GVMMR0.cpp@ 49966

Last change on this file since 49966 was 48395, checked in by vboxsync, 11 years ago
VMM/GVMMR0: Todo question.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 87.3 KB

Line
1	/* $Id: GVMMR0.cpp 48395 2013-09-09 16:37:41Z vboxsync $ */
2	/** @file
3	* GVMM - Global VM Manager.
4	*/
5
6	/*
7	* Copyright (C) 2007-2012 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
18
19	/** @page pg_gvmm GVMM - The Global VM Manager
20	*
21	* The Global VM Manager lives in ring-0. Its main function at the moment is
22	* to manage a list of all running VMs, keep a ring-0 only structure (GVM) for
23	* each of them, and assign them unique identifiers (so GMM can track page
24	* owners). The GVMM also manage some of the host CPU resources, like the
25	* periodic preemption timer.
26	*
27	* The GVMM will create a ring-0 object for each VM when it is registered, this
28	* is both for session cleanup purposes and for having a point where it is
29	* possible to implement usage polices later (in SUPR0ObjRegister).
30	*
31	*
32	* @section sec_gvmm_ppt Periodic Preemption Timer (PPT)
33	*
34	* On system that sports a high resolution kernel timer API, we use per-cpu
35	* timers to generate interrupts that preempts VT-x, AMD-V and raw-mode guest
36	* execution. The timer frequency is calculating by taking the max
37	* TMCalcHostTimerFrequency for all VMs running on a CPU for the last ~160 ms
38	* (RT_ELEMENTS((PGVMMHOSTCPU)0, Ppt.aHzHistory) *
39	* GVMMHOSTCPU_PPT_HIST_INTERVAL_NS).
40	*
41	* The TMCalcHostTimerFrequency() part of the things gets its takes the max
42	* TMTimerSetFrequencyHint() value and adjusts by the current catch-up percent,
43	* warp drive percent and some fudge factors. VMMR0.cpp reports the result via
44	* GVMMR0SchedUpdatePeriodicPreemptionTimer() before switching to the VT-x,
45	* AMD-V and raw-mode execution environments.
46	*/
47
48
49	/*******************************************************************************
50	* Header Files *
51	*******************************************************************************/
52	#define LOG_GROUP LOG_GROUP_GVMM
53	#include <VBox/vmm/gvmm.h>
54	#include <VBox/vmm/gmm.h>
55	#include "GVMMR0Internal.h"
56	#include <VBox/vmm/gvm.h>
57	#include <VBox/vmm/vm.h>
58	#include <VBox/vmm/vmcpuset.h>
59	#include <VBox/vmm/vmm.h>
60	#include <VBox/param.h>
61	#include <VBox/err.h>
62
63	#include <iprt/asm.h>
64	#include <iprt/asm-amd64-x86.h>
65	#include <iprt/mem.h>
66	#include <iprt/semaphore.h>
67	#include <iprt/time.h>
68	#include <VBox/log.h>
69	#include <iprt/thread.h>
70	#include <iprt/process.h>
71	#include <iprt/param.h>
72	#include <iprt/string.h>
73	#include <iprt/assert.h>
74	#include <iprt/mem.h>
75	#include <iprt/memobj.h>
76	#include <iprt/mp.h>
77	#include <iprt/cpuset.h>
78	#include <iprt/spinlock.h>
79	#include <iprt/timer.h>
80
81	#include "dtrace/VBoxVMM.h"
82
83
84	/*******************************************************************************
85	* Defined Constants And Macros *
86	*******************************************************************************/
87	#if defined(RT_OS_LINUX) \|\| defined(DOXYGEN_RUNNING)
88	/** Define this to enable the periodic preemption timer. */
89	# define GVMM_SCHED_WITH_PPT
90	#endif
91
92
93	/*******************************************************************************
94	* Structures and Typedefs *
95	*******************************************************************************/
96
97	/**
98	* Global VM handle.
99	*/
100	typedef struct GVMHANDLE
101	{
102	/** The index of the next handle in the list (free or used). (0 is nil.) */
103	uint16_t volatile iNext;
104	/** Our own index / handle value. */
105	uint16_t iSelf;
106	/** The process ID of the handle owner.
107	* This is used for access checks. */
108	RTPROCESS ProcId;
109	/** The pointer to the ring-0 only (aka global) VM structure. */
110	PGVM pGVM;
111	/** The ring-0 mapping of the shared VM instance data. */
112	PVM pVM;
113	/** The virtual machine object. */
114	void *pvObj;
115	/** The session this VM is associated with. */
116	PSUPDRVSESSION pSession;
117	/** The ring-0 handle of the EMT0 thread.
118	* This is used for ownership checks as well as looking up a VM handle by thread
119	* at times like assertions. */
120	RTNATIVETHREAD hEMT0;
121	} GVMHANDLE;
122	/** Pointer to a global VM handle. */
123	typedef GVMHANDLE *PGVMHANDLE;
124
125	/** Number of GVM handles (including the NIL handle). */
126	#if HC_ARCH_BITS == 64
127	# define GVMM_MAX_HANDLES 8192
128	#else
129	# define GVMM_MAX_HANDLES 128
130	#endif
131
132	/**
133	* Per host CPU GVMM data.
134	*/
135	typedef struct GVMMHOSTCPU
136	{
137	/** Magic number (GVMMHOSTCPU_MAGIC). */
138	uint32_t volatile u32Magic;
139	/** The CPU ID. */
140	RTCPUID idCpu;
141	/** The CPU set index. */
142	uint32_t idxCpuSet;
143
144	#ifdef GVMM_SCHED_WITH_PPT
145	/** Periodic preemption timer data. */
146	struct
147	{
148	/** The handle to the periodic preemption timer. */
149	PRTTIMER pTimer;
150	/** Spinlock protecting the data below. */
151	RTSPINLOCK hSpinlock;
152	/** The smalles Hz that we need to care about. (static) */
153	uint32_t uMinHz;
154	/** The number of ticks between each historization. */
155	uint32_t cTicksHistoriziationInterval;
156	/** The current historization tick (counting up to
157	* cTicksHistoriziationInterval and then resetting). */
158	uint32_t iTickHistorization;
159	/** The current timer interval. This is set to 0 when inactive. */
160	uint32_t cNsInterval;
161	/** The current timer frequency. This is set to 0 when inactive. */
162	uint32_t uTimerHz;
163	/** The current max frequency reported by the EMTs.
164	* This gets historicize and reset by the timer callback. This is
165	* read without holding the spinlock, so needs atomic updating. */
166	uint32_t volatile uDesiredHz;
167	/** Whether the timer was started or not. */
168	bool volatile fStarted;
169	/** Set if we're starting timer. */
170	bool volatile fStarting;
171	/** The index of the next history entry (mod it). */
172	uint32_t iHzHistory;
173	/** Historicized uDesiredHz values. The array wraps around, new entries
174	* are added at iHzHistory. This is updated approximately every
175	* GVMMHOSTCPU_PPT_HIST_INTERVAL_NS by the timer callback. */
176	uint32_t aHzHistory[8];
177	/** Statistics counter for recording the number of interval changes. */
178	uint32_t cChanges;
179	/** Statistics counter for recording the number of timer starts. */
180	uint32_t cStarts;
181	} Ppt;
182	#endif /* GVMM_SCHED_WITH_PPT */
183
184	} GVMMHOSTCPU;
185	/** Pointer to the per host CPU GVMM data. */
186	typedef GVMMHOSTCPU *PGVMMHOSTCPU;
187	/** The GVMMHOSTCPU::u32Magic value (Petra, Tanya & Rachel Haden). */
188	#define GVMMHOSTCPU_MAGIC UINT32_C(0x19711011)
189	/** The interval on history entry should cover (approximately) give in
190	* nanoseconds. */
191	#define GVMMHOSTCPU_PPT_HIST_INTERVAL_NS UINT32_C(20000000)
192
193
194	/**
195	* The GVMM instance data.
196	*/
197	typedef struct GVMM
198	{
199	/** Eyecatcher / magic. */
200	uint32_t u32Magic;
201	/** The index of the head of the free handle chain. (0 is nil.) */
202	uint16_t volatile iFreeHead;
203	/** The index of the head of the active handle chain. (0 is nil.) */
204	uint16_t volatile iUsedHead;
205	/** The number of VMs. */
206	uint16_t volatile cVMs;
207	/** Alignment padding. */
208	uint16_t u16Reserved;
209	/** The number of EMTs. */
210	uint32_t volatile cEMTs;
211	/** The number of EMTs that have halted in GVMMR0SchedHalt. */
212	uint32_t volatile cHaltedEMTs;
213	/** Alignment padding. */
214	uint32_t u32Alignment;
215	/** When the next halted or sleeping EMT will wake up.
216	* This is set to 0 when it needs recalculating and to UINT64_MAX when
217	* there are no halted or sleeping EMTs in the GVMM. */
218	uint64_t uNsNextEmtWakeup;
219	/** The lock used to serialize VM creation, destruction and associated events that
220	* isn't performance critical. Owners may acquire the list lock. */
221	RTSEMFASTMUTEX CreateDestroyLock;
222	/** The lock used to serialize used list updates and accesses.
223	* This indirectly includes scheduling since the scheduler will have to walk the
224	* used list to examin running VMs. Owners may not acquire any other locks. */
225	RTSEMFASTMUTEX UsedLock;
226	/** The handle array.
227	* The size of this array defines the maximum number of currently running VMs.
228	* The first entry is unused as it represents the NIL handle. */
229	GVMHANDLE aHandles[GVMM_MAX_HANDLES];
230
231	/** @gcfgm{/GVMM/cEMTsMeansCompany, 32-bit, 0, UINT32_MAX, 1}
232	* The number of EMTs that means we no longer consider ourselves alone on a
233	* CPU/Core.
234	*/
235	uint32_t cEMTsMeansCompany;
236	/** @gcfgm{/GVMM/MinSleepAlone,32-bit, 0, 100000000, 750000, ns}
237	* The minimum sleep time for when we're alone, in nano seconds.
238	*/
239	uint32_t nsMinSleepAlone;
240	/** @gcfgm{/GVMM/MinSleepCompany,32-bit,0, 100000000, 15000, ns}
241	* The minimum sleep time for when we've got company, in nano seconds.
242	*/
243	uint32_t nsMinSleepCompany;
244	/** @gcfgm{/GVMM/EarlyWakeUp1, 32-bit, 0, 100000000, 25000, ns}
245	* The limit for the first round of early wakeups, given in nano seconds.
246	*/
247	uint32_t nsEarlyWakeUp1;
248	/** @gcfgm{/GVMM/EarlyWakeUp2, 32-bit, 0, 100000000, 50000, ns}
249	* The limit for the second round of early wakeups, given in nano seconds.
250	*/
251	uint32_t nsEarlyWakeUp2;
252
253	/** The number of entries in the host CPU array (aHostCpus). */
254	uint32_t cHostCpus;
255	/** Per host CPU data (variable length). */
256	GVMMHOSTCPU aHostCpus[1];
257	} GVMM;
258	/** Pointer to the GVMM instance data. */
259	typedef GVMM *PGVMM;
260
261	/** The GVMM::u32Magic value (Charlie Haden). */
262	#define GVMM_MAGIC UINT32_C(0x19370806)
263
264
265
266	/*******************************************************************************
267	* Global Variables *
268	*******************************************************************************/
269	/** Pointer to the GVMM instance data.
270	* (Just my general dislike for global variables.) */
271	static PGVMM g_pGVMM = NULL;
272
273	/** Macro for obtaining and validating the g_pGVMM pointer.
274	* On failure it will return from the invoking function with the specified return value.
275	*
276	* @param pGVMM The name of the pGVMM variable.
277	* @param rc The return value on failure. Use VERR_GVMM_INSTANCE for VBox
278	* status codes.
279	*/
280	#define GVMM_GET_VALID_INSTANCE(pGVMM, rc) \
281	do { \
282	(pGVMM) = g_pGVMM;\
283	AssertPtrReturn((pGVMM), (rc)); \
284	AssertMsgReturn((pGVMM)->u32Magic == GVMM_MAGIC, ("%p - %#x\n", (pGVMM), (pGVMM)->u32Magic), (rc)); \
285	} while (0)
286
287	/** Macro for obtaining and validating the g_pGVMM pointer, void function variant.
288	* On failure it will return from the invoking function.
289	*
290	* @param pGVMM The name of the pGVMM variable.
291	*/
292	#define GVMM_GET_VALID_INSTANCE_VOID(pGVMM) \
293	do { \
294	(pGVMM) = g_pGVMM;\
295	AssertPtrReturnVoid((pGVMM)); \
296	AssertMsgReturnVoid((pGVMM)->u32Magic == GVMM_MAGIC, ("%p - %#x\n", (pGVMM), (pGVMM)->u32Magic)); \
297	} while (0)
298
299
300	/*******************************************************************************
301	* Internal Functions *
302	*******************************************************************************/
303	static void gvmmR0InitPerVMData(PGVM pGVM);
304	static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void pvObj, void pvGVMM, void *pvHandle);
305	static int gvmmR0ByVM(PVM pVM, PGVM ppGVM, PGVMM ppGVMM, bool fTakeUsedLock);
306	static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM ppGVM, PGVMM ppGVMM);
307	#ifdef GVMM_SCHED_WITH_PPT
308	static DECLCALLBACK(void) gvmmR0SchedPeriodicPreemptionTimerCallback(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
309	#endif
310
311
312	/**
313	* Initializes the GVMM.
314	*
315	* This is called while owning the loader semaphore (see supdrvIOCtl_LdrLoad()).
316	*
317	* @returns VBox status code.
318	*/
319	GVMMR0DECL(int) GVMMR0Init(void)
320	{
321	LogFlow(("GVMMR0Init:\n"));
322
323	/*
324	* Allocate and initialize the instance data.
325	*/
326	uint32_t cHostCpus = RTMpGetArraySize();
327	AssertMsgReturn(cHostCpus > 0 && cHostCpus < _64K, ("%d", (int)cHostCpus), VERR_GVMM_HOST_CPU_RANGE);
328
329	PGVMM pGVMM = (PGVMM)RTMemAllocZ(RT_UOFFSETOF(GVMM, aHostCpus[cHostCpus]));
330	if (!pGVMM)
331	return VERR_NO_MEMORY;
332	int rc = RTSemFastMutexCreate(&pGVMM->CreateDestroyLock);
333	if (RT_SUCCESS(rc))
334	{
335	rc = RTSemFastMutexCreate(&pGVMM->UsedLock);
336	if (RT_SUCCESS(rc))
337	{
338	pGVMM->u32Magic = GVMM_MAGIC;
339	pGVMM->iUsedHead = 0;
340	pGVMM->iFreeHead = 1;
341
342	/* the nil handle */
343	pGVMM->aHandles[0].iSelf = 0;
344	pGVMM->aHandles[0].iNext = 0;
345
346	/* the tail */
347	unsigned i = RT_ELEMENTS(pGVMM->aHandles) - 1;
348	pGVMM->aHandles[i].iSelf = i;
349	pGVMM->aHandles[i].iNext = 0; /* nil */
350
351	/* the rest */
352	while (i-- > 1)
353	{
354	pGVMM->aHandles[i].iSelf = i;
355	pGVMM->aHandles[i].iNext = i + 1;
356	}
357
358	/* The default configuration values. */
359	uint32_t cNsResolution = RTSemEventMultiGetResolution();
360	pGVMM->cEMTsMeansCompany = 1; /** @todo should be adjusted to relative to the cpu count or something... */
361	if (cNsResolution >= 5*RT_NS_100US)
362	{
363	pGVMM->nsMinSleepAlone = 750000 /* ns (0.750 ms) /; /* @todo this should be adjusted to be 75% (or something) of the scheduler granularity... */
364	pGVMM->nsMinSleepCompany = 15000 /* ns (0.015 ms) */;
365	pGVMM->nsEarlyWakeUp1 = 25000 /* ns (0.025 ms) */;
366	pGVMM->nsEarlyWakeUp2 = 50000 /* ns (0.050 ms) */;
367	}
368	else if (cNsResolution > RT_NS_100US)
369	{
370	pGVMM->nsMinSleepAlone = cNsResolution / 2;
371	pGVMM->nsMinSleepCompany = cNsResolution / 4;
372	pGVMM->nsEarlyWakeUp1 = 0;
373	pGVMM->nsEarlyWakeUp2 = 0;
374	}
375	else
376	{
377	pGVMM->nsMinSleepAlone = 2000;
378	pGVMM->nsMinSleepCompany = 2000;
379	pGVMM->nsEarlyWakeUp1 = 0;
380	pGVMM->nsEarlyWakeUp2 = 0;
381	}
382
383	/* The host CPU data. */
384	pGVMM->cHostCpus = cHostCpus;
385	uint32_t iCpu = cHostCpus;
386	RTCPUSET PossibleSet;
387	RTMpGetSet(&PossibleSet);
388	while (iCpu-- > 0)
389	{
390	pGVMM->aHostCpus[iCpu].idxCpuSet = iCpu;
391	#ifdef GVMM_SCHED_WITH_PPT
392	pGVMM->aHostCpus[iCpu].Ppt.pTimer = NULL;
393	pGVMM->aHostCpus[iCpu].Ppt.hSpinlock = NIL_RTSPINLOCK;
394	pGVMM->aHostCpus[iCpu].Ppt.uMinHz = 5; /** @todo Add some API which figures this one out. (not that important) */
395	pGVMM->aHostCpus[iCpu].Ppt.cTicksHistoriziationInterval = 1;
396	//pGVMM->aHostCpus[iCpu].Ppt.iTickHistorization = 0;
397	//pGVMM->aHostCpus[iCpu].Ppt.cNsInterval = 0;
398	//pGVMM->aHostCpus[iCpu].Ppt.uTimerHz = 0;
399	//pGVMM->aHostCpus[iCpu].Ppt.uDesiredHz = 0;
400	//pGVMM->aHostCpus[iCpu].Ppt.fStarted = false;
401	//pGVMM->aHostCpus[iCpu].Ppt.fStarting = false;
402	//pGVMM->aHostCpus[iCpu].Ppt.iHzHistory = 0;
403	//pGVMM->aHostCpus[iCpu].Ppt.aHzHistory = {0};
404	#endif
405
406	if (RTCpuSetIsMember(&PossibleSet, iCpu))
407	{
408	pGVMM->aHostCpus[iCpu].idCpu = RTMpCpuIdFromSetIndex(iCpu);
409	pGVMM->aHostCpus[iCpu].u32Magic = GVMMHOSTCPU_MAGIC;
410
411	#ifdef GVMM_SCHED_WITH_PPT
412	rc = RTTimerCreateEx(&pGVMM->aHostCpus[iCpu].Ppt.pTimer,
413	5010001000 /* whatever */,
414	RTTIMER_FLAGS_CPU(iCpu) \| RTTIMER_FLAGS_HIGH_RES,
415	gvmmR0SchedPeriodicPreemptionTimerCallback,
416	&pGVMM->aHostCpus[iCpu]);
417	if (RT_SUCCESS(rc))
418	rc = RTSpinlockCreate(&pGVMM->aHostCpus[iCpu].Ppt.hSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "GVMM/CPU");
419	if (RT_FAILURE(rc))
420	{
421	while (iCpu < cHostCpus)
422	{
423	RTTimerDestroy(pGVMM->aHostCpus[iCpu].Ppt.pTimer);
424	RTSpinlockDestroy(pGVMM->aHostCpus[iCpu].Ppt.hSpinlock);
425	pGVMM->aHostCpus[iCpu].Ppt.hSpinlock = NIL_RTSPINLOCK;
426	iCpu++;
427	}
428	break;
429	}
430	#endif
431	}
432	else
433	{
434	pGVMM->aHostCpus[iCpu].idCpu = NIL_RTCPUID;
435	pGVMM->aHostCpus[iCpu].u32Magic = 0;
436	}
437	}
438	if (RT_SUCCESS(rc))
439	{
440	g_pGVMM = pGVMM;
441	LogFlow(("GVMMR0Init: pGVMM=%p cHostCpus=%u\n", pGVMM, cHostCpus));
442	return VINF_SUCCESS;
443	}
444
445	/* bail out. */
446	RTSemFastMutexDestroy(pGVMM->UsedLock);
447	pGVMM->UsedLock = NIL_RTSEMFASTMUTEX;
448	}
449	RTSemFastMutexDestroy(pGVMM->CreateDestroyLock);
450	pGVMM->CreateDestroyLock = NIL_RTSEMFASTMUTEX;
451	}
452
453	RTMemFree(pGVMM);
454	return rc;
455	}
456
457
458	/**
459	* Terminates the GVM.
460	*
461	* This is called while owning the loader semaphore (see supdrvLdrFree()).
462	* And unless something is wrong, there should be absolutely no VMs
463	* registered at this point.
464	*/
465	GVMMR0DECL(void) GVMMR0Term(void)
466	{
467	LogFlow(("GVMMR0Term:\n"));
468
469	PGVMM pGVMM = g_pGVMM;
470	g_pGVMM = NULL;
471	if (RT_UNLIKELY(!VALID_PTR(pGVMM)))
472	{
473	SUPR0Printf("GVMMR0Term: pGVMM=%p\n", pGVMM);
474	return;
475	}
476
477	/*
478	* First of all, stop all active timers.
479	*/
480	uint32_t cActiveTimers = 0;
481	uint32_t iCpu = pGVMM->cHostCpus;
482	while (iCpu-- > 0)
483	{
484	ASMAtomicWriteU32(&pGVMM->aHostCpus[iCpu].u32Magic, ~GVMMHOSTCPU_MAGIC);
485	#ifdef GVMM_SCHED_WITH_PPT
486	if ( pGVMM->aHostCpus[iCpu].Ppt.pTimer != NULL
487	&& RT_SUCCESS(RTTimerStop(pGVMM->aHostCpus[iCpu].Ppt.pTimer)))
488	cActiveTimers++;
489	#endif
490	}
491	if (cActiveTimers)
492	RTThreadSleep(1); /* fudge */
493
494	/*
495	* Invalidate the and free resources.
496	*/
497	pGVMM->u32Magic = ~GVMM_MAGIC;
498	RTSemFastMutexDestroy(pGVMM->UsedLock);
499	pGVMM->UsedLock = NIL_RTSEMFASTMUTEX;
500	RTSemFastMutexDestroy(pGVMM->CreateDestroyLock);
501	pGVMM->CreateDestroyLock = NIL_RTSEMFASTMUTEX;
502
503	pGVMM->iFreeHead = 0;
504	if (pGVMM->iUsedHead)
505	{
506	SUPR0Printf("GVMMR0Term: iUsedHead=%#x! (cVMs=%#x cEMTs=%#x)\n", pGVMM->iUsedHead, pGVMM->cVMs, pGVMM->cEMTs);
507	pGVMM->iUsedHead = 0;
508	}
509
510	#ifdef GVMM_SCHED_WITH_PPT
511	iCpu = pGVMM->cHostCpus;
512	while (iCpu-- > 0)
513	{
514	RTTimerDestroy(pGVMM->aHostCpus[iCpu].Ppt.pTimer);
515	pGVMM->aHostCpus[iCpu].Ppt.pTimer = NULL;
516	RTSpinlockDestroy(pGVMM->aHostCpus[iCpu].Ppt.hSpinlock);
517	pGVMM->aHostCpus[iCpu].Ppt.hSpinlock = NIL_RTSPINLOCK;
518	}
519	#endif
520
521	RTMemFree(pGVMM);
522	}
523
524
525	/**
526	* A quick hack for setting global config values.
527	*
528	* @returns VBox status code.
529	*
530	* @param pSession The session handle. Used for authentication.
531	* @param pszName The variable name.
532	* @param u64Value The new value.
533	*/
534	GVMMR0DECL(int) GVMMR0SetConfig(PSUPDRVSESSION pSession, const char *pszName, uint64_t u64Value)
535	{
536	/*
537	* Validate input.
538	*/
539	PGVMM pGVMM;
540	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
541	AssertPtrReturn(pSession, VERR_INVALID_HANDLE);
542	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
543
544	/*
545	* String switch time!
546	*/
547	if (strncmp(pszName, RT_STR_TUPLE("/GVMM/")))
548	return VERR_CFGM_VALUE_NOT_FOUND; /* borrow status codes from CFGM... */
549	int rc = VINF_SUCCESS;
550	pszName += sizeof("/GVMM/") - 1;
551	if (!strcmp(pszName, "cEMTsMeansCompany"))
552	{
553	if (u64Value <= UINT32_MAX)
554	pGVMM->cEMTsMeansCompany = u64Value;
555	else
556	rc = VERR_OUT_OF_RANGE;
557	}
558	else if (!strcmp(pszName, "MinSleepAlone"))
559	{
560	if (u64Value <= RT_NS_100MS)
561	pGVMM->nsMinSleepAlone = u64Value;
562	else
563	rc = VERR_OUT_OF_RANGE;
564	}
565	else if (!strcmp(pszName, "MinSleepCompany"))
566	{
567	if (u64Value <= RT_NS_100MS)
568	pGVMM->nsMinSleepCompany = u64Value;
569	else
570	rc = VERR_OUT_OF_RANGE;
571	}
572	else if (!strcmp(pszName, "EarlyWakeUp1"))
573	{
574	if (u64Value <= RT_NS_100MS)
575	pGVMM->nsEarlyWakeUp1 = u64Value;
576	else
577	rc = VERR_OUT_OF_RANGE;
578	}
579	else if (!strcmp(pszName, "EarlyWakeUp2"))
580	{
581	if (u64Value <= RT_NS_100MS)
582	pGVMM->nsEarlyWakeUp2 = u64Value;
583	else
584	rc = VERR_OUT_OF_RANGE;
585	}
586	else
587	rc = VERR_CFGM_VALUE_NOT_FOUND;
588	return rc;
589	}
590
591
592	/**
593	* A quick hack for getting global config values.
594	*
595	* @returns VBox status code.
596	*
597	* @param pSession The session handle. Used for authentication.
598	* @param pszName The variable name.
599	* @param u64Value The new value.
600	*/
601	GVMMR0DECL(int) GVMMR0QueryConfig(PSUPDRVSESSION pSession, const char pszName, uint64_t pu64Value)
602	{
603	/*
604	* Validate input.
605	*/
606	PGVMM pGVMM;
607	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
608	AssertPtrReturn(pSession, VERR_INVALID_HANDLE);
609	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
610	AssertPtrReturn(pu64Value, VERR_INVALID_POINTER);
611
612	/*
613	* String switch time!
614	*/
615	if (strncmp(pszName, RT_STR_TUPLE("/GVMM/")))
616	return VERR_CFGM_VALUE_NOT_FOUND; /* borrow status codes from CFGM... */
617	int rc = VINF_SUCCESS;
618	pszName += sizeof("/GVMM/") - 1;
619	if (!strcmp(pszName, "cEMTsMeansCompany"))
620	*pu64Value = pGVMM->cEMTsMeansCompany;
621	else if (!strcmp(pszName, "MinSleepAlone"))
622	*pu64Value = pGVMM->nsMinSleepAlone;
623	else if (!strcmp(pszName, "MinSleepCompany"))
624	*pu64Value = pGVMM->nsMinSleepCompany;
625	else if (!strcmp(pszName, "EarlyWakeUp1"))
626	*pu64Value = pGVMM->nsEarlyWakeUp1;
627	else if (!strcmp(pszName, "EarlyWakeUp2"))
628	*pu64Value = pGVMM->nsEarlyWakeUp2;
629	else
630	rc = VERR_CFGM_VALUE_NOT_FOUND;
631	return rc;
632	}
633
634
635	/**
636	* Try acquire the 'used' lock.
637	*
638	* @returns IPRT status code, see RTSemFastMutexRequest.
639	* @param pGVMM The GVMM instance data.
640	*/
641	DECLINLINE(int) gvmmR0UsedLock(PGVMM pGVMM)
642	{
643	LogFlow(("++gvmmR0UsedLock(%p)\n", pGVMM));
644	int rc = RTSemFastMutexRequest(pGVMM->UsedLock);
645	LogFlow(("gvmmR0UsedLock(%p)->%Rrc\n", pGVMM, rc));
646	return rc;
647	}
648
649
650	/**
651	* Release the 'used' lock.
652	*
653	* @returns IPRT status code, see RTSemFastMutexRelease.
654	* @param pGVMM The GVMM instance data.
655	*/
656	DECLINLINE(int) gvmmR0UsedUnlock(PGVMM pGVMM)
657	{
658	LogFlow(("--gvmmR0UsedUnlock(%p)\n", pGVMM));
659	int rc = RTSemFastMutexRelease(pGVMM->UsedLock);
660	AssertRC(rc);
661	return rc;
662	}
663
664
665	/**
666	* Try acquire the 'create & destroy' lock.
667	*
668	* @returns IPRT status code, see RTSemFastMutexRequest.
669	* @param pGVMM The GVMM instance data.
670	*/
671	DECLINLINE(int) gvmmR0CreateDestroyLock(PGVMM pGVMM)
672	{
673	LogFlow(("++gvmmR0CreateDestroyLock(%p)\n", pGVMM));
674	int rc = RTSemFastMutexRequest(pGVMM->CreateDestroyLock);
675	LogFlow(("gvmmR0CreateDestroyLock(%p)->%Rrc\n", pGVMM, rc));
676	return rc;
677	}
678
679
680	/**
681	* Release the 'create & destroy' lock.
682	*
683	* @returns IPRT status code, see RTSemFastMutexRequest.
684	* @param pGVMM The GVMM instance data.
685	*/
686	DECLINLINE(int) gvmmR0CreateDestroyUnlock(PGVMM pGVMM)
687	{
688	LogFlow(("--gvmmR0CreateDestroyUnlock(%p)\n", pGVMM));
689	int rc = RTSemFastMutexRelease(pGVMM->CreateDestroyLock);
690	AssertRC(rc);
691	return rc;
692	}
693
694
695	/**
696	* Request wrapper for the GVMMR0CreateVM API.
697	*
698	* @returns VBox status code.
699	* @param pReq The request buffer.
700	*/
701	GVMMR0DECL(int) GVMMR0CreateVMReq(PGVMMCREATEVMREQ pReq)
702	{
703	/*
704	* Validate the request.
705	*/
706	if (!VALID_PTR(pReq))
707	return VERR_INVALID_POINTER;
708	if (pReq->Hdr.cbReq != sizeof(*pReq))
709	return VERR_INVALID_PARAMETER;
710	if (!VALID_PTR(pReq->pSession))
711	return VERR_INVALID_POINTER;
712
713	/*
714	* Execute it.
715	*/
716	PVM pVM;
717	pReq->pVMR0 = NULL;
718	pReq->pVMR3 = NIL_RTR3PTR;
719	int rc = GVMMR0CreateVM(pReq->pSession, pReq->cCpus, &pVM);
720	if (RT_SUCCESS(rc))
721	{
722	pReq->pVMR0 = pVM;
723	pReq->pVMR3 = pVM->pVMR3;
724	}
725	return rc;
726	}
727
728
729	/**
730	* Allocates the VM structure and registers it with GVM.
731	*
732	* The caller will become the VM owner and there by the EMT.
733	*
734	* @returns VBox status code.
735	* @param pSession The support driver session.
736	* @param cCpus Number of virtual CPUs for the new VM.
737	* @param ppVM Where to store the pointer to the VM structure.
738	*
739	* @thread EMT.
740	*/
741	GVMMR0DECL(int) GVMMR0CreateVM(PSUPDRVSESSION pSession, uint32_t cCpus, PVM *ppVM)
742	{
743	LogFlow(("GVMMR0CreateVM: pSession=%p\n", pSession));
744	PGVMM pGVMM;
745	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
746
747	AssertPtrReturn(ppVM, VERR_INVALID_POINTER);
748	*ppVM = NULL;
749
750	if ( cCpus == 0
751	\|\| cCpus > VMM_MAX_CPU_COUNT)
752	return VERR_INVALID_PARAMETER;
753
754	RTNATIVETHREAD hEMT0 = RTThreadNativeSelf();
755	AssertReturn(hEMT0 != NIL_RTNATIVETHREAD, VERR_GVMM_BROKEN_IPRT);
756	RTPROCESS ProcId = RTProcSelf();
757	AssertReturn(ProcId != NIL_RTPROCESS, VERR_GVMM_BROKEN_IPRT);
758
759	/*
760	* The whole allocation process is protected by the lock.
761	*/
762	int rc = gvmmR0CreateDestroyLock(pGVMM);
763	AssertRCReturn(rc, rc);
764
765	/*
766	* Allocate a handle first so we don't waste resources unnecessarily.
767	*/
768	uint16_t iHandle = pGVMM->iFreeHead;
769	if (iHandle)
770	{
771	PGVMHANDLE pHandle = &pGVMM->aHandles[iHandle];
772
773	/* consistency checks, a bit paranoid as always. */
774	if ( !pHandle->pVM
775	&& !pHandle->pGVM
776	&& !pHandle->pvObj
777	&& pHandle->iSelf == iHandle)
778	{
779	pHandle->pvObj = SUPR0ObjRegister(pSession, SUPDRVOBJTYPE_VM, gvmmR0HandleObjDestructor, pGVMM, pHandle);
780	if (pHandle->pvObj)
781	{
782	/*
783	* Move the handle from the free to used list and perform permission checks.
784	*/
785	rc = gvmmR0UsedLock(pGVMM);
786	AssertRC(rc);
787
788	pGVMM->iFreeHead = pHandle->iNext;
789	pHandle->iNext = pGVMM->iUsedHead;
790	pGVMM->iUsedHead = iHandle;
791	pGVMM->cVMs++;
792
793	pHandle->pVM = NULL;
794	pHandle->pGVM = NULL;
795	pHandle->pSession = pSession;
796	pHandle->hEMT0 = NIL_RTNATIVETHREAD;
797	pHandle->ProcId = NIL_RTPROCESS;
798
799	gvmmR0UsedUnlock(pGVMM);
800
801	rc = SUPR0ObjVerifyAccess(pHandle->pvObj, pSession, NULL);
802	if (RT_SUCCESS(rc))
803	{
804	/*
805	* Allocate the global VM structure (GVM) and initialize it.
806	*/
807	PGVM pGVM = (PGVM)RTMemAllocZ(RT_UOFFSETOF(GVM, aCpus[cCpus]));
808	if (pGVM)
809	{
810	pGVM->u32Magic = GVM_MAGIC;
811	pGVM->hSelf = iHandle;
812	pGVM->pVM = NULL;
813	pGVM->cCpus = cCpus;
814
815	gvmmR0InitPerVMData(pGVM);
816	GMMR0InitPerVMData(pGVM);
817
818	/*
819	* Allocate the shared VM structure and associated page array.
820	*/
821	const uint32_t cbVM = RT_UOFFSETOF(VM, aCpus[cCpus]);
822	const uint32_t cPages = RT_ALIGN_32(cbVM, PAGE_SIZE) >> PAGE_SHIFT;
823	rc = RTR0MemObjAllocLow(&pGVM->gvmm.s.VMMemObj, cPages << PAGE_SHIFT, false /* fExecutable */);
824	if (RT_SUCCESS(rc))
825	{
826	PVM pVM = (PVM)RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj); AssertPtr(pVM);
827	memset(pVM, 0, cPages << PAGE_SHIFT);
828	pVM->enmVMState = VMSTATE_CREATING;
829	pVM->pVMR0 = pVM;
830	pVM->pSession = pSession;
831	pVM->hSelf = iHandle;
832	pVM->cbSelf = cbVM;
833	pVM->cCpus = cCpus;
834	pVM->uCpuExecutionCap = 100; /* default is no cap. */
835	pVM->offVMCPU = RT_UOFFSETOF(VM, aCpus);
836	AssertCompileMemberAlignment(VM, cpum, 64);
837	AssertCompileMemberAlignment(VM, tm, 64);
838	AssertCompileMemberAlignment(VM, aCpus, PAGE_SIZE);
839
840	rc = RTR0MemObjAllocPage(&pGVM->gvmm.s.VMPagesMemObj, cPages * sizeof(SUPPAGE), false /* fExecutable */);
841	if (RT_SUCCESS(rc))
842	{
843	PSUPPAGE paPages = (PSUPPAGE)RTR0MemObjAddress(pGVM->gvmm.s.VMPagesMemObj); AssertPtr(paPages);
844	for (uint32_t iPage = 0; iPage < cPages; iPage++)
845	{
846	paPages[iPage].uReserved = 0;
847	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pGVM->gvmm.s.VMMemObj, iPage);
848	Assert(paPages[iPage].Phys != NIL_RTHCPHYS);
849	}
850
851	/*
852	* Map them into ring-3.
853	*/
854	rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMMapObj, pGVM->gvmm.s.VMMemObj, (RTR3PTR)-1, 0,
855	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
856	if (RT_SUCCESS(rc))
857	{
858	pVM->pVMR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMMapObj);
859	AssertPtr((void *)pVM->pVMR3);
860
861	/* Initialize all the VM pointers. */
862	for (uint32_t i = 0; i < cCpus; i++)
863	{
864	pVM->aCpus[i].pVMR0 = pVM;
865	pVM->aCpus[i].pVMR3 = pVM->pVMR3;
866	pVM->aCpus[i].idHostCpu = NIL_RTCPUID;
867	pVM->aCpus[i].hNativeThreadR0 = NIL_RTNATIVETHREAD;
868	}
869
870	rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMPagesMapObj, pGVM->gvmm.s.VMPagesMemObj, (RTR3PTR)-1,
871	0 /* uAlignment */, RTMEM_PROT_READ \| RTMEM_PROT_WRITE,
872	NIL_RTR0PROCESS);
873	if (RT_SUCCESS(rc))
874	{
875	pVM->paVMPagesR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMPagesMapObj);
876	AssertPtr((void *)pVM->paVMPagesR3);
877
878	/* complete the handle - take the UsedLock sem just to be careful. */
879	rc = gvmmR0UsedLock(pGVMM);
880	AssertRC(rc);
881
882	pHandle->pVM = pVM;
883	pHandle->pGVM = pGVM;
884	pHandle->hEMT0 = hEMT0;
885	pHandle->ProcId = ProcId;
886	pGVM->pVM = pVM;
887	pGVM->aCpus[0].hEMT = hEMT0;
888	pVM->aCpus[0].hNativeThreadR0 = hEMT0;
889	pGVMM->cEMTs += cCpus;
890
891	rc = VMMR0ThreadCtxHooksCreate(&pVM->aCpus[0]);
892	if (RT_SUCCESS(rc))
893	{
894	VBOXVMM_R0_GVMM_VM_CREATED(pGVM, pVM, ProcId, (void *)hEMT0, cCpus);
895
896	gvmmR0UsedUnlock(pGVMM);
897	gvmmR0CreateDestroyUnlock(pGVMM);
898
899	*ppVM = pVM;
900	Log(("GVMMR0CreateVM: pVM=%p pVMR3=%p pGVM=%p hGVM=%d\n", pVM, pVM->pVMR3, pGVM, iHandle));
901	return VINF_SUCCESS;
902	}
903	}
904
905	RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */);
906	pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
907	}
908	RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */);
909	pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
910	}
911	RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */);
912	pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
913	}
914	}
915	}
916	/* else: The user wasn't permitted to create this VM. */
917
918	/*
919	* The handle will be freed by gvmmR0HandleObjDestructor as we release the
920	* object reference here. A little extra mess because of non-recursive lock.
921	*/
922	void *pvObj = pHandle->pvObj;
923	pHandle->pvObj = NULL;
924	gvmmR0CreateDestroyUnlock(pGVMM);
925
926	SUPR0ObjRelease(pvObj, pSession);
927
928	SUPR0Printf("GVMMR0CreateVM: failed, rc=%d\n", rc);
929	return rc;
930	}
931
932	rc = VERR_NO_MEMORY;
933	}
934	else
935	rc = VERR_GVMM_IPE_1;
936	}
937	else
938	rc = VERR_GVM_TOO_MANY_VMS;
939
940	gvmmR0CreateDestroyUnlock(pGVMM);
941	return rc;
942	}
943
944
945	/**
946	* Initializes the per VM data belonging to GVMM.
947	*
948	* @param pGVM Pointer to the global VM structure.
949	*/
950	static void gvmmR0InitPerVMData(PGVM pGVM)
951	{
952	AssertCompile(RT_SIZEOFMEMB(GVM,gvmm.s) <= RT_SIZEOFMEMB(GVM,gvmm.padding));
953	AssertCompile(RT_SIZEOFMEMB(GVMCPU,gvmm.s) <= RT_SIZEOFMEMB(GVMCPU,gvmm.padding));
954	pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
955	pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
956	pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
957	pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
958	pGVM->gvmm.s.fDoneVMMR0Init = false;
959	pGVM->gvmm.s.fDoneVMMR0Term = false;
960
961	for (VMCPUID i = 0; i < pGVM->cCpus; i++)
962	{
963	pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
964	pGVM->aCpus[i].hEMT = NIL_RTNATIVETHREAD;
965	}
966	}
967
968
969	/**
970	* Does the VM initialization.
971	*
972	* @returns VBox status code.
973	* @param pVM Pointer to the VM.
974	*/
975	GVMMR0DECL(int) GVMMR0InitVM(PVM pVM)
976	{
977	LogFlow(("GVMMR0InitVM: pVM=%p\n", pVM));
978
979	/*
980	* Validate the VM structure, state and handle.
981	*/
982	PGVM pGVM;
983	PGVMM pGVMM;
984	int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
985	if (RT_SUCCESS(rc))
986	{
987	if ( !pGVM->gvmm.s.fDoneVMMR0Init
988	&& pGVM->aCpus[0].gvmm.s.HaltEventMulti == NIL_RTSEMEVENTMULTI)
989	{
990	for (VMCPUID i = 0; i < pGVM->cCpus; i++)
991	{
992	rc = RTSemEventMultiCreate(&pGVM->aCpus[i].gvmm.s.HaltEventMulti);
993	if (RT_FAILURE(rc))
994	{
995	pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
996	break;
997	}
998	}
999	}
1000	else
1001	rc = VERR_WRONG_ORDER;
1002	}
1003
1004	LogFlow(("GVMMR0InitVM: returns %Rrc\n", rc));
1005	return rc;
1006	}
1007
1008
1009	/**
1010	* Indicates that we're done with the ring-0 initialization
1011	* of the VM.
1012	*
1013	* @param pVM Pointer to the VM.
1014	* @thread EMT(0)
1015	*/
1016	GVMMR0DECL(void) GVMMR0DoneInitVM(PVM pVM)
1017	{
1018	/* Validate the VM structure, state and handle. */
1019	PGVM pGVM;
1020	PGVMM pGVMM;
1021	int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
1022	AssertRCReturnVoid(rc);
1023
1024	/* Set the indicator. */
1025	pGVM->gvmm.s.fDoneVMMR0Init = true;
1026	}
1027
1028
1029	/**
1030	* Indicates that we're doing the ring-0 termination of the VM.
1031	*
1032	* @returns true if termination hasn't been done already, false if it has.
1033	* @param pVM Pointer to the VM.
1034	* @param pGVM Pointer to the global VM structure. Optional.
1035	* @thread EMT(0)
1036	*/
1037	GVMMR0DECL(bool) GVMMR0DoingTermVM(PVM pVM, PGVM pGVM)
1038	{
1039	/* Validate the VM structure, state and handle. */
1040	AssertPtrNullReturn(pGVM, false);
1041	AssertReturn(!pGVM \|\| pGVM->u32Magic == GVM_MAGIC, false);
1042	if (!pGVM)
1043	{
1044	PGVMM pGVMM;
1045	int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
1046	AssertRCReturn(rc, false);
1047	}
1048
1049	/* Set the indicator. */
1050	if (pGVM->gvmm.s.fDoneVMMR0Term)
1051	return false;
1052	pGVM->gvmm.s.fDoneVMMR0Term = true;
1053	return true;
1054	}
1055
1056
1057	/**
1058	* Destroys the VM, freeing all associated resources (the ring-0 ones anyway).
1059	*
1060	* This is call from the vmR3DestroyFinalBit and from a error path in VMR3Create,
1061	* and the caller is not the EMT thread, unfortunately. For security reasons, it
1062	* would've been nice if the caller was actually the EMT thread or that we somehow
1063	* could've associated the calling thread with the VM up front.
1064	*
1065	* @returns VBox status code.
1066	* @param pVM Pointer to the VM.
1067	*
1068	* @thread EMT(0) if it's associated with the VM, otherwise any thread.
1069	*/
1070	GVMMR0DECL(int) GVMMR0DestroyVM(PVM pVM)
1071	{
1072	LogFlow(("GVMMR0DestroyVM: pVM=%p\n", pVM));
1073	PGVMM pGVMM;
1074	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
1075
1076	/*
1077	* Validate the VM structure, state and caller.
1078	*/
1079	AssertPtrReturn(pVM, VERR_INVALID_POINTER);
1080	AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
1081	AssertMsgReturn(pVM->enmVMState >= VMSTATE_CREATING && pVM->enmVMState <= VMSTATE_TERMINATED, ("%d\n", pVM->enmVMState),
1082	VERR_WRONG_ORDER);
1083
1084	uint32_t hGVM = pVM->hSelf;
1085	AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
1086	AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
1087
1088	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1089	AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
1090
1091	RTPROCESS ProcId = RTProcSelf();
1092	RTNATIVETHREAD hSelf = RTThreadNativeSelf();
1093	AssertReturn( ( pHandle->hEMT0 == hSelf
1094	&& pHandle->ProcId == ProcId)
1095	\|\| pHandle->hEMT0 == NIL_RTNATIVETHREAD, VERR_NOT_OWNER);
1096
1097	/*
1098	* Lookup the handle and destroy the object.
1099	* Since the lock isn't recursive and we'll have to leave it before dereferencing the
1100	* object, we take some precautions against racing callers just in case...
1101	*/
1102	int rc = gvmmR0CreateDestroyLock(pGVMM);
1103	AssertRC(rc);
1104
1105	/* Be careful here because we might theoretically be racing someone else cleaning up. */
1106	if ( pHandle->pVM == pVM
1107	&& ( ( pHandle->hEMT0 == hSelf
1108	&& pHandle->ProcId == ProcId)
1109	\|\| pHandle->hEMT0 == NIL_RTNATIVETHREAD)
1110	&& VALID_PTR(pHandle->pvObj)
1111	&& VALID_PTR(pHandle->pSession)
1112	&& VALID_PTR(pHandle->pGVM)
1113	&& pHandle->pGVM->u32Magic == GVM_MAGIC)
1114	{
1115	void *pvObj = pHandle->pvObj;
1116	pHandle->pvObj = NULL;
1117	gvmmR0CreateDestroyUnlock(pGVMM);
1118
1119	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1120	{
1121	/** @todo Can we busy wait here for all thread-context hooks to be
1122	* deregistered before releasing (destroying) it? Only until we find a
1123	* solution for not deregistering hooks everytime we're leaving HMR0
1124	* context. */
1125	VMMR0ThreadCtxHooksRelease(&pVM->aCpus[idCpu]);
1126	}
1127
1128	SUPR0ObjRelease(pvObj, pHandle->pSession);
1129	}
1130	else
1131	{
1132	SUPR0Printf("GVMMR0DestroyVM: pHandle=%p:{.pVM=%p, .hEMT0=%p, .ProcId=%u, .pvObj=%p} pVM=%p hSelf=%p\n",
1133	pHandle, pHandle->pVM, pHandle->hEMT0, pHandle->ProcId, pHandle->pvObj, pVM, hSelf);
1134	gvmmR0CreateDestroyUnlock(pGVMM);
1135	rc = VERR_GVMM_IPE_2;
1136	}
1137
1138	return rc;
1139	}
1140
1141
1142	/**
1143	* Performs VM cleanup task as part of object destruction.
1144	*
1145	* @param pGVM The GVM pointer.
1146	*/
1147	static void gvmmR0CleanupVM(PGVM pGVM)
1148	{
1149	if ( pGVM->gvmm.s.fDoneVMMR0Init
1150	&& !pGVM->gvmm.s.fDoneVMMR0Term)
1151	{
1152	if ( pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ
1153	&& RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj) == pGVM->pVM)
1154	{
1155	LogFlow(("gvmmR0CleanupVM: Calling VMMR0TermVM\n"));
1156	VMMR0TermVM(pGVM->pVM, pGVM);
1157	}
1158	else
1159	AssertMsgFailed(("gvmmR0CleanupVM: VMMemObj=%p pVM=%p\n", pGVM->gvmm.s.VMMemObj, pGVM->pVM));
1160	}
1161
1162	GMMR0CleanupVM(pGVM);
1163	}
1164
1165
1166	/**
1167	* Handle destructor.
1168	*
1169	* @param pvGVMM The GVM instance pointer.
1170	* @param pvHandle The handle pointer.
1171	*/
1172	static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void pvObj, void pvGVMM, void *pvHandle)
1173	{
1174	LogFlow(("gvmmR0HandleObjDestructor: %p %p %p\n", pvObj, pvGVMM, pvHandle));
1175
1176	/*
1177	* Some quick, paranoid, input validation.
1178	*/
1179	PGVMHANDLE pHandle = (PGVMHANDLE)pvHandle;
1180	AssertPtr(pHandle);
1181	PGVMM pGVMM = (PGVMM)pvGVMM;
1182	Assert(pGVMM == g_pGVMM);
1183	const uint16_t iHandle = pHandle - &pGVMM->aHandles[0];
1184	if ( !iHandle
1185	\|\| iHandle >= RT_ELEMENTS(pGVMM->aHandles)
1186	\|\| iHandle != pHandle->iSelf)
1187	{
1188	SUPR0Printf("GVM: handle %d is out of range or corrupt (iSelf=%d)!\n", iHandle, pHandle->iSelf);
1189	return;
1190	}
1191
1192	int rc = gvmmR0CreateDestroyLock(pGVMM);
1193	AssertRC(rc);
1194	rc = gvmmR0UsedLock(pGVMM);
1195	AssertRC(rc);
1196
1197	/*
1198	* This is a tad slow but a doubly linked list is too much hassle.
1199	*/
1200	if (RT_UNLIKELY(pHandle->iNext >= RT_ELEMENTS(pGVMM->aHandles)))
1201	{
1202	SUPR0Printf("GVM: used list index %d is out of range!\n", pHandle->iNext);
1203	gvmmR0UsedUnlock(pGVMM);
1204	gvmmR0CreateDestroyUnlock(pGVMM);
1205	return;
1206	}
1207
1208	if (pGVMM->iUsedHead == iHandle)
1209	pGVMM->iUsedHead = pHandle->iNext;
1210	else
1211	{
1212	uint16_t iPrev = pGVMM->iUsedHead;
1213	int c = RT_ELEMENTS(pGVMM->aHandles) + 2;
1214	while (iPrev)
1215	{
1216	if (RT_UNLIKELY(iPrev >= RT_ELEMENTS(pGVMM->aHandles)))
1217	{
1218	SUPR0Printf("GVM: used list index %d is out of range!\n", iPrev);
1219	gvmmR0UsedUnlock(pGVMM);
1220	gvmmR0CreateDestroyUnlock(pGVMM);
1221	return;
1222	}
1223	if (RT_UNLIKELY(c-- <= 0))
1224	{
1225	iPrev = 0;
1226	break;
1227	}
1228
1229	if (pGVMM->aHandles[iPrev].iNext == iHandle)
1230	break;
1231	iPrev = pGVMM->aHandles[iPrev].iNext;
1232	}
1233	if (!iPrev)
1234	{
1235	SUPR0Printf("GVM: can't find the handle previous previous of %d!\n", pHandle->iSelf);
1236	gvmmR0UsedUnlock(pGVMM);
1237	gvmmR0CreateDestroyUnlock(pGVMM);
1238	return;
1239	}
1240
1241	Assert(pGVMM->aHandles[iPrev].iNext == iHandle);
1242	pGVMM->aHandles[iPrev].iNext = pHandle->iNext;
1243	}
1244	pHandle->iNext = 0;
1245	pGVMM->cVMs--;
1246
1247	/*
1248	* Do the global cleanup round.
1249	*/
1250	PGVM pGVM = pHandle->pGVM;
1251	if ( VALID_PTR(pGVM)
1252	&& pGVM->u32Magic == GVM_MAGIC)
1253	{
1254	pGVMM->cEMTs -= pGVM->cCpus;
1255	gvmmR0UsedUnlock(pGVMM);
1256
1257	gvmmR0CleanupVM(pGVM);
1258
1259	/*
1260	* Do the GVMM cleanup - must be done last.
1261	*/
1262	/* The VM and VM pages mappings/allocations. */
1263	if (pGVM->gvmm.s.VMPagesMapObj != NIL_RTR0MEMOBJ)
1264	{
1265	rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMapObj, false /* fFreeMappings */); AssertRC(rc);
1266	pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
1267	}
1268
1269	if (pGVM->gvmm.s.VMMapObj != NIL_RTR0MEMOBJ)
1270	{
1271	rc = RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */); AssertRC(rc);
1272	pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
1273	}
1274
1275	if (pGVM->gvmm.s.VMPagesMemObj != NIL_RTR0MEMOBJ)
1276	{
1277	rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */); AssertRC(rc);
1278	pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
1279	}
1280
1281	if (pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ)
1282	{
1283	rc = RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */); AssertRC(rc);
1284	pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
1285	}
1286
1287	for (VMCPUID i = 0; i < pGVM->cCpus; i++)
1288	{
1289	if (pGVM->aCpus[i].gvmm.s.HaltEventMulti != NIL_RTSEMEVENTMULTI)
1290	{
1291	rc = RTSemEventMultiDestroy(pGVM->aCpus[i].gvmm.s.HaltEventMulti); AssertRC(rc);
1292	pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
1293	}
1294	}
1295
1296	/* the GVM structure itself. */
1297	pGVM->u32Magic \|= UINT32_C(0x80000000);
1298	RTMemFree(pGVM);
1299
1300	/* Re-acquire the UsedLock before freeing the handle since we're updating handle fields. */
1301	rc = gvmmR0UsedLock(pGVMM);
1302	AssertRC(rc);
1303	}
1304	/* else: GVMMR0CreateVM cleanup. */
1305
1306	/*
1307	* Free the handle.
1308	*/
1309	pHandle->iNext = pGVMM->iFreeHead;
1310	pGVMM->iFreeHead = iHandle;
1311	ASMAtomicWriteNullPtr(&pHandle->pGVM);
1312	ASMAtomicWriteNullPtr(&pHandle->pVM);
1313	ASMAtomicWriteNullPtr(&pHandle->pvObj);
1314	ASMAtomicWriteNullPtr(&pHandle->pSession);
1315	ASMAtomicWriteHandle(&pHandle->hEMT0, NIL_RTNATIVETHREAD);
1316	ASMAtomicWriteU32(&pHandle->ProcId, NIL_RTPROCESS);
1317
1318	gvmmR0UsedUnlock(pGVMM);
1319	gvmmR0CreateDestroyUnlock(pGVMM);
1320	LogFlow(("gvmmR0HandleObjDestructor: returns\n"));
1321	}
1322
1323
1324	/**
1325	* Registers the calling thread as the EMT of a Virtual CPU.
1326	*
1327	* Note that VCPU 0 is automatically registered during VM creation.
1328	*
1329	* @returns VBox status code
1330	* @param pVM Pointer to the VM.
1331	* @param idCpu VCPU id.
1332	*/
1333	GVMMR0DECL(int) GVMMR0RegisterVCpu(PVM pVM, VMCPUID idCpu)
1334	{
1335	AssertReturn(idCpu != 0, VERR_NOT_OWNER);
1336
1337	/*
1338	* Validate the VM structure, state and handle.
1339	*/
1340	PGVM pGVM;
1341	PGVMM pGVMM;
1342	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, false /* fTakeUsedLock */);
1343	if (RT_FAILURE(rc))
1344	return rc;
1345
1346	AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
1347	AssertReturn(pGVM->aCpus[idCpu].hEMT == NIL_RTNATIVETHREAD, VERR_ACCESS_DENIED);
1348	Assert(pGVM->cCpus == pVM->cCpus);
1349	Assert(pVM->aCpus[idCpu].hNativeThreadR0 == NIL_RTNATIVETHREAD);
1350
1351	pVM->aCpus[idCpu].hNativeThreadR0 = pGVM->aCpus[idCpu].hEMT = RTThreadNativeSelf();
1352
1353	rc = VMMR0ThreadCtxHooksCreate(&pVM->aCpus[idCpu]);
1354	return rc;
1355	}
1356
1357
1358	/**
1359	* Lookup a GVM structure by its handle.
1360	*
1361	* @returns The GVM pointer on success, NULL on failure.
1362	* @param hGVM The global VM handle. Asserts on bad handle.
1363	*/
1364	GVMMR0DECL(PGVM) GVMMR0ByHandle(uint32_t hGVM)
1365	{
1366	PGVMM pGVMM;
1367	GVMM_GET_VALID_INSTANCE(pGVMM, NULL);
1368
1369	/*
1370	* Validate.
1371	*/
1372	AssertReturn(hGVM != NIL_GVM_HANDLE, NULL);
1373	AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), NULL);
1374
1375	/*
1376	* Look it up.
1377	*/
1378	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1379	AssertPtrReturn(pHandle->pVM, NULL);
1380	AssertPtrReturn(pHandle->pvObj, NULL);
1381	PGVM pGVM = pHandle->pGVM;
1382	AssertPtrReturn(pGVM, NULL);
1383	AssertReturn(pGVM->pVM == pHandle->pVM, NULL);
1384
1385	return pHandle->pGVM;
1386	}
1387
1388
1389	/**
1390	* Lookup a GVM structure by the shared VM structure.
1391	*
1392	* The calling thread must be in the same process as the VM. All current lookups
1393	* are by threads inside the same process, so this will not be an issue.
1394	*
1395	* @returns VBox status code.
1396	* @param pVM Pointer to the VM.
1397	* @param ppGVM Where to store the GVM pointer.
1398	* @param ppGVMM Where to store the pointer to the GVMM instance data.
1399	* @param fTakeUsedLock Whether to take the used lock or not.
1400	* Be very careful if not taking the lock as it's possible that
1401	* the VM will disappear then.
1402	*
1403	* @remark This will not assert on an invalid pVM but try return silently.
1404	*/
1405	static int gvmmR0ByVM(PVM pVM, PGVM ppGVM, PGVMM ppGVMM, bool fTakeUsedLock)
1406	{
1407	RTPROCESS ProcId = RTProcSelf();
1408	PGVMM pGVMM;
1409	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
1410
1411	/*
1412	* Validate.
1413	*/
1414	if (RT_UNLIKELY( !VALID_PTR(pVM)
1415	\|\| ((uintptr_t)pVM & PAGE_OFFSET_MASK)))
1416	return VERR_INVALID_POINTER;
1417	if (RT_UNLIKELY( pVM->enmVMState < VMSTATE_CREATING
1418	\|\| pVM->enmVMState >= VMSTATE_TERMINATED))
1419	return VERR_INVALID_POINTER;
1420
1421	uint16_t hGVM = pVM->hSelf;
1422	if (RT_UNLIKELY( hGVM == NIL_GVM_HANDLE
1423	\|\| hGVM >= RT_ELEMENTS(pGVMM->aHandles)))
1424	return VERR_INVALID_HANDLE;
1425
1426	/*
1427	* Look it up.
1428	*/
1429	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1430	PGVM pGVM;
1431	if (fTakeUsedLock)
1432	{
1433	int rc = gvmmR0UsedLock(pGVMM);
1434	AssertRCReturn(rc, rc);
1435
1436	pGVM = pHandle->pGVM;
1437	if (RT_UNLIKELY( pHandle->pVM != pVM
1438	\|\| pHandle->ProcId != ProcId
1439	\|\| !VALID_PTR(pHandle->pvObj)
1440	\|\| !VALID_PTR(pGVM)
1441	\|\| pGVM->pVM != pVM))
1442	{
1443	gvmmR0UsedUnlock(pGVMM);
1444	return VERR_INVALID_HANDLE;
1445	}
1446	}
1447	else
1448	{
1449	if (RT_UNLIKELY(pHandle->pVM != pVM))
1450	return VERR_INVALID_HANDLE;
1451	if (RT_UNLIKELY(pHandle->ProcId != ProcId))
1452	return VERR_INVALID_HANDLE;
1453	if (RT_UNLIKELY(!VALID_PTR(pHandle->pvObj)))
1454	return VERR_INVALID_HANDLE;
1455
1456	pGVM = pHandle->pGVM;
1457	if (RT_UNLIKELY(!VALID_PTR(pGVM)))
1458	return VERR_INVALID_HANDLE;
1459	if (RT_UNLIKELY(pGVM->pVM != pVM))
1460	return VERR_INVALID_HANDLE;
1461	}
1462
1463	*ppGVM = pGVM;
1464	*ppGVMM = pGVMM;
1465	return VINF_SUCCESS;
1466	}
1467
1468
1469	/**
1470	* Lookup a GVM structure by the shared VM structure.
1471	*
1472	* @returns VBox status code.
1473	* @param pVM Pointer to the VM.
1474	* @param ppGVM Where to store the GVM pointer.
1475	*
1476	* @remark This will not take the 'used'-lock because it doesn't do
1477	* nesting and this function will be used from under the lock.
1478	*/
1479	GVMMR0DECL(int) GVMMR0ByVM(PVM pVM, PGVM *ppGVM)
1480	{
1481	PGVMM pGVMM;
1482	return gvmmR0ByVM(pVM, ppGVM, &pGVMM, false /* fTakeUsedLock */);
1483	}
1484
1485
1486	/**
1487	* Lookup a GVM structure by the shared VM structure and ensuring that the
1488	* caller is an EMT thread.
1489	*
1490	* @returns VBox status code.
1491	* @param pVM Pointer to the VM.
1492	* @param idCpu The Virtual CPU ID of the calling EMT.
1493	* @param ppGVM Where to store the GVM pointer.
1494	* @param ppGVMM Where to store the pointer to the GVMM instance data.
1495	* @thread EMT
1496	*
1497	* @remark This will assert in all failure paths.
1498	*/
1499	static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM ppGVM, PGVMM ppGVMM)
1500	{
1501	PGVMM pGVMM;
1502	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
1503
1504	/*
1505	* Validate.
1506	*/
1507	AssertPtrReturn(pVM, VERR_INVALID_POINTER);
1508	AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
1509
1510	uint16_t hGVM = pVM->hSelf;
1511	AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
1512	AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
1513
1514	/*
1515	* Look it up.
1516	*/
1517	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1518	AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
1519	RTPROCESS ProcId = RTProcSelf();
1520	AssertReturn(pHandle->ProcId == ProcId, VERR_NOT_OWNER);
1521	AssertPtrReturn(pHandle->pvObj, VERR_NOT_OWNER);
1522
1523	PGVM pGVM = pHandle->pGVM;
1524	AssertPtrReturn(pGVM, VERR_NOT_OWNER);
1525	AssertReturn(pGVM->pVM == pVM, VERR_NOT_OWNER);
1526	RTNATIVETHREAD hAllegedEMT = RTThreadNativeSelf();
1527	AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
1528	AssertReturn(pGVM->aCpus[idCpu].hEMT == hAllegedEMT, VERR_NOT_OWNER);
1529
1530	*ppGVM = pGVM;
1531	*ppGVMM = pGVMM;
1532	return VINF_SUCCESS;
1533	}
1534
1535
1536	/**
1537	* Lookup a GVM structure by the shared VM structure
1538	* and ensuring that the caller is the EMT thread.
1539	*
1540	* @returns VBox status code.
1541	* @param pVM Pointer to the VM.
1542	* @param idCpu The Virtual CPU ID of the calling EMT.
1543	* @param ppGVM Where to store the GVM pointer.
1544	* @thread EMT
1545	*/
1546	GVMMR0DECL(int) GVMMR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM)
1547	{
1548	AssertPtrReturn(ppGVM, VERR_INVALID_POINTER);
1549	PGVMM pGVMM;
1550	return gvmmR0ByVMAndEMT(pVM, idCpu, ppGVM, &pGVMM);
1551	}
1552
1553
1554	/**
1555	* Lookup a VM by its global handle.
1556	*
1557	* @returns Pointer to the VM on success, NULL on failure.
1558	* @param hGVM The global VM handle. Asserts on bad handle.
1559	*/
1560	GVMMR0DECL(PVM) GVMMR0GetVMByHandle(uint32_t hGVM)
1561	{
1562	PGVM pGVM = GVMMR0ByHandle(hGVM);
1563	return pGVM ? pGVM->pVM : NULL;
1564	}
1565
1566
1567	/**
1568	* Looks up the VM belonging to the specified EMT thread.
1569	*
1570	* This is used by the assertion machinery in VMMR0.cpp to avoid causing
1571	* unnecessary kernel panics when the EMT thread hits an assertion. The
1572	* call may or not be an EMT thread.
1573	*
1574	* @returns Pointer to the VM on success, NULL on failure.
1575	* @param hEMT The native thread handle of the EMT.
1576	* NIL_RTNATIVETHREAD means the current thread
1577	*/
1578	GVMMR0DECL(PVM) GVMMR0GetVMByEMT(RTNATIVETHREAD hEMT)
1579	{
1580	/*
1581	* No Assertions here as we're usually called in a AssertMsgN or
1582	* RTAssert* context.
1583	*/
1584	PGVMM pGVMM = g_pGVMM;
1585	if ( !VALID_PTR(pGVMM)
1586	\|\| pGVMM->u32Magic != GVMM_MAGIC)
1587	return NULL;
1588
1589	if (hEMT == NIL_RTNATIVETHREAD)
1590	hEMT = RTThreadNativeSelf();
1591	RTPROCESS ProcId = RTProcSelf();
1592
1593	/*
1594	* Search the handles in a linear fashion as we don't dare to take the lock (assert).
1595	*/
1596	for (unsigned i = 1; i < RT_ELEMENTS(pGVMM->aHandles); i++)
1597	{
1598	if ( pGVMM->aHandles[i].iSelf == i
1599	&& pGVMM->aHandles[i].ProcId == ProcId
1600	&& VALID_PTR(pGVMM->aHandles[i].pvObj)
1601	&& VALID_PTR(pGVMM->aHandles[i].pVM)
1602	&& VALID_PTR(pGVMM->aHandles[i].pGVM))
1603	{
1604	if (pGVMM->aHandles[i].hEMT0 == hEMT)
1605	return pGVMM->aHandles[i].pVM;
1606
1607	/* This is fearly safe with the current process per VM approach. */
1608	PGVM pGVM = pGVMM->aHandles[i].pGVM;
1609	VMCPUID const cCpus = pGVM->cCpus;
1610	if ( cCpus < 1
1611	\|\| cCpus > VMM_MAX_CPU_COUNT)
1612	continue;
1613	for (VMCPUID idCpu = 1; idCpu < cCpus; idCpu++)
1614	if (pGVM->aCpus[idCpu].hEMT == hEMT)
1615	return pGVMM->aHandles[i].pVM;
1616	}
1617	}
1618	return NULL;
1619	}
1620
1621
1622	/**
1623	* This is will wake up expired and soon-to-be expired VMs.
1624	*
1625	* @returns Number of VMs that has been woken up.
1626	* @param pGVMM Pointer to the GVMM instance data.
1627	* @param u64Now The current time.
1628	*/
1629	static unsigned gvmmR0SchedDoWakeUps(PGVMM pGVMM, uint64_t u64Now)
1630	{
1631	/*
1632	* Skip this if we've got disabled because of high resolution wakeups or by
1633	* the user.
1634	*/
1635	if ( !pGVMM->nsEarlyWakeUp1
1636	&& !pGVMM->nsEarlyWakeUp2)
1637	return 0;
1638
1639	/** @todo Rewrite this algorithm. See performance defect XYZ. */
1640
1641	/*
1642	* A cheap optimization to stop wasting so much time here on big setups.
1643	*/
1644	const uint64_t uNsEarlyWakeUp2 = u64Now + pGVMM->nsEarlyWakeUp2;
1645	if ( pGVMM->cHaltedEMTs == 0
1646	\|\| uNsEarlyWakeUp2 > pGVMM->uNsNextEmtWakeup)
1647	return 0;
1648
1649	/*
1650	* The first pass will wake up VMs which have actually expired
1651	* and look for VMs that should be woken up in the 2nd and 3rd passes.
1652	*/
1653	const uint64_t uNsEarlyWakeUp1 = u64Now + pGVMM->nsEarlyWakeUp1;
1654	uint64_t u64Min = UINT64_MAX;
1655	unsigned cWoken = 0;
1656	unsigned cHalted = 0;
1657	unsigned cTodo2nd = 0;
1658	unsigned cTodo3rd = 0;
1659	for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1660	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1661	i = pGVMM->aHandles[i].iNext)
1662	{
1663	PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1664	if ( VALID_PTR(pCurGVM)
1665	&& pCurGVM->u32Magic == GVM_MAGIC)
1666	{
1667	for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1668	{
1669	PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1670	uint64_t u64 = ASMAtomicUoReadU64(&pCurGVCpu->gvmm.s.u64HaltExpire);
1671	if (u64)
1672	{
1673	if (u64 <= u64Now)
1674	{
1675	if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1676	{
1677	int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1678	AssertRC(rc);
1679	cWoken++;
1680	}
1681	}
1682	else
1683	{
1684	cHalted++;
1685	if (u64 <= uNsEarlyWakeUp1)
1686	cTodo2nd++;
1687	else if (u64 <= uNsEarlyWakeUp2)
1688	cTodo3rd++;
1689	else if (u64 < u64Min)
1690	u64 = u64Min;
1691	}
1692	}
1693	}
1694	}
1695	AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1696	}
1697
1698	if (cTodo2nd)
1699	{
1700	for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1701	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1702	i = pGVMM->aHandles[i].iNext)
1703	{
1704	PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1705	if ( VALID_PTR(pCurGVM)
1706	&& pCurGVM->u32Magic == GVM_MAGIC)
1707	{
1708	for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1709	{
1710	PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1711	uint64_t u64 = ASMAtomicUoReadU64(&pCurGVCpu->gvmm.s.u64HaltExpire);
1712	if ( u64
1713	&& u64 <= uNsEarlyWakeUp1)
1714	{
1715	if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1716	{
1717	int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1718	AssertRC(rc);
1719	cWoken++;
1720	}
1721	}
1722	}
1723	}
1724	AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1725	}
1726	}
1727
1728	if (cTodo3rd)
1729	{
1730	for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1731	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1732	i = pGVMM->aHandles[i].iNext)
1733	{
1734	PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1735	if ( VALID_PTR(pCurGVM)
1736	&& pCurGVM->u32Magic == GVM_MAGIC)
1737	{
1738	for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1739	{
1740	PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1741	uint64_t u64 = ASMAtomicUoReadU64(&pCurGVCpu->gvmm.s.u64HaltExpire);
1742	if ( u64
1743	&& u64 <= uNsEarlyWakeUp2)
1744	{
1745	if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1746	{
1747	int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1748	AssertRC(rc);
1749	cWoken++;
1750	}
1751	}
1752	}
1753	}
1754	AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1755	}
1756	}
1757
1758	/*
1759	* Set the minimum value.
1760	*/
1761	pGVMM->uNsNextEmtWakeup = u64Min;
1762
1763	return cWoken;
1764	}
1765
1766
1767	/**
1768	* Halt the EMT thread.
1769	*
1770	* @returns VINF_SUCCESS normal wakeup (timeout or kicked by other thread).
1771	* VERR_INTERRUPTED if a signal was scheduled for the thread.
1772	* @param pVM Pointer to the VM.
1773	* @param idCpu The Virtual CPU ID of the calling EMT.
1774	* @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
1775	* @thread EMT(idCpu).
1776	*/
1777	GVMMR0DECL(int) GVMMR0SchedHalt(PVM pVM, VMCPUID idCpu, uint64_t u64ExpireGipTime)
1778	{
1779	LogFlow(("GVMMR0SchedHalt: pVM=%p\n", pVM));
1780
1781	/*
1782	* Validate the VM structure, state and handle.
1783	*/
1784	PGVM pGVM;
1785	PGVMM pGVMM;
1786	int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
1787	if (RT_FAILURE(rc))
1788	return rc;
1789	pGVM->gvmm.s.StatsSched.cHaltCalls++;
1790
1791	PGVMCPU pCurGVCpu = &pGVM->aCpus[idCpu];
1792	Assert(!pCurGVCpu->gvmm.s.u64HaltExpire);
1793
1794	/*
1795	* Take the UsedList semaphore, get the current time
1796	* and check if anyone needs waking up.
1797	* Interrupts must NOT be disabled at this point because we ask for GIP time!
1798	*/
1799	rc = gvmmR0UsedLock(pGVMM);
1800	AssertRC(rc);
1801
1802	pCurGVCpu->gvmm.s.iCpuEmt = ASMGetApicId();
1803
1804	/* GIP hack: We might are frequently sleeping for short intervals where the
1805	difference between GIP and system time matters on systems with high resolution
1806	system time. So, convert the input from GIP to System time in that case. */
1807	Assert(ASMGetFlags() & X86_EFL_IF);
1808	const uint64_t u64NowSys = RTTimeSystemNanoTS();
1809	const uint64_t u64NowGip = RTTimeNanoTS();
1810	pGVM->gvmm.s.StatsSched.cHaltWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64NowGip);
1811
1812	/*
1813	* Go to sleep if we must...
1814	* Cap the sleep time to 1 second to be on the safe side.
1815	*/
1816	uint64_t cNsInterval = u64ExpireGipTime - u64NowGip;
1817	if ( u64NowGip < u64ExpireGipTime
1818	&& cNsInterval >= (pGVMM->cEMTs > pGVMM->cEMTsMeansCompany
1819	? pGVMM->nsMinSleepCompany
1820	: pGVMM->nsMinSleepAlone))
1821	{
1822	pGVM->gvmm.s.StatsSched.cHaltBlocking++;
1823	if (cNsInterval > RT_NS_1SEC)
1824	u64ExpireGipTime = u64NowGip + RT_NS_1SEC;
1825	if (u64ExpireGipTime < pGVMM->uNsNextEmtWakeup)
1826	pGVMM->uNsNextEmtWakeup = u64ExpireGipTime;
1827	ASMAtomicWriteU64(&pCurGVCpu->gvmm.s.u64HaltExpire, u64ExpireGipTime);
1828	ASMAtomicIncU32(&pGVMM->cHaltedEMTs);
1829	gvmmR0UsedUnlock(pGVMM);
1830
1831	rc = RTSemEventMultiWaitEx(pCurGVCpu->gvmm.s.HaltEventMulti,
1832	RTSEMWAIT_FLAGS_ABSOLUTE \| RTSEMWAIT_FLAGS_NANOSECS \| RTSEMWAIT_FLAGS_INTERRUPTIBLE,
1833	u64NowGip > u64NowSys ? u64ExpireGipTime : u64NowSys + cNsInterval);
1834
1835	ASMAtomicWriteU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0);
1836	ASMAtomicDecU32(&pGVMM->cHaltedEMTs);
1837
1838	/* Reset the semaphore to try prevent a few false wake-ups. */
1839	if (rc == VINF_SUCCESS)
1840	RTSemEventMultiReset(pCurGVCpu->gvmm.s.HaltEventMulti);
1841	else if (rc == VERR_TIMEOUT)
1842	{
1843	pGVM->gvmm.s.StatsSched.cHaltTimeouts++;
1844	rc = VINF_SUCCESS;
1845	}
1846	}
1847	else
1848	{
1849	pGVM->gvmm.s.StatsSched.cHaltNotBlocking++;
1850	gvmmR0UsedUnlock(pGVMM);
1851	RTSemEventMultiReset(pCurGVCpu->gvmm.s.HaltEventMulti);
1852	}
1853
1854	return rc;
1855	}
1856
1857
1858	/**
1859	* Worker for GVMMR0SchedWakeUp and GVMMR0SchedWakeUpAndPokeCpus that wakes up
1860	* the a sleeping EMT.
1861	*
1862	* @retval VINF_SUCCESS if successfully woken up.
1863	* @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1864	*
1865	* @param pGVM The global (ring-0) VM structure.
1866	* @param pGVCpu The global (ring-0) VCPU structure.
1867	*/
1868	DECLINLINE(int) gvmmR0SchedWakeUpOne(PGVM pGVM, PGVMCPU pGVCpu)
1869	{
1870	pGVM->gvmm.s.StatsSched.cWakeUpCalls++;
1871
1872	/*
1873	* Signal the semaphore regardless of whether it's current blocked on it.
1874	*
1875	* The reason for this is that there is absolutely no way we can be 100%
1876	* certain that it isn't about go to go to sleep on it and just got
1877	* delayed a bit en route. So, we will always signal the semaphore when
1878	* the it is flagged as halted in the VMM.
1879	*/
1880	/** @todo we can optimize some of that by means of the pVCpu->enmState now. */
1881	int rc;
1882	if (pGVCpu->gvmm.s.u64HaltExpire)
1883	{
1884	rc = VINF_SUCCESS;
1885	ASMAtomicWriteU64(&pGVCpu->gvmm.s.u64HaltExpire, 0);
1886	}
1887	else
1888	{
1889	rc = VINF_GVM_NOT_BLOCKED;
1890	pGVM->gvmm.s.StatsSched.cWakeUpNotHalted++;
1891	}
1892
1893	int rc2 = RTSemEventMultiSignal(pGVCpu->gvmm.s.HaltEventMulti);
1894	AssertRC(rc2);
1895
1896	return rc;
1897	}
1898
1899
1900	/**
1901	* Wakes up the halted EMT thread so it can service a pending request.
1902	*
1903	* @returns VBox status code.
1904	* @retval VINF_SUCCESS if successfully woken up.
1905	* @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1906	*
1907	* @param pVM Pointer to the VM.
1908	* @param idCpu The Virtual CPU ID of the EMT to wake up.
1909	* @param fTakeUsedLock Take the used lock or not
1910	* @thread Any but EMT.
1911	*/
1912	GVMMR0DECL(int) GVMMR0SchedWakeUpEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
1913	{
1914	/*
1915	* Validate input and take the UsedLock.
1916	*/
1917	PGVM pGVM;
1918	PGVMM pGVMM;
1919	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
1920	if (RT_SUCCESS(rc))
1921	{
1922	if (idCpu < pGVM->cCpus)
1923	{
1924	/*
1925	* Do the actual job.
1926	*/
1927	rc = gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
1928
1929	if (fTakeUsedLock)
1930	{
1931	/*
1932	* While we're here, do a round of scheduling.
1933	*/
1934	Assert(ASMGetFlags() & X86_EFL_IF);
1935	const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1936	pGVM->gvmm.s.StatsSched.cWakeUpWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1937	}
1938	}
1939	else
1940	rc = VERR_INVALID_CPU_ID;
1941
1942	if (fTakeUsedLock)
1943	{
1944	int rc2 = gvmmR0UsedUnlock(pGVMM);
1945	AssertRC(rc2);
1946	}
1947	}
1948
1949	LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
1950	return rc;
1951	}
1952
1953
1954	/**
1955	* Wakes up the halted EMT thread so it can service a pending request.
1956	*
1957	* @returns VBox status code.
1958	* @retval VINF_SUCCESS if successfully woken up.
1959	* @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1960	*
1961	* @param pVM Pointer to the VM.
1962	* @param idCpu The Virtual CPU ID of the EMT to wake up.
1963	* @thread Any but EMT.
1964	*/
1965	GVMMR0DECL(int) GVMMR0SchedWakeUp(PVM pVM, VMCPUID idCpu)
1966	{
1967	return GVMMR0SchedWakeUpEx(pVM, idCpu, true /* fTakeUsedLock */);
1968	}
1969
1970	/**
1971	* Worker common to GVMMR0SchedPoke and GVMMR0SchedWakeUpAndPokeCpus that pokes
1972	* the Virtual CPU if it's still busy executing guest code.
1973	*
1974	* @returns VBox status code.
1975	* @retval VINF_SUCCESS if poked successfully.
1976	* @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1977	*
1978	* @param pGVM The global (ring-0) VM structure.
1979	* @param pVCpu Pointer to the VMCPU.
1980	*/
1981	DECLINLINE(int) gvmmR0SchedPokeOne(PGVM pGVM, PVMCPU pVCpu)
1982	{
1983	pGVM->gvmm.s.StatsSched.cPokeCalls++;
1984
1985	RTCPUID idHostCpu = pVCpu->idHostCpu;
1986	if ( idHostCpu == NIL_RTCPUID
1987	\|\| VMCPU_GET_STATE(pVCpu) != VMCPUSTATE_STARTED_EXEC)
1988	{
1989	pGVM->gvmm.s.StatsSched.cPokeNotBusy++;
1990	return VINF_GVM_NOT_BUSY_IN_GC;
1991	}
1992
1993	/* Note: this function is not implemented on Darwin and Linux (kernel < 2.6.19) */
1994	RTMpPokeCpu(idHostCpu);
1995	return VINF_SUCCESS;
1996	}
1997
1998	/**
1999	* Pokes an EMT if it's still busy running guest code.
2000	*
2001	* @returns VBox status code.
2002	* @retval VINF_SUCCESS if poked successfully.
2003	* @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
2004	*
2005	* @param pVM Pointer to the VM.
2006	* @param idCpu The ID of the virtual CPU to poke.
2007	* @param fTakeUsedLock Take the used lock or not
2008	*/
2009	GVMMR0DECL(int) GVMMR0SchedPokeEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
2010	{
2011	/*
2012	* Validate input and take the UsedLock.
2013	*/
2014	PGVM pGVM;
2015	PGVMM pGVMM;
2016	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
2017	if (RT_SUCCESS(rc))
2018	{
2019	if (idCpu < pGVM->cCpus)
2020	rc = gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
2021	else
2022	rc = VERR_INVALID_CPU_ID;
2023
2024	if (fTakeUsedLock)
2025	{
2026	int rc2 = gvmmR0UsedUnlock(pGVMM);
2027	AssertRC(rc2);
2028	}
2029	}
2030
2031	LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
2032	return rc;
2033	}
2034
2035
2036	/**
2037	* Pokes an EMT if it's still busy running guest code.
2038	*
2039	* @returns VBox status code.
2040	* @retval VINF_SUCCESS if poked successfully.
2041	* @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
2042	*
2043	* @param pVM Pointer to the VM.
2044	* @param idCpu The ID of the virtual CPU to poke.
2045	*/
2046	GVMMR0DECL(int) GVMMR0SchedPoke(PVM pVM, VMCPUID idCpu)
2047	{
2048	return GVMMR0SchedPokeEx(pVM, idCpu, true /* fTakeUsedLock */);
2049	}
2050
2051
2052	/**
2053	* Wakes up a set of halted EMT threads so they can service pending request.
2054	*
2055	* @returns VBox status code, no informational stuff.
2056	*
2057	* @param pVM Pointer to the VM.
2058	* @param pSleepSet The set of sleepers to wake up.
2059	* @param pPokeSet The set of CPUs to poke.
2060	*/
2061	GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpus(PVM pVM, PCVMCPUSET pSleepSet, PCVMCPUSET pPokeSet)
2062	{
2063	AssertPtrReturn(pSleepSet, VERR_INVALID_POINTER);
2064	AssertPtrReturn(pPokeSet, VERR_INVALID_POINTER);
2065	RTNATIVETHREAD hSelf = RTThreadNativeSelf();
2066
2067	/*
2068	* Validate input and take the UsedLock.
2069	*/
2070	PGVM pGVM;
2071	PGVMM pGVMM;
2072	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /* fTakeUsedLock */);
2073	if (RT_SUCCESS(rc))
2074	{
2075	rc = VINF_SUCCESS;
2076	VMCPUID idCpu = pGVM->cCpus;
2077	while (idCpu-- > 0)
2078	{
2079	/* Don't try poke or wake up ourselves. */
2080	if (pGVM->aCpus[idCpu].hEMT == hSelf)
2081	continue;
2082
2083	/* just ignore errors for now. */
2084	if (VMCPUSET_IS_PRESENT(pSleepSet, idCpu))
2085	gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
2086	else if (VMCPUSET_IS_PRESENT(pPokeSet, idCpu))
2087	gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
2088	}
2089
2090	int rc2 = gvmmR0UsedUnlock(pGVMM);
2091	AssertRC(rc2);
2092	}
2093
2094	LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
2095	return rc;
2096	}
2097
2098
2099	/**
2100	* VMMR0 request wrapper for GVMMR0SchedWakeUpAndPokeCpus.
2101	*
2102	* @returns see GVMMR0SchedWakeUpAndPokeCpus.
2103	* @param pVM Pointer to the VM.
2104	* @param pReq Pointer to the request packet.
2105	*/
2106	GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpusReq(PVM pVM, PGVMMSCHEDWAKEUPANDPOKECPUSREQ pReq)
2107	{
2108	/*
2109	* Validate input and pass it on.
2110	*/
2111	AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2112	AssertMsgReturn(pReq->Hdr.cbReq == sizeof(pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(pReq)), VERR_INVALID_PARAMETER);
2113
2114	return GVMMR0SchedWakeUpAndPokeCpus(pVM, &pReq->SleepSet, &pReq->PokeSet);
2115	}
2116
2117
2118
2119	/**
2120	* Poll the schedule to see if someone else should get a chance to run.
2121	*
2122	* This is a bit hackish and will not work too well if the machine is
2123	* under heavy load from non-VM processes.
2124	*
2125	* @returns VINF_SUCCESS if not yielded.
2126	* VINF_GVM_YIELDED if an attempt to switch to a different VM task was made.
2127	* @param pVM Pointer to the VM.
2128	* @param idCpu The Virtual CPU ID of the calling EMT.
2129	* @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
2130	* @param fYield Whether to yield or not.
2131	* This is for when we're spinning in the halt loop.
2132	* @thread EMT(idCpu).
2133	*/
2134	GVMMR0DECL(int) GVMMR0SchedPoll(PVM pVM, VMCPUID idCpu, bool fYield)
2135	{
2136	/*
2137	* Validate input.
2138	*/
2139	PGVM pGVM;
2140	PGVMM pGVMM;
2141	int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
2142	if (RT_SUCCESS(rc))
2143	{
2144	rc = gvmmR0UsedLock(pGVMM);
2145	AssertRC(rc);
2146	pGVM->gvmm.s.StatsSched.cPollCalls++;
2147
2148	Assert(ASMGetFlags() & X86_EFL_IF);
2149	const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
2150
2151	if (!fYield)
2152	pGVM->gvmm.s.StatsSched.cPollWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
2153	else
2154	{
2155	/** @todo implement this... */
2156	rc = VERR_NOT_IMPLEMENTED;
2157	}
2158
2159	gvmmR0UsedUnlock(pGVMM);
2160	}
2161
2162	LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
2163	return rc;
2164	}
2165
2166
2167	#ifdef GVMM_SCHED_WITH_PPT
2168	/**
2169	* Timer callback for the periodic preemption timer.
2170	*
2171	* @param pTimer The timer handle.
2172	* @param pvUser Pointer to the per cpu structure.
2173	* @param iTick The current tick.
2174	*/
2175	static DECLCALLBACK(void) gvmmR0SchedPeriodicPreemptionTimerCallback(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
2176	{
2177	PGVMMHOSTCPU pCpu = (PGVMMHOSTCPU)pvUser;
2178	NOREF(pTimer); NOREF(iTick);
2179
2180	/*
2181	* Termination check
2182	*/
2183	if (pCpu->u32Magic != GVMMHOSTCPU_MAGIC)
2184	return;
2185
2186	/*
2187	* Do the house keeping.
2188	*/
2189	RTSpinlockAcquire(pCpu->Ppt.hSpinlock);
2190
2191	if (++pCpu->Ppt.iTickHistorization >= pCpu->Ppt.cTicksHistoriziationInterval)
2192	{
2193	/*
2194	* Historicize the max frequency.
2195	*/
2196	uint32_t iHzHistory = ++pCpu->Ppt.iHzHistory % RT_ELEMENTS(pCpu->Ppt.aHzHistory);
2197	pCpu->Ppt.aHzHistory[iHzHistory] = pCpu->Ppt.uDesiredHz;
2198	pCpu->Ppt.iTickHistorization = 0;
2199	pCpu->Ppt.uDesiredHz = 0;
2200
2201	/*
2202	* Check if the current timer frequency.
2203	*/
2204	uint32_t uHistMaxHz = 0;
2205	for (uint32_t i = 0; i < RT_ELEMENTS(pCpu->Ppt.aHzHistory); i++)
2206	if (pCpu->Ppt.aHzHistory[i] > uHistMaxHz)
2207	uHistMaxHz = pCpu->Ppt.aHzHistory[i];
2208	if (uHistMaxHz == pCpu->Ppt.uTimerHz)
2209	RTSpinlockReleaseNoInts(pCpu->Ppt.hSpinlock);
2210	else if (uHistMaxHz)
2211	{
2212	/*
2213	* Reprogram it.
2214	*/
2215	pCpu->Ppt.cChanges++;
2216	pCpu->Ppt.iTickHistorization = 0;
2217	pCpu->Ppt.uTimerHz = uHistMaxHz;
2218	uint32_t const cNsInterval = RT_NS_1SEC / uHistMaxHz;
2219	pCpu->Ppt.cNsInterval = cNsInterval;
2220	if (cNsInterval < GVMMHOSTCPU_PPT_HIST_INTERVAL_NS)
2221	pCpu->Ppt.cTicksHistoriziationInterval = ( GVMMHOSTCPU_PPT_HIST_INTERVAL_NS
2222	+ GVMMHOSTCPU_PPT_HIST_INTERVAL_NS / 2 - 1)
2223	/ cNsInterval;
2224	else
2225	pCpu->Ppt.cTicksHistoriziationInterval = 1;
2226	RTSpinlockReleaseNoInts(pCpu->Ppt.hSpinlock);
2227
2228	/SUPR0Printf("Cpu%u: change to %u Hz / %u ns\n", pCpu->idxCpuSet, uHistMaxHz, cNsInterval);/
2229	RTTimerChangeInterval(pTimer, cNsInterval);
2230	}
2231	else
2232	{
2233	/*
2234	* Stop it.
2235	*/
2236	pCpu->Ppt.fStarted = false;
2237	pCpu->Ppt.uTimerHz = 0;
2238	pCpu->Ppt.cNsInterval = 0;
2239	RTSpinlockReleaseNoInts(pCpu->Ppt.hSpinlock);
2240
2241	/SUPR0Printf("Cpu%u: stopping (%u Hz)\n", pCpu->idxCpuSet, uHistMaxHz);/
2242	RTTimerStop(pTimer);
2243	}
2244	}
2245	else
2246	RTSpinlockReleaseNoInts(pCpu->Ppt.hSpinlock);
2247	}
2248	#endif /* GVMM_SCHED_WITH_PPT */
2249
2250
2251	/**
2252	* Updates the periodic preemption timer for the calling CPU.
2253	*
2254	* The caller must have disabled preemption!
2255	* The caller must check that the host can do high resolution timers.
2256	*
2257	* @param pVM Pointer to the VM.
2258	* @param idHostCpu The current host CPU id.
2259	* @param uHz The desired frequency.
2260	*/
2261	GVMMR0DECL(void) GVMMR0SchedUpdatePeriodicPreemptionTimer(PVM pVM, RTCPUID idHostCpu, uint32_t uHz)
2262	{
2263	NOREF(pVM);
2264	#ifdef GVMM_SCHED_WITH_PPT
2265	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2266	Assert(RTTimerCanDoHighResolution());
2267
2268	/*
2269	* Resolve the per CPU data.
2270	*/
2271	uint32_t iCpu = RTMpCpuIdToSetIndex(idHostCpu);
2272	PGVMM pGVMM = g_pGVMM;
2273	if ( !VALID_PTR(pGVMM)
2274	\|\| pGVMM->u32Magic != GVMM_MAGIC)
2275	return;
2276	AssertMsgReturnVoid(iCpu < pGVMM->cHostCpus, ("iCpu=%d cHostCpus=%d\n", iCpu, pGVMM->cHostCpus));
2277	PGVMMHOSTCPU pCpu = &pGVMM->aHostCpus[iCpu];
2278	AssertMsgReturnVoid( pCpu->u32Magic == GVMMHOSTCPU_MAGIC
2279	&& pCpu->idCpu == idHostCpu,
2280	("u32Magic=%#x idCpu=% idHostCpu=%d\n", pCpu->u32Magic, pCpu->idCpu, idHostCpu));
2281
2282	/*
2283	* Check whether we need to do anything about the timer.
2284	* We have to be a little bit careful since we might be race the timer
2285	* callback here.
2286	*/
2287	if (uHz > 16384)
2288	uHz = 16384; /** @todo add a query method for this! */
2289	if (RT_UNLIKELY( uHz > ASMAtomicReadU32(&pCpu->Ppt.uDesiredHz)
2290	&& uHz >= pCpu->Ppt.uMinHz
2291	&& !pCpu->Ppt.fStarting /* solaris paranoia */))
2292	{
2293	RTSpinlockAcquire(pCpu->Ppt.hSpinlock);
2294
2295	pCpu->Ppt.uDesiredHz = uHz;
2296	uint32_t cNsInterval = 0;
2297	if (!pCpu->Ppt.fStarted)
2298	{
2299	pCpu->Ppt.cStarts++;
2300	pCpu->Ppt.fStarted = true;
2301	pCpu->Ppt.fStarting = true;
2302	pCpu->Ppt.iTickHistorization = 0;
2303	pCpu->Ppt.uTimerHz = uHz;
2304	pCpu->Ppt.cNsInterval = cNsInterval = RT_NS_1SEC / uHz;
2305	if (cNsInterval < GVMMHOSTCPU_PPT_HIST_INTERVAL_NS)
2306	pCpu->Ppt.cTicksHistoriziationInterval = ( GVMMHOSTCPU_PPT_HIST_INTERVAL_NS
2307	+ GVMMHOSTCPU_PPT_HIST_INTERVAL_NS / 2 - 1)
2308	/ cNsInterval;
2309	else
2310	pCpu->Ppt.cTicksHistoriziationInterval = 1;
2311	}
2312
2313	RTSpinlockReleaseNoInts(pCpu->Ppt.hSpinlock);
2314
2315	if (cNsInterval)
2316	{
2317	RTTimerChangeInterval(pCpu->Ppt.pTimer, cNsInterval);
2318	int rc = RTTimerStart(pCpu->Ppt.pTimer, cNsInterval);
2319	AssertRC(rc);
2320
2321	RTSpinlockAcquire(pCpu->Ppt.hSpinlock);
2322	if (RT_FAILURE(rc))
2323	pCpu->Ppt.fStarted = false;
2324	pCpu->Ppt.fStarting = false;
2325	RTSpinlockReleaseNoInts(pCpu->Ppt.hSpinlock);
2326	}
2327	}
2328	#else /* !GVMM_SCHED_WITH_PPT */
2329	NOREF(idHostCpu); NOREF(uHz);
2330	#endif /* !GVMM_SCHED_WITH_PPT */
2331	}
2332
2333
2334	/**
2335	* Retrieves the GVMM statistics visible to the caller.
2336	*
2337	* @returns VBox status code.
2338	*
2339	* @param pStats Where to put the statistics.
2340	* @param pSession The current session.
2341	* @param pVM The VM to obtain statistics for. Optional.
2342	*/
2343	GVMMR0DECL(int) GVMMR0QueryStatistics(PGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
2344	{
2345	LogFlow(("GVMMR0QueryStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
2346
2347	/*
2348	* Validate input.
2349	*/
2350	AssertPtrReturn(pSession, VERR_INVALID_POINTER);
2351	AssertPtrReturn(pStats, VERR_INVALID_POINTER);
2352	pStats->cVMs = 0; /* (crash before taking the sem...) */
2353
2354	/*
2355	* Take the lock and get the VM statistics.
2356	*/
2357	PGVMM pGVMM;
2358	if (pVM)
2359	{
2360	PGVM pGVM;
2361	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /fTakeUsedLock/);
2362	if (RT_FAILURE(rc))
2363	return rc;
2364	pStats->SchedVM = pGVM->gvmm.s.StatsSched;
2365	}
2366	else
2367	{
2368	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
2369	memset(&pStats->SchedVM, 0, sizeof(pStats->SchedVM));
2370
2371	int rc = gvmmR0UsedLock(pGVMM);
2372	AssertRCReturn(rc, rc);
2373	}
2374
2375	/*
2376	* Enumerate the VMs and add the ones visible to the statistics.
2377	*/
2378	pStats->cVMs = 0;
2379	pStats->cEMTs = 0;
2380	memset(&pStats->SchedSum, 0, sizeof(pStats->SchedSum));
2381
2382	for (unsigned i = pGVMM->iUsedHead;
2383	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
2384	i = pGVMM->aHandles[i].iNext)
2385	{
2386	PGVM pGVM = pGVMM->aHandles[i].pGVM;
2387	void *pvObj = pGVMM->aHandles[i].pvObj;
2388	if ( VALID_PTR(pvObj)
2389	&& VALID_PTR(pGVM)
2390	&& pGVM->u32Magic == GVM_MAGIC
2391	&& RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
2392	{
2393	pStats->cVMs++;
2394	pStats->cEMTs += pGVM->cCpus;
2395
2396	pStats->SchedSum.cHaltCalls += pGVM->gvmm.s.StatsSched.cHaltCalls;
2397	pStats->SchedSum.cHaltBlocking += pGVM->gvmm.s.StatsSched.cHaltBlocking;
2398	pStats->SchedSum.cHaltTimeouts += pGVM->gvmm.s.StatsSched.cHaltTimeouts;
2399	pStats->SchedSum.cHaltNotBlocking += pGVM->gvmm.s.StatsSched.cHaltNotBlocking;
2400	pStats->SchedSum.cHaltWakeUps += pGVM->gvmm.s.StatsSched.cHaltWakeUps;
2401
2402	pStats->SchedSum.cWakeUpCalls += pGVM->gvmm.s.StatsSched.cWakeUpCalls;
2403	pStats->SchedSum.cWakeUpNotHalted += pGVM->gvmm.s.StatsSched.cWakeUpNotHalted;
2404	pStats->SchedSum.cWakeUpWakeUps += pGVM->gvmm.s.StatsSched.cWakeUpWakeUps;
2405
2406	pStats->SchedSum.cPokeCalls += pGVM->gvmm.s.StatsSched.cPokeCalls;
2407	pStats->SchedSum.cPokeNotBusy += pGVM->gvmm.s.StatsSched.cPokeNotBusy;
2408
2409	pStats->SchedSum.cPollCalls += pGVM->gvmm.s.StatsSched.cPollCalls;
2410	pStats->SchedSum.cPollHalts += pGVM->gvmm.s.StatsSched.cPollHalts;
2411	pStats->SchedSum.cPollWakeUps += pGVM->gvmm.s.StatsSched.cPollWakeUps;
2412	}
2413	}
2414
2415	/*
2416	* Copy out the per host CPU statistics.
2417	*/
2418	uint32_t iDstCpu = 0;
2419	uint32_t cSrcCpus = pGVMM->cHostCpus;
2420	for (uint32_t iSrcCpu = 0; iSrcCpu < cSrcCpus; iSrcCpu++)
2421	{
2422	if (pGVMM->aHostCpus[iSrcCpu].idCpu != NIL_RTCPUID)
2423	{
2424	pStats->aHostCpus[iDstCpu].idCpu = pGVMM->aHostCpus[iSrcCpu].idCpu;
2425	pStats->aHostCpus[iDstCpu].idxCpuSet = pGVMM->aHostCpus[iSrcCpu].idxCpuSet;
2426	#ifdef GVMM_SCHED_WITH_PPT
2427	pStats->aHostCpus[iDstCpu].uDesiredHz = pGVMM->aHostCpus[iSrcCpu].Ppt.uDesiredHz;
2428	pStats->aHostCpus[iDstCpu].uTimerHz = pGVMM->aHostCpus[iSrcCpu].Ppt.uTimerHz;
2429	pStats->aHostCpus[iDstCpu].cChanges = pGVMM->aHostCpus[iSrcCpu].Ppt.cChanges;
2430	pStats->aHostCpus[iDstCpu].cStarts = pGVMM->aHostCpus[iSrcCpu].Ppt.cStarts;
2431	#else
2432	pStats->aHostCpus[iDstCpu].uDesiredHz = 0;
2433	pStats->aHostCpus[iDstCpu].uTimerHz = 0;
2434	pStats->aHostCpus[iDstCpu].cChanges = 0;
2435	pStats->aHostCpus[iDstCpu].cStarts = 0;
2436	#endif
2437	iDstCpu++;
2438	if (iDstCpu >= RT_ELEMENTS(pStats->aHostCpus))
2439	break;
2440	}
2441	}
2442	pStats->cHostCpus = iDstCpu;
2443
2444	gvmmR0UsedUnlock(pGVMM);
2445
2446	return VINF_SUCCESS;
2447	}
2448
2449
2450	/**
2451	* VMMR0 request wrapper for GVMMR0QueryStatistics.
2452	*
2453	* @returns see GVMMR0QueryStatistics.
2454	* @param pVM Pointer to the VM. Optional.
2455	* @param pReq Pointer to the request packet.
2456	*/
2457	GVMMR0DECL(int) GVMMR0QueryStatisticsReq(PVM pVM, PGVMMQUERYSTATISTICSSREQ pReq)
2458	{
2459	/*
2460	* Validate input and pass it on.
2461	*/
2462	AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2463	AssertMsgReturn(pReq->Hdr.cbReq == sizeof(pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(pReq)), VERR_INVALID_PARAMETER);
2464
2465	return GVMMR0QueryStatistics(&pReq->Stats, pReq->pSession, pVM);
2466	}
2467
2468
2469	/**
2470	* Resets the specified GVMM statistics.
2471	*
2472	* @returns VBox status code.
2473	*
2474	* @param pStats Which statistics to reset, that is, non-zero fields indicates which to reset.
2475	* @param pSession The current session.
2476	* @param pVM The VM to reset statistics for. Optional.
2477	*/
2478	GVMMR0DECL(int) GVMMR0ResetStatistics(PCGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
2479	{
2480	LogFlow(("GVMMR0ResetStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
2481
2482	/*
2483	* Validate input.
2484	*/
2485	AssertPtrReturn(pSession, VERR_INVALID_POINTER);
2486	AssertPtrReturn(pStats, VERR_INVALID_POINTER);
2487
2488	/*
2489	* Take the lock and get the VM statistics.
2490	*/
2491	PGVMM pGVMM;
2492	if (pVM)
2493	{
2494	PGVM pGVM;
2495	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /fTakeUsedLock/);
2496	if (RT_FAILURE(rc))
2497	return rc;
2498	# define MAYBE_RESET_FIELD(field) \
2499	do { if (pStats->SchedVM. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2500	MAYBE_RESET_FIELD(cHaltCalls);
2501	MAYBE_RESET_FIELD(cHaltBlocking);
2502	MAYBE_RESET_FIELD(cHaltTimeouts);
2503	MAYBE_RESET_FIELD(cHaltNotBlocking);
2504	MAYBE_RESET_FIELD(cHaltWakeUps);
2505	MAYBE_RESET_FIELD(cWakeUpCalls);
2506	MAYBE_RESET_FIELD(cWakeUpNotHalted);
2507	MAYBE_RESET_FIELD(cWakeUpWakeUps);
2508	MAYBE_RESET_FIELD(cPokeCalls);
2509	MAYBE_RESET_FIELD(cPokeNotBusy);
2510	MAYBE_RESET_FIELD(cPollCalls);
2511	MAYBE_RESET_FIELD(cPollHalts);
2512	MAYBE_RESET_FIELD(cPollWakeUps);
2513	# undef MAYBE_RESET_FIELD
2514	}
2515	else
2516	{
2517	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
2518
2519	int rc = gvmmR0UsedLock(pGVMM);
2520	AssertRCReturn(rc, rc);
2521	}
2522
2523	/*
2524	* Enumerate the VMs and add the ones visible to the statistics.
2525	*/
2526	if (ASMMemIsAll8(&pStats->SchedSum, sizeof(pStats->SchedSum), 0))
2527	{
2528	for (unsigned i = pGVMM->iUsedHead;
2529	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
2530	i = pGVMM->aHandles[i].iNext)
2531	{
2532	PGVM pGVM = pGVMM->aHandles[i].pGVM;
2533	void *pvObj = pGVMM->aHandles[i].pvObj;
2534	if ( VALID_PTR(pvObj)
2535	&& VALID_PTR(pGVM)
2536	&& pGVM->u32Magic == GVM_MAGIC
2537	&& RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
2538	{
2539	# define MAYBE_RESET_FIELD(field) \
2540	do { if (pStats->SchedSum. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2541	MAYBE_RESET_FIELD(cHaltCalls);
2542	MAYBE_RESET_FIELD(cHaltBlocking);
2543	MAYBE_RESET_FIELD(cHaltTimeouts);
2544	MAYBE_RESET_FIELD(cHaltNotBlocking);
2545	MAYBE_RESET_FIELD(cHaltWakeUps);
2546	MAYBE_RESET_FIELD(cWakeUpCalls);
2547	MAYBE_RESET_FIELD(cWakeUpNotHalted);
2548	MAYBE_RESET_FIELD(cWakeUpWakeUps);
2549	MAYBE_RESET_FIELD(cPokeCalls);
2550	MAYBE_RESET_FIELD(cPokeNotBusy);
2551	MAYBE_RESET_FIELD(cPollCalls);
2552	MAYBE_RESET_FIELD(cPollHalts);
2553	MAYBE_RESET_FIELD(cPollWakeUps);
2554	# undef MAYBE_RESET_FIELD
2555	}
2556	}
2557	}
2558
2559	gvmmR0UsedUnlock(pGVMM);
2560
2561	return VINF_SUCCESS;
2562	}
2563
2564
2565	/**
2566	* VMMR0 request wrapper for GVMMR0ResetStatistics.
2567	*
2568	* @returns see GVMMR0ResetStatistics.
2569	* @param pVM Pointer to the VM. Optional.
2570	* @param pReq Pointer to the request packet.
2571	*/
2572	GVMMR0DECL(int) GVMMR0ResetStatisticsReq(PVM pVM, PGVMMRESETSTATISTICSSREQ pReq)
2573	{
2574	/*
2575	* Validate input and pass it on.
2576	*/
2577	AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2578	AssertMsgReturn(pReq->Hdr.cbReq == sizeof(pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(pReq)), VERR_INVALID_PARAMETER);
2579
2580	return GVMMR0ResetStatistics(&pReq->Stats, pReq->pSession, pVM);
2581	}
2582

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

source: vbox/trunk/src/VBox/VMM/VMMR0/GVMMR0.cpp@ 49966

Download in other formats: