDATEHZ_RECALC_FREQ) / u64Delta);
2393	if (u32UpdateHz <= 2000 && u32UpdateHz >= 30)
2394	{
2395	/** @todo r=ramshankar: Changing u32UpdateHz might screw up TSC frequency
2396	* calculation on non-invariant hosts if it changes the history decision
2397	* taken in supdrvGipDoUpdateCpu(). */
2398	uint64_t u64Interval = u64Delta / GIP_UPDATEHZ_RECALC_FREQ;
2399	ASMAtomicWriteU32(&pGip->u32UpdateHz, u32UpdateHz);
2400	ASMAtomicWriteU32(&pGip->u32UpdateIntervalNS, (uint32_t)u64Interval);
2401	}
2402	#endif
2403	}
2404	ASMAtomicWriteU64(&pGip->u64NanoTSLastUpdateHz, u64NanoTS \| 1);
2405	}
2406
2407	/*
2408	* Update the data.
2409	*/
2410	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
2411
2412	/*
2413	* Complete transaction.
2414	*/
2415	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2416	}
2417
2418
2419	/**
2420	* Updates the per cpu GIP data for the calling cpu.
2421	*
2422	* @param pDevExt The device extension.
2423	* @param u64NanoTS The current nanosecond timestamp.
2424	* @param u64TSC The current TSC timesaver.
2425	* @param idCpu The CPU ID.
2426	* @param idApic The APIC id for the CPU index.
2427	* @param iTick The current timer tick.
2428	*
2429	* @remarks Can be called with interrupts disabled!
2430	*/
2431	static void supdrvGipUpdatePerCpu(PSUPDRVDEVEXT pDevExt, uint64_t u64NanoTS, uint64_t u64TSC,
2432	RTCPUID idCpu, uint8_t idApic, uint64_t iTick)
2433	{
2434	uint32_t iCpu;
2435	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2436
2437	/*
2438	* Avoid a potential race when a CPU online notification doesn't fire on
2439	* the onlined CPU but the tick creeps in before the event notification is
2440	* run.
2441	*/
2442	if (RT_LIKELY(iTick != 1))
2443	{ /* likely*/ }
2444	else
2445	{
2446	iCpu = supdrvGipFindOrAllocCpuIndexForCpuId(pGip, idCpu);
2447	if (pGip->aCPUs[iCpu].enmState == SUPGIPCPUSTATE_OFFLINE)
2448	supdrvGipMpEventOnlineOrInitOnCpu(pDevExt, idCpu);
2449	}
2450
2451	iCpu = pGip->aiCpuFromApicId[idApic];
2452	if (RT_LIKELY(iCpu < pGip->cCpus))
2453	{
2454	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
2455	if (pGipCpu->idCpu == idCpu)
2456	{
2457	/*
2458	* Start update transaction.
2459	*/
2460	if (!(ASMAtomicIncU32(&pGipCpu->u32TransactionId) & 1))
2461	{
2462	AssertMsgFailed(("Invalid transaction id, %#x, not odd!\n", pGipCpu->u32TransactionId));
2463	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2464	pGipCpu->cErrors++;
2465	return;
2466	}
2467
2468	/*
2469	* Update the data.
2470	*/
2471	supdrvGipDoUpdateCpu(pDevExt, pGipCpu, u64NanoTS, u64TSC, iTick);
2472
2473	/*
2474	* Complete transaction.
2475	*/
2476	ASMAtomicIncU32(&pGipCpu->u32TransactionId);
2477	}
2478	}
2479	}
2480
2481
2482	/**
2483	* Timer callback function for the sync and invariant GIP modes.
2484	*
2485	* @param pTimer The timer.
2486	* @param pvUser Opaque pointer to the device extension.
2487	* @param iTick The timer tick.
2488	*/
2489	static DECLCALLBACK(void) supdrvGipSyncAndInvariantTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
2490	{
2491	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
2492	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
2493	RTCCUINTREG fEFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
2494	uint64_t u64TSC = ASMReadTSC();
2495	uint64_t u64NanoTS = RTTimeSystemNanoTS();
2496	RT_NOREF1(pTimer);
2497
2498	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_PRACTICALLY_ZERO)
2499	{
2500	/*
2501	* The calculations in supdrvGipUpdate() is somewhat timing sensitive,
2502	* missing timer ticks is not an option for GIP because the GIP users
2503	* will end up incrementing the time in 1ns per time getter call until
2504	* there is a complete timer update. So, if the delta has yet to be
2505	* calculated, we just pretend it is zero for now (the GIP users
2506	* probably won't have it for a wee while either and will do the same).
2507	*
2508	* We could maybe on some platforms try cross calling a CPU with a
2509	* working delta here, but it's not worth the hassle since the
2510	* likelihood of this happening is really low. On Windows, Linux, and
2511	* Solaris timers fire on the CPU they were registered/started on.
2512	* Darwin timers doesn't necessarily (they are high priority threads).
2513	*/
2514	uint32_t iCpuSet = RTMpCpuIdToSetIndex(RTMpCpuId());
2515	uint16_t iGipCpu = RT_LIKELY(iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx))
2516	? pGip->aiCpuFromCpuSetIdx[iCpuSet] : UINT16_MAX;
2517	Assert(!ASMIntAreEnabled());
2518	if (RT_LIKELY(iGipCpu < pGip->cCpus))
2519	{
2520	int64_t iTscDelta = pGip->aCPUs[iGipCpu].i64TSCDelta;
2521	if (iTscDelta != INT64_MAX)
2522	u64TSC -= iTscDelta;
2523	}
2524	}
2525
2526	supdrvGipUpdate(pDevExt, u64NanoTS, u64TSC, NIL_RTCPUID, iTick);
2527
2528	ASMSetFlags(fEFlags);
2529	}
2530
2531
2532	/**
2533	* Timer callback function for async GIP mode.
2534	* @param pTimer The timer.
2535	* @param pvUser Opaque pointer to the device extension.
2536	* @param iTick The timer tick.
2537	*/
2538	static DECLCALLBACK(void) supdrvGipAsyncTimer(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
2539	{
2540	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
2541	RTCCUINTREG fEFlags = ASMIntDisableFlags(); /* No interruptions please (real problem on S10). */
2542	RTCPUID idCpu = RTMpCpuId();
2543	uint64_t u64TSC = ASMReadTSC();
2544	uint64_t NanoTS = RTTimeSystemNanoTS();
2545	RT_NOREF1(pTimer);
2546
2547	/** @todo reset the transaction number and whatnot when iTick == 1. */
2548	if (pDevExt->idGipMaster == idCpu)
2549	supdrvGipUpdate(pDevExt, NanoTS, u64TSC, idCpu, iTick);
2550	else
2551	supdrvGipUpdatePerCpu(pDevExt, NanoTS, u64TSC, idCpu, ASMGetApicId(), iTick);
2552
2553	ASMSetFlags(fEFlags);
2554	}
2555
2556
2557
2558
2559	/*
2560	*
2561	*
2562	* TSC Delta Measurements And Related Code
2563	* TSC Delta Measurements And Related Code
2564	* TSC Delta Measurements And Related Code
2565	*
2566	*
2567	*/
2568
2569
2570	/*
2571	* Select TSC delta measurement algorithm.
2572	*/
2573	#if 0
2574	# define GIP_TSC_DELTA_METHOD_1
2575	#else
2576	# define GIP_TSC_DELTA_METHOD_2
2577	#endif
2578
2579	/** For padding variables to keep them away from other cache lines. Better too
2580	* large than too small!
2581	* @remarks Current AMD64 and x86 CPUs seems to use 64 bytes. There are claims
2582	* that NetBurst had 128 byte cache lines while the 486 thru Pentium
2583	* III had 32 bytes cache lines. */
2584	#define GIP_TSC_DELTA_CACHE_LINE_SIZE 128
2585
2586
2587	/**
2588	* TSC delta measurement algorithm \#2 result entry.
2589	*/
2590	typedef struct SUPDRVTSCDELTAMETHOD2ENTRY
2591	{
2592	uint32_t iSeqMine;
2593	uint32_t iSeqOther;
2594	uint64_t uTsc;
2595	} SUPDRVTSCDELTAMETHOD2ENTRY;
2596
2597	/**
2598	* TSC delta measurement algorithm \#2 Data.
2599	*/
2600	typedef struct SUPDRVTSCDELTAMETHOD2
2601	{
2602	/** Padding to make sure the iCurSeqNo is in its own cache line. */
2603	uint64_t au64CacheLinePaddingBefore[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2604	/** The current sequence number of this worker. */
2605	uint32_t volatile iCurSeqNo;
2606	/** Padding to make sure the iCurSeqNo is in its own cache line. */
2607	uint32_t au64CacheLinePaddingAfter[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint32_t) - 1];
2608	/** Result table. */
2609	SUPDRVTSCDELTAMETHOD2ENTRY aResults[64];
2610	} SUPDRVTSCDELTAMETHOD2;
2611	/** Pointer to the data for TSC delta measurement algorithm \#2 .*/
2612	typedef SUPDRVTSCDELTAMETHOD2 *PSUPDRVTSCDELTAMETHOD2;
2613
2614
2615	/**
2616	* The TSC delta synchronization struct, version 2.
2617	*
2618	* The synchronization variable is completely isolated in its own cache line
2619	* (provided our max cache line size estimate is correct).
2620	*/
2621	typedef struct SUPTSCDELTASYNC2
2622	{
2623	/** Padding to make sure the uVar1 is in its own cache line. */
2624	uint64_t au64CacheLinePaddingBefore[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2625
2626	/** The synchronization variable, holds values GIP_TSC_DELTA_SYNC_. /
2627	volatile uint32_t uSyncVar;
2628	/** Sequence synchronizing variable used for post 'GO' synchronization. */
2629	volatile uint32_t uSyncSeq;
2630
2631	/** Padding to make sure the uVar1 is in its own cache line. */
2632	uint64_t au64CacheLinePaddingAfter[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t) - 2];
2633
2634	/** Start RDTSC value. Put here mainly to save stack space. */
2635	uint64_t uTscStart;
2636	/** Copy of SUPDRVGIPTSCDELTARGS::cMaxTscTicks. */
2637	uint64_t cMaxTscTicks;
2638	} SUPTSCDELTASYNC2;
2639	AssertCompileSize(SUPTSCDELTASYNC2, GIP_TSC_DELTA_CACHE_LINE_SIZE * 2 + sizeof(uint64_t));
2640	typedef SUPTSCDELTASYNC2 *PSUPTSCDELTASYNC2;
2641
2642	/** Prestart wait. */
2643	#define GIP_TSC_DELTA_SYNC2_PRESTART_WAIT UINT32_C(0x0ffe)
2644	/** Prestart aborted. */
2645	#define GIP_TSC_DELTA_SYNC2_PRESTART_ABORT UINT32_C(0x0fff)
2646	/** Ready (on your mark). */
2647	#define GIP_TSC_DELTA_SYNC2_READY UINT32_C(0x1000)
2648	/** Steady (get set). */
2649	#define GIP_TSC_DELTA_SYNC2_STEADY UINT32_C(0x1001)
2650	/** Go! */
2651	#define GIP_TSC_DELTA_SYNC2_GO UINT32_C(0x1002)
2652	/** Used by the verification test. */
2653	#define GIP_TSC_DELTA_SYNC2_GO_GO UINT32_C(0x1003)
2654
2655	/** We reached the time limit. */
2656	#define GIP_TSC_DELTA_SYNC2_TIMEOUT UINT32_C(0x1ffe)
2657	/** The other party won't touch the sync struct ever again. */
2658	#define GIP_TSC_DELTA_SYNC2_FINAL UINT32_C(0x1fff)
2659
2660
2661	/**
2662	* Argument package/state passed by supdrvTscMeasureDeltaOne() to the RTMpOn
2663	* callback worker.
2664	* @todo add
2665	*/
2666	typedef struct SUPDRVGIPTSCDELTARGS
2667	{
2668	/** The device extension. */
2669	PSUPDRVDEVEXT pDevExt;
2670	/** Pointer to the GIP CPU array entry for the worker. */
2671	PSUPGIPCPU pWorker;
2672	/** Pointer to the GIP CPU array entry for the master. */
2673	PSUPGIPCPU pMaster;
2674	/** The maximum number of ticks to spend in supdrvTscMeasureDeltaCallback.
2675	* (This is what we need a rough TSC frequency for.) */
2676	uint64_t cMaxTscTicks;
2677	/** Used to abort synchronization setup. */
2678	bool volatile fAbortSetup;
2679
2680	/** Padding to make sure the master variables live in its own cache lines. */
2681	uint64_t au64CacheLinePaddingBefore[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2682
2683	/** @name Master
2684	* @{ */
2685	/** The time the master spent in the MP worker. */
2686	uint64_t cElapsedMasterTscTicks;
2687	/** The iTry value when stopped at. */
2688	uint32_t iTry;
2689	/** Set if the run timed out. */
2690	bool volatile fTimedOut;
2691	/** Pointer to the master's synchronization struct (on stack). */
2692	PSUPTSCDELTASYNC2 volatile pSyncMaster;
2693	/** Master data union. */
2694	union
2695	{
2696	/** Data (master) for delta verification. */
2697	struct
2698	{
2699	/** Verification test TSC values for the master. */
2700	uint64_t volatile auTscs[32];
2701	} Verify;
2702	/** Data (master) for measurement method \#2. */
2703	struct
2704	{
2705	/** Data and sequence number. */
2706	SUPDRVTSCDELTAMETHOD2 Data;
2707	/** The lag setting for the next run. */
2708	bool fLag;
2709	/** Number of hits. */
2710	uint32_t cHits;
2711	} M2;
2712	} uMaster;
2713	/** The verifier verdict, VINF_SUCCESS if ok, VERR_OUT_OF_RANGE if not,
2714	* VERR_TRY_AGAIN on timeout. */
2715	int32_t rcVerify;
2716	#ifdef TSCDELTA_VERIFY_WITH_STATS
2717	/** The maximum difference between TSC read during delta verification. */
2718	int64_t cMaxVerifyTscTicks;
2719	/** The minimum difference between two TSC reads during verification. */
2720	int64_t cMinVerifyTscTicks;
2721	/** The bad TSC diff, worker relative to master (= worker - master).
2722	* Negative value means the worker is behind the master. */
2723	int64_t iVerifyBadTscDiff;
2724	#endif
2725	/** @} */
2726
2727	/** Padding to make sure the worker variables live is in its own cache line. */
2728	uint64_t au64CacheLinePaddingBetween[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2729
2730	/** @name Proletarian
2731	* @{ */
2732	/** Pointer to the worker's synchronization struct (on stack). */
2733	PSUPTSCDELTASYNC2 volatile pSyncWorker;
2734	/** The time the worker spent in the MP worker. */
2735	uint64_t cElapsedWorkerTscTicks;
2736	/** Worker data union. */
2737	union
2738	{
2739	/** Data (worker) for delta verification. */
2740	struct
2741	{
2742	/** Verification test TSC values for the worker. */
2743	uint64_t volatile auTscs[32];
2744	} Verify;
2745	/** Data (worker) for measurement method \#2. */
2746	struct
2747	{
2748	/** Data and sequence number. */
2749	SUPDRVTSCDELTAMETHOD2 Data;
2750	/** The lag setting for the next run (set by master). */
2751	bool fLag;
2752	} M2;
2753	} uWorker;
2754	/** @} */
2755
2756	/** Padding to make sure the above is in its own cache line. */
2757	uint64_t au64CacheLinePaddingAfter[GIP_TSC_DELTA_CACHE_LINE_SIZE / sizeof(uint64_t)];
2758	} SUPDRVGIPTSCDELTARGS;
2759	typedef SUPDRVGIPTSCDELTARGS *PSUPDRVGIPTSCDELTARGS;
2760
2761
2762	/** @name Macros that implements the basic synchronization steps common to
2763	* the algorithms.
2764	*
2765	* Must be used from loop as the timeouts are implemented via 'break' statements
2766	* at the moment.
2767	*
2768	* @{
2769	*/
2770	#if defined(DEBUG_bird) /* \|\| defined(VBOX_STRICT) */
2771	# define TSCDELTA_DBG_VARS() uint32_t iDbgCounter
2772	# define TSCDELTA_DBG_START_LOOP() do { iDbgCounter = 0; } while (0)
2773	# define TSCDELTA_DBG_CHECK_LOOP() \
2774	do { iDbgCounter++; if ((iDbgCounter & UINT32_C(0x01ffffff)) == 0) RT_BREAKPOINT(); } while (0)
2775	#else
2776	# define TSCDELTA_DBG_VARS() ((void)0)
2777	# define TSCDELTA_DBG_START_LOOP() ((void)0)
2778	# define TSCDELTA_DBG_CHECK_LOOP() ((void)0)
2779	#endif
2780	#if 0
2781	# define TSCDELTA_DBG_SYNC_MSG(a_Args) SUPR0Printf a_Args
2782	#else
2783	# define TSCDELTA_DBG_SYNC_MSG(a_Args) ((void)0)
2784	#endif
2785	#if 0
2786	# define TSCDELTA_DBG_SYNC_MSG2(a_Args) SUPR0Printf a_Args
2787	#else
2788	# define TSCDELTA_DBG_SYNC_MSG2(a_Args) ((void)0)
2789	#endif
2790	#if 0
2791	# define TSCDELTA_DBG_SYNC_MSG9(a_Args) SUPR0Printf a_Args
2792	#else
2793	# define TSCDELTA_DBG_SYNC_MSG9(a_Args) ((void)0)
2794	#endif
2795
2796
2797	static bool supdrvTscDeltaSync2_Before(PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
2798	bool fIsMaster, PRTCCUINTREG pfEFlags, PSUPDRVGIPTSCDELTARGS pArgs)
2799	{
2800	uint32_t iMySeq = fIsMaster ? 0 : 256;
2801	uint32_t const iMaxSeq = iMySeq + 16; /* For the last loop, darn linux/freebsd C-ishness. */
2802	uint32_t u32Tmp;
2803	uint32_t iSync2Loops = 0;
2804	RTCCUINTREG fEFlags;
2805	TSCDELTA_DBG_VARS();
2806
2807	pfEFlags = X86_EFL_IF \| X86_EFL_1; / should shut up most nagging compilers. */
2808
2809	/*
2810	* The master tells the worker to get on it's mark.
2811	*/
2812	if (fIsMaster)
2813	{
2814	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_STEADY, GIP_TSC_DELTA_SYNC2_READY)))
2815	{ /* likely*/ }
2816	else
2817	{
2818	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #1 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2819	return false;
2820	}
2821	}
2822
2823	/*
2824	* Wait for the on your mark signal (ack in the master case). We process timeouts here.
2825	*/
2826	ASMAtomicWriteU32(&(pMySync)->uSyncSeq, 0);
2827	for (;;)
2828	{
2829	fEFlags = ASMIntDisableFlags();
2830	u32Tmp = ASMAtomicReadU32(&pMySync->uSyncVar);
2831	if (u32Tmp == GIP_TSC_DELTA_SYNC2_STEADY)
2832	break;
2833	ASMSetFlags(fEFlags);
2834	ASMNopPause();
2835
2836	/* Abort? */
2837	if (u32Tmp != GIP_TSC_DELTA_SYNC2_READY)
2838	{
2839	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #2 u32Tmp=%#x\n", fIsMaster ? "master" : "worker", u32Tmp));
2840	return false;
2841	}
2842
2843	/* Check for timeouts every so often (not every loop in case RDTSC is
2844	trapping or something). Must check the first time around. */
2845	#if 0 /* For debugging the timeout paths. */
2846	static uint32_t volatile xxx;
2847	#endif
2848	if ( ( (iSync2Loops & 0x3ff) == 0
2849	&& ASMReadTSC() - pMySync->uTscStart > pMySync->cMaxTscTicks)
2850	#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). */
2851	\|\| (!fIsMaster && (++xxx & 0xf) == 0)
2852	#endif
2853	)
2854	{
2855	/* Try switch our own state into timeout mode so the master cannot tell us to 'GO',
2856	ignore the timeout if we've got the go ahead already (simpler). */
2857	if (ASMAtomicCmpXchgU32(&pMySync->uSyncVar, GIP_TSC_DELTA_SYNC2_TIMEOUT, GIP_TSC_DELTA_SYNC2_READY))
2858	{
2859	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: timeout\n", fIsMaster ? "master" : "worker"));
2860	ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_TIMEOUT, GIP_TSC_DELTA_SYNC2_STEADY);
2861	ASMAtomicWriteBool(&pArgs->fTimedOut, true);
2862	return false;
2863	}
2864	}
2865	iSync2Loops++;
2866	}
2867
2868	/*
2869	* Interrupts are now disabled and will remain disabled until we do
2870	* TSCDELTA_MASTER_SYNC_AFTER / TSCDELTA_OTHER_SYNC_AFTER.
2871	*/
2872	*pfEFlags = fEFlags;
2873
2874	/*
2875	* The worker tells the master that it is on its mark and that the master
2876	* need to get into position as well.
2877	*/
2878	if (!fIsMaster)
2879	{
2880	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_STEADY, GIP_TSC_DELTA_SYNC2_READY)))
2881	{ /* likely */ }
2882	else
2883	{
2884	ASMSetFlags(fEFlags);
2885	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #3 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2886	return false;
2887	}
2888	}
2889
2890	/*
2891	* The master sends the 'go' to the worker and wait for ACK.
2892	*/
2893	if (fIsMaster)
2894	{
2895	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO, GIP_TSC_DELTA_SYNC2_STEADY)))
2896	{ /* likely */ }
2897	else
2898	{
2899	ASMSetFlags(fEFlags);
2900	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #4 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2901	return false;
2902	}
2903	}
2904
2905	/*
2906	* Wait for the 'go' signal (ack in the master case).
2907	*/
2908	TSCDELTA_DBG_START_LOOP();
2909	for (;;)
2910	{
2911	u32Tmp = ASMAtomicReadU32(&pMySync->uSyncVar);
2912	if (u32Tmp == GIP_TSC_DELTA_SYNC2_GO)
2913	break;
2914	if (RT_LIKELY(u32Tmp == GIP_TSC_DELTA_SYNC2_STEADY))
2915	{ /* likely */ }
2916	else
2917	{
2918	ASMSetFlags(fEFlags);
2919	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #5 u32Tmp=%#x\n", fIsMaster ? "master" : "worker", u32Tmp));
2920	return false;
2921	}
2922
2923	TSCDELTA_DBG_CHECK_LOOP();
2924	ASMNopPause();
2925	}
2926
2927	/*
2928	* The worker acks the 'go' (shouldn't fail).
2929	*/
2930	if (!fIsMaster)
2931	{
2932	if (RT_LIKELY(ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO, GIP_TSC_DELTA_SYNC2_STEADY)))
2933	{ /* likely */ }
2934	else
2935	{
2936	ASMSetFlags(fEFlags);
2937	TSCDELTA_DBG_SYNC_MSG(("sync/before/%s: #6 uSyncVar=%#x\n", fIsMaster ? "master" : "worker", pOtherSync->uSyncVar));
2938	return false;
2939	}
2940	}
2941
2942	/*
2943	* Try enter mostly lockstep execution with it.
2944	*/
2945	for (;;)
2946	{
2947	uint32_t iOtherSeq1, iOtherSeq2;
2948	ASMCompilerBarrier();
2949	ASMSerializeInstruction();
2950
2951	ASMAtomicWriteU32(&pMySync->uSyncSeq, iMySeq);
2952	ASMNopPause();
2953	iOtherSeq1 = ASMAtomicXchgU32(&pOtherSync->uSyncSeq, iMySeq);
2954	ASMNopPause();
2955	iOtherSeq2 = ASMAtomicReadU32(&pMySync->uSyncSeq);
2956
2957	ASMCompilerBarrier();
2958	if (iOtherSeq1 == iOtherSeq2)
2959	return true;
2960
2961	/* Did the other guy give up? Should we give up? */
2962	if ( iOtherSeq1 == UINT32_MAX
2963	\|\| iOtherSeq2 == UINT32_MAX)
2964	return true;
2965	if (++iMySeq >= iMaxSeq)
2966	{
2967	ASMAtomicWriteU32(&pMySync->uSyncSeq, UINT32_MAX);
2968	return true;
2969	}
2970	ASMNopPause();
2971	}
2972	}
2973
2974	#define TSCDELTA_MASTER_SYNC_BEFORE(a_pMySync, a_pOtherSync, a_pfEFlags, a_pArgs) \
2975	if (RT_LIKELY(supdrvTscDeltaSync2_Before(a_pMySync, a_pOtherSync, true /fIsMaster/, a_pfEFlags, a_pArgs))) \
2976	{ /likely/ } \
2977	else if (true) \
2978	{ \
2979	TSCDELTA_DBG_SYNC_MSG9(("sync/before/master: #89\n")); \
2980	break; \
2981	} else do {} while (0)
2982	#define TSCDELTA_OTHER_SYNC_BEFORE(a_pMySync, a_pOtherSync, a_pfEFlags, a_pArgs) \
2983	if (RT_LIKELY(supdrvTscDeltaSync2_Before(a_pMySync, a_pOtherSync, false /fIsMaster/, a_pfEFlags, a_pArgs))) \
2984	{ /likely/ } \
2985	else if (true) \
2986	{ \
2987	TSCDELTA_DBG_SYNC_MSG9(("sync/before/other: #89\n")); \
2988	break; \
2989	} else do {} while (0)
2990
2991
2992	static bool supdrvTscDeltaSync2_After(PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
2993	bool fIsMaster, RTCCUINTREG fEFlags)
2994	{
2995	TSCDELTA_DBG_VARS();
2996	RT_NOREF1(pOtherSync);
2997
2998	/*
2999	* Wait for the 'ready' signal. In the master's case, this means the
3000	* worker has completed its data collection, while in the worker's case it
3001	* means the master is done processing the data and it's time for the next
3002	* loop iteration (or whatever).
3003	*/
3004	ASMSetFlags(fEFlags);
3005	TSCDELTA_DBG_START_LOOP();
3006	for (;;)
3007	{
3008	uint32_t u32Tmp = ASMAtomicReadU32(&pMySync->uSyncVar);
3009	if ( u32Tmp == GIP_TSC_DELTA_SYNC2_READY
3010	\|\| (u32Tmp == GIP_TSC_DELTA_SYNC2_STEADY && !fIsMaster) /* kicked twice => race */ )
3011	return true;
3012	ASMNopPause();
3013	if (RT_LIKELY(u32Tmp == GIP_TSC_DELTA_SYNC2_GO))
3014	{ /* likely */}
3015	else
3016	{
3017	TSCDELTA_DBG_SYNC_MSG(("sync/after/other: #1 u32Tmp=%#x\n", u32Tmp));
3018	return false; /* shouldn't ever happen! */
3019	}
3020	TSCDELTA_DBG_CHECK_LOOP();
3021	ASMNopPause();
3022	}
3023	}
3024
3025	#define TSCDELTA_MASTER_SYNC_AFTER(a_pMySync, a_pOtherSync, a_fEFlags) \
3026	if (RT_LIKELY(supdrvTscDeltaSync2_After(a_pMySync, a_pOtherSync, true /fIsMaster/, a_fEFlags))) \
3027	{ /* likely */ } \
3028	else if (true) \
3029	{ \
3030	TSCDELTA_DBG_SYNC_MSG9(("sync/after/master: #97\n")); \
3031	break; \
3032	} else do {} while (0)
3033
3034	#define TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(a_pMySync, a_pOtherSync) \
3035	/* \
3036	* Tell the worker that we're done processing the data and ready for the next round. \
3037	*/ \
3038	if (RT_LIKELY(ASMAtomicCmpXchgU32(&(a_pOtherSync)->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_GO))) \
3039	{ /* likely */ } \
3040	else if (true)\
3041	{ \
3042	TSCDELTA_DBG_SYNC_MSG(("sync/after/master: #99 uSyncVar=%#x\n", (a_pOtherSync)->uSyncVar)); \
3043	break; \
3044	} else do {} while (0)
3045
3046	#define TSCDELTA_OTHER_SYNC_AFTER(a_pMySync, a_pOtherSync, a_fEFlags) \
3047	if (true) { \
3048	/* \
3049	* Tell the master that we're done collecting data and wait for the next round to start. \
3050	*/ \
3051	if (RT_LIKELY(ASMAtomicCmpXchgU32(&(a_pOtherSync)->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_GO))) \
3052	{ /* likely */ } \
3053	else \
3054	{ \
3055	ASMSetFlags(a_fEFlags); \
3056	TSCDELTA_DBG_SYNC_MSG(("sync/after/other: #0 uSyncVar=%#x\n", (a_pOtherSync)->uSyncVar)); \
3057	break; \
3058	} \
3059	if (RT_LIKELY(supdrvTscDeltaSync2_After(a_pMySync, a_pOtherSync, false /fIsMaster/, a_fEFlags))) \
3060	{ /* likely */ } \
3061	else \
3062	{ \
3063	TSCDELTA_DBG_SYNC_MSG9(("sync/after/other: #98\n")); \
3064	break; \
3065	} \
3066	} else do {} while (0)
3067	/** @} */
3068
3069
3070	#ifdef GIP_TSC_DELTA_METHOD_1
3071	/**
3072	* TSC delta measurement algorithm \#1 (GIP_TSC_DELTA_METHOD_1).
3073	*
3074	*
3075	* We ignore the first few runs of the loop in order to prime the
3076	* cache. Also, we need to be careful about using 'pause' instruction
3077	* in critical busy-wait loops in this code - it can cause undesired
3078	* behaviour with hyperthreading.
3079	*
3080	* We try to minimize the measurement error by computing the minimum
3081	* read time of the compare statement in the worker by taking TSC
3082	* measurements across it.
3083	*
3084	* It must be noted that the computed minimum read time is mostly to
3085	* eliminate huge deltas when the worker is too early and doesn't by
3086	* itself help produce more accurate deltas. We allow two times the
3087	* computed minimum as an arbitrary acceptable threshold. Therefore,
3088	* it is still possible to get negative deltas where there are none
3089	* when the worker is earlier. As long as these occasional negative
3090	* deltas are lower than the time it takes to exit guest-context and
3091	* the OS to reschedule EMT on a different CPU, we won't expose a TSC
3092	* that jumped backwards. It is due to the existence of the negative
3093	* deltas that we don't recompute the delta with the master and
3094	* worker interchanged to eliminate the remaining measurement error.
3095	*
3096	*
3097	* @param pArgs The argument/state data.
3098	* @param pMySync My synchronization structure.
3099	* @param pOtherSync My partner's synchronization structure.
3100	* @param fIsMaster Set if master, clear if worker.
3101	* @param iTry The attempt number.
3102	*/
3103	static void supdrvTscDeltaMethod1Loop(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
3104	bool fIsMaster, uint32_t iTry)
3105	{
3106	PSUPGIPCPU pGipCpuWorker = pArgs->pWorker;
3107	PSUPGIPCPU pGipCpuMaster = pArgs->pMaster;
3108	uint64_t uMinCmpReadTime = UINT64_MAX;
3109	unsigned iLoop;
3110	NOREF(iTry);
3111
3112	for (iLoop = 0; iLoop < GIP_TSC_DELTA_LOOPS; iLoop++)
3113	{
3114	RTCCUINTREG fEFlags;
3115	if (fIsMaster)
3116	{
3117	/*
3118	* The master.
3119	*/
3120	AssertMsg(pGipCpuMaster->u64TSCSample == GIP_TSC_DELTA_RSVD,
3121	("%#llx idMaster=%#x idWorker=%#x (idGipMaster=%#x)\n",
3122	pGipCpuMaster->u64TSCSample, pGipCpuMaster->idCpu, pGipCpuWorker->idCpu, pArgs->pDevExt->idGipMaster));
3123	TSCDELTA_MASTER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3124
3125	do
3126	{
3127	ASMSerializeInstruction();
3128	ASMAtomicWriteU64(&pGipCpuMaster->u64TSCSample, ASMReadTSC());
3129	} while (pGipCpuMaster->u64TSCSample == GIP_TSC_DELTA_RSVD);
3130
3131	TSCDELTA_MASTER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3132
3133	/* Process the data. */
3134	if (iLoop > GIP_TSC_DELTA_PRIMER_LOOPS + GIP_TSC_DELTA_READ_TIME_LOOPS)
3135	{
3136	if (pGipCpuWorker->u64TSCSample != GIP_TSC_DELTA_RSVD)
3137	{
3138	int64_t iDelta = pGipCpuWorker->u64TSCSample
3139	- (pGipCpuMaster->u64TSCSample - pGipCpuMaster->i64TSCDelta);
3140	if ( iDelta >= GIP_TSC_DELTA_INITIAL_MASTER_VALUE
3141	? iDelta < pGipCpuWorker->i64TSCDelta
3142	: iDelta > pGipCpuWorker->i64TSCDelta \|\| pGipCpuWorker->i64TSCDelta == INT64_MAX)
3143	pGipCpuWorker->i64TSCDelta = iDelta;
3144	}
3145	}
3146
3147	/* Reset our TSC sample and tell the worker to move on. */
3148	ASMAtomicWriteU64(&pGipCpuMaster->u64TSCSample, GIP_TSC_DELTA_RSVD);
3149	TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(pMySync, pOtherSync);
3150	}
3151	else
3152	{
3153	/*
3154	* The worker.
3155	*/
3156	uint64_t uTscWorker;
3157	uint64_t uTscWorkerFlushed;
3158	uint64_t uCmpReadTime;
3159
3160	ASMAtomicReadU64(&pGipCpuMaster->u64TSCSample); /* Warm the cache line. */
3161	TSCDELTA_OTHER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3162
3163	/*
3164	* Keep reading the TSC until we notice that the master has read his. Reading
3165	* the TSC -after- the master has updated the memory is way too late. We thus
3166	* compensate by trying to measure how long it took for the worker to notice
3167	* the memory flushed from the master.
3168	*/
3169	do
3170	{
3171	ASMSerializeInstruction();
3172	uTscWorker = ASMReadTSC();
3173	} while (pGipCpuMaster->u64TSCSample == GIP_TSC_DELTA_RSVD);
3174	ASMSerializeInstruction();
3175	uTscWorkerFlushed = ASMReadTSC();
3176
3177	uCmpReadTime = uTscWorkerFlushed - uTscWorker;
3178	if (iLoop > GIP_TSC_DELTA_PRIMER_LOOPS + GIP_TSC_DELTA_READ_TIME_LOOPS)
3179	{
3180	/* This is totally arbitrary a.k.a I don't like it but I have no better ideas for now. */
3181	if (uCmpReadTime < (uMinCmpReadTime << 1))
3182	{
3183	ASMAtomicWriteU64(&pGipCpuWorker->u64TSCSample, uTscWorker);
3184	if (uCmpReadTime < uMinCmpReadTime)
3185	uMinCmpReadTime = uCmpReadTime;
3186	}
3187	else
3188	ASMAtomicWriteU64(&pGipCpuWorker->u64TSCSample, GIP_TSC_DELTA_RSVD);
3189	}
3190	else if (iLoop > GIP_TSC_DELTA_PRIMER_LOOPS)
3191	{
3192	if (uCmpReadTime < uMinCmpReadTime)
3193	uMinCmpReadTime = uCmpReadTime;
3194	}
3195
3196	TSCDELTA_OTHER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3197	}
3198	}
3199
3200	TSCDELTA_DBG_SYNC_MSG9(("sync/method1loop/%s: #92 iLoop=%u MyState=%#x\n", fIsMaster ? "master" : "worker", iLoop,
3201	pMySync->uSyncVar));
3202
3203	/*
3204	* We must reset the worker TSC sample value in case it gets picked as a
3205	* GIP master later on (it's trashed above, naturally).
3206	*/
3207	if (!fIsMaster)
3208	ASMAtomicWriteU64(&pGipCpuWorker->u64TSCSample, GIP_TSC_DELTA_RSVD);
3209	}
3210	#endif /* GIP_TSC_DELTA_METHOD_1 */
3211
3212
3213	#ifdef GIP_TSC_DELTA_METHOD_2
3214	/*
3215	* TSC delta measurement algorithm \#2 configuration and code - Experimental!!
3216	*/
3217
3218	# define GIP_TSC_DELTA_M2_LOOPS (7 + GIP_TSC_DELTA_M2_PRIMER_LOOPS)
3219	# define GIP_TSC_DELTA_M2_PRIMER_LOOPS 0
3220
3221
3222	static void supdrvTscDeltaMethod2ProcessDataOnMaster(PSUPDRVGIPTSCDELTARGS pArgs)
3223	{
3224	int64_t iMasterTscDelta = pArgs->pMaster->i64TSCDelta;
3225	int64_t iBestDelta = pArgs->pWorker->i64TSCDelta;
3226	uint32_t idxResult;
3227	uint32_t cHits = 0;
3228
3229	/*
3230	* Look for matching entries in the master and worker tables.
3231	*/
3232	for (idxResult = 0; idxResult < RT_ELEMENTS(pArgs->uMaster.M2.Data.aResults); idxResult++)
3233	{
3234	uint32_t idxOther = pArgs->uMaster.M2.Data.aResults[idxResult].iSeqOther;
3235	if (idxOther & 1)
3236	{
3237	idxOther >>= 1;
3238	if (idxOther < RT_ELEMENTS(pArgs->uWorker.M2.Data.aResults))
3239	{
3240	if (pArgs->uWorker.M2.Data.aResults[idxOther].iSeqOther == pArgs->uMaster.M2.Data.aResults[idxResult].iSeqMine)
3241	{
3242	int64_t iDelta;
3243	iDelta = pArgs->uWorker.M2.Data.aResults[idxOther].uTsc
3244	- (pArgs->uMaster.M2.Data.aResults[idxResult].uTsc - iMasterTscDelta);
3245	if ( iDelta >= GIP_TSC_DELTA_INITIAL_MASTER_VALUE
3246	? iDelta < iBestDelta
3247	: iDelta > iBestDelta \|\| iBestDelta == INT64_MAX)
3248	iBestDelta = iDelta;
3249	cHits++;
3250	}
3251	}
3252	}
3253	}
3254
3255	/*
3256	* Save the results.
3257	*/
3258	if (cHits > 2)
3259	pArgs->pWorker->i64TSCDelta = iBestDelta;
3260	pArgs->uMaster.M2.cHits += cHits;
3261	}
3262
3263
3264	/**
3265	* The core function of the 2nd TSC delta measurement algorithm.
3266	*
3267	* The idea here is that we have the two CPUs execute the exact same code
3268	* collecting a largish set of TSC samples. The code has one data dependency on
3269	* the other CPU which intention it is to synchronize the execution as well as
3270	* help cross references the two sets of TSC samples (the sequence numbers).
3271	*
3272	* The @a fLag parameter is used to modify the execution a tiny bit on one or
3273	* both of the CPUs. When @a fLag differs between the CPUs, it is thought that
3274	* it will help with making the CPUs enter lock step execution occasionally.
3275	*
3276	*/
3277	static void supdrvTscDeltaMethod2CollectData(PSUPDRVTSCDELTAMETHOD2 pMyData, uint32_t volatile *piOtherSeqNo, bool fLag)
3278	{
3279	SUPDRVTSCDELTAMETHOD2ENTRY *pEntry = &pMyData->aResults[0];
3280	uint32_t cLeft = RT_ELEMENTS(pMyData->aResults);
3281
3282	ASMAtomicWriteU32(&pMyData->iCurSeqNo, 0);
3283	ASMSerializeInstruction();
3284	while (cLeft-- > 0)
3285	{
3286	uint64_t uTsc;
3287	uint32_t iSeqMine = ASMAtomicIncU32(&pMyData->iCurSeqNo);
3288	uint32_t iSeqOther = ASMAtomicReadU32(piOtherSeqNo);
3289	ASMCompilerBarrier();
3290	ASMSerializeInstruction(); /* Way better result than with ASMMemoryFenceSSE2() in this position! */
3291	uTsc = ASMReadTSC();
3292	ASMAtomicIncU32(&pMyData->iCurSeqNo);
3293	ASMCompilerBarrier();
3294	ASMSerializeInstruction();
3295	pEntry->iSeqMine = iSeqMine;
3296	pEntry->iSeqOther = iSeqOther;
3297	pEntry->uTsc = uTsc;
3298	pEntry++;
3299	ASMSerializeInstruction();
3300	if (fLag)
3301	ASMNopPause();
3302	}
3303	}
3304
3305
3306	/**
3307	* TSC delta measurement algorithm \#2 (GIP_TSC_DELTA_METHOD_2).
3308	*
3309	* See supdrvTscDeltaMethod2CollectData for algorithm details.
3310	*
3311	* @param pArgs The argument/state data.
3312	* @param pMySync My synchronization structure.
3313	* @param pOtherSync My partner's synchronization structure.
3314	* @param fIsMaster Set if master, clear if worker.
3315	* @param iTry The attempt number.
3316	*/
3317	static void supdrvTscDeltaMethod2Loop(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync, PSUPTSCDELTASYNC2 pOtherSync,
3318	bool fIsMaster, uint32_t iTry)
3319	{
3320	unsigned iLoop;
3321	RT_NOREF1(iTry);
3322
3323	for (iLoop = 0; iLoop < GIP_TSC_DELTA_M2_LOOPS; iLoop++)
3324	{
3325	RTCCUINTREG fEFlags;
3326	if (fIsMaster)
3327	{
3328	/*
3329	* Adjust the loop lag fudge.
3330	*/
3331	# if GIP_TSC_DELTA_M2_PRIMER_LOOPS > 0
3332	if (iLoop < GIP_TSC_DELTA_M2_PRIMER_LOOPS)
3333	{
3334	/* Lag during the priming to be nice to everyone.. */
3335	pArgs->uMaster.M2.fLag = true;
3336	pArgs->uWorker.M2.fLag = true;
3337	}
3338	else
3339	# endif
3340	if (iLoop < (GIP_TSC_DELTA_M2_LOOPS - GIP_TSC_DELTA_M2_PRIMER_LOOPS) / 4)
3341	{
3342	/* 25 % of the body without lagging. */
3343	pArgs->uMaster.M2.fLag = false;
3344	pArgs->uWorker.M2.fLag = false;
3345	}
3346	else if (iLoop < (GIP_TSC_DELTA_M2_LOOPS - GIP_TSC_DELTA_M2_PRIMER_LOOPS) / 4 * 2)
3347	{
3348	/* 25 % of the body with both lagging. */
3349	pArgs->uMaster.M2.fLag = true;
3350	pArgs->uWorker.M2.fLag = true;
3351	}
3352	else
3353	{
3354	/* 50% of the body with alternating lag. */
3355	pArgs->uMaster.M2.fLag = (iLoop & 1) == 0;
3356	pArgs->uWorker.M2.fLag= (iLoop & 1) == 1;
3357	}
3358
3359	/*
3360	* Sync up with the worker and collect data.
3361	*/
3362	TSCDELTA_MASTER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3363	supdrvTscDeltaMethod2CollectData(&pArgs->uMaster.M2.Data, &pArgs->uWorker.M2.Data.iCurSeqNo, pArgs->uMaster.M2.fLag);
3364	TSCDELTA_MASTER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3365
3366	/*
3367	* Process the data.
3368	*/
3369	# if GIP_TSC_DELTA_M2_PRIMER_LOOPS > 0
3370	if (iLoop >= GIP_TSC_DELTA_M2_PRIMER_LOOPS)
3371	# endif
3372	supdrvTscDeltaMethod2ProcessDataOnMaster(pArgs);
3373
3374	TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(pMySync, pOtherSync);
3375	}
3376	else
3377	{
3378	/*
3379	* The worker.
3380	*/
3381	TSCDELTA_OTHER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3382	supdrvTscDeltaMethod2CollectData(&pArgs->uWorker.M2.Data, &pArgs->uMaster.M2.Data.iCurSeqNo, pArgs->uWorker.M2.fLag);
3383	TSCDELTA_OTHER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3384	}
3385	}
3386	}
3387
3388	#endif /* GIP_TSC_DELTA_METHOD_2 */
3389
3390
3391
3392	static int supdrvTscDeltaVerify(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync,
3393	PSUPTSCDELTASYNC2 pOtherSync, bool fIsMaster, int64_t iWorkerTscDelta)
3394	{
3395	/PSUPGIPCPU pGipCpuWorker = pArgs->pWorker; - unused /
3396	PSUPGIPCPU pGipCpuMaster = pArgs->pMaster;
3397	uint32_t i;
3398	TSCDELTA_DBG_VARS();
3399
3400	for (;;)
3401	{
3402	RTCCUINTREG fEFlags;
3403	AssertCompile((RT_ELEMENTS(pArgs->uMaster.Verify.auTscs) & 1) == 0);
3404	AssertCompile(RT_ELEMENTS(pArgs->uMaster.Verify.auTscs) == RT_ELEMENTS(pArgs->uWorker.Verify.auTscs));
3405
3406	if (fIsMaster)
3407	{
3408	uint64_t uTscWorker;
3409	TSCDELTA_MASTER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3410
3411	/*
3412	* Collect TSC, master goes first.
3413	*/
3414	for (i = 0; i < RT_ELEMENTS(pArgs->uMaster.Verify.auTscs); i += 2)
3415	{
3416	/* Read, kick & wait #1. */
3417	uint64_t register uTsc = ASMReadTSC();
3418	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO_GO);
3419	ASMSerializeInstruction();
3420	pArgs->uMaster.Verify.auTscs[i] = uTsc;
3421	TSCDELTA_DBG_START_LOOP();
3422	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO)
3423	{
3424	TSCDELTA_DBG_CHECK_LOOP();
3425	ASMNopPause();
3426	}
3427
3428	/* Read, kick & wait #2. */
3429	uTsc = ASMReadTSC();
3430	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO);
3431	ASMSerializeInstruction();
3432	pArgs->uMaster.Verify.auTscs[i + 1] = uTsc;
3433	TSCDELTA_DBG_START_LOOP();
3434	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO_GO)
3435	{
3436	TSCDELTA_DBG_CHECK_LOOP();
3437	ASMNopPause();
3438	}
3439	}
3440
3441	TSCDELTA_MASTER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3442
3443	/*
3444	* Process the data.
3445	*/
3446	#ifdef TSCDELTA_VERIFY_WITH_STATS
3447	pArgs->cMaxVerifyTscTicks = INT64_MIN;
3448	pArgs->cMinVerifyTscTicks = INT64_MAX;
3449	pArgs->iVerifyBadTscDiff = 0;
3450	#endif
3451	ASMAtomicWriteS32(&pArgs->rcVerify, VINF_SUCCESS);
3452	uTscWorker = 0;
3453	for (i = 0; i < RT_ELEMENTS(pArgs->uMaster.Verify.auTscs); i++)
3454	{
3455	/* Master vs previous worker entry. */
3456	uint64_t uTscMaster = pArgs->uMaster.Verify.auTscs[i] - pGipCpuMaster->i64TSCDelta;
3457	int64_t iDiff;
3458	if (i > 0)
3459	{
3460	iDiff = uTscMaster - uTscWorker;
3461	#ifdef TSCDELTA_VERIFY_WITH_STATS
3462	if (iDiff > pArgs->cMaxVerifyTscTicks)
3463	pArgs->cMaxVerifyTscTicks = iDiff;
3464	if (iDiff < pArgs->cMinVerifyTscTicks)
3465	pArgs->cMinVerifyTscTicks = iDiff;
3466	#endif
3467	if (iDiff < 0)
3468	{
3469	#ifdef TSCDELTA_VERIFY_WITH_STATS
3470	pArgs->iVerifyBadTscDiff = -iDiff;
3471	#endif
3472	ASMAtomicWriteS32(&pArgs->rcVerify, VERR_OUT_OF_RANGE);
3473	break;
3474	}
3475	}
3476
3477	/* Worker vs master. */
3478	uTscWorker = pArgs->uWorker.Verify.auTscs[i] - iWorkerTscDelta;
3479	iDiff = uTscWorker - uTscMaster;
3480	#ifdef TSCDELTA_VERIFY_WITH_STATS
3481	if (iDiff > pArgs->cMaxVerifyTscTicks)
3482	pArgs->cMaxVerifyTscTicks = iDiff;
3483	if (iDiff < pArgs->cMinVerifyTscTicks)
3484	pArgs->cMinVerifyTscTicks = iDiff;
3485	#endif
3486	if (iDiff < 0)
3487	{
3488	#ifdef TSCDELTA_VERIFY_WITH_STATS
3489	pArgs->iVerifyBadTscDiff = iDiff;
3490	#endif
3491	ASMAtomicWriteS32(&pArgs->rcVerify, VERR_OUT_OF_RANGE);
3492	break;
3493	}
3494	}
3495
3496	/* Done. */
3497	TSCDELTA_MASTER_KICK_OTHER_OUT_OF_AFTER(pMySync, pOtherSync);
3498	}
3499	else
3500	{
3501	/*
3502	* The worker, master leads.
3503	*/
3504	TSCDELTA_OTHER_SYNC_BEFORE(pMySync, pOtherSync, &fEFlags, pArgs);
3505
3506	for (i = 0; i < RT_ELEMENTS(pArgs->uWorker.Verify.auTscs); i += 2)
3507	{
3508	uint64_t register uTsc;
3509
3510	/* Wait, Read and Kick #1. */
3511	TSCDELTA_DBG_START_LOOP();
3512	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO)
3513	{
3514	TSCDELTA_DBG_CHECK_LOOP();
3515	ASMNopPause();
3516	}
3517	uTsc = ASMReadTSC();
3518	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO_GO);
3519	ASMSerializeInstruction();
3520	pArgs->uWorker.Verify.auTscs[i] = uTsc;
3521
3522	/* Wait, Read and Kick #2. */
3523	TSCDELTA_DBG_START_LOOP();
3524	while (ASMAtomicReadU32(&pMySync->uSyncVar) == GIP_TSC_DELTA_SYNC2_GO_GO)
3525	{
3526	TSCDELTA_DBG_CHECK_LOOP();
3527	ASMNopPause();
3528	}
3529	uTsc = ASMReadTSC();
3530	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_GO);
3531	ASMSerializeInstruction();
3532	pArgs->uWorker.Verify.auTscs[i + 1] = uTsc;
3533	}
3534
3535	TSCDELTA_OTHER_SYNC_AFTER(pMySync, pOtherSync, fEFlags);
3536	}
3537	return pArgs->rcVerify;
3538	}
3539
3540	/*
3541	* Timed out, please retry.
3542	*/
3543	ASMAtomicWriteS32(&pArgs->rcVerify, VERR_TRY_AGAIN);
3544	return VERR_TIMEOUT;
3545	}
3546
3547
3548
3549	/**
3550	* Handles the special abort procedure during synchronization setup in
3551	* supdrvTscMeasureDeltaCallbackUnwrapped().
3552	*
3553	* @returns 0 (dummy, ignored)
3554	* @param pArgs Pointer to argument/state data.
3555	* @param pMySync Pointer to my sync structure.
3556	* @param fIsMaster Set if we're the master, clear if worker.
3557	* @param fTimeout Set if it's a timeout.
3558	*/
3559	DECL_NO_INLINE(static, int)
3560	supdrvTscMeasureDeltaCallbackAbortSyncSetup(PSUPDRVGIPTSCDELTARGS pArgs, PSUPTSCDELTASYNC2 pMySync, bool fIsMaster, bool fTimeout)
3561	{
3562	PSUPTSCDELTASYNC2 volatile *ppMySync = fIsMaster ? &pArgs->pSyncMaster : &pArgs->pSyncWorker;
3563	PSUPTSCDELTASYNC2 volatile *ppOtherSync = fIsMaster ? &pArgs->pSyncWorker : &pArgs->pSyncMaster;
3564	TSCDELTA_DBG_VARS();
3565	RT_NOREF1(pMySync);
3566
3567	/*
3568	* Clear our sync pointer and make sure the abort flag is set.
3569	*/
3570	ASMAtomicWriteNullPtr(ppMySync);
3571	ASMAtomicWriteBool(&pArgs->fAbortSetup, true);
3572	if (fTimeout)
3573	ASMAtomicWriteBool(&pArgs->fTimedOut, true);
3574
3575	/*
3576	* Make sure the other party is out of there and won't be touching our
3577	* sync state again (would cause stack corruption).
3578	*/
3579	TSCDELTA_DBG_START_LOOP();
3580	while (ASMAtomicReadPtrT(ppOtherSync, PSUPTSCDELTASYNC2) != NULL)
3581	{
3582	ASMNopPause();
3583	ASMNopPause();
3584	ASMNopPause();
3585	TSCDELTA_DBG_CHECK_LOOP();
3586	}
3587
3588	return 0;
3589	}
3590
3591
3592	/**
3593	* This is used by supdrvTscMeasureInitialDeltas() to read the TSC on two CPUs
3594	* and compute the delta between them.
3595	*
3596	* To reduce code size a good when timeout handling was added, a dummy return
3597	* value had to be added (saves 1-3 lines per timeout case), thus this
3598	* 'Unwrapped' function and the dummy 0 return value.
3599	*
3600	* @returns 0 (dummy, ignored)
3601	* @param idCpu The CPU we are current scheduled on.
3602	* @param pArgs Pointer to a parameter package.
3603	*
3604	* @remarks Measuring TSC deltas between the CPUs is tricky because we need to
3605	* read the TSC at exactly the same time on both the master and the
3606	* worker CPUs. Due to DMA, bus arbitration, cache locality,
3607	* contention, SMI, pipelining etc. there is no guaranteed way of
3608	* doing this on x86 CPUs.
3609	*/
3610	static int supdrvTscMeasureDeltaCallbackUnwrapped(RTCPUID idCpu, PSUPDRVGIPTSCDELTARGS pArgs)
3611	{
3612	PSUPDRVDEVEXT pDevExt = pArgs->pDevExt;
3613	PSUPGIPCPU pGipCpuWorker = pArgs->pWorker;
3614	PSUPGIPCPU pGipCpuMaster = pArgs->pMaster;
3615	bool const fIsMaster = idCpu == pGipCpuMaster->idCpu;
3616	uint32_t iTry;
3617	PSUPTSCDELTASYNC2 volatile *ppMySync = fIsMaster ? &pArgs->pSyncMaster : &pArgs->pSyncWorker;
3618	PSUPTSCDELTASYNC2 volatile *ppOtherSync = fIsMaster ? &pArgs->pSyncWorker : &pArgs->pSyncMaster;
3619	SUPTSCDELTASYNC2 MySync;
3620	PSUPTSCDELTASYNC2 pOtherSync;
3621	int rc;
3622	TSCDELTA_DBG_VARS();
3623
3624	/* A bit of paranoia first. */
3625	if (!pGipCpuMaster \|\| !pGipCpuWorker)
3626	return 0;
3627
3628	/*
3629	* If the CPU isn't part of the measurement, return immediately.
3630	*/
3631	if ( !fIsMaster
3632	&& idCpu != pGipCpuWorker->idCpu)
3633	return 0;
3634
3635	/*
3636	* Set up my synchronization stuff and wait for the other party to show up.
3637	*
3638	* We don't wait forever since the other party may be off fishing (offline,
3639	* spinning with ints disables, whatever), we must play nice to the rest of
3640	* the system as this context generally isn't one in which we will get
3641	* preempted and we may hold up a number of lower priority interrupts.
3642	*/
3643	ASMAtomicWriteU32(&MySync.uSyncVar, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT);
3644	ASMAtomicWritePtr(ppMySync, &MySync);
3645	MySync.uTscStart = ASMReadTSC();
3646	MySync.cMaxTscTicks = pArgs->cMaxTscTicks;
3647
3648	/* Look for the partner, might not be here yet... Special abort considerations. */
3649	iTry = 0;
3650	TSCDELTA_DBG_START_LOOP();
3651	while ((pOtherSync = ASMAtomicReadPtrT(ppOtherSync, PSUPTSCDELTASYNC2)) == NULL)
3652	{
3653	ASMNopPause();
3654	if ( ASMAtomicReadBool(&pArgs->fAbortSetup)
3655	\|\| !RTMpIsCpuOnline(fIsMaster ? pGipCpuWorker->idCpu : pGipCpuMaster->idCpu) )
3656	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3657	if ( (iTry++ & 0xff) == 0
3658	&& ASMReadTSC() - MySync.uTscStart > pArgs->cMaxTscTicks)
3659	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, true /fTimeout/);
3660	TSCDELTA_DBG_CHECK_LOOP();
3661	ASMNopPause();
3662	}
3663
3664	/* I found my partner, waiting to be found... Special abort considerations. */
3665	if (fIsMaster)
3666	if (!ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT)) /* parnaoia */
3667	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3668
3669	iTry = 0;
3670	TSCDELTA_DBG_START_LOOP();
3671	while (ASMAtomicReadU32(&MySync.uSyncVar) == GIP_TSC_DELTA_SYNC2_PRESTART_WAIT)
3672	{
3673	ASMNopPause();
3674	if (ASMAtomicReadBool(&pArgs->fAbortSetup))
3675	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3676	if ( (iTry++ & 0xff) == 0
3677	&& ASMReadTSC() - MySync.uTscStart > pArgs->cMaxTscTicks)
3678	{
3679	if ( fIsMaster
3680	&& !ASMAtomicCmpXchgU32(&MySync.uSyncVar, GIP_TSC_DELTA_SYNC2_PRESTART_ABORT, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT))
3681	break; /* race #1: slave has moved on, handle timeout in loop instead. */
3682	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, true /fTimeout/);
3683	}
3684	TSCDELTA_DBG_CHECK_LOOP();
3685	}
3686
3687	if (!fIsMaster)
3688	if (!ASMAtomicCmpXchgU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_READY, GIP_TSC_DELTA_SYNC2_PRESTART_WAIT)) /* race #1 */
3689	return supdrvTscMeasureDeltaCallbackAbortSyncSetup(pArgs, &MySync, fIsMaster, false /fTimeout/);
3690
3691	/** @todo Add a resumable state to pArgs so we don't waste time if we time
3692	* out or something. Timeouts are legit, any of the two CPUs may get
3693	* interrupted. */
3694
3695	/*
3696	* Start by seeing if we have a zero delta between the two CPUs.
3697	* This should normally be the case.
3698	*/
3699	rc = supdrvTscDeltaVerify(pArgs, &MySync, pOtherSync, fIsMaster, GIP_TSC_DELTA_INITIAL_MASTER_VALUE);
3700	if (RT_SUCCESS(rc))
3701	{
3702	if (fIsMaster)
3703	{
3704	ASMAtomicWriteS64(&pGipCpuWorker->i64TSCDelta, GIP_TSC_DELTA_INITIAL_MASTER_VALUE);
3705	RTCpuSetDelByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet);
3706	RTCpuSetAddByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpuWorker->iCpuSet);
3707	}
3708	}
3709	/*
3710	* If the verification didn't time out, do regular delta measurements.
3711	* We retry this until we get a reasonable value.
3712	*/
3713	else if (rc != VERR_TIMEOUT)
3714	{
3715	Assert(pGipCpuWorker->i64TSCDelta == INT64_MAX);
3716	for (iTry = 0; iTry < 12; iTry++)
3717	{
3718	/*
3719	* Check the state before we start.
3720	*/
3721	uint32_t u32Tmp = ASMAtomicReadU32(&MySync.uSyncVar);
3722	if ( u32Tmp != GIP_TSC_DELTA_SYNC2_READY
3723	&& (fIsMaster \|\| u32Tmp != GIP_TSC_DELTA_SYNC2_STEADY) /* worker may be late prepping for the next round */ )
3724	{
3725	TSCDELTA_DBG_SYNC_MSG(("sync/loop/%s: #0 iTry=%u MyState=%#x\n", fIsMaster ? "master" : "worker", iTry, u32Tmp));
3726	break;
3727	}
3728
3729	/*
3730	* Do the measurements.
3731	*/
3732	#ifdef GIP_TSC_DELTA_METHOD_1
3733	supdrvTscDeltaMethod1Loop(pArgs, &MySync, pOtherSync, fIsMaster, iTry);
3734	#elif defined(GIP_TSC_DELTA_METHOD_2)
3735	supdrvTscDeltaMethod2Loop(pArgs, &MySync, pOtherSync, fIsMaster, iTry);
3736	#else
3737	# error "huh??"
3738	#endif
3739
3740	/*
3741	* Check the state.
3742	*/
3743	u32Tmp = ASMAtomicReadU32(&MySync.uSyncVar);
3744	if ( u32Tmp != GIP_TSC_DELTA_SYNC2_READY
3745	&& (fIsMaster \|\| u32Tmp != GIP_TSC_DELTA_SYNC2_STEADY) /* worker may be late prepping for the next round */ )
3746	{
3747	if (fIsMaster)
3748	TSCDELTA_DBG_SYNC_MSG(("sync/loop/master: #1 iTry=%u MyState=%#x\n", iTry, u32Tmp));
3749	else
3750	TSCDELTA_DBG_SYNC_MSG2(("sync/loop/worker: #1 iTry=%u MyState=%#x\n", iTry, u32Tmp));
3751	break;
3752	}
3753
3754	/*
3755	* Success? If so, stop trying. Master decides.
3756	*/
3757	if (fIsMaster)
3758	{
3759	if (pGipCpuWorker->i64TSCDelta != INT64_MAX)
3760	{
3761	RTCpuSetDelByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet);
3762	RTCpuSetAddByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpuWorker->iCpuSet);
3763	TSCDELTA_DBG_SYNC_MSG2(("sync/loop/master: #9 iTry=%u MyState=%#x\n", iTry, MySync.uSyncVar));
3764	break;
3765	}
3766	}
3767	}
3768	if (fIsMaster)
3769	pArgs->iTry = iTry;
3770	}
3771
3772	/*
3773	* End the synchronization dance. We tell the other that we're done,
3774	* then wait for the same kind of reply.
3775	*/
3776	ASMAtomicWriteU32(&pOtherSync->uSyncVar, GIP_TSC_DELTA_SYNC2_FINAL);
3777	ASMAtomicWriteNullPtr(ppMySync);
3778	iTry = 0;
3779	TSCDELTA_DBG_START_LOOP();
3780	while (ASMAtomicReadU32(&MySync.uSyncVar) != GIP_TSC_DELTA_SYNC2_FINAL)
3781	{
3782	iTry++;
3783	if ( iTry == 0
3784	&& !RTMpIsCpuOnline(fIsMaster ? pGipCpuWorker->idCpu : pGipCpuMaster->idCpu))
3785	break; /* this really shouldn't happen. */
3786	TSCDELTA_DBG_CHECK_LOOP();
3787	ASMNopPause();
3788	}
3789
3790	/*
3791	* Collect some runtime stats.
3792	*/
3793	if (fIsMaster)
3794	pArgs->cElapsedMasterTscTicks = ASMReadTSC() - MySync.uTscStart;
3795	else
3796	pArgs->cElapsedWorkerTscTicks = ASMReadTSC() - MySync.uTscStart;
3797	return 0;
3798	}
3799
3800	/**
3801	* Callback used by supdrvTscMeasureInitialDeltas() to read the TSC on two CPUs
3802	* and compute the delta between them.
3803	*
3804	* @param idCpu The CPU we are current scheduled on.
3805	* @param pvUser1 Pointer to a parameter package (SUPDRVGIPTSCDELTARGS).
3806	* @param pvUser2 Unused.
3807	*/
3808	static DECLCALLBACK(void) supdrvTscMeasureDeltaCallback(RTCPUID idCpu, void pvUser1, void pvUser2)
3809	{
3810	supdrvTscMeasureDeltaCallbackUnwrapped(idCpu, (PSUPDRVGIPTSCDELTARGS)pvUser1);
3811	RT_NOREF1(pvUser2);
3812	}
3813
3814
3815	/**
3816	* Measures the TSC delta between the master GIP CPU and one specified worker
3817	* CPU.
3818	*
3819	* @returns VBox status code.
3820	* @retval VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED on pure measurement
3821	* failure.
3822	* @param pDevExt Pointer to the device instance data.
3823	* @param idxWorker The index of the worker CPU from the GIP's array of
3824	* CPUs.
3825	*
3826	* @remarks This must be called with preemption enabled!
3827	*/
3828	static int supdrvTscMeasureDeltaOne(PSUPDRVDEVEXT pDevExt, uint32_t idxWorker)
3829	{
3830	int rc;
3831	int rc2;
3832	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
3833	RTCPUID idMaster = pDevExt->idGipMaster;
3834	PSUPGIPCPU pGipCpuWorker = &pGip->aCPUs[idxWorker];
3835	PSUPGIPCPU pGipCpuMaster;
3836	uint32_t iGipCpuMaster;
3837	uint32_t u32Tmp;
3838
3839	/* Validate input a bit. */
3840	AssertReturn(pGip, VERR_INVALID_PARAMETER);
3841	Assert(pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED);
3842	Assert(RTThreadPreemptIsEnabled(NIL_RTTHREAD));
3843
3844	/*
3845	* Don't attempt measuring the delta for the GIP master.
3846	*/
3847	if (pGipCpuWorker->idCpu == idMaster)
3848	{
3849	if (pGipCpuWorker->i64TSCDelta == INT64_MAX) /* This shouldn't happen, but just in case. */
3850	ASMAtomicWriteS64(&pGipCpuWorker->i64TSCDelta, GIP_TSC_DELTA_INITIAL_MASTER_VALUE);
3851	return VINF_SUCCESS;
3852	}
3853
3854	/*
3855	* One measurement at a time, at least for now. We might be using
3856	* broadcast IPIs so, so be nice to the rest of the system.
3857	*/
3858	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3859	rc = RTSemMutexRequest(pDevExt->mtxTscDelta, RT_INDEFINITE_WAIT);
3860	#else
3861	rc = RTSemFastMutexRequest(pDevExt->mtxTscDelta);
3862	#endif
3863	if (RT_FAILURE(rc))
3864	return rc;
3865
3866	/*
3867	* If the CPU has hyper-threading and the APIC IDs of the master and worker are adjacent,
3868	* try pick a different master. (This fudge only works with multi core systems.)
3869	* ASSUMES related threads have adjacent APIC IDs. ASSUMES two threads per core.
3870	*
3871	* We skip this on AMDs for now as their HTT is different from Intel's and
3872	* it doesn't seem to have any favorable effect on the results.
3873	*
3874	* If the master is offline, we need a new master too, so share the code.
3875	*/
3876	iGipCpuMaster = supdrvGipFindCpuIndexForCpuId(pGip, idMaster);
3877	AssertReturn(iGipCpuMaster < pGip->cCpus, VERR_INVALID_CPU_ID);
3878	pGipCpuMaster = &pGip->aCPUs[iGipCpuMaster];
3879	if ( ( (pGipCpuMaster->idApic & ~1) == (pGipCpuWorker->idApic & ~1)
3880	&& pGip->cOnlineCpus > 2
3881	&& ASMHasCpuId()
3882	&& ASMIsValidStdRange(ASMCpuId_EAX(0))
3883	&& (ASMCpuId_EDX(1) & X86_CPUID_FEATURE_EDX_HTT)
3884	&& ( !ASMIsAmdCpu()
3885	\|\| ASMGetCpuFamily(u32Tmp = ASMCpuId_EAX(1)) > 0x15
3886	\|\| ( ASMGetCpuFamily(u32Tmp) == 0x15 /* Piledriver+, not bulldozer (FX-4150 didn't like it). */
3887	&& ASMGetCpuModelAMD(u32Tmp) >= 0x02) ) )
3888	\|\| !RTMpIsCpuOnline(idMaster) )
3889	{
3890	uint32_t i;
3891	for (i = 0; i < pGip->cCpus; i++)
3892	if ( i != iGipCpuMaster
3893	&& i != idxWorker
3894	&& pGip->aCPUs[i].enmState == SUPGIPCPUSTATE_ONLINE
3895	&& pGip->aCPUs[i].i64TSCDelta != INT64_MAX
3896	&& pGip->aCPUs[i].idCpu != NIL_RTCPUID
3897	&& pGip->aCPUs[i].idCpu != idMaster /* paranoia starts here... */
3898	&& pGip->aCPUs[i].idCpu != pGipCpuWorker->idCpu
3899	&& pGip->aCPUs[i].idApic != pGipCpuWorker->idApic
3900	&& pGip->aCPUs[i].idApic != pGipCpuMaster->idApic
3901	&& RTMpIsCpuOnline(pGip->aCPUs[i].idCpu))
3902	{
3903	iGipCpuMaster = i;
3904	pGipCpuMaster = &pGip->aCPUs[i];
3905	idMaster = pGipCpuMaster->idCpu;
3906	break;
3907	}
3908	}
3909
3910	if (RTCpuSetIsMemberByIndex(&pGip->OnlineCpuSet, pGipCpuWorker->iCpuSet))
3911	{
3912	/*
3913	* Initialize data package for the RTMpOnPair callback.
3914	*/
3915	PSUPDRVGIPTSCDELTARGS pArgs = (PSUPDRVGIPTSCDELTARGS)RTMemAllocZ(sizeof(*pArgs));
3916	if (pArgs)
3917	{
3918	pArgs->pWorker = pGipCpuWorker;
3919	pArgs->pMaster = pGipCpuMaster;
3920	pArgs->pDevExt = pDevExt;
3921	pArgs->pSyncMaster = NULL;
3922	pArgs->pSyncWorker = NULL;
3923	pArgs->cMaxTscTicks = ASMAtomicReadU64(&pGip->u64CpuHz) / 512; /* 1953 us */
3924
3925	/*
3926	* Do the RTMpOnPair call. We reset i64TSCDelta first so we
3927	* and supdrvTscMeasureDeltaCallback can use it as a success check.
3928	*/
3929	/** @todo Store the i64TSCDelta result in pArgs first? Perhaps deals with
3930	* that when doing the restart loop reorg. */
3931	ASMAtomicWriteS64(&pGipCpuWorker->i64TSCDelta, INT64_MAX);
3932	rc = RTMpOnPair(pGipCpuMaster->idCpu, pGipCpuWorker->idCpu, RTMPON_F_CONCURRENT_EXEC,
3933	supdrvTscMeasureDeltaCallback, pArgs, NULL);
3934	if (RT_SUCCESS(rc))
3935	{
3936	#if 0
3937	SUPR0Printf("mponpair ticks: %9llu %9llu max: %9llu iTry: %u%s\n", pArgs->cElapsedMasterTscTicks,
3938	pArgs->cElapsedWorkerTscTicks, pArgs->cMaxTscTicks, pArgs->iTry,
3939	pArgs->fTimedOut ? " timed out" :"");
3940	#endif
3941	#if 0
3942	SUPR0Printf("rcVerify=%d iVerifyBadTscDiff=%lld cMinVerifyTscTicks=%lld cMaxVerifyTscTicks=%lld\n",
3943	pArgs->rcVerify, pArgs->iVerifyBadTscDiff, pArgs->cMinVerifyTscTicks, pArgs->cMaxVerifyTscTicks);
3944	#endif
3945	if (RT_LIKELY(pGipCpuWorker->i64TSCDelta != INT64_MAX))
3946	{
3947	/*
3948	* Work the TSC delta applicability rating. It starts
3949	* optimistic in supdrvGipInit, we downgrade it here.
3950	*/
3951	SUPGIPUSETSCDELTA enmRating;
3952	if ( pGipCpuWorker->i64TSCDelta > GIP_TSC_DELTA_THRESHOLD_ROUGHLY_ZERO
3953	\|\| pGipCpuWorker->i64TSCDelta < -GIP_TSC_DELTA_THRESHOLD_ROUGHLY_ZERO)
3954	enmRating = SUPGIPUSETSCDELTA_NOT_ZERO;
3955	else if ( pGipCpuWorker->i64TSCDelta > GIP_TSC_DELTA_THRESHOLD_PRACTICALLY_ZERO
3956	\|\| pGipCpuWorker->i64TSCDelta < -GIP_TSC_DELTA_THRESHOLD_PRACTICALLY_ZERO)
3957	enmRating = SUPGIPUSETSCDELTA_ROUGHLY_ZERO;
3958	else
3959	enmRating = SUPGIPUSETSCDELTA_PRACTICALLY_ZERO;
3960	if (pGip->enmUseTscDelta < enmRating)
3961	{
3962	AssertCompile(sizeof(pGip->enmUseTscDelta) == sizeof(uint32_t));
3963	ASMAtomicWriteU32((uint32_t volatile *)&pGip->enmUseTscDelta, enmRating);
3964	}
3965	}
3966	else
3967	rc = VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED;
3968	}
3969	/** @todo return try-again if we get an offline CPU error. */
3970
3971	RTMemFree(pArgs);
3972	}
3973	else
3974	rc = VERR_NO_MEMORY;
3975	}
3976	else
3977	rc = VERR_CPU_OFFLINE;
3978
3979	/*
3980	* We're done now.
3981	*/
3982	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
3983	rc2 = RTSemMutexRelease(pDevExt->mtxTscDelta); AssertRC(rc2);
3984	#else
3985	rc2 = RTSemFastMutexRelease(pDevExt->mtxTscDelta); AssertRC(rc2);
3986	#endif
3987	return rc;
3988	}
3989
3990
3991	/**
3992	* Resets the TSC-delta related TSC samples and optionally the deltas
3993	* themselves.
3994	*
3995	* @param pDevExt Pointer to the device instance data.
3996	* @param fResetTscDeltas Whether the TSC-deltas are also to be reset.
3997	*
3998	* @remarks This might be called while holding a spinlock!
3999	*/
4000	static void supdrvTscResetSamples(PSUPDRVDEVEXT pDevExt, bool fResetTscDeltas)
4001	{
4002	unsigned iCpu;
4003	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4004	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
4005	{
4006	PSUPGIPCPU pGipCpu = &pGip->aCPUs[iCpu];
4007	ASMAtomicWriteU64(&pGipCpu->u64TSCSample, GIP_TSC_DELTA_RSVD);
4008	if (fResetTscDeltas)
4009	{
4010	RTCpuSetDelByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpu->iCpuSet);
4011	ASMAtomicWriteS64(&pGipCpu->i64TSCDelta, INT64_MAX);
4012	}
4013	}
4014	}
4015
4016
4017	/**
4018	* Picks an online CPU as the master TSC for TSC-delta computations.
4019	*
4020	* @returns VBox status code.
4021	* @param pDevExt Pointer to the device instance data.
4022	* @param pidxMaster Where to store the CPU array index of the chosen
4023	* master. Optional, can be NULL.
4024	*/
4025	static int supdrvTscPickMaster(PSUPDRVDEVEXT pDevExt, uint32_t *pidxMaster)
4026	{
4027	/*
4028	* Pick the first CPU online as the master TSC and make it the new GIP master based
4029	* on the APIC ID.
4030	*
4031	* Technically we can simply use "idGipMaster" but doing this gives us master as CPU 0
4032	* in most cases making it nicer/easier for comparisons. It is safe to update the GIP
4033	* master as this point since the sync/async timer isn't created yet.
4034	*/
4035	unsigned iCpu;
4036	uint32_t idxMaster = UINT32_MAX;
4037	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4038	for (iCpu = 0; iCpu < RT_ELEMENTS(pGip->aiCpuFromApicId); iCpu++)
4039	{
4040	uint16_t idxCpu = pGip->aiCpuFromApicId[iCpu];
4041	if (idxCpu != UINT16_MAX)
4042	{
4043	PSUPGIPCPU pGipCpu = &pGip->aCPUs[idxCpu];
4044	if (RTCpuSetIsMemberByIndex(&pGip->OnlineCpuSet, pGipCpu->iCpuSet))
4045	{
4046	idxMaster = idxCpu;
4047	pGipCpu->i64TSCDelta = GIP_TSC_DELTA_INITIAL_MASTER_VALUE;
4048	ASMAtomicWriteSize(&pDevExt->idGipMaster, pGipCpu->idCpu);
4049	if (pidxMaster)
4050	*pidxMaster = idxMaster;
4051	return VINF_SUCCESS;
4052	}
4053	}
4054	}
4055	return VERR_CPU_OFFLINE;
4056	}
4057
4058
4059	/**
4060	* Performs the initial measurements of the TSC deltas between CPUs.
4061	*
4062	* This is called by supdrvGipCreate(), supdrvGipPowerNotificationCallback() or
4063	* triggered by it if threaded.
4064	*
4065	* @returns VBox status code.
4066	* @param pDevExt Pointer to the device instance data.
4067	*
4068	* @remarks Must be called only after supdrvGipInitOnCpu() as this function uses
4069	* idCpu, GIP's online CPU set which are populated in
4070	* supdrvGipInitOnCpu().
4071	*/
4072	static int supdrvTscMeasureInitialDeltas(PSUPDRVDEVEXT pDevExt)
4073	{
4074	PSUPGIPCPU pGipCpuMaster;
4075	unsigned iCpu;
4076	unsigned iOddEven;
4077	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4078	uint32_t idxMaster = UINT32_MAX;
4079	uint32_t cMpOnOffEvents = ASMAtomicReadU32(&pDevExt->cMpOnOffEvents);
4080
4081	Assert(pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED);
4082	supdrvTscResetSamples(pDevExt, true /* fClearDeltas */);
4083	int rc = supdrvTscPickMaster(pDevExt, &idxMaster);
4084	if (RT_FAILURE(rc))
4085	{
4086	SUPR0Printf("Failed to pick a CPU master for TSC-delta measurements rc=%Rrc\n", rc);
4087	return rc;
4088	}
4089	AssertReturn(idxMaster < pGip->cCpus, VERR_INVALID_CPU_INDEX);
4090	pGipCpuMaster = &pGip->aCPUs[idxMaster];
4091	Assert(pDevExt->idGipMaster == pGipCpuMaster->idCpu);
4092
4093	/*
4094	* If there is only a single CPU online we have nothing to do.
4095	*/
4096	if (pGip->cOnlineCpus <= 1)
4097	{
4098	AssertReturn(pGip->cOnlineCpus > 0, VERR_INTERNAL_ERROR_5);
4099	return VINF_SUCCESS;
4100	}
4101
4102	/*
4103	* Loop thru the GIP CPU array and get deltas for each CPU (except the
4104	* master). We do the CPUs with the even numbered APIC IDs first so that
4105	* we've got alternative master CPUs to pick from on hyper-threaded systems.
4106	*/
4107	for (iOddEven = 0; iOddEven < 2; iOddEven++)
4108	{
4109	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
4110	{
4111	PSUPGIPCPU pGipCpuWorker = &pGip->aCPUs[iCpu];
4112	if ( iCpu != idxMaster
4113	&& (iOddEven > 0 \|\| (pGipCpuWorker->idApic & 1) == 0)
4114	&& RTCpuSetIsMemberByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet))
4115	{
4116	rc = supdrvTscMeasureDeltaOne(pDevExt, iCpu);
4117	if (RT_FAILURE(rc))
4118	{
4119	SUPR0Printf("supdrvTscMeasureDeltaOne failed. rc=%d CPU[%u].idCpu=%u Master[%u].idCpu=%u\n", rc, iCpu,
4120	pGipCpuWorker->idCpu, idxMaster, pDevExt->idGipMaster, pGipCpuMaster->idCpu);
4121	break;
4122	}
4123
4124	if (ASMAtomicReadU32(&pDevExt->cMpOnOffEvents) != cMpOnOffEvents)
4125	{
4126	SUPR0Printf("One or more CPUs transitioned between online & offline states. I'm confused, retry...\n");
4127	rc = VERR_TRY_AGAIN;
4128	break;
4129	}
4130	}
4131	}
4132	}
4133
4134	return rc;
4135	}
4136
4137
4138	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
4139
4140	/**
4141	* Switches the TSC-delta measurement thread into the butchered state.
4142	*
4143	* @returns VBox status code.
4144	* @param pDevExt Pointer to the device instance data.
4145	* @param fSpinlockHeld Whether the TSC-delta spinlock is held or not.
4146	* @param pszFailed An error message to log.
4147	* @param rcFailed The error code to exit the thread with.
4148	*/
4149	static int supdrvTscDeltaThreadButchered(PSUPDRVDEVEXT pDevExt, bool fSpinlockHeld, const char *pszFailed, int rcFailed)
4150	{
4151	if (!fSpinlockHeld)
4152	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4153
4154	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Butchered;
4155	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4156	OSDBGPRINT(("supdrvTscDeltaThreadButchered: %s. rc=%Rrc\n", pszFailed, rcFailed));
4157	return rcFailed;
4158	}
4159
4160
4161	/**
4162	* The TSC-delta measurement thread.
4163	*
4164	* @returns VBox status code.
4165	* @param hThread The thread handle.
4166	* @param pvUser Opaque pointer to the device instance data.
4167	*/
4168	static DECLCALLBACK(int) supdrvTscDeltaThread(RTTHREAD hThread, void *pvUser)
4169	{
4170	PSUPDRVDEVEXT pDevExt = (PSUPDRVDEVEXT)pvUser;
4171	uint32_t cConsecutiveTimeouts = 0;
4172	int rc = VERR_INTERNAL_ERROR_2;
4173	for (;;)
4174	{
4175	/*
4176	* Switch on the current state.
4177	*/
4178	SUPDRVTSCDELTATHREADSTATE enmState;
4179	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4180	enmState = pDevExt->enmTscDeltaThreadState;
4181	switch (enmState)
4182	{
4183	case kTscDeltaThreadState_Creating:
4184	{
4185	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Listening;
4186	rc = RTSemEventSignal(pDevExt->hTscDeltaEvent);
4187	if (RT_FAILURE(rc))
4188	return supdrvTscDeltaThreadButchered(pDevExt, true /* fSpinlockHeld */, "RTSemEventSignal", rc);
4189	/* fall thru */
4190	}
4191
4192	case kTscDeltaThreadState_Listening:
4193	{
4194	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4195
4196	/*
4197	* Linux counts uninterruptible sleeps as load, hence we shall do a
4198	* regular, interruptible sleep here and ignore wake ups due to signals.
4199	* See task_contributes_to_load() in include/linux/sched.h in the Linux sources.
4200	*/
4201	rc = RTThreadUserWaitNoResume(hThread, pDevExt->cMsTscDeltaTimeout);
4202	if ( RT_FAILURE(rc)
4203	&& rc != VERR_TIMEOUT
4204	&& rc != VERR_INTERRUPTED)
4205	return supdrvTscDeltaThreadButchered(pDevExt, false /* fSpinlockHeld */, "RTThreadUserWait", rc);
4206	RTThreadUserReset(hThread);
4207	break;
4208	}
4209
4210	case kTscDeltaThreadState_WaitAndMeasure:
4211	{
4212	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Measuring;
4213	rc = RTSemEventSignal(pDevExt->hTscDeltaEvent); /* (Safe on windows as long as spinlock isn't IRQ safe.) */
4214	if (RT_FAILURE(rc))
4215	return supdrvTscDeltaThreadButchered(pDevExt, true /* fSpinlockHeld */, "RTSemEventSignal", rc);
4216	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4217	RTThreadSleep(1);
4218	/* fall thru */
4219	}
4220
4221	case kTscDeltaThreadState_Measuring:
4222	{
4223	cConsecutiveTimeouts = 0;
4224	if (pDevExt->fTscThreadRecomputeAllDeltas)
4225	{
4226	int cTries = 8;
4227	int cMsWaitPerTry = 10;
4228	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4229	Assert(pGip);
4230	do
4231	{
4232	RTCpuSetCopy(&pDevExt->TscDeltaCpuSet, &pGip->OnlineCpuSet);
4233	rc = supdrvTscMeasureInitialDeltas(pDevExt);
4234	if ( RT_SUCCESS(rc)
4235	\|\| ( RT_FAILURE(rc)
4236	&& rc != VERR_TRY_AGAIN
4237	&& rc != VERR_CPU_OFFLINE))
4238	{
4239	break;
4240	}
4241	RTThreadSleep(cMsWaitPerTry);
4242	} while (cTries-- > 0);
4243	pDevExt->fTscThreadRecomputeAllDeltas = false;
4244	}
4245	else
4246	{
4247	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4248	unsigned iCpu;
4249
4250	/* Measure TSC-deltas only for the CPUs that are in the set. */
4251	rc = VINF_SUCCESS;
4252	for (iCpu = 0; iCpu < pGip->cCpus; iCpu++)
4253	{
4254	PSUPGIPCPU pGipCpuWorker = &pGip->aCPUs[iCpu];
4255	if (RTCpuSetIsMemberByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet))
4256	{
4257	if (pGipCpuWorker->i64TSCDelta == INT64_MAX)
4258	{
4259	int rc2 = supdrvTscMeasureDeltaOne(pDevExt, iCpu);
4260	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
4261	rc = rc2;
4262	}
4263	else
4264	{
4265	/*
4266	* The thread/someone must've called SUPR0TscDeltaMeasureBySetIndex(),
4267	* mark the delta as fine to get the timer thread off our back.
4268	*/
4269	RTCpuSetDelByIndex(&pDevExt->TscDeltaCpuSet, pGipCpuWorker->iCpuSet);
4270	RTCpuSetAddByIndex(&pDevExt->TscDeltaObtainedCpuSet, pGipCpuWorker->iCpuSet);
4271	}
4272	}
4273	}
4274	}
4275	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4276	if (pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Measuring)
4277	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Listening;
4278	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4279	Assert(rc != VERR_NOT_AVAILABLE); /* VERR_NOT_AVAILABLE is used as init value, see supdrvTscDeltaThreadInit(). */
4280	ASMAtomicWriteS32(&pDevExt->rcTscDelta, rc);
4281	break;
4282	}
4283
4284	case kTscDeltaThreadState_Terminating:
4285	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Destroyed;
4286	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4287	return VINF_SUCCESS;
4288
4289	case kTscDeltaThreadState_Butchered:
4290	default:
4291	return supdrvTscDeltaThreadButchered(pDevExt, true /* fSpinlockHeld */, "Invalid state", VERR_INVALID_STATE);
4292	}
4293	}
4294	/* not reached */
4295	}
4296
4297
4298	/**
4299	* Waits for the TSC-delta measurement thread to respond to a state change.
4300	*
4301	* @returns VINF_SUCCESS on success, VERR_TIMEOUT if it doesn't respond in time,
4302	* other error code on internal error.
4303	*
4304	* @param pDevExt The device instance data.
4305	* @param enmCurState The current state.
4306	* @param enmNewState The new state we're waiting for it to enter.
4307	*/
4308	static int supdrvTscDeltaThreadWait(PSUPDRVDEVEXT pDevExt, SUPDRVTSCDELTATHREADSTATE enmCurState,
4309	SUPDRVTSCDELTATHREADSTATE enmNewState)
4310	{
4311	SUPDRVTSCDELTATHREADSTATE enmActualState;
4312	int rc;
4313
4314	/*
4315	* Wait a short while for the expected state transition.
4316	*/
4317	RTSemEventWait(pDevExt->hTscDeltaEvent, RT_MS_1SEC);
4318	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4319	enmActualState = pDevExt->enmTscDeltaThreadState;
4320	if (enmActualState == enmNewState)
4321	{
4322	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4323	rc = VINF_SUCCESS;
4324	}
4325	else if (enmActualState == enmCurState)
4326	{
4327	/*
4328	* Wait longer if the state has not yet transitioned to the one we want.
4329	*/
4330	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4331	rc = RTSemEventWait(pDevExt->hTscDeltaEvent, 50 * RT_MS_1SEC);
4332	if ( RT_SUCCESS(rc)
4333	\|\| rc == VERR_TIMEOUT)
4334	{
4335	/*
4336	* Check the state whether we've succeeded.
4337	*/
4338	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4339	enmActualState = pDevExt->enmTscDeltaThreadState;
4340	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4341	if (enmActualState == enmNewState)
4342	rc = VINF_SUCCESS;
4343	else if (enmActualState == enmCurState)
4344	{
4345	rc = VERR_TIMEOUT;
4346	OSDBGPRINT(("supdrvTscDeltaThreadWait: timed out state transition. enmActualState=%d enmNewState=%d\n",
4347	enmActualState, enmNewState));
4348	}
4349	else
4350	{
4351	rc = VERR_INTERNAL_ERROR;
4352	OSDBGPRINT(("supdrvTscDeltaThreadWait: invalid state transition from %d to %d, expected %d\n", enmCurState,
4353	enmActualState, enmNewState));
4354	}
4355	}
4356	else
4357	OSDBGPRINT(("supdrvTscDeltaThreadWait: RTSemEventWait failed. rc=%Rrc\n", rc));
4358	}
4359	else
4360	{
4361	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4362	OSDBGPRINT(("supdrvTscDeltaThreadWait: invalid state %d when transitioning from %d to %d\n",
4363	enmActualState, enmCurState, enmNewState));
4364	rc = VERR_INTERNAL_ERROR;
4365	}
4366
4367	return rc;
4368	}
4369
4370
4371	/**
4372	* Signals the TSC-delta thread to start measuring TSC-deltas.
4373	*
4374	* @param pDevExt Pointer to the device instance data.
4375	* @param fForceAll Force re-calculating TSC-deltas on all CPUs.
4376	*/
4377	static void supdrvTscDeltaThreadStartMeasurement(PSUPDRVDEVEXT pDevExt, bool fForceAll)
4378	{
4379	if (pDevExt->hTscDeltaThread != NIL_RTTHREAD)
4380	{
4381	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4382	if ( pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Listening
4383	\|\| pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Measuring)
4384	{
4385	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_WaitAndMeasure;
4386	if (fForceAll)
4387	pDevExt->fTscThreadRecomputeAllDeltas = true;
4388	}
4389	else if ( pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_WaitAndMeasure
4390	&& fForceAll)
4391	pDevExt->fTscThreadRecomputeAllDeltas = true;
4392	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4393	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4394	}
4395	}
4396
4397
4398	/**
4399	* Terminates the actual thread running supdrvTscDeltaThread().
4400	*
4401	* This is an internal worker function for supdrvTscDeltaThreadInit() and
4402	* supdrvTscDeltaTerm().
4403	*
4404	* @param pDevExt Pointer to the device instance data.
4405	*/
4406	static void supdrvTscDeltaThreadTerminate(PSUPDRVDEVEXT pDevExt)
4407	{
4408	int rc;
4409	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4410	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Terminating;
4411	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4412	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4413	rc = RTThreadWait(pDevExt->hTscDeltaThread, 50 * RT_MS_1SEC, NULL /* prc */);
4414	if (RT_FAILURE(rc))
4415	{
4416	/* Signal a few more times before giving up. */
4417	int cTriesLeft = 5;
4418	while (--cTriesLeft > 0)
4419	{
4420	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4421	rc = RTThreadWait(pDevExt->hTscDeltaThread, 2 * RT_MS_1SEC, NULL /* prc */);
4422	if (rc != VERR_TIMEOUT)
4423	break;
4424	}
4425	}
4426	}
4427
4428
4429	/**
4430	* Initializes and spawns the TSC-delta measurement thread.
4431	*
4432	* A thread is required for servicing re-measurement requests from events like
4433	* CPUs coming online, suspend/resume etc. as it cannot be done synchronously
4434	* under all contexts on all OSs.
4435	*
4436	* @returns VBox status code.
4437	* @param pDevExt Pointer to the device instance data.
4438	*
4439	* @remarks Must only be called -after- initializing GIP and setting up MP
4440	* notifications!
4441	*/
4442	static int supdrvTscDeltaThreadInit(PSUPDRVDEVEXT pDevExt)
4443	{
4444	int rc;
4445	Assert(pDevExt->pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED);
4446	rc = RTSpinlockCreate(&pDevExt->hTscDeltaSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "VBoxTscSpnLck");
4447	if (RT_SUCCESS(rc))
4448	{
4449	rc = RTSemEventCreate(&pDevExt->hTscDeltaEvent);
4450	if (RT_SUCCESS(rc))
4451	{
4452	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_Creating;
4453	pDevExt->cMsTscDeltaTimeout = 60000;
4454	rc = RTThreadCreate(&pDevExt->hTscDeltaThread, supdrvTscDeltaThread, pDevExt, 0 /* cbStack */,
4455	RTTHREADTYPE_DEFAULT, RTTHREADFLAGS_WAITABLE, "VBoxTscThread");
4456	if (RT_SUCCESS(rc))
4457	{
4458	rc = supdrvTscDeltaThreadWait(pDevExt, kTscDeltaThreadState_Creating, kTscDeltaThreadState_Listening);
4459	if (RT_SUCCESS(rc))
4460	{
4461	ASMAtomicWriteS32(&pDevExt->rcTscDelta, VERR_NOT_AVAILABLE);
4462	return rc;
4463	}
4464
4465	OSDBGPRINT(("supdrvTscDeltaInit: supdrvTscDeltaThreadWait failed. rc=%Rrc\n", rc));
4466	supdrvTscDeltaThreadTerminate(pDevExt);
4467	}
4468	else
4469	OSDBGPRINT(("supdrvTscDeltaInit: RTThreadCreate failed. rc=%Rrc\n", rc));
4470	RTSemEventDestroy(pDevExt->hTscDeltaEvent);
4471	pDevExt->hTscDeltaEvent = NIL_RTSEMEVENT;
4472	}
4473	else
4474	OSDBGPRINT(("supdrvTscDeltaInit: RTSemEventCreate failed. rc=%Rrc\n", rc));
4475	RTSpinlockDestroy(pDevExt->hTscDeltaSpinlock);
4476	pDevExt->hTscDeltaSpinlock = NIL_RTSPINLOCK;
4477	}
4478	else
4479	OSDBGPRINT(("supdrvTscDeltaInit: RTSpinlockCreate failed. rc=%Rrc\n", rc));
4480
4481	return rc;
4482	}
4483
4484
4485	/**
4486	* Terminates the TSC-delta measurement thread and cleanup.
4487	*
4488	* @param pDevExt Pointer to the device instance data.
4489	*/
4490	static void supdrvTscDeltaTerm(PSUPDRVDEVEXT pDevExt)
4491	{
4492	if ( pDevExt->hTscDeltaSpinlock != NIL_RTSPINLOCK
4493	&& pDevExt->hTscDeltaEvent != NIL_RTSEMEVENT)
4494	{
4495	supdrvTscDeltaThreadTerminate(pDevExt);
4496	}
4497
4498	if (pDevExt->hTscDeltaSpinlock != NIL_RTSPINLOCK)
4499	{
4500	RTSpinlockDestroy(pDevExt->hTscDeltaSpinlock);
4501	pDevExt->hTscDeltaSpinlock = NIL_RTSPINLOCK;
4502	}
4503
4504	if (pDevExt->hTscDeltaEvent != NIL_RTSEMEVENT)
4505	{
4506	RTSemEventDestroy(pDevExt->hTscDeltaEvent);
4507	pDevExt->hTscDeltaEvent = NIL_RTSEMEVENT;
4508	}
4509
4510	ASMAtomicWriteS32(&pDevExt->rcTscDelta, VERR_NOT_AVAILABLE);
4511	}
4512
4513	#endif /* SUPDRV_USE_TSC_DELTA_THREAD */
4514
4515	/**
4516	* Measure the TSC delta for the CPU given by its CPU set index.
4517	*
4518	* @returns VBox status code.
4519	* @retval VERR_INTERRUPTED if interrupted while waiting.
4520	* @retval VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED if we were unable to get a
4521	* measurement.
4522	* @retval VERR_CPU_OFFLINE if the specified CPU is offline.
4523	*
4524	* @param pSession The caller's session. GIP must've been mapped.
4525	* @param iCpuSet The CPU set index of the CPU to measure.
4526	* @param fFlags Flags, SUP_TSCDELTA_MEASURE_F_XXX.
4527	* @param cMsWaitRetry Number of milliseconds to wait between each retry.
4528	* @param cMsWaitThread Number of milliseconds to wait for the thread to get
4529	* ready.
4530	* @param cTries Number of times to try, pass 0 for the default.
4531	*/
4532	SUPR0DECL(int) SUPR0TscDeltaMeasureBySetIndex(PSUPDRVSESSION pSession, uint32_t iCpuSet, uint32_t fFlags,
4533	RTMSINTERVAL cMsWaitRetry, RTMSINTERVAL cMsWaitThread, uint32_t cTries)
4534	{
4535	PSUPDRVDEVEXT pDevExt;
4536	PSUPGLOBALINFOPAGE pGip;
4537	uint16_t iGipCpu;
4538	int rc;
4539	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
4540	uint64_t msTsStartWait;
4541	uint32_t iWaitLoop;
4542	#endif
4543
4544	/*
4545	* Validate and adjust the input.
4546	*/
4547	AssertReturn(SUP_IS_SESSION_VALID(pSession), VERR_INVALID_PARAMETER);
4548	if (!pSession->fGipReferenced)
4549	return VERR_WRONG_ORDER;
4550
4551	pDevExt = pSession->pDevExt;
4552	AssertReturn(SUP_IS_DEVEXT_VALID(pDevExt), VERR_INVALID_PARAMETER);
4553
4554	pGip = pDevExt->pGip;
4555	AssertPtrReturn(pGip, VERR_INTERNAL_ERROR_2);
4556
4557	AssertReturn(iCpuSet < RTCPUSET_MAX_CPUS, VERR_INVALID_CPU_INDEX);
4558	AssertReturn(iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx), VERR_INVALID_CPU_INDEX);
4559	iGipCpu = pGip->aiCpuFromCpuSetIdx[iCpuSet];
4560	AssertReturn(iGipCpu < pGip->cCpus, VERR_INVALID_CPU_INDEX);
4561
4562	if (fFlags & ~SUP_TSCDELTA_MEASURE_F_VALID_MASK)
4563	return VERR_INVALID_FLAGS;
4564
4565	/*
4566	* The request is a noop if the TSC delta isn't being used.
4567	*/
4568	if (pGip->enmUseTscDelta <= SUPGIPUSETSCDELTA_ZERO_CLAIMED)
4569	return VINF_SUCCESS;
4570
4571	if (cTries == 0)
4572	cTries = 12;
4573	else if (cTries > 256)
4574	cTries = 256;
4575
4576	if (cMsWaitRetry == 0)
4577	cMsWaitRetry = 2;
4578	else if (cMsWaitRetry > 1000)
4579	cMsWaitRetry = 1000;
4580
4581	#ifdef SUPDRV_USE_TSC_DELTA_THREAD
4582	/*
4583	* Has the TSC already been measured and we're not forced to redo it?
4584	*/
4585	if ( pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX
4586	&& !(fFlags & SUP_TSCDELTA_MEASURE_F_FORCE))
4587	return VINF_SUCCESS;
4588
4589	/*
4590	* Asynchronous request? Forward it to the thread, no waiting.
4591	*/
4592	if (fFlags & SUP_TSCDELTA_MEASURE_F_ASYNC)
4593	{
4594	/** @todo Async. doesn't implement options like retries, waiting. We'll need
4595	* to pass those options to the thread somehow and implement it in the
4596	* thread. Check if anyone uses/needs fAsync before implementing this. */
4597	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4598	RTCpuSetAddByIndex(&pDevExt->TscDeltaCpuSet, iCpuSet);
4599	if ( pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Listening
4600	\|\| pDevExt->enmTscDeltaThreadState == kTscDeltaThreadState_Measuring)
4601	{
4602	pDevExt->enmTscDeltaThreadState = kTscDeltaThreadState_WaitAndMeasure;
4603	rc = VINF_SUCCESS;
4604	}
4605	else if (pDevExt->enmTscDeltaThreadState != kTscDeltaThreadState_WaitAndMeasure)
4606	rc = VERR_THREAD_IS_DEAD;
4607	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4608	RTThreadUserSignal(pDevExt->hTscDeltaThread);
4609	return VINF_SUCCESS;
4610	}
4611
4612	/*
4613	* If a TSC-delta measurement request is already being serviced by the thread,
4614	* wait 'cTries' times if a retry-timeout is provided, otherwise bail as busy.
4615	*/
4616	msTsStartWait = RTTimeSystemMilliTS();
4617	for (iWaitLoop = 0;; iWaitLoop++)
4618	{
4619	uint64_t cMsElapsed;
4620	SUPDRVTSCDELTATHREADSTATE enmState;
4621	RTSpinlockAcquire(pDevExt->hTscDeltaSpinlock);
4622	enmState = pDevExt->enmTscDeltaThreadState;
4623	RTSpinlockRelease(pDevExt->hTscDeltaSpinlock);
4624
4625	if (enmState == kTscDeltaThreadState_Measuring)
4626	{ /* Must wait, the thread is busy. */ }
4627	else if (enmState == kTscDeltaThreadState_WaitAndMeasure)
4628	{ /* Must wait, this state only says what will happen next. */ }
4629	else if (enmState == kTscDeltaThreadState_Terminating)
4630	{ /* Must wait, this state only says what should happen next. */ }
4631	else
4632	break; /* All other states, the thread is either idly listening or dead. */
4633
4634	/* Wait or fail. */
4635	if (cMsWaitThread == 0)
4636	return VERR_SUPDRV_TSC_DELTA_MEASUREMENT_BUSY;
4637	cMsElapsed = RTTimeSystemMilliTS() - msTsStartWait;
4638	if (cMsElapsed >= cMsWaitThread)
4639	return VERR_SUPDRV_TSC_DELTA_MEASUREMENT_BUSY;
4640
4641	rc = RTThreadSleep(RT_MIN((RTMSINTERVAL)(cMsWaitThread - cMsElapsed), RT_MIN(iWaitLoop + 1, 10)));
4642	if (rc == VERR_INTERRUPTED)
4643	return rc;
4644	}
4645	#endif /* SUPDRV_USE_TSC_DELTA_THREAD */
4646
4647	/*
4648	* Try measure the TSC delta the given number of times.
4649	*/
4650	for (;;)
4651	{
4652	/* Unless we're forced to measure the delta, check whether it's done already. */
4653	if ( !(fFlags & SUP_TSCDELTA_MEASURE_F_FORCE)
4654	&& pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX)
4655	{
4656	rc = VINF_SUCCESS;
4657	break;
4658	}
4659
4660	/* Measure it. */
4661	rc = supdrvTscMeasureDeltaOne(pDevExt, iGipCpu);
4662	if (rc != VERR_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED)
4663	{
4664	Assert(pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX \|\| RT_FAILURE_NP(rc));
4665	break;
4666	}
4667
4668	/* Retry? */
4669	if (cTries <= 1)
4670	break;
4671	cTries--;
4672
4673	/* Always delay between retries (be nice to the rest of the system
4674	and avoid the BSOD hounds). */
4675	rc = RTThreadSleep(cMsWaitRetry);
4676	if (rc == VERR_INTERRUPTED)
4677	break;
4678	}
4679
4680	return rc;
4681	}
4682
4683
4684	/**
4685	* Service a TSC-delta measurement request.
4686	*
4687	* @returns VBox status code.
4688	* @param pDevExt Pointer to the device instance data.
4689	* @param pSession The support driver session.
4690	* @param pReq Pointer to the TSC-delta measurement request.
4691	*/
4692	int VBOXCALL supdrvIOCtl_TscDeltaMeasure(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPTSCDELTAMEASURE pReq)
4693	{
4694	uint32_t cTries;
4695	uint32_t iCpuSet;
4696	uint32_t fFlags;
4697	RTMSINTERVAL cMsWaitRetry;
4698	RT_NOREF1(pDevExt);
4699
4700	/*
4701	* Validate and adjust/resolve the input so they can be passed onto SUPR0TscDeltaMeasureBySetIndex.
4702	*/
4703	AssertPtr(pDevExt); AssertPtr(pSession); AssertPtr(pReq); /* paranoia^2 */
4704
4705	if (pReq->u.In.idCpu == NIL_RTCPUID)
4706	return VERR_INVALID_CPU_ID;
4707	iCpuSet = RTMpCpuIdToSetIndex(pReq->u.In.idCpu);
4708	if (iCpuSet >= RTCPUSET_MAX_CPUS)
4709	return VERR_INVALID_CPU_ID;
4710
4711	cTries = pReq->u.In.cRetries == 0 ? 0 : (uint32_t)pReq->u.In.cRetries + 1;
4712
4713	cMsWaitRetry = RT_MAX(pReq->u.In.cMsWaitRetry, 5);
4714
4715	fFlags = 0;
4716	if (pReq->u.In.fAsync)
4717	fFlags \|= SUP_TSCDELTA_MEASURE_F_ASYNC;
4718	if (pReq->u.In.fForce)
4719	fFlags \|= SUP_TSCDELTA_MEASURE_F_FORCE;
4720
4721	return SUPR0TscDeltaMeasureBySetIndex(pSession, iCpuSet, fFlags, cMsWaitRetry,
4722	cTries == 0 ? 5 * RT_MS_1SEC : cMsWaitRetry * cTries /cMsWaitThread/,
4723	cTries);
4724	}
4725
4726
4727	/**
4728	* Reads TSC with delta applied.
4729	*
4730	* Will try to resolve delta value INT64_MAX before applying it. This is the
4731	* main purpose of this function, to handle the case where the delta needs to be
4732	* determined.
4733	*
4734	* @returns VBox status code.
4735	* @param pDevExt Pointer to the device instance data.
4736	* @param pSession The support driver session.
4737	* @param pReq Pointer to the TSC-read request.
4738	*/
4739	int VBOXCALL supdrvIOCtl_TscRead(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, PSUPTSCREAD pReq)
4740	{
4741	PSUPGLOBALINFOPAGE pGip;
4742	int rc;
4743
4744	/*
4745	* Validate. We require the client to have mapped GIP (no asserting on
4746	* ring-3 preconditions).
4747	*/
4748	AssertPtr(pDevExt); AssertPtr(pReq); AssertPtr(pSession); /* paranoia^2 */
4749	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
4750	return VERR_WRONG_ORDER;
4751	pGip = pDevExt->pGip;
4752	AssertReturn(pGip, VERR_INTERNAL_ERROR_2);
4753
4754	/*
4755	* We're usually here because we need to apply delta, but we shouldn't be
4756	* upset if the GIP is some different mode.
4757	*/
4758	if (pGip->enmUseTscDelta > SUPGIPUSETSCDELTA_ZERO_CLAIMED)
4759	{
4760	uint32_t cTries = 0;
4761	for (;;)
4762	{
4763	/*
4764	* Start by gathering the data, using CLI for disabling preemption
4765	* while we do that.
4766	*/
4767	RTCCUINTREG fEFlags = ASMIntDisableFlags();
4768	int iCpuSet = RTMpCpuIdToSetIndex(RTMpCpuId());
4769	int iGipCpu;
4770	if (RT_LIKELY( (unsigned)iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
4771	&& (iGipCpu = pGip->aiCpuFromCpuSetIdx[iCpuSet]) < pGip->cCpus ))
4772	{
4773	int64_t i64Delta = pGip->aCPUs[iGipCpu].i64TSCDelta;
4774	pReq->u.Out.idApic = pGip->aCPUs[iGipCpu].idApic;
4775	pReq->u.Out.u64AdjustedTsc = ASMReadTSC();
4776	ASMSetFlags(fEFlags);
4777
4778	/*
4779	* If we're lucky we've got a delta, but no predictions here
4780	* as this I/O control is normally only used when the TSC delta
4781	* is set to INT64_MAX.
4782	*/
4783	if (i64Delta != INT64_MAX)
4784	{
4785	pReq->u.Out.u64AdjustedTsc -= i64Delta;
4786	rc = VINF_SUCCESS;
4787	break;
4788	}
4789
4790	/* Give up after a few times. */
4791	if (cTries >= 4)
4792	{
4793	rc = VWRN_SUPDRV_TSC_DELTA_MEASUREMENT_FAILED;
4794	break;
4795	}
4796
4797	/* Need to measure the delta an try again. */
4798	rc = supdrvTscMeasureDeltaOne(pDevExt, iGipCpu);
4799	Assert(pGip->aCPUs[iGipCpu].i64TSCDelta != INT64_MAX \|\| RT_FAILURE_NP(rc));
4800	/** @todo should probably delay on failure... dpc watchdogs */
4801	}
4802	else
4803	{
4804	/* This really shouldn't happen. */
4805	AssertMsgFailed(("idCpu=%#x iCpuSet=%#x (%d)\n", RTMpCpuId(), iCpuSet, iCpuSet));
4806	pReq->u.Out.idApic = ASMGetApicId();
4807	pReq->u.Out.u64AdjustedTsc = ASMReadTSC();
4808	ASMSetFlags(fEFlags);
4809	rc = VERR_INTERNAL_ERROR_5; /** @todo change to warning. */
4810	break;
4811	}
4812	}
4813	}
4814	else
4815	{
4816	/*
4817	* No delta to apply. Easy. Deal with preemption the lazy way.
4818	*/
4819	RTCCUINTREG fEFlags = ASMIntDisableFlags();
4820	int iCpuSet = RTMpCpuIdToSetIndex(RTMpCpuId());
4821	int iGipCpu;
4822	if (RT_LIKELY( (unsigned)iCpuSet < RT_ELEMENTS(pGip->aiCpuFromCpuSetIdx)
4823	&& (iGipCpu = pGip->aiCpuFromCpuSetIdx[iCpuSet]) < pGip->cCpus ))
4824	pReq->u.Out.idApic = pGip->aCPUs[iGipCpu].idApic;
4825	else
4826	pReq->u.Out.idApic = ASMGetApicId();
4827	pReq->u.Out.u64AdjustedTsc = ASMReadTSC();
4828	ASMSetFlags(fEFlags);
4829	rc = VINF_SUCCESS;
4830	}
4831
4832	return rc;
4833	}
4834
4835
4836	/**
4837	* Worker for supdrvIOCtl_GipSetFlags.
4838	*
4839	* @returns VBox status code.
4840	* @retval VERR_WRONG_ORDER if an enable-once-per-session flag is set again for
4841	* a session.
4842	*
4843	* @param pDevExt Pointer to the device instance data.
4844	* @param pSession The support driver session.
4845	* @param fOrMask The OR mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4846	* @param fAndMask The AND mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4847	*
4848	* @remarks Caller must own the GIP mutex.
4849	*
4850	* @remarks This function doesn't validate any of the flags.
4851	*/
4852	static int supdrvGipSetFlags(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, uint32_t fOrMask, uint32_t fAndMask)
4853	{
4854	uint32_t cRefs;
4855	PSUPGLOBALINFOPAGE pGip = pDevExt->pGip;
4856	AssertMsg((fOrMask & fAndMask) == fOrMask, ("%#x & %#x\n", fOrMask, fAndMask)); /* ASSUMED by code below */
4857
4858	/*
4859	* Compute GIP test-mode flags.
4860	*/
4861	if (fOrMask & SUPGIP_FLAGS_TESTING_ENABLE)
4862	{
4863	if (!pSession->fGipTestMode)
4864	{
4865	Assert(pDevExt->cGipTestModeRefs < _64K);
4866	pSession->fGipTestMode = true;
4867	cRefs = ++pDevExt->cGipTestModeRefs;
4868	if (cRefs == 1)
4869	{
4870	fOrMask \|= SUPGIP_FLAGS_TESTING \| SUPGIP_FLAGS_TESTING_START;
4871	fAndMask &= ~SUPGIP_FLAGS_TESTING_STOP;
4872	}
4873	}
4874	else
4875	{
4876	LogRelMax(10, ("supdrvGipSetFlags: SUPGIP_FLAGS_TESTING_ENABLE already set for this session\n"));
4877	return VERR_WRONG_ORDER;
4878	}
4879	}
4880	else if ( !(fAndMask & SUPGIP_FLAGS_TESTING_ENABLE)
4881	&& pSession->fGipTestMode)
4882	{
4883	Assert(pDevExt->cGipTestModeRefs > 0);
4884	Assert(pDevExt->cGipTestModeRefs < _64K);
4885	pSession->fGipTestMode = false;
4886	cRefs = --pDevExt->cGipTestModeRefs;
4887	if (!cRefs)
4888	fOrMask \|= SUPGIP_FLAGS_TESTING_STOP;
4889	else
4890	fAndMask \|= SUPGIP_FLAGS_TESTING_ENABLE;
4891	}
4892
4893	/*
4894	* Commit the flags. This should be done as atomically as possible
4895	* since the flag consumers won't be holding the GIP mutex.
4896	*/
4897	ASMAtomicOrU32(&pGip->fFlags, fOrMask);
4898	ASMAtomicAndU32(&pGip->fFlags, fAndMask);
4899
4900	return VINF_SUCCESS;
4901	}
4902
4903
4904	/**
4905	* Sets GIP test mode parameters.
4906	*
4907	* @returns VBox status code.
4908	* @param pDevExt Pointer to the device instance data.
4909	* @param pSession The support driver session.
4910	* @param fOrMask The OR mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4911	* @param fAndMask The AND mask of the GIP flags, see SUPGIP_FLAGS_XXX.
4912	*/
4913	int VBOXCALL supdrvIOCtl_GipSetFlags(PSUPDRVDEVEXT pDevExt, PSUPDRVSESSION pSession, uint32_t fOrMask, uint32_t fAndMask)
4914	{
4915	PSUPGLOBALINFOPAGE pGip;
4916	int rc;
4917
4918	/*
4919	* Validate. We require the client to have mapped GIP (no asserting on
4920	* ring-3 preconditions).
4921	*/
4922	AssertPtr(pDevExt); AssertPtr(pSession); /* paranoia^2 */
4923	if (pSession->GipMapObjR3 == NIL_RTR0MEMOBJ)
4924	return VERR_WRONG_ORDER;
4925	pGip = pDevExt->pGip;
4926	AssertReturn(pGip, VERR_INTERNAL_ERROR_3);
4927
4928	if (fOrMask & ~SUPGIP_FLAGS_VALID_MASK)
4929	return VERR_INVALID_PARAMETER;
4930	if ((fAndMask & ~SUPGIP_FLAGS_VALID_MASK) != ~SUPGIP_FLAGS_VALID_MASK)
4931	return VERR_INVALID_PARAMETER;
4932
4933	/*
4934	* Don't confuse supdrvGipSetFlags or anyone else by both setting
4935	* and clearing the same flags. AND takes precedence.
4936	*/
4937	fOrMask &= fAndMask;
4938
4939	/*
4940	* Take the loader lock to avoid having to think about races between two
4941	* clients changing the flags at the same time (state is not simple).
4942	*/
4943	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4944	RTSemMutexRequest(pDevExt->mtxGip, RT_INDEFINITE_WAIT);
4945	#else
4946	RTSemFastMutexRequest(pDevExt->mtxGip);
4947	#endif
4948
4949	rc = supdrvGipSetFlags(pDevExt, pSession, fOrMask, fAndMask);
4950
4951	#ifdef SUPDRV_USE_MUTEX_FOR_GIP
4952	RTSemMutexRelease(pDevExt->mtxGip);
4953	#else
4954	RTSemFastMutexRelease(pDevExt->mtxGip);
4955	#endif
4956	return rc;
4957	}
4958

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

Download in other formats:

© 2023 Oracle
Contact – Privacy policy – Terms of Use