GVMMR0.cpp@ 61632

Last change on this file since 61632 was 61144, checked in by vboxsync, 9 years ago
CPUM,HM,GVMM,TRPM,VMM: Next part of the FPU state handling for IEM. This is a little bit risky change as we now leave CR0.TS+EM cleared after saving the host state, they only get restored when we restore the host state. On Windows, Darwin, and later on Linux (needs testing) we will rely on #NM handling of the host OS and not our own CR.TS/EM handy work. This means we won't be saving the host state but rather the ring-3 state of our own thread. This change also introduces a CPUM force flag that we're using for restoring CR0.TS/EM in raw-mode (it may be extended with other uses later if we need to).
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 94.0 KB

Line
1	/* $Id: GVMMR0.cpp 61144 2016-05-23 22:16:26Z vboxsync $ */
2	/** @file
3	* GVMM - Global VM Manager.
4	*/
5
6	/*
7	* Copyright (C) 2007-2015 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/critsect.h>
66	#include <iprt/mem.h>
67	#include <iprt/semaphore.h>
68	#include <iprt/time.h>
69	#include <VBox/log.h>
70	#include <iprt/thread.h>
71	#include <iprt/process.h>
72	#include <iprt/param.h>
73	#include <iprt/string.h>
74	#include <iprt/assert.h>
75	#include <iprt/mem.h>
76	#include <iprt/memobj.h>
77	#include <iprt/mp.h>
78	#include <iprt/cpuset.h>
79	#include <iprt/spinlock.h>
80	#include <iprt/timer.h>
81
82	#include "dtrace/VBoxVMM.h"
83
84
85	/*********************************************************************************************************************************
86	* Defined Constants And Macros *
87	*********************************************************************************************************************************/
88	#if defined(RT_OS_LINUX) \|\| defined(RT_OS_SOLARIS) \|\| defined(DOXYGEN_RUNNING)
89	/** Define this to enable the periodic preemption timer. */
90	# define GVMM_SCHED_WITH_PPT
91	#endif
92
93
94	/** @def GVMM_CHECK_SMAP_SETUP
95	* SMAP check setup. */
96	/** @def GVMM_CHECK_SMAP_CHECK
97	* Checks that the AC flag is set if SMAP is enabled. If AC is not set,
98	* it will be logged and @a a_BadExpr is executed. */
99	/** @def GVMM_CHECK_SMAP_CHECK2
100	* Checks that the AC flag is set if SMAP is enabled. If AC is not set, it will
101	* be logged, written to the VMs assertion text buffer, and @a a_BadExpr is
102	* executed. */
103	#if defined(VBOX_STRICT) \|\| 1
104	# define GVMM_CHECK_SMAP_SETUP() uint32_t const fKernelFeatures = SUPR0GetKernelFeatures()
105	# define GVMM_CHECK_SMAP_CHECK(a_BadExpr) \
106	do { \
107	if (fKernelFeatures & SUPKERNELFEATURES_SMAP) \
108	{ \
109	RTCCUINTREG fEflCheck = ASMGetFlags(); \
110	if (RT_LIKELY(fEflCheck & X86_EFL_AC)) \
111	{ /* likely */ } \
112	else \
113	{ \
114	SUPR0Printf("%s, line %d: EFLAGS.AC is clear! (%#x)\n", __FUNCTION__, __LINE__, (uint32_t)fEflCheck); \
115	a_BadExpr; \
116	} \
117	} \
118	} while (0)
119	# define GVMM_CHECK_SMAP_CHECK2(a_pVM, a_BadExpr) \
120	do { \
121	if (fKernelFeatures & SUPKERNELFEATURES_SMAP) \
122	{ \
123	RTCCUINTREG fEflCheck = ASMGetFlags(); \
124	if (RT_LIKELY(fEflCheck & X86_EFL_AC)) \
125	{ /* likely */ } \
126	else \
127	{ \
128	SUPR0BadContext((a_pVM) ? (a_pVM)->pSession : NULL, __FILE__, __LINE__, "EFLAGS.AC is zero!"); \
129	a_BadExpr; \
130	} \
131	} \
132	} while (0)
133	#else
134	# define GVMM_CHECK_SMAP_SETUP() uint32_t const fKernelFeatures = 0
135	# define GVMM_CHECK_SMAP_CHECK(a_BadExpr) NOREF(fKernelFeatures)
136	# define GVMM_CHECK_SMAP_CHECK2(a_pVM, a_BadExpr) NOREF(fKernelFeatures)
137	#endif
138
139
140
141	/*********************************************************************************************************************************
142	* Structures and Typedefs *
143	*********************************************************************************************************************************/
144
145	/**
146	* Global VM handle.
147	*/
148	typedef struct GVMHANDLE
149	{
150	/** The index of the next handle in the list (free or used). (0 is nil.) */
151	uint16_t volatile iNext;
152	/** Our own index / handle value. */
153	uint16_t iSelf;
154	/** The process ID of the handle owner.
155	* This is used for access checks. */
156	RTPROCESS ProcId;
157	/** The pointer to the ring-0 only (aka global) VM structure. */
158	PGVM pGVM;
159	/** The ring-0 mapping of the shared VM instance data. */
160	PVM pVM;
161	/** The virtual machine object. */
162	void *pvObj;
163	/** The session this VM is associated with. */
164	PSUPDRVSESSION pSession;
165	/** The ring-0 handle of the EMT0 thread.
166	* This is used for ownership checks as well as looking up a VM handle by thread
167	* at times like assertions. */
168	RTNATIVETHREAD hEMT0;
169	} GVMHANDLE;
170	/** Pointer to a global VM handle. */
171	typedef GVMHANDLE *PGVMHANDLE;
172
173	/** Number of GVM handles (including the NIL handle). */
174	#if HC_ARCH_BITS == 64
175	# define GVMM_MAX_HANDLES 8192
176	#else
177	# define GVMM_MAX_HANDLES 128
178	#endif
179
180	/**
181	* Per host CPU GVMM data.
182	*/
183	typedef struct GVMMHOSTCPU
184	{
185	/** Magic number (GVMMHOSTCPU_MAGIC). */
186	uint32_t volatile u32Magic;
187	/** The CPU ID. */
188	RTCPUID idCpu;
189	/** The CPU set index. */
190	uint32_t idxCpuSet;
191
192	#ifdef GVMM_SCHED_WITH_PPT
193	/** Periodic preemption timer data. */
194	struct
195	{
196	/** The handle to the periodic preemption timer. */
197	PRTTIMER pTimer;
198	/** Spinlock protecting the data below. */
199	RTSPINLOCK hSpinlock;
200	/** The smalles Hz that we need to care about. (static) */
201	uint32_t uMinHz;
202	/** The number of ticks between each historization. */
203	uint32_t cTicksHistoriziationInterval;
204	/** The current historization tick (counting up to
205	* cTicksHistoriziationInterval and then resetting). */
206	uint32_t iTickHistorization;
207	/** The current timer interval. This is set to 0 when inactive. */
208	uint32_t cNsInterval;
209	/** The current timer frequency. This is set to 0 when inactive. */
210	uint32_t uTimerHz;
211	/** The current max frequency reported by the EMTs.
212	* This gets historicize and reset by the timer callback. This is
213	* read without holding the spinlock, so needs atomic updating. */
214	uint32_t volatile uDesiredHz;
215	/** Whether the timer was started or not. */
216	bool volatile fStarted;
217	/** Set if we're starting timer. */
218	bool volatile fStarting;
219	/** The index of the next history entry (mod it). */
220	uint32_t iHzHistory;
221	/** Historicized uDesiredHz values. The array wraps around, new entries
222	* are added at iHzHistory. This is updated approximately every
223	* GVMMHOSTCPU_PPT_HIST_INTERVAL_NS by the timer callback. */
224	uint32_t aHzHistory[8];
225	/** Statistics counter for recording the number of interval changes. */
226	uint32_t cChanges;
227	/** Statistics counter for recording the number of timer starts. */
228	uint32_t cStarts;
229	} Ppt;
230	#endif /* GVMM_SCHED_WITH_PPT */
231
232	} GVMMHOSTCPU;
233	/** Pointer to the per host CPU GVMM data. */
234	typedef GVMMHOSTCPU *PGVMMHOSTCPU;
235	/** The GVMMHOSTCPU::u32Magic value (Petra, Tanya & Rachel Haden). */
236	#define GVMMHOSTCPU_MAGIC UINT32_C(0x19711011)
237	/** The interval on history entry should cover (approximately) give in
238	* nanoseconds. */
239	#define GVMMHOSTCPU_PPT_HIST_INTERVAL_NS UINT32_C(20000000)
240
241
242	/**
243	* The GVMM instance data.
244	*/
245	typedef struct GVMM
246	{
247	/** Eyecatcher / magic. */
248	uint32_t u32Magic;
249	/** The index of the head of the free handle chain. (0 is nil.) */
250	uint16_t volatile iFreeHead;
251	/** The index of the head of the active handle chain. (0 is nil.) */
252	uint16_t volatile iUsedHead;
253	/** The number of VMs. */
254	uint16_t volatile cVMs;
255	/** Alignment padding. */
256	uint16_t u16Reserved;
257	/** The number of EMTs. */
258	uint32_t volatile cEMTs;
259	/** The number of EMTs that have halted in GVMMR0SchedHalt. */
260	uint32_t volatile cHaltedEMTs;
261	/** Mini lock for restricting early wake-ups to one thread. */
262	bool volatile fDoingEarlyWakeUps;
263	bool afPadding[3]; /*< explicit alignment padding. /
264	/** When the next halted or sleeping EMT will wake up.
265	* This is set to 0 when it needs recalculating and to UINT64_MAX when
266	* there are no halted or sleeping EMTs in the GVMM. */
267	uint64_t uNsNextEmtWakeup;
268	/** The lock used to serialize VM creation, destruction and associated events that
269	* isn't performance critical. Owners may acquire the list lock. */
270	RTCRITSECT CreateDestroyLock;
271	/** The lock used to serialize used list updates and accesses.
272	* This indirectly includes scheduling since the scheduler will have to walk the
273	* used list to examin running VMs. Owners may not acquire any other locks. */
274	RTCRITSECTRW UsedLock;
275	/** The handle array.
276	* The size of this array defines the maximum number of currently running VMs.
277	* The first entry is unused as it represents the NIL handle. */
278	GVMHANDLE aHandles[GVMM_MAX_HANDLES];
279
280	/** @gcfgm{/GVMM/cEMTsMeansCompany, 32-bit, 0, UINT32_MAX, 1}
281	* The number of EMTs that means we no longer consider ourselves alone on a
282	* CPU/Core.
283	*/
284	uint32_t cEMTsMeansCompany;
285	/** @gcfgm{/GVMM/MinSleepAlone,32-bit, 0, 100000000, 750000, ns}
286	* The minimum sleep time for when we're alone, in nano seconds.
287	*/
288	uint32_t nsMinSleepAlone;
289	/** @gcfgm{/GVMM/MinSleepCompany,32-bit,0, 100000000, 15000, ns}
290	* The minimum sleep time for when we've got company, in nano seconds.
291	*/
292	uint32_t nsMinSleepCompany;
293	/** @gcfgm{/GVMM/EarlyWakeUp1, 32-bit, 0, 100000000, 25000, ns}
294	* The limit for the first round of early wake-ups, given in nano seconds.
295	*/
296	uint32_t nsEarlyWakeUp1;
297	/** @gcfgm{/GVMM/EarlyWakeUp2, 32-bit, 0, 100000000, 50000, ns}
298	* The limit for the second round of early wake-ups, given in nano seconds.
299	*/
300	uint32_t nsEarlyWakeUp2;
301
302	/** Set if we're doing early wake-ups.
303	* This reflects nsEarlyWakeUp1 and nsEarlyWakeUp2. */
304	bool volatile fDoEarlyWakeUps;
305
306	/** The number of entries in the host CPU array (aHostCpus). */
307	uint32_t cHostCpus;
308	/** Per host CPU data (variable length). */
309	GVMMHOSTCPU aHostCpus[1];
310	} GVMM;
311	AssertCompileMemberAlignment(GVMM, CreateDestroyLock, 8);
312	AssertCompileMemberAlignment(GVMM, UsedLock, 8);
313	AssertCompileMemberAlignment(GVMM, uNsNextEmtWakeup, 8);
314	/** Pointer to the GVMM instance data. */
315	typedef GVMM *PGVMM;
316
317	/** The GVMM::u32Magic value (Charlie Haden). */
318	#define GVMM_MAGIC UINT32_C(0x19370806)
319
320
321
322	/*********************************************************************************************************************************
323	* Global Variables *
324	*********************************************************************************************************************************/
325	/** Pointer to the GVMM instance data.
326	* (Just my general dislike for global variables.) */
327	static PGVMM g_pGVMM = NULL;
328
329	/** Macro for obtaining and validating the g_pGVMM pointer.
330	* On failure it will return from the invoking function with the specified return value.
331	*
332	* @param pGVMM The name of the pGVMM variable.
333	* @param rc The return value on failure. Use VERR_GVMM_INSTANCE for VBox
334	* status codes.
335	*/
336	#define GVMM_GET_VALID_INSTANCE(pGVMM, rc) \
337	do { \
338	(pGVMM) = g_pGVMM;\
339	AssertPtrReturn((pGVMM), (rc)); \
340	AssertMsgReturn((pGVMM)->u32Magic == GVMM_MAGIC, ("%p - %#x\n", (pGVMM), (pGVMM)->u32Magic), (rc)); \
341	} while (0)
342
343	/** Macro for obtaining and validating the g_pGVMM pointer, void function variant.
344	* On failure it will return from the invoking function.
345	*
346	* @param pGVMM The name of the pGVMM variable.
347	*/
348	#define GVMM_GET_VALID_INSTANCE_VOID(pGVMM) \
349	do { \
350	(pGVMM) = g_pGVMM;\
351	AssertPtrReturnVoid((pGVMM)); \
352	AssertMsgReturnVoid((pGVMM)->u32Magic == GVMM_MAGIC, ("%p - %#x\n", (pGVMM), (pGVMM)->u32Magic)); \
353	} while (0)
354
355
356	/*********************************************************************************************************************************
357	* Internal Functions *
358	*********************************************************************************************************************************/
359	static void gvmmR0InitPerVMData(PGVM pGVM);
360	static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void pvObj, void pvGVMM, void *pvHandle);
361	static int gvmmR0ByVM(PVM pVM, PGVM ppGVM, PGVMM ppGVMM, bool fTakeUsedLock);
362	static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM ppGVM, PGVMM ppGVMM);
363	#ifdef GVMM_SCHED_WITH_PPT
364	static DECLCALLBACK(void) gvmmR0SchedPeriodicPreemptionTimerCallback(PRTTIMER pTimer, void *pvUser, uint64_t iTick);
365	#endif
366
367
368	/**
369	* Initializes the GVMM.
370	*
371	* This is called while owning the loader semaphore (see supdrvIOCtl_LdrLoad()).
372	*
373	* @returns VBox status code.
374	*/
375	GVMMR0DECL(int) GVMMR0Init(void)
376	{
377	LogFlow(("GVMMR0Init:\n"));
378
379	/*
380	* Allocate and initialize the instance data.
381	*/
382	uint32_t cHostCpus = RTMpGetArraySize();
383	AssertMsgReturn(cHostCpus > 0 && cHostCpus < _64K, ("%d", (int)cHostCpus), VERR_GVMM_HOST_CPU_RANGE);
384
385	PGVMM pGVMM = (PGVMM)RTMemAllocZ(RT_UOFFSETOF(GVMM, aHostCpus[cHostCpus]));
386	if (!pGVMM)
387	return VERR_NO_MEMORY;
388	int rc = RTCritSectInitEx(&pGVMM->CreateDestroyLock, 0, NIL_RTLOCKVALCLASS, RTLOCKVAL_SUB_CLASS_NONE,
389	"GVMM-CreateDestroyLock");
390	if (RT_SUCCESS(rc))
391	{
392	rc = RTCritSectRwInitEx(&pGVMM->UsedLock, 0, NIL_RTLOCKVALCLASS, RTLOCKVAL_SUB_CLASS_NONE, "GVMM-UsedLock");
393	if (RT_SUCCESS(rc))
394	{
395	pGVMM->u32Magic = GVMM_MAGIC;
396	pGVMM->iUsedHead = 0;
397	pGVMM->iFreeHead = 1;
398
399	/* the nil handle */
400	pGVMM->aHandles[0].iSelf = 0;
401	pGVMM->aHandles[0].iNext = 0;
402
403	/* the tail */
404	unsigned i = RT_ELEMENTS(pGVMM->aHandles) - 1;
405	pGVMM->aHandles[i].iSelf = i;
406	pGVMM->aHandles[i].iNext = 0; /* nil */
407
408	/* the rest */
409	while (i-- > 1)
410	{
411	pGVMM->aHandles[i].iSelf = i;
412	pGVMM->aHandles[i].iNext = i + 1;
413	}
414
415	/* The default configuration values. */
416	uint32_t cNsResolution = RTSemEventMultiGetResolution();
417	pGVMM->cEMTsMeansCompany = 1; /** @todo should be adjusted to relative to the cpu count or something... */
418	if (cNsResolution >= 5*RT_NS_100US)
419	{
420	pGVMM->nsMinSleepAlone = 750000 /* ns (0.750 ms) /; /* @todo this should be adjusted to be 75% (or something) of the scheduler granularity... */
421	pGVMM->nsMinSleepCompany = 15000 /* ns (0.015 ms) */;
422	pGVMM->nsEarlyWakeUp1 = 25000 /* ns (0.025 ms) */;
423	pGVMM->nsEarlyWakeUp2 = 50000 /* ns (0.050 ms) */;
424	}
425	else if (cNsResolution > RT_NS_100US)
426	{
427	pGVMM->nsMinSleepAlone = cNsResolution / 2;
428	pGVMM->nsMinSleepCompany = cNsResolution / 4;
429	pGVMM->nsEarlyWakeUp1 = 0;
430	pGVMM->nsEarlyWakeUp2 = 0;
431	}
432	else
433	{
434	pGVMM->nsMinSleepAlone = 2000;
435	pGVMM->nsMinSleepCompany = 2000;
436	pGVMM->nsEarlyWakeUp1 = 0;
437	pGVMM->nsEarlyWakeUp2 = 0;
438	}
439	pGVMM->fDoEarlyWakeUps = pGVMM->nsEarlyWakeUp1 > 0 && pGVMM->nsEarlyWakeUp2 > 0;
440
441	/* The host CPU data. */
442	pGVMM->cHostCpus = cHostCpus;
443	uint32_t iCpu = cHostCpus;
444	RTCPUSET PossibleSet;
445	RTMpGetSet(&PossibleSet);
446	while (iCpu-- > 0)
447	{
448	pGVMM->aHostCpus[iCpu].idxCpuSet = iCpu;
449	#ifdef GVMM_SCHED_WITH_PPT
450	pGVMM->aHostCpus[iCpu].Ppt.pTimer = NULL;
451	pGVMM->aHostCpus[iCpu].Ppt.hSpinlock = NIL_RTSPINLOCK;
452	pGVMM->aHostCpus[iCpu].Ppt.uMinHz = 5; /** @todo Add some API which figures this one out. (not that important) */
453	pGVMM->aHostCpus[iCpu].Ppt.cTicksHistoriziationInterval = 1;
454	//pGVMM->aHostCpus[iCpu].Ppt.iTickHistorization = 0;
455	//pGVMM->aHostCpus[iCpu].Ppt.cNsInterval = 0;
456	//pGVMM->aHostCpus[iCpu].Ppt.uTimerHz = 0;
457	//pGVMM->aHostCpus[iCpu].Ppt.uDesiredHz = 0;
458	//pGVMM->aHostCpus[iCpu].Ppt.fStarted = false;
459	//pGVMM->aHostCpus[iCpu].Ppt.fStarting = false;
460	//pGVMM->aHostCpus[iCpu].Ppt.iHzHistory = 0;
461	//pGVMM->aHostCpus[iCpu].Ppt.aHzHistory = {0};
462	#endif
463
464	if (RTCpuSetIsMember(&PossibleSet, iCpu))
465	{
466	pGVMM->aHostCpus[iCpu].idCpu = RTMpCpuIdFromSetIndex(iCpu);
467	pGVMM->aHostCpus[iCpu].u32Magic = GVMMHOSTCPU_MAGIC;
468
469	#ifdef GVMM_SCHED_WITH_PPT
470	rc = RTTimerCreateEx(&pGVMM->aHostCpus[iCpu].Ppt.pTimer,
471	5010001000 /* whatever */,
472	RTTIMER_FLAGS_CPU(iCpu) \| RTTIMER_FLAGS_HIGH_RES,
473	gvmmR0SchedPeriodicPreemptionTimerCallback,
474	&pGVMM->aHostCpus[iCpu]);
475	if (RT_SUCCESS(rc))
476	rc = RTSpinlockCreate(&pGVMM->aHostCpus[iCpu].Ppt.hSpinlock, RTSPINLOCK_FLAGS_INTERRUPT_SAFE, "GVMM/CPU");
477	if (RT_FAILURE(rc))
478	{
479	while (iCpu < cHostCpus)
480	{
481	RTTimerDestroy(pGVMM->aHostCpus[iCpu].Ppt.pTimer);
482	RTSpinlockDestroy(pGVMM->aHostCpus[iCpu].Ppt.hSpinlock);
483	pGVMM->aHostCpus[iCpu].Ppt.hSpinlock = NIL_RTSPINLOCK;
484	iCpu++;
485	}
486	break;
487	}
488	#endif
489	}
490	else
491	{
492	pGVMM->aHostCpus[iCpu].idCpu = NIL_RTCPUID;
493	pGVMM->aHostCpus[iCpu].u32Magic = 0;
494	}
495	}
496	if (RT_SUCCESS(rc))
497	{
498	g_pGVMM = pGVMM;
499	LogFlow(("GVMMR0Init: pGVMM=%p cHostCpus=%u\n", pGVMM, cHostCpus));
500	return VINF_SUCCESS;
501	}
502
503	/* bail out. */
504	RTCritSectRwDelete(&pGVMM->UsedLock);
505	}
506	RTCritSectDelete(&pGVMM->CreateDestroyLock);
507	}
508
509	RTMemFree(pGVMM);
510	return rc;
511	}
512
513
514	/**
515	* Terminates the GVM.
516	*
517	* This is called while owning the loader semaphore (see supdrvLdrFree()).
518	* And unless something is wrong, there should be absolutely no VMs
519	* registered at this point.
520	*/
521	GVMMR0DECL(void) GVMMR0Term(void)
522	{
523	LogFlow(("GVMMR0Term:\n"));
524
525	PGVMM pGVMM = g_pGVMM;
526	g_pGVMM = NULL;
527	if (RT_UNLIKELY(!VALID_PTR(pGVMM)))
528	{
529	SUPR0Printf("GVMMR0Term: pGVMM=%p\n", pGVMM);
530	return;
531	}
532
533	/*
534	* First of all, stop all active timers.
535	*/
536	uint32_t cActiveTimers = 0;
537	uint32_t iCpu = pGVMM->cHostCpus;
538	while (iCpu-- > 0)
539	{
540	ASMAtomicWriteU32(&pGVMM->aHostCpus[iCpu].u32Magic, ~GVMMHOSTCPU_MAGIC);
541	#ifdef GVMM_SCHED_WITH_PPT
542	if ( pGVMM->aHostCpus[iCpu].Ppt.pTimer != NULL
543	&& RT_SUCCESS(RTTimerStop(pGVMM->aHostCpus[iCpu].Ppt.pTimer)))
544	cActiveTimers++;
545	#endif
546	}
547	if (cActiveTimers)
548	RTThreadSleep(1); /* fudge */
549
550	/*
551	* Invalidate the and free resources.
552	*/
553	pGVMM->u32Magic = ~GVMM_MAGIC;
554	RTCritSectRwDelete(&pGVMM->UsedLock);
555	RTCritSectDelete(&pGVMM->CreateDestroyLock);
556
557	pGVMM->iFreeHead = 0;
558	if (pGVMM->iUsedHead)
559	{
560	SUPR0Printf("GVMMR0Term: iUsedHead=%#x! (cVMs=%#x cEMTs=%#x)\n", pGVMM->iUsedHead, pGVMM->cVMs, pGVMM->cEMTs);
561	pGVMM->iUsedHead = 0;
562	}
563
564	#ifdef GVMM_SCHED_WITH_PPT
565	iCpu = pGVMM->cHostCpus;
566	while (iCpu-- > 0)
567	{
568	RTTimerDestroy(pGVMM->aHostCpus[iCpu].Ppt.pTimer);
569	pGVMM->aHostCpus[iCpu].Ppt.pTimer = NULL;
570	RTSpinlockDestroy(pGVMM->aHostCpus[iCpu].Ppt.hSpinlock);
571	pGVMM->aHostCpus[iCpu].Ppt.hSpinlock = NIL_RTSPINLOCK;
572	}
573	#endif
574
575	RTMemFree(pGVMM);
576	}
577
578
579	/**
580	* A quick hack for setting global config values.
581	*
582	* @returns VBox status code.
583	*
584	* @param pSession The session handle. Used for authentication.
585	* @param pszName The variable name.
586	* @param u64Value The new value.
587	*/
588	GVMMR0DECL(int) GVMMR0SetConfig(PSUPDRVSESSION pSession, const char *pszName, uint64_t u64Value)
589	{
590	/*
591	* Validate input.
592	*/
593	PGVMM pGVMM;
594	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
595	AssertPtrReturn(pSession, VERR_INVALID_HANDLE);
596	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
597
598	/*
599	* String switch time!
600	*/
601	if (strncmp(pszName, RT_STR_TUPLE("/GVMM/")))
602	return VERR_CFGM_VALUE_NOT_FOUND; /* borrow status codes from CFGM... */
603	int rc = VINF_SUCCESS;
604	pszName += sizeof("/GVMM/") - 1;
605	if (!strcmp(pszName, "cEMTsMeansCompany"))
606	{
607	if (u64Value <= UINT32_MAX)
608	pGVMM->cEMTsMeansCompany = u64Value;
609	else
610	rc = VERR_OUT_OF_RANGE;
611	}
612	else if (!strcmp(pszName, "MinSleepAlone"))
613	{
614	if (u64Value <= RT_NS_100MS)
615	pGVMM->nsMinSleepAlone = u64Value;
616	else
617	rc = VERR_OUT_OF_RANGE;
618	}
619	else if (!strcmp(pszName, "MinSleepCompany"))
620	{
621	if (u64Value <= RT_NS_100MS)
622	pGVMM->nsMinSleepCompany = u64Value;
623	else
624	rc = VERR_OUT_OF_RANGE;
625	}
626	else if (!strcmp(pszName, "EarlyWakeUp1"))
627	{
628	if (u64Value <= RT_NS_100MS)
629	{
630	pGVMM->nsEarlyWakeUp1 = u64Value;
631	pGVMM->fDoEarlyWakeUps = pGVMM->nsEarlyWakeUp1 > 0 && pGVMM->nsEarlyWakeUp2 > 0;
632	}
633	else
634	rc = VERR_OUT_OF_RANGE;
635	}
636	else if (!strcmp(pszName, "EarlyWakeUp2"))
637	{
638	if (u64Value <= RT_NS_100MS)
639	{
640	pGVMM->nsEarlyWakeUp2 = u64Value;
641	pGVMM->fDoEarlyWakeUps = pGVMM->nsEarlyWakeUp1 > 0 && pGVMM->nsEarlyWakeUp2 > 0;
642	}
643	else
644	rc = VERR_OUT_OF_RANGE;
645	}
646	else
647	rc = VERR_CFGM_VALUE_NOT_FOUND;
648	return rc;
649	}
650
651
652	/**
653	* A quick hack for getting global config values.
654	*
655	* @returns VBox status code.
656	*
657	* @param pSession The session handle. Used for authentication.
658	* @param pszName The variable name.
659	* @param pu64Value Where to return the value.
660	*/
661	GVMMR0DECL(int) GVMMR0QueryConfig(PSUPDRVSESSION pSession, const char pszName, uint64_t pu64Value)
662	{
663	/*
664	* Validate input.
665	*/
666	PGVMM pGVMM;
667	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
668	AssertPtrReturn(pSession, VERR_INVALID_HANDLE);
669	AssertPtrReturn(pszName, VERR_INVALID_POINTER);
670	AssertPtrReturn(pu64Value, VERR_INVALID_POINTER);
671
672	/*
673	* String switch time!
674	*/
675	if (strncmp(pszName, RT_STR_TUPLE("/GVMM/")))
676	return VERR_CFGM_VALUE_NOT_FOUND; /* borrow status codes from CFGM... */
677	int rc = VINF_SUCCESS;
678	pszName += sizeof("/GVMM/") - 1;
679	if (!strcmp(pszName, "cEMTsMeansCompany"))
680	*pu64Value = pGVMM->cEMTsMeansCompany;
681	else if (!strcmp(pszName, "MinSleepAlone"))
682	*pu64Value = pGVMM->nsMinSleepAlone;
683	else if (!strcmp(pszName, "MinSleepCompany"))
684	*pu64Value = pGVMM->nsMinSleepCompany;
685	else if (!strcmp(pszName, "EarlyWakeUp1"))
686	*pu64Value = pGVMM->nsEarlyWakeUp1;
687	else if (!strcmp(pszName, "EarlyWakeUp2"))
688	*pu64Value = pGVMM->nsEarlyWakeUp2;
689	else
690	rc = VERR_CFGM_VALUE_NOT_FOUND;
691	return rc;
692	}
693
694
695	/**
696	* Acquire the 'used' lock in shared mode.
697	*
698	* This prevents destruction of the VM while we're in ring-0.
699	*
700	* @returns IPRT status code, see RTSemFastMutexRequest.
701	* @param a_pGVMM The GVMM instance data.
702	* @sa GVMMR0_USED_SHARED_UNLOCK, GVMMR0_USED_EXCLUSIVE_LOCK
703	*/
704	#define GVMMR0_USED_SHARED_LOCK(a_pGVMM) RTCritSectRwEnterShared(&(a_pGVMM)->UsedLock)
705
706	/**
707	* Release the 'used' lock in when owning it in shared mode.
708	*
709	* @returns IPRT status code, see RTSemFastMutexRequest.
710	* @param a_pGVMM The GVMM instance data.
711	* @sa GVMMR0_USED_SHARED_LOCK
712	*/
713	#define GVMMR0_USED_SHARED_UNLOCK(a_pGVMM) RTCritSectRwLeaveShared(&(a_pGVMM)->UsedLock)
714
715	/**
716	* Acquire the 'used' lock in exclusive mode.
717	*
718	* Only use this function when making changes to the used list.
719	*
720	* @returns IPRT status code, see RTSemFastMutexRequest.
721	* @param a_pGVMM The GVMM instance data.
722	* @sa GVMMR0_USED_EXCLUSIVE_UNLOCK
723	*/
724	#define GVMMR0_USED_EXCLUSIVE_LOCK(a_pGVMM) RTCritSectRwEnterExcl(&(a_pGVMM)->UsedLock)
725
726	/**
727	* Release the 'used' lock when owning it in exclusive mode.
728	*
729	* @returns IPRT status code, see RTSemFastMutexRelease.
730	* @param a_pGVMM The GVMM instance data.
731	* @sa GVMMR0_USED_EXCLUSIVE_LOCK, GVMMR0_USED_SHARED_UNLOCK
732	*/
733	#define GVMMR0_USED_EXCLUSIVE_UNLOCK(a_pGVMM) RTCritSectRwLeaveExcl(&(a_pGVMM)->UsedLock)
734
735
736	/**
737	* Try acquire the 'create & destroy' lock.
738	*
739	* @returns IPRT status code, see RTSemFastMutexRequest.
740	* @param pGVMM The GVMM instance data.
741	*/
742	DECLINLINE(int) gvmmR0CreateDestroyLock(PGVMM pGVMM)
743	{
744	LogFlow(("++gvmmR0CreateDestroyLock(%p)\n", pGVMM));
745	int rc = RTCritSectEnter(&pGVMM->CreateDestroyLock);
746	LogFlow(("gvmmR0CreateDestroyLock(%p)->%Rrc\n", pGVMM, rc));
747	return rc;
748	}
749
750
751	/**
752	* Release the 'create & destroy' lock.
753	*
754	* @returns IPRT status code, see RTSemFastMutexRequest.
755	* @param pGVMM The GVMM instance data.
756	*/
757	DECLINLINE(int) gvmmR0CreateDestroyUnlock(PGVMM pGVMM)
758	{
759	LogFlow(("--gvmmR0CreateDestroyUnlock(%p)\n", pGVMM));
760	int rc = RTCritSectLeave(&pGVMM->CreateDestroyLock);
761	AssertRC(rc);
762	return rc;
763	}
764
765
766	/**
767	* Request wrapper for the GVMMR0CreateVM API.
768	*
769	* @returns VBox status code.
770	* @param pReq The request buffer.
771	*/
772	GVMMR0DECL(int) GVMMR0CreateVMReq(PGVMMCREATEVMREQ pReq)
773	{
774	/*
775	* Validate the request.
776	*/
777	if (!VALID_PTR(pReq))
778	return VERR_INVALID_POINTER;
779	if (pReq->Hdr.cbReq != sizeof(*pReq))
780	return VERR_INVALID_PARAMETER;
781	if (!VALID_PTR(pReq->pSession))
782	return VERR_INVALID_POINTER;
783
784	/*
785	* Execute it.
786	*/
787	PVM pVM;
788	pReq->pVMR0 = NULL;
789	pReq->pVMR3 = NIL_RTR3PTR;
790	int rc = GVMMR0CreateVM(pReq->pSession, pReq->cCpus, &pVM);
791	if (RT_SUCCESS(rc))
792	{
793	pReq->pVMR0 = pVM;
794	pReq->pVMR3 = pVM->pVMR3;
795	}
796	return rc;
797	}
798
799
800	/**
801	* Allocates the VM structure and registers it with GVM.
802	*
803	* The caller will become the VM owner and there by the EMT.
804	*
805	* @returns VBox status code.
806	* @param pSession The support driver session.
807	* @param cCpus Number of virtual CPUs for the new VM.
808	* @param ppVM Where to store the pointer to the VM structure.
809	*
810	* @thread EMT.
811	*/
812	GVMMR0DECL(int) GVMMR0CreateVM(PSUPDRVSESSION pSession, uint32_t cCpus, PVM *ppVM)
813	{
814	LogFlow(("GVMMR0CreateVM: pSession=%p\n", pSession));
815	PGVMM pGVMM;
816	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
817
818	AssertPtrReturn(ppVM, VERR_INVALID_POINTER);
819	*ppVM = NULL;
820
821	if ( cCpus == 0
822	\|\| cCpus > VMM_MAX_CPU_COUNT)
823	return VERR_INVALID_PARAMETER;
824
825	RTNATIVETHREAD hEMT0 = RTThreadNativeSelf();
826	AssertReturn(hEMT0 != NIL_RTNATIVETHREAD, VERR_GVMM_BROKEN_IPRT);
827	RTPROCESS ProcId = RTProcSelf();
828	AssertReturn(ProcId != NIL_RTPROCESS, VERR_GVMM_BROKEN_IPRT);
829
830	/*
831	* The whole allocation process is protected by the lock.
832	*/
833	int rc = gvmmR0CreateDestroyLock(pGVMM);
834	AssertRCReturn(rc, rc);
835
836	/*
837	* Allocate a handle first so we don't waste resources unnecessarily.
838	*/
839	uint16_t iHandle = pGVMM->iFreeHead;
840	if (iHandle)
841	{
842	PGVMHANDLE pHandle = &pGVMM->aHandles[iHandle];
843
844	/* consistency checks, a bit paranoid as always. */
845	if ( !pHandle->pVM
846	&& !pHandle->pGVM
847	&& !pHandle->pvObj
848	&& pHandle->iSelf == iHandle)
849	{
850	pHandle->pvObj = SUPR0ObjRegister(pSession, SUPDRVOBJTYPE_VM, gvmmR0HandleObjDestructor, pGVMM, pHandle);
851	if (pHandle->pvObj)
852	{
853	/*
854	* Move the handle from the free to used list and perform permission checks.
855	*/
856	rc = GVMMR0_USED_EXCLUSIVE_LOCK(pGVMM);
857	AssertRC(rc);
858
859	pGVMM->iFreeHead = pHandle->iNext;
860	pHandle->iNext = pGVMM->iUsedHead;
861	pGVMM->iUsedHead = iHandle;
862	pGVMM->cVMs++;
863
864	pHandle->pVM = NULL;
865	pHandle->pGVM = NULL;
866	pHandle->pSession = pSession;
867	pHandle->hEMT0 = NIL_RTNATIVETHREAD;
868	pHandle->ProcId = NIL_RTPROCESS;
869
870	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
871
872	rc = SUPR0ObjVerifyAccess(pHandle->pvObj, pSession, NULL);
873	if (RT_SUCCESS(rc))
874	{
875	/*
876	* Allocate the global VM structure (GVM) and initialize it.
877	*/
878	PGVM pGVM = (PGVM)RTMemAllocZ(RT_UOFFSETOF(GVM, aCpus[cCpus]));
879	if (pGVM)
880	{
881	pGVM->u32Magic = GVM_MAGIC;
882	pGVM->hSelf = iHandle;
883	pGVM->pVM = NULL;
884	pGVM->cCpus = cCpus;
885
886	gvmmR0InitPerVMData(pGVM);
887	GMMR0InitPerVMData(pGVM);
888
889	/*
890	* Allocate the shared VM structure and associated page array.
891	*/
892	const uint32_t cbVM = RT_UOFFSETOF(VM, aCpus[cCpus]);
893	const uint32_t cPages = RT_ALIGN_32(cbVM, PAGE_SIZE) >> PAGE_SHIFT;
894	rc = RTR0MemObjAllocLow(&pGVM->gvmm.s.VMMemObj, cPages << PAGE_SHIFT, false /* fExecutable */);
895	if (RT_SUCCESS(rc))
896	{
897	PVM pVM = (PVM)RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj); AssertPtr(pVM);
898	memset(pVM, 0, cPages << PAGE_SHIFT);
899	pVM->enmVMState = VMSTATE_CREATING;
900	pVM->pVMR0 = pVM;
901	pVM->pSession = pSession;
902	pVM->hSelf = iHandle;
903	pVM->cbSelf = cbVM;
904	pVM->cCpus = cCpus;
905	pVM->uCpuExecutionCap = 100; /* default is no cap. */
906	pVM->offVMCPU = RT_UOFFSETOF(VM, aCpus);
907	AssertCompileMemberAlignment(VM, cpum, 64);
908	AssertCompileMemberAlignment(VM, tm, 64);
909	AssertCompileMemberAlignment(VM, aCpus, PAGE_SIZE);
910
911	rc = RTR0MemObjAllocPage(&pGVM->gvmm.s.VMPagesMemObj, cPages * sizeof(SUPPAGE), false /* fExecutable */);
912	if (RT_SUCCESS(rc))
913	{
914	PSUPPAGE paPages = (PSUPPAGE)RTR0MemObjAddress(pGVM->gvmm.s.VMPagesMemObj); AssertPtr(paPages);
915	for (uint32_t iPage = 0; iPage < cPages; iPage++)
916	{
917	paPages[iPage].uReserved = 0;
918	paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pGVM->gvmm.s.VMMemObj, iPage);
919	Assert(paPages[iPage].Phys != NIL_RTHCPHYS);
920	}
921
922	/*
923	* Map them into ring-3.
924	*/
925	rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMMapObj, pGVM->gvmm.s.VMMemObj, (RTR3PTR)-1, 0,
926	RTMEM_PROT_READ \| RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
927	if (RT_SUCCESS(rc))
928	{
929	pVM->pVMR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMMapObj);
930	AssertPtr((void *)pVM->pVMR3);
931
932	/* Initialize all the VM pointers. */
933	for (uint32_t i = 0; i < cCpus; i++)
934	{
935	pVM->aCpus[i].pVMR0 = pVM;
936	pVM->aCpus[i].pVMR3 = pVM->pVMR3;
937	pVM->aCpus[i].idHostCpu = NIL_RTCPUID;
938	pVM->aCpus[i].hNativeThreadR0 = NIL_RTNATIVETHREAD;
939	}
940
941	rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMPagesMapObj, pGVM->gvmm.s.VMPagesMemObj, (RTR3PTR)-1,
942	0 /* uAlignment */, RTMEM_PROT_READ \| RTMEM_PROT_WRITE,
943	NIL_RTR0PROCESS);
944	if (RT_SUCCESS(rc))
945	{
946	pVM->paVMPagesR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMPagesMapObj);
947	AssertPtr((void *)pVM->paVMPagesR3);
948
949	/* complete the handle - take the UsedLock sem just to be careful. */
950	rc = GVMMR0_USED_EXCLUSIVE_LOCK(pGVMM);
951	AssertRC(rc);
952
953	pHandle->pVM = pVM;
954	pHandle->pGVM = pGVM;
955	pHandle->hEMT0 = hEMT0;
956	pHandle->ProcId = ProcId;
957	pGVM->pVM = pVM;
958	pGVM->aCpus[0].hEMT = hEMT0;
959	pVM->aCpus[0].hNativeThreadR0 = hEMT0;
960	pGVMM->cEMTs += cCpus;
961
962	rc = VMMR0ThreadCtxHookCreateForEmt(&pVM->aCpus[0]);
963	if (RT_SUCCESS(rc))
964	{
965	VBOXVMM_R0_GVMM_VM_CREATED(pGVM, pVM, ProcId, (void *)hEMT0, cCpus);
966
967	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
968	gvmmR0CreateDestroyUnlock(pGVMM);
969
970	CPUMR0RegisterVCpuThread(&pVM->aCpus[0]);
971
972	*ppVM = pVM;
973	Log(("GVMMR0CreateVM: pVM=%p pVMR3=%p pGVM=%p hGVM=%d\n", pVM, pVM->pVMR3, pGVM, iHandle));
974	return VINF_SUCCESS;
975	}
976
977	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
978	}
979
980	RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */);
981	pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
982	}
983	RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */);
984	pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
985	}
986	RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */);
987	pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
988	}
989	}
990	}
991	/* else: The user wasn't permitted to create this VM. */
992
993	/*
994	* The handle will be freed by gvmmR0HandleObjDestructor as we release the
995	* object reference here. A little extra mess because of non-recursive lock.
996	*/
997	void *pvObj = pHandle->pvObj;
998	pHandle->pvObj = NULL;
999	gvmmR0CreateDestroyUnlock(pGVMM);
1000
1001	SUPR0ObjRelease(pvObj, pSession);
1002
1003	SUPR0Printf("GVMMR0CreateVM: failed, rc=%d\n", rc);
1004	return rc;
1005	}
1006
1007	rc = VERR_NO_MEMORY;
1008	}
1009	else
1010	rc = VERR_GVMM_IPE_1;
1011	}
1012	else
1013	rc = VERR_GVM_TOO_MANY_VMS;
1014
1015	gvmmR0CreateDestroyUnlock(pGVMM);
1016	return rc;
1017	}
1018
1019
1020	/**
1021	* Initializes the per VM data belonging to GVMM.
1022	*
1023	* @param pGVM Pointer to the global VM structure.
1024	*/
1025	static void gvmmR0InitPerVMData(PGVM pGVM)
1026	{
1027	AssertCompile(RT_SIZEOFMEMB(GVM,gvmm.s) <= RT_SIZEOFMEMB(GVM,gvmm.padding));
1028	AssertCompile(RT_SIZEOFMEMB(GVMCPU,gvmm.s) <= RT_SIZEOFMEMB(GVMCPU,gvmm.padding));
1029	pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
1030	pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
1031	pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
1032	pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
1033	pGVM->gvmm.s.fDoneVMMR0Init = false;
1034	pGVM->gvmm.s.fDoneVMMR0Term = false;
1035
1036	for (VMCPUID i = 0; i < pGVM->cCpus; i++)
1037	{
1038	pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
1039	pGVM->aCpus[i].hEMT = NIL_RTNATIVETHREAD;
1040	}
1041	}
1042
1043
1044	/**
1045	* Does the VM initialization.
1046	*
1047	* @returns VBox status code.
1048	* @param pVM The cross context VM structure.
1049	*/
1050	GVMMR0DECL(int) GVMMR0InitVM(PVM pVM)
1051	{
1052	LogFlow(("GVMMR0InitVM: pVM=%p\n", pVM));
1053
1054	/*
1055	* Validate the VM structure, state and handle.
1056	*/
1057	PGVM pGVM;
1058	PGVMM pGVMM;
1059	int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
1060	if (RT_SUCCESS(rc))
1061	{
1062	if ( !pGVM->gvmm.s.fDoneVMMR0Init
1063	&& pGVM->aCpus[0].gvmm.s.HaltEventMulti == NIL_RTSEMEVENTMULTI)
1064	{
1065	for (VMCPUID i = 0; i < pGVM->cCpus; i++)
1066	{
1067	rc = RTSemEventMultiCreate(&pGVM->aCpus[i].gvmm.s.HaltEventMulti);
1068	if (RT_FAILURE(rc))
1069	{
1070	pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
1071	break;
1072	}
1073	}
1074	}
1075	else
1076	rc = VERR_WRONG_ORDER;
1077	}
1078
1079	LogFlow(("GVMMR0InitVM: returns %Rrc\n", rc));
1080	return rc;
1081	}
1082
1083
1084	/**
1085	* Indicates that we're done with the ring-0 initialization
1086	* of the VM.
1087	*
1088	* @param pVM The cross context VM structure.
1089	* @thread EMT(0)
1090	*/
1091	GVMMR0DECL(void) GVMMR0DoneInitVM(PVM pVM)
1092	{
1093	/* Validate the VM structure, state and handle. */
1094	PGVM pGVM;
1095	PGVMM pGVMM;
1096	int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
1097	AssertRCReturnVoid(rc);
1098
1099	/* Set the indicator. */
1100	pGVM->gvmm.s.fDoneVMMR0Init = true;
1101	}
1102
1103
1104	/**
1105	* Indicates that we're doing the ring-0 termination of the VM.
1106	*
1107	* @returns true if termination hasn't been done already, false if it has.
1108	* @param pVM The cross context VM structure.
1109	* @param pGVM Pointer to the global VM structure. Optional.
1110	* @thread EMT(0)
1111	*/
1112	GVMMR0DECL(bool) GVMMR0DoingTermVM(PVM pVM, PGVM pGVM)
1113	{
1114	/* Validate the VM structure, state and handle. */
1115	AssertPtrNullReturn(pGVM, false);
1116	AssertReturn(!pGVM \|\| pGVM->u32Magic == GVM_MAGIC, false);
1117	if (!pGVM)
1118	{
1119	PGVMM pGVMM;
1120	int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
1121	AssertRCReturn(rc, false);
1122	}
1123
1124	/* Set the indicator. */
1125	if (pGVM->gvmm.s.fDoneVMMR0Term)
1126	return false;
1127	pGVM->gvmm.s.fDoneVMMR0Term = true;
1128	return true;
1129	}
1130
1131
1132	/**
1133	* Destroys the VM, freeing all associated resources (the ring-0 ones anyway).
1134	*
1135	* This is call from the vmR3DestroyFinalBit and from a error path in VMR3Create,
1136	* and the caller is not the EMT thread, unfortunately. For security reasons, it
1137	* would've been nice if the caller was actually the EMT thread or that we somehow
1138	* could've associated the calling thread with the VM up front.
1139	*
1140	* @returns VBox status code.
1141	* @param pVM The cross context VM structure.
1142	*
1143	* @thread EMT(0) if it's associated with the VM, otherwise any thread.
1144	*/
1145	GVMMR0DECL(int) GVMMR0DestroyVM(PVM pVM)
1146	{
1147	LogFlow(("GVMMR0DestroyVM: pVM=%p\n", pVM));
1148	PGVMM pGVMM;
1149	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
1150
1151	/*
1152	* Validate the VM structure, state and caller.
1153	*/
1154	AssertPtrReturn(pVM, VERR_INVALID_POINTER);
1155	AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
1156	AssertMsgReturn(pVM->enmVMState >= VMSTATE_CREATING && pVM->enmVMState <= VMSTATE_TERMINATED, ("%d\n", pVM->enmVMState),
1157	VERR_WRONG_ORDER);
1158
1159	uint32_t hGVM = pVM->hSelf;
1160	AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
1161	AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
1162
1163	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1164	AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
1165
1166	RTPROCESS ProcId = RTProcSelf();
1167	RTNATIVETHREAD hSelf = RTThreadNativeSelf();
1168	AssertReturn( ( pHandle->hEMT0 == hSelf
1169	&& pHandle->ProcId == ProcId)
1170	\|\| pHandle->hEMT0 == NIL_RTNATIVETHREAD, VERR_NOT_OWNER);
1171
1172	/*
1173	* Lookup the handle and destroy the object.
1174	* Since the lock isn't recursive and we'll have to leave it before dereferencing the
1175	* object, we take some precautions against racing callers just in case...
1176	*/
1177	int rc = gvmmR0CreateDestroyLock(pGVMM);
1178	AssertRC(rc);
1179
1180	/* Be careful here because we might theoretically be racing someone else cleaning up. */
1181	if ( pHandle->pVM == pVM
1182	&& ( ( pHandle->hEMT0 == hSelf
1183	&& pHandle->ProcId == ProcId)
1184	\|\| pHandle->hEMT0 == NIL_RTNATIVETHREAD)
1185	&& VALID_PTR(pHandle->pvObj)
1186	&& VALID_PTR(pHandle->pSession)
1187	&& VALID_PTR(pHandle->pGVM)
1188	&& pHandle->pGVM->u32Magic == GVM_MAGIC)
1189	{
1190	void *pvObj = pHandle->pvObj;
1191	pHandle->pvObj = NULL;
1192	gvmmR0CreateDestroyUnlock(pGVMM);
1193
1194	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1195	{
1196	/** @todo Can we busy wait here for all thread-context hooks to be
1197	* deregistered before releasing (destroying) it? Only until we find a
1198	* solution for not deregistering hooks everytime we're leaving HMR0
1199	* context. */
1200	VMMR0ThreadCtxHookDestroyForEmt(&pVM->aCpus[idCpu]);
1201	}
1202
1203	SUPR0ObjRelease(pvObj, pHandle->pSession);
1204	}
1205	else
1206	{
1207	SUPR0Printf("GVMMR0DestroyVM: pHandle=%p:{.pVM=%p, .hEMT0=%p, .ProcId=%u, .pvObj=%p} pVM=%p hSelf=%p\n",
1208	pHandle, pHandle->pVM, pHandle->hEMT0, pHandle->ProcId, pHandle->pvObj, pVM, hSelf);
1209	gvmmR0CreateDestroyUnlock(pGVMM);
1210	rc = VERR_GVMM_IPE_2;
1211	}
1212
1213	return rc;
1214	}
1215
1216
1217	/**
1218	* Performs VM cleanup task as part of object destruction.
1219	*
1220	* @param pGVM The GVM pointer.
1221	*/
1222	static void gvmmR0CleanupVM(PGVM pGVM)
1223	{
1224	if ( pGVM->gvmm.s.fDoneVMMR0Init
1225	&& !pGVM->gvmm.s.fDoneVMMR0Term)
1226	{
1227	if ( pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ
1228	&& RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj) == pGVM->pVM)
1229	{
1230	LogFlow(("gvmmR0CleanupVM: Calling VMMR0TermVM\n"));
1231	VMMR0TermVM(pGVM->pVM, pGVM);
1232	}
1233	else
1234	AssertMsgFailed(("gvmmR0CleanupVM: VMMemObj=%p pVM=%p\n", pGVM->gvmm.s.VMMemObj, pGVM->pVM));
1235	}
1236
1237	GMMR0CleanupVM(pGVM);
1238	}
1239
1240
1241	/**
1242	* @callback_method_impl{FNSUPDRVDESTRUCTOR,VM handle destructor}
1243	*
1244	* pvUser1 is the GVM instance pointer.
1245	* pvUser2 is the handle pointer.
1246	*/
1247	static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void pvObj, void pvUser1, void *pvUser2)
1248	{
1249	LogFlow(("gvmmR0HandleObjDestructor: %p %p %p\n", pvObj, pvUser1, pvUser2));
1250
1251	NOREF(pvObj);
1252
1253	/*
1254	* Some quick, paranoid, input validation.
1255	*/
1256	PGVMHANDLE pHandle = (PGVMHANDLE)pvUser2;
1257	AssertPtr(pHandle);
1258	PGVMM pGVMM = (PGVMM)pvUser1;
1259	Assert(pGVMM == g_pGVMM);
1260	const uint16_t iHandle = pHandle - &pGVMM->aHandles[0];
1261	if ( !iHandle
1262	\|\| iHandle >= RT_ELEMENTS(pGVMM->aHandles)
1263	\|\| iHandle != pHandle->iSelf)
1264	{
1265	SUPR0Printf("GVM: handle %d is out of range or corrupt (iSelf=%d)!\n", iHandle, pHandle->iSelf);
1266	return;
1267	}
1268
1269	int rc = gvmmR0CreateDestroyLock(pGVMM);
1270	AssertRC(rc);
1271	rc = GVMMR0_USED_EXCLUSIVE_LOCK(pGVMM);
1272	AssertRC(rc);
1273
1274	/*
1275	* This is a tad slow but a doubly linked list is too much hassle.
1276	*/
1277	if (RT_UNLIKELY(pHandle->iNext >= RT_ELEMENTS(pGVMM->aHandles)))
1278	{
1279	SUPR0Printf("GVM: used list index %d is out of range!\n", pHandle->iNext);
1280	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
1281	gvmmR0CreateDestroyUnlock(pGVMM);
1282	return;
1283	}
1284
1285	if (pGVMM->iUsedHead == iHandle)
1286	pGVMM->iUsedHead = pHandle->iNext;
1287	else
1288	{
1289	uint16_t iPrev = pGVMM->iUsedHead;
1290	int c = RT_ELEMENTS(pGVMM->aHandles) + 2;
1291	while (iPrev)
1292	{
1293	if (RT_UNLIKELY(iPrev >= RT_ELEMENTS(pGVMM->aHandles)))
1294	{
1295	SUPR0Printf("GVM: used list index %d is out of range!\n", iPrev);
1296	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
1297	gvmmR0CreateDestroyUnlock(pGVMM);
1298	return;
1299	}
1300	if (RT_UNLIKELY(c-- <= 0))
1301	{
1302	iPrev = 0;
1303	break;
1304	}
1305
1306	if (pGVMM->aHandles[iPrev].iNext == iHandle)
1307	break;
1308	iPrev = pGVMM->aHandles[iPrev].iNext;
1309	}
1310	if (!iPrev)
1311	{
1312	SUPR0Printf("GVM: can't find the handle previous previous of %d!\n", pHandle->iSelf);
1313	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
1314	gvmmR0CreateDestroyUnlock(pGVMM);
1315	return;
1316	}
1317
1318	Assert(pGVMM->aHandles[iPrev].iNext == iHandle);
1319	pGVMM->aHandles[iPrev].iNext = pHandle->iNext;
1320	}
1321	pHandle->iNext = 0;
1322	pGVMM->cVMs--;
1323
1324	/*
1325	* Do the global cleanup round.
1326	*/
1327	PGVM pGVM = pHandle->pGVM;
1328	if ( VALID_PTR(pGVM)
1329	&& pGVM->u32Magic == GVM_MAGIC)
1330	{
1331	pGVMM->cEMTs -= pGVM->cCpus;
1332	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
1333
1334	gvmmR0CleanupVM(pGVM);
1335
1336	/*
1337	* Do the GVMM cleanup - must be done last.
1338	*/
1339	/* The VM and VM pages mappings/allocations. */
1340	if (pGVM->gvmm.s.VMPagesMapObj != NIL_RTR0MEMOBJ)
1341	{
1342	rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMapObj, false /* fFreeMappings */); AssertRC(rc);
1343	pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
1344	}
1345
1346	if (pGVM->gvmm.s.VMMapObj != NIL_RTR0MEMOBJ)
1347	{
1348	rc = RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */); AssertRC(rc);
1349	pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
1350	}
1351
1352	if (pGVM->gvmm.s.VMPagesMemObj != NIL_RTR0MEMOBJ)
1353	{
1354	rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */); AssertRC(rc);
1355	pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
1356	}
1357
1358	if (pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ)
1359	{
1360	rc = RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */); AssertRC(rc);
1361	pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
1362	}
1363
1364	for (VMCPUID i = 0; i < pGVM->cCpus; i++)
1365	{
1366	if (pGVM->aCpus[i].gvmm.s.HaltEventMulti != NIL_RTSEMEVENTMULTI)
1367	{
1368	rc = RTSemEventMultiDestroy(pGVM->aCpus[i].gvmm.s.HaltEventMulti); AssertRC(rc);
1369	pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
1370	}
1371	}
1372
1373	/* the GVM structure itself. */
1374	pGVM->u32Magic \|= UINT32_C(0x80000000);
1375	RTMemFree(pGVM);
1376
1377	/* Re-acquire the UsedLock before freeing the handle since we're updating handle fields. */
1378	rc = GVMMR0_USED_EXCLUSIVE_LOCK(pGVMM);
1379	AssertRC(rc);
1380	}
1381	/* else: GVMMR0CreateVM cleanup. */
1382
1383	/*
1384	* Free the handle.
1385	*/
1386	pHandle->iNext = pGVMM->iFreeHead;
1387	pGVMM->iFreeHead = iHandle;
1388	ASMAtomicWriteNullPtr(&pHandle->pGVM);
1389	ASMAtomicWriteNullPtr(&pHandle->pVM);
1390	ASMAtomicWriteNullPtr(&pHandle->pvObj);
1391	ASMAtomicWriteNullPtr(&pHandle->pSession);
1392	ASMAtomicWriteHandle(&pHandle->hEMT0, NIL_RTNATIVETHREAD);
1393	ASMAtomicWriteU32(&pHandle->ProcId, NIL_RTPROCESS);
1394
1395	GVMMR0_USED_EXCLUSIVE_UNLOCK(pGVMM);
1396	gvmmR0CreateDestroyUnlock(pGVMM);
1397	LogFlow(("gvmmR0HandleObjDestructor: returns\n"));
1398	}
1399
1400
1401	/**
1402	* Registers the calling thread as the EMT of a Virtual CPU.
1403	*
1404	* Note that VCPU 0 is automatically registered during VM creation.
1405	*
1406	* @returns VBox status code
1407	* @param pVM The cross context VM structure.
1408	* @param idCpu VCPU id.
1409	*/
1410	GVMMR0DECL(int) GVMMR0RegisterVCpu(PVM pVM, VMCPUID idCpu)
1411	{
1412	AssertReturn(idCpu != 0, VERR_NOT_OWNER);
1413
1414	/*
1415	* Validate the VM structure, state and handle.
1416	*/
1417	PGVM pGVM;
1418	PGVMM pGVMM;
1419	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, false /* fTakeUsedLock /); /* @todo take lock here. */
1420	if (RT_FAILURE(rc))
1421	return rc;
1422
1423	AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
1424	AssertReturn(pGVM->aCpus[idCpu].hEMT == NIL_RTNATIVETHREAD, VERR_ACCESS_DENIED);
1425	Assert(pGVM->cCpus == pVM->cCpus);
1426	Assert(pVM->aCpus[idCpu].hNativeThreadR0 == NIL_RTNATIVETHREAD);
1427
1428	pVM->aCpus[idCpu].hNativeThreadR0 = pGVM->aCpus[idCpu].hEMT = RTThreadNativeSelf();
1429
1430	rc = VMMR0ThreadCtxHookCreateForEmt(&pVM->aCpus[idCpu]);
1431	if (RT_SUCCESS(rc))
1432	CPUMR0RegisterVCpuThread(&pVM->aCpus[idCpu]);
1433	return rc;
1434	}
1435
1436
1437	/**
1438	* Lookup a GVM structure by its handle.
1439	*
1440	* @returns The GVM pointer on success, NULL on failure.
1441	* @param hGVM The global VM handle. Asserts on bad handle.
1442	*/
1443	GVMMR0DECL(PGVM) GVMMR0ByHandle(uint32_t hGVM)
1444	{
1445	PGVMM pGVMM;
1446	GVMM_GET_VALID_INSTANCE(pGVMM, NULL);
1447
1448	/*
1449	* Validate.
1450	*/
1451	AssertReturn(hGVM != NIL_GVM_HANDLE, NULL);
1452	AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), NULL);
1453
1454	/*
1455	* Look it up.
1456	*/
1457	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1458	AssertPtrReturn(pHandle->pVM, NULL);
1459	AssertPtrReturn(pHandle->pvObj, NULL);
1460	PGVM pGVM = pHandle->pGVM;
1461	AssertPtrReturn(pGVM, NULL);
1462	AssertReturn(pGVM->pVM == pHandle->pVM, NULL);
1463
1464	return pHandle->pGVM;
1465	}
1466
1467
1468	/**
1469	* Lookup a GVM structure by the shared VM structure.
1470	*
1471	* The calling thread must be in the same process as the VM. All current lookups
1472	* are by threads inside the same process, so this will not be an issue.
1473	*
1474	* @returns VBox status code.
1475	* @param pVM The cross context VM structure.
1476	* @param ppGVM Where to store the GVM pointer.
1477	* @param ppGVMM Where to store the pointer to the GVMM instance data.
1478	* @param fTakeUsedLock Whether to take the used lock or not. We take it in
1479	* shared mode when requested.
1480	*
1481	* Be very careful if not taking the lock as it's
1482	* possible that the VM will disappear then!
1483	*
1484	* @remark This will not assert on an invalid pVM but try return silently.
1485	*/
1486	static int gvmmR0ByVM(PVM pVM, PGVM ppGVM, PGVMM ppGVMM, bool fTakeUsedLock)
1487	{
1488	RTPROCESS ProcId = RTProcSelf();
1489	PGVMM pGVMM;
1490	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
1491
1492	/*
1493	* Validate.
1494	*/
1495	if (RT_UNLIKELY( !VALID_PTR(pVM)
1496	\|\| ((uintptr_t)pVM & PAGE_OFFSET_MASK)))
1497	return VERR_INVALID_POINTER;
1498	if (RT_UNLIKELY( pVM->enmVMState < VMSTATE_CREATING
1499	\|\| pVM->enmVMState >= VMSTATE_TERMINATED))
1500	return VERR_INVALID_POINTER;
1501
1502	uint16_t hGVM = pVM->hSelf;
1503	if (RT_UNLIKELY( hGVM == NIL_GVM_HANDLE
1504	\|\| hGVM >= RT_ELEMENTS(pGVMM->aHandles)))
1505	return VERR_INVALID_HANDLE;
1506
1507	/*
1508	* Look it up.
1509	*/
1510	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1511	PGVM pGVM;
1512	if (fTakeUsedLock)
1513	{
1514	int rc = GVMMR0_USED_SHARED_LOCK(pGVMM);
1515	AssertRCReturn(rc, rc);
1516
1517	pGVM = pHandle->pGVM;
1518	if (RT_UNLIKELY( pHandle->pVM != pVM
1519	\|\| pHandle->ProcId != ProcId
1520	\|\| !VALID_PTR(pHandle->pvObj)
1521	\|\| !VALID_PTR(pGVM)
1522	\|\| pGVM->pVM != pVM))
1523	{
1524	GVMMR0_USED_SHARED_UNLOCK(pGVMM);
1525	return VERR_INVALID_HANDLE;
1526	}
1527	}
1528	else
1529	{
1530	if (RT_UNLIKELY(pHandle->pVM != pVM))
1531	return VERR_INVALID_HANDLE;
1532	if (RT_UNLIKELY(pHandle->ProcId != ProcId))
1533	return VERR_INVALID_HANDLE;
1534	if (RT_UNLIKELY(!VALID_PTR(pHandle->pvObj)))
1535	return VERR_INVALID_HANDLE;
1536
1537	pGVM = pHandle->pGVM;
1538	if (RT_UNLIKELY(!VALID_PTR(pGVM)))
1539	return VERR_INVALID_HANDLE;
1540	if (RT_UNLIKELY(pGVM->pVM != pVM))
1541	return VERR_INVALID_HANDLE;
1542	}
1543
1544	*ppGVM = pGVM;
1545	*ppGVMM = pGVMM;
1546	return VINF_SUCCESS;
1547	}
1548
1549
1550	/**
1551	* Lookup a GVM structure by the shared VM structure.
1552	*
1553	* @returns VBox status code.
1554	* @param pVM The cross context VM structure.
1555	* @param ppGVM Where to store the GVM pointer.
1556	*
1557	* @remark This will not take the 'used'-lock because it doesn't do
1558	* nesting and this function will be used from under the lock.
1559	* Update: This is no longer true. Consider taking the lock in shared
1560	* mode!
1561	*/
1562	GVMMR0DECL(int) GVMMR0ByVM(PVM pVM, PGVM *ppGVM)
1563	{
1564	PGVMM pGVMM;
1565	return gvmmR0ByVM(pVM, ppGVM, &pGVMM, false /* fTakeUsedLock */);
1566	}
1567
1568
1569	/**
1570	* Lookup a GVM structure by the shared VM structure and ensuring that the
1571	* caller is an EMT thread.
1572	*
1573	* @returns VBox status code.
1574	* @param pVM The cross context VM structure.
1575	* @param idCpu The Virtual CPU ID of the calling EMT.
1576	* @param ppGVM Where to store the GVM pointer.
1577	* @param ppGVMM Where to store the pointer to the GVMM instance data.
1578	* @thread EMT
1579	*
1580	* @remark This will assert in all failure paths.
1581	*/
1582	static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM ppGVM, PGVMM ppGVMM)
1583	{
1584	PGVMM pGVMM;
1585	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
1586
1587	/*
1588	* Validate.
1589	*/
1590	AssertPtrReturn(pVM, VERR_INVALID_POINTER);
1591	AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
1592
1593	uint16_t hGVM = pVM->hSelf;
1594	AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
1595	AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
1596
1597	/*
1598	* Look it up.
1599	*/
1600	PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1601	AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
1602	RTPROCESS ProcId = RTProcSelf();
1603	AssertReturn(pHandle->ProcId == ProcId, VERR_NOT_OWNER);
1604	AssertPtrReturn(pHandle->pvObj, VERR_NOT_OWNER);
1605
1606	PGVM pGVM = pHandle->pGVM;
1607	AssertPtrReturn(pGVM, VERR_NOT_OWNER);
1608	AssertReturn(pGVM->pVM == pVM, VERR_NOT_OWNER);
1609	RTNATIVETHREAD hAllegedEMT = RTThreadNativeSelf();
1610	AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
1611	AssertReturn(pGVM->aCpus[idCpu].hEMT == hAllegedEMT, VERR_NOT_OWNER);
1612
1613	*ppGVM = pGVM;
1614	*ppGVMM = pGVMM;
1615	return VINF_SUCCESS;
1616	}
1617
1618
1619	/**
1620	* Lookup a GVM structure by the shared VM structure
1621	* and ensuring that the caller is the EMT thread.
1622	*
1623	* @returns VBox status code.
1624	* @param pVM The cross context VM structure.
1625	* @param idCpu The Virtual CPU ID of the calling EMT.
1626	* @param ppGVM Where to store the GVM pointer.
1627	* @thread EMT
1628	*/
1629	GVMMR0DECL(int) GVMMR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM)
1630	{
1631	AssertPtrReturn(ppGVM, VERR_INVALID_POINTER);
1632	PGVMM pGVMM;
1633	return gvmmR0ByVMAndEMT(pVM, idCpu, ppGVM, &pGVMM);
1634	}
1635
1636
1637	/**
1638	* Lookup a VM by its global handle.
1639	*
1640	* @returns Pointer to the VM on success, NULL on failure.
1641	* @param hGVM The global VM handle. Asserts on bad handle.
1642	*/
1643	GVMMR0DECL(PVM) GVMMR0GetVMByHandle(uint32_t hGVM)
1644	{
1645	PGVM pGVM = GVMMR0ByHandle(hGVM);
1646	return pGVM ? pGVM->pVM : NULL;
1647	}
1648
1649
1650	/**
1651	* Looks up the VM belonging to the specified EMT thread.
1652	*
1653	* This is used by the assertion machinery in VMMR0.cpp to avoid causing
1654	* unnecessary kernel panics when the EMT thread hits an assertion. The
1655	* call may or not be an EMT thread.
1656	*
1657	* @returns Pointer to the VM on success, NULL on failure.
1658	* @param hEMT The native thread handle of the EMT.
1659	* NIL_RTNATIVETHREAD means the current thread
1660	*/
1661	GVMMR0DECL(PVM) GVMMR0GetVMByEMT(RTNATIVETHREAD hEMT)
1662	{
1663	/*
1664	* No Assertions here as we're usually called in a AssertMsgN or
1665	* RTAssert* context.
1666	*/
1667	PGVMM pGVMM = g_pGVMM;
1668	if ( !VALID_PTR(pGVMM)
1669	\|\| pGVMM->u32Magic != GVMM_MAGIC)
1670	return NULL;
1671
1672	if (hEMT == NIL_RTNATIVETHREAD)
1673	hEMT = RTThreadNativeSelf();
1674	RTPROCESS ProcId = RTProcSelf();
1675
1676	/*
1677	* Search the handles in a linear fashion as we don't dare to take the lock (assert).
1678	*/
1679	for (unsigned i = 1; i < RT_ELEMENTS(pGVMM->aHandles); i++)
1680	{
1681	if ( pGVMM->aHandles[i].iSelf == i
1682	&& pGVMM->aHandles[i].ProcId == ProcId
1683	&& VALID_PTR(pGVMM->aHandles[i].pvObj)
1684	&& VALID_PTR(pGVMM->aHandles[i].pVM)
1685	&& VALID_PTR(pGVMM->aHandles[i].pGVM))
1686	{
1687	if (pGVMM->aHandles[i].hEMT0 == hEMT)
1688	return pGVMM->aHandles[i].pVM;
1689
1690	/* This is fearly safe with the current process per VM approach. */
1691	PGVM pGVM = pGVMM->aHandles[i].pGVM;
1692	VMCPUID const cCpus = pGVM->cCpus;
1693	if ( cCpus < 1
1694	\|\| cCpus > VMM_MAX_CPU_COUNT)
1695	continue;
1696	for (VMCPUID idCpu = 1; idCpu < cCpus; idCpu++)
1697	if (pGVM->aCpus[idCpu].hEMT == hEMT)
1698	return pGVMM->aHandles[i].pVM;
1699	}
1700	}
1701	return NULL;
1702	}
1703
1704
1705	/**
1706	* This is will wake up expired and soon-to-be expired VMs.
1707	*
1708	* @returns Number of VMs that has been woken up.
1709	* @param pGVMM Pointer to the GVMM instance data.
1710	* @param u64Now The current time.
1711	*/
1712	static unsigned gvmmR0SchedDoWakeUps(PGVMM pGVMM, uint64_t u64Now)
1713	{
1714	/*
1715	* Skip this if we've got disabled because of high resolution wakeups or by
1716	* the user.
1717	*/
1718	if (!pGVMM->fDoEarlyWakeUps)
1719	return 0;
1720
1721	/** @todo Rewrite this algorithm. See performance defect XYZ. */
1722
1723	/*
1724	* A cheap optimization to stop wasting so much time here on big setups.
1725	*/
1726	const uint64_t uNsEarlyWakeUp2 = u64Now + pGVMM->nsEarlyWakeUp2;
1727	if ( pGVMM->cHaltedEMTs == 0
1728	\|\| uNsEarlyWakeUp2 > pGVMM->uNsNextEmtWakeup)
1729	return 0;
1730
1731	/*
1732	* Only one thread doing this at a time.
1733	*/
1734	if (!ASMAtomicCmpXchgBool(&pGVMM->fDoingEarlyWakeUps, true, false))
1735	return 0;
1736
1737	/*
1738	* The first pass will wake up VMs which have actually expired
1739	* and look for VMs that should be woken up in the 2nd and 3rd passes.
1740	*/
1741	const uint64_t uNsEarlyWakeUp1 = u64Now + pGVMM->nsEarlyWakeUp1;
1742	uint64_t u64Min = UINT64_MAX;
1743	unsigned cWoken = 0;
1744	unsigned cHalted = 0;
1745	unsigned cTodo2nd = 0;
1746	unsigned cTodo3rd = 0;
1747	for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1748	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1749	i = pGVMM->aHandles[i].iNext)
1750	{
1751	PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1752	if ( VALID_PTR(pCurGVM)
1753	&& pCurGVM->u32Magic == GVM_MAGIC)
1754	{
1755	for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1756	{
1757	PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1758	uint64_t u64 = ASMAtomicUoReadU64(&pCurGVCpu->gvmm.s.u64HaltExpire);
1759	if (u64)
1760	{
1761	if (u64 <= u64Now)
1762	{
1763	if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1764	{
1765	int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1766	AssertRC(rc);
1767	cWoken++;
1768	}
1769	}
1770	else
1771	{
1772	cHalted++;
1773	if (u64 <= uNsEarlyWakeUp1)
1774	cTodo2nd++;
1775	else if (u64 <= uNsEarlyWakeUp2)
1776	cTodo3rd++;
1777	else if (u64 < u64Min)
1778	u64 = u64Min;
1779	}
1780	}
1781	}
1782	}
1783	AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1784	}
1785
1786	if (cTodo2nd)
1787	{
1788	for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1789	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1790	i = pGVMM->aHandles[i].iNext)
1791	{
1792	PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1793	if ( VALID_PTR(pCurGVM)
1794	&& pCurGVM->u32Magic == GVM_MAGIC)
1795	{
1796	for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1797	{
1798	PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1799	uint64_t u64 = ASMAtomicUoReadU64(&pCurGVCpu->gvmm.s.u64HaltExpire);
1800	if ( u64
1801	&& u64 <= uNsEarlyWakeUp1)
1802	{
1803	if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1804	{
1805	int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1806	AssertRC(rc);
1807	cWoken++;
1808	}
1809	}
1810	}
1811	}
1812	AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1813	}
1814	}
1815
1816	if (cTodo3rd)
1817	{
1818	for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1819	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1820	i = pGVMM->aHandles[i].iNext)
1821	{
1822	PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1823	if ( VALID_PTR(pCurGVM)
1824	&& pCurGVM->u32Magic == GVM_MAGIC)
1825	{
1826	for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1827	{
1828	PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1829	uint64_t u64 = ASMAtomicUoReadU64(&pCurGVCpu->gvmm.s.u64HaltExpire);
1830	if ( u64
1831	&& u64 <= uNsEarlyWakeUp2)
1832	{
1833	if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1834	{
1835	int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1836	AssertRC(rc);
1837	cWoken++;
1838	}
1839	}
1840	}
1841	}
1842	AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1843	}
1844	}
1845
1846	/*
1847	* Set the minimum value.
1848	*/
1849	pGVMM->uNsNextEmtWakeup = u64Min;
1850
1851	ASMAtomicWriteBool(&pGVMM->fDoingEarlyWakeUps, false);
1852	return cWoken;
1853	}
1854
1855
1856	/**
1857	* Halt the EMT thread.
1858	*
1859	* @returns VINF_SUCCESS normal wakeup (timeout or kicked by other thread).
1860	* VERR_INTERRUPTED if a signal was scheduled for the thread.
1861	* @param pVM The cross context VM structure.
1862	* @param idCpu The Virtual CPU ID of the calling EMT.
1863	* @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
1864	* @thread EMT(idCpu).
1865	*/
1866	GVMMR0DECL(int) GVMMR0SchedHalt(PVM pVM, VMCPUID idCpu, uint64_t u64ExpireGipTime)
1867	{
1868	LogFlow(("GVMMR0SchedHalt: pVM=%p\n", pVM));
1869	GVMM_CHECK_SMAP_SETUP();
1870	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1871
1872	/*
1873	* Validate the VM structure, state and handle.
1874	*/
1875	PGVM pGVM;
1876	PGVMM pGVMM;
1877	int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
1878	if (RT_FAILURE(rc))
1879	return rc;
1880	pGVM->gvmm.s.StatsSched.cHaltCalls++;
1881	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1882
1883	PGVMCPU pCurGVCpu = &pGVM->aCpus[idCpu];
1884	Assert(!pCurGVCpu->gvmm.s.u64HaltExpire);
1885
1886	/*
1887	* If we're doing early wake-ups, we must take the UsedList lock before we
1888	* start querying the current time.
1889	* Note! Interrupts must NOT be disabled at this point because we ask for GIP time!
1890	*/
1891	bool const fDoEarlyWakeUps = pGVMM->fDoEarlyWakeUps;
1892	if (fDoEarlyWakeUps)
1893	{
1894	rc = GVMMR0_USED_SHARED_LOCK(pGVMM); AssertRC(rc);
1895	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1896	}
1897
1898	pCurGVCpu->gvmm.s.iCpuEmt = ASMGetApicId();
1899
1900	/* GIP hack: We might are frequently sleeping for short intervals where the
1901	difference between GIP and system time matters on systems with high resolution
1902	system time. So, convert the input from GIP to System time in that case. */
1903	Assert(ASMGetFlags() & X86_EFL_IF);
1904	const uint64_t u64NowSys = RTTimeSystemNanoTS();
1905	const uint64_t u64NowGip = RTTimeNanoTS();
1906	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1907
1908	if (fDoEarlyWakeUps)
1909	{
1910	pGVM->gvmm.s.StatsSched.cHaltWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64NowGip);
1911	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1912	}
1913
1914	/*
1915	* Go to sleep if we must...
1916	* Cap the sleep time to 1 second to be on the safe side.
1917	*/
1918	uint64_t cNsInterval = u64ExpireGipTime - u64NowGip;
1919	if ( u64NowGip < u64ExpireGipTime
1920	&& cNsInterval >= (pGVMM->cEMTs > pGVMM->cEMTsMeansCompany
1921	? pGVMM->nsMinSleepCompany
1922	: pGVMM->nsMinSleepAlone))
1923	{
1924	pGVM->gvmm.s.StatsSched.cHaltBlocking++;
1925	if (cNsInterval > RT_NS_1SEC)
1926	u64ExpireGipTime = u64NowGip + RT_NS_1SEC;
1927	ASMAtomicWriteU64(&pCurGVCpu->gvmm.s.u64HaltExpire, u64ExpireGipTime);
1928	ASMAtomicIncU32(&pGVMM->cHaltedEMTs);
1929	if (fDoEarlyWakeUps)
1930	{
1931	if (u64ExpireGipTime < pGVMM->uNsNextEmtWakeup)
1932	pGVMM->uNsNextEmtWakeup = u64ExpireGipTime;
1933	GVMMR0_USED_SHARED_UNLOCK(pGVMM);
1934	}
1935	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1936
1937	rc = RTSemEventMultiWaitEx(pCurGVCpu->gvmm.s.HaltEventMulti,
1938	RTSEMWAIT_FLAGS_ABSOLUTE \| RTSEMWAIT_FLAGS_NANOSECS \| RTSEMWAIT_FLAGS_INTERRUPTIBLE,
1939	u64NowGip > u64NowSys ? u64ExpireGipTime : u64NowSys + cNsInterval);
1940	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1941
1942	ASMAtomicWriteU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0);
1943	ASMAtomicDecU32(&pGVMM->cHaltedEMTs);
1944
1945	/* Reset the semaphore to try prevent a few false wake-ups. */
1946	if (rc == VINF_SUCCESS)
1947	{
1948	RTSemEventMultiReset(pCurGVCpu->gvmm.s.HaltEventMulti);
1949	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1950	}
1951	else if (rc == VERR_TIMEOUT)
1952	{
1953	pGVM->gvmm.s.StatsSched.cHaltTimeouts++;
1954	rc = VINF_SUCCESS;
1955	}
1956	}
1957	else
1958	{
1959	pGVM->gvmm.s.StatsSched.cHaltNotBlocking++;
1960	if (fDoEarlyWakeUps)
1961	GVMMR0_USED_SHARED_UNLOCK(pGVMM);
1962	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1963	RTSemEventMultiReset(pCurGVCpu->gvmm.s.HaltEventMulti);
1964	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
1965	}
1966
1967	return rc;
1968	}
1969
1970
1971	/**
1972	* Worker for GVMMR0SchedWakeUp and GVMMR0SchedWakeUpAndPokeCpus that wakes up
1973	* the a sleeping EMT.
1974	*
1975	* @retval VINF_SUCCESS if successfully woken up.
1976	* @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1977	*
1978	* @param pGVM The global (ring-0) VM structure.
1979	* @param pGVCpu The global (ring-0) VCPU structure.
1980	*/
1981	DECLINLINE(int) gvmmR0SchedWakeUpOne(PGVM pGVM, PGVMCPU pGVCpu)
1982	{
1983	pGVM->gvmm.s.StatsSched.cWakeUpCalls++;
1984
1985	/*
1986	* Signal the semaphore regardless of whether it's current blocked on it.
1987	*
1988	* The reason for this is that there is absolutely no way we can be 100%
1989	* certain that it isn't about go to go to sleep on it and just got
1990	* delayed a bit en route. So, we will always signal the semaphore when
1991	* the it is flagged as halted in the VMM.
1992	*/
1993	/** @todo we can optimize some of that by means of the pVCpu->enmState now. */
1994	int rc;
1995	if (pGVCpu->gvmm.s.u64HaltExpire)
1996	{
1997	rc = VINF_SUCCESS;
1998	ASMAtomicWriteU64(&pGVCpu->gvmm.s.u64HaltExpire, 0);
1999	}
2000	else
2001	{
2002	rc = VINF_GVM_NOT_BLOCKED;
2003	pGVM->gvmm.s.StatsSched.cWakeUpNotHalted++;
2004	}
2005
2006	int rc2 = RTSemEventMultiSignal(pGVCpu->gvmm.s.HaltEventMulti);
2007	AssertRC(rc2);
2008
2009	return rc;
2010	}
2011
2012
2013	/**
2014	* Wakes up the halted EMT thread so it can service a pending request.
2015	*
2016	* @returns VBox status code.
2017	* @retval VINF_SUCCESS if successfully woken up.
2018	* @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
2019	*
2020	* @param pVM The cross context VM structure.
2021	* @param idCpu The Virtual CPU ID of the EMT to wake up.
2022	* @param fTakeUsedLock Take the used lock or not
2023	* @thread Any but EMT.
2024	*/
2025	GVMMR0DECL(int) GVMMR0SchedWakeUpEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
2026	{
2027	GVMM_CHECK_SMAP_SETUP();
2028	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2029
2030	/*
2031	* Validate input and take the UsedLock.
2032	*/
2033	PGVM pGVM;
2034	PGVMM pGVMM;
2035	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
2036	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2037	if (RT_SUCCESS(rc))
2038	{
2039	if (idCpu < pGVM->cCpus)
2040	{
2041	/*
2042	* Do the actual job.
2043	*/
2044	rc = gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
2045	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2046
2047	if (fTakeUsedLock && pGVMM->fDoEarlyWakeUps)
2048	{
2049	/*
2050	* While we're here, do a round of scheduling.
2051	*/
2052	Assert(ASMGetFlags() & X86_EFL_IF);
2053	const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
2054	pGVM->gvmm.s.StatsSched.cWakeUpWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
2055	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2056	}
2057	}
2058	else
2059	rc = VERR_INVALID_CPU_ID;
2060
2061	if (fTakeUsedLock)
2062	{
2063	int rc2 = GVMMR0_USED_SHARED_UNLOCK(pGVMM);
2064	AssertRC(rc2);
2065	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2066	}
2067	}
2068
2069	LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
2070	return rc;
2071	}
2072
2073
2074	/**
2075	* Wakes up the halted EMT thread so it can service a pending request.
2076	*
2077	* @returns VBox status code.
2078	* @retval VINF_SUCCESS if successfully woken up.
2079	* @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
2080	*
2081	* @param pVM The cross context VM structure.
2082	* @param idCpu The Virtual CPU ID of the EMT to wake up.
2083	* @thread Any but EMT.
2084	*/
2085	GVMMR0DECL(int) GVMMR0SchedWakeUp(PVM pVM, VMCPUID idCpu)
2086	{
2087	return GVMMR0SchedWakeUpEx(pVM, idCpu, true /* fTakeUsedLock */);
2088	}
2089
2090	/**
2091	* Worker common to GVMMR0SchedPoke and GVMMR0SchedWakeUpAndPokeCpus that pokes
2092	* the Virtual CPU if it's still busy executing guest code.
2093	*
2094	* @returns VBox status code.
2095	* @retval VINF_SUCCESS if poked successfully.
2096	* @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
2097	*
2098	* @param pGVM The global (ring-0) VM structure.
2099	* @param pVCpu The cross context virtual CPU structure.
2100	*/
2101	DECLINLINE(int) gvmmR0SchedPokeOne(PGVM pGVM, PVMCPU pVCpu)
2102	{
2103	pGVM->gvmm.s.StatsSched.cPokeCalls++;
2104
2105	RTCPUID idHostCpu = pVCpu->idHostCpu;
2106	if ( idHostCpu == NIL_RTCPUID
2107	\|\| VMCPU_GET_STATE(pVCpu) != VMCPUSTATE_STARTED_EXEC)
2108	{
2109	pGVM->gvmm.s.StatsSched.cPokeNotBusy++;
2110	return VINF_GVM_NOT_BUSY_IN_GC;
2111	}
2112
2113	/* Note: this function is not implemented on Darwin and Linux (kernel < 2.6.19) */
2114	RTMpPokeCpu(idHostCpu);
2115	return VINF_SUCCESS;
2116	}
2117
2118	/**
2119	* Pokes an EMT if it's still busy running guest code.
2120	*
2121	* @returns VBox status code.
2122	* @retval VINF_SUCCESS if poked successfully.
2123	* @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
2124	*
2125	* @param pVM The cross context VM structure.
2126	* @param idCpu The ID of the virtual CPU to poke.
2127	* @param fTakeUsedLock Take the used lock or not
2128	*/
2129	GVMMR0DECL(int) GVMMR0SchedPokeEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
2130	{
2131	/*
2132	* Validate input and take the UsedLock.
2133	*/
2134	PGVM pGVM;
2135	PGVMM pGVMM;
2136	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
2137	if (RT_SUCCESS(rc))
2138	{
2139	if (idCpu < pGVM->cCpus)
2140	rc = gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
2141	else
2142	rc = VERR_INVALID_CPU_ID;
2143
2144	if (fTakeUsedLock)
2145	{
2146	int rc2 = GVMMR0_USED_SHARED_UNLOCK(pGVMM);
2147	AssertRC(rc2);
2148	}
2149	}
2150
2151	LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
2152	return rc;
2153	}
2154
2155
2156	/**
2157	* Pokes an EMT if it's still busy running guest code.
2158	*
2159	* @returns VBox status code.
2160	* @retval VINF_SUCCESS if poked successfully.
2161	* @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
2162	*
2163	* @param pVM The cross context VM structure.
2164	* @param idCpu The ID of the virtual CPU to poke.
2165	*/
2166	GVMMR0DECL(int) GVMMR0SchedPoke(PVM pVM, VMCPUID idCpu)
2167	{
2168	return GVMMR0SchedPokeEx(pVM, idCpu, true /* fTakeUsedLock */);
2169	}
2170
2171
2172	/**
2173	* Wakes up a set of halted EMT threads so they can service pending request.
2174	*
2175	* @returns VBox status code, no informational stuff.
2176	*
2177	* @param pVM The cross context VM structure.
2178	* @param pSleepSet The set of sleepers to wake up.
2179	* @param pPokeSet The set of CPUs to poke.
2180	*/
2181	GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpus(PVM pVM, PCVMCPUSET pSleepSet, PCVMCPUSET pPokeSet)
2182	{
2183	AssertPtrReturn(pSleepSet, VERR_INVALID_POINTER);
2184	AssertPtrReturn(pPokeSet, VERR_INVALID_POINTER);
2185	GVMM_CHECK_SMAP_SETUP();
2186	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2187	RTNATIVETHREAD hSelf = RTThreadNativeSelf();
2188
2189	/*
2190	* Validate input and take the UsedLock.
2191	*/
2192	PGVM pGVM;
2193	PGVMM pGVMM;
2194	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /* fTakeUsedLock */);
2195	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2196	if (RT_SUCCESS(rc))
2197	{
2198	rc = VINF_SUCCESS;
2199	VMCPUID idCpu = pGVM->cCpus;
2200	while (idCpu-- > 0)
2201	{
2202	/* Don't try poke or wake up ourselves. */
2203	if (pGVM->aCpus[idCpu].hEMT == hSelf)
2204	continue;
2205
2206	/* just ignore errors for now. */
2207	if (VMCPUSET_IS_PRESENT(pSleepSet, idCpu))
2208	{
2209	gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
2210	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2211	}
2212	else if (VMCPUSET_IS_PRESENT(pPokeSet, idCpu))
2213	{
2214	gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
2215	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2216	}
2217	}
2218
2219	int rc2 = GVMMR0_USED_SHARED_UNLOCK(pGVMM);
2220	AssertRC(rc2);
2221	GVMM_CHECK_SMAP_CHECK2(pVM, RT_NOTHING);
2222	}
2223
2224	LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
2225	return rc;
2226	}
2227
2228
2229	/**
2230	* VMMR0 request wrapper for GVMMR0SchedWakeUpAndPokeCpus.
2231	*
2232	* @returns see GVMMR0SchedWakeUpAndPokeCpus.
2233	* @param pVM The cross context VM structure.
2234	* @param pReq Pointer to the request packet.
2235	*/
2236	GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpusReq(PVM pVM, PGVMMSCHEDWAKEUPANDPOKECPUSREQ pReq)
2237	{
2238	/*
2239	* Validate input and pass it on.
2240	*/
2241	AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2242	AssertMsgReturn(pReq->Hdr.cbReq == sizeof(pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(pReq)), VERR_INVALID_PARAMETER);
2243
2244	return GVMMR0SchedWakeUpAndPokeCpus(pVM, &pReq->SleepSet, &pReq->PokeSet);
2245	}
2246
2247
2248
2249	/**
2250	* Poll the schedule to see if someone else should get a chance to run.
2251	*
2252	* This is a bit hackish and will not work too well if the machine is
2253	* under heavy load from non-VM processes.
2254	*
2255	* @returns VINF_SUCCESS if not yielded.
2256	* VINF_GVM_YIELDED if an attempt to switch to a different VM task was made.
2257	* @param pVM The cross context VM structure.
2258	* @param idCpu The Virtual CPU ID of the calling EMT.
2259	* @param fYield Whether to yield or not.
2260	* This is for when we're spinning in the halt loop.
2261	* @thread EMT(idCpu).
2262	*/
2263	GVMMR0DECL(int) GVMMR0SchedPoll(PVM pVM, VMCPUID idCpu, bool fYield)
2264	{
2265	/*
2266	* Validate input.
2267	*/
2268	PGVM pGVM;
2269	PGVMM pGVMM;
2270	int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
2271	if (RT_SUCCESS(rc))
2272	{
2273	/*
2274	* We currently only implement helping doing wakeups (fYield = false), so don't
2275	* bother taking the lock if gvmmR0SchedDoWakeUps is not going to do anything.
2276	*/
2277	if (!fYield && pGVMM->fDoEarlyWakeUps)
2278	{
2279	rc = GVMMR0_USED_SHARED_LOCK(pGVMM); AssertRC(rc);
2280	pGVM->gvmm.s.StatsSched.cPollCalls++;
2281
2282	Assert(ASMGetFlags() & X86_EFL_IF);
2283	const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
2284
2285	pGVM->gvmm.s.StatsSched.cPollWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
2286
2287	GVMMR0_USED_SHARED_UNLOCK(pGVMM);
2288	}
2289	/*
2290	* Not quite sure what we could do here...
2291	*/
2292	else if (fYield)
2293	rc = VERR_NOT_IMPLEMENTED; /** @todo implement this... */
2294	else
2295	rc = VINF_SUCCESS;
2296	}
2297
2298	LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
2299	return rc;
2300	}
2301
2302
2303	#ifdef GVMM_SCHED_WITH_PPT
2304	/**
2305	* Timer callback for the periodic preemption timer.
2306	*
2307	* @param pTimer The timer handle.
2308	* @param pvUser Pointer to the per cpu structure.
2309	* @param iTick The current tick.
2310	*/
2311	static DECLCALLBACK(void) gvmmR0SchedPeriodicPreemptionTimerCallback(PRTTIMER pTimer, void *pvUser, uint64_t iTick)
2312	{
2313	PGVMMHOSTCPU pCpu = (PGVMMHOSTCPU)pvUser;
2314	NOREF(pTimer); NOREF(iTick);
2315
2316	/*
2317	* Termination check
2318	*/
2319	if (pCpu->u32Magic != GVMMHOSTCPU_MAGIC)
2320	return;
2321
2322	/*
2323	* Do the house keeping.
2324	*/
2325	RTSpinlockAcquire(pCpu->Ppt.hSpinlock);
2326
2327	if (++pCpu->Ppt.iTickHistorization >= pCpu->Ppt.cTicksHistoriziationInterval)
2328	{
2329	/*
2330	* Historicize the max frequency.
2331	*/
2332	uint32_t iHzHistory = ++pCpu->Ppt.iHzHistory % RT_ELEMENTS(pCpu->Ppt.aHzHistory);
2333	pCpu->Ppt.aHzHistory[iHzHistory] = pCpu->Ppt.uDesiredHz;
2334	pCpu->Ppt.iTickHistorization = 0;
2335	pCpu->Ppt.uDesiredHz = 0;
2336
2337	/*
2338	* Check if the current timer frequency.
2339	*/
2340	uint32_t uHistMaxHz = 0;
2341	for (uint32_t i = 0; i < RT_ELEMENTS(pCpu->Ppt.aHzHistory); i++)
2342	if (pCpu->Ppt.aHzHistory[i] > uHistMaxHz)
2343	uHistMaxHz = pCpu->Ppt.aHzHistory[i];
2344	if (uHistMaxHz == pCpu->Ppt.uTimerHz)
2345	RTSpinlockRelease(pCpu->Ppt.hSpinlock);
2346	else if (uHistMaxHz)
2347	{
2348	/*
2349	* Reprogram it.
2350	*/
2351	pCpu->Ppt.cChanges++;
2352	pCpu->Ppt.iTickHistorization = 0;
2353	pCpu->Ppt.uTimerHz = uHistMaxHz;
2354	uint32_t const cNsInterval = RT_NS_1SEC / uHistMaxHz;
2355	pCpu->Ppt.cNsInterval = cNsInterval;
2356	if (cNsInterval < GVMMHOSTCPU_PPT_HIST_INTERVAL_NS)
2357	pCpu->Ppt.cTicksHistoriziationInterval = ( GVMMHOSTCPU_PPT_HIST_INTERVAL_NS
2358	+ GVMMHOSTCPU_PPT_HIST_INTERVAL_NS / 2 - 1)
2359	/ cNsInterval;
2360	else
2361	pCpu->Ppt.cTicksHistoriziationInterval = 1;
2362	RTSpinlockRelease(pCpu->Ppt.hSpinlock);
2363
2364	/SUPR0Printf("Cpu%u: change to %u Hz / %u ns\n", pCpu->idxCpuSet, uHistMaxHz, cNsInterval);/
2365	RTTimerChangeInterval(pTimer, cNsInterval);
2366	}
2367	else
2368	{
2369	/*
2370	* Stop it.
2371	*/
2372	pCpu->Ppt.fStarted = false;
2373	pCpu->Ppt.uTimerHz = 0;
2374	pCpu->Ppt.cNsInterval = 0;
2375	RTSpinlockRelease(pCpu->Ppt.hSpinlock);
2376
2377	/SUPR0Printf("Cpu%u: stopping (%u Hz)\n", pCpu->idxCpuSet, uHistMaxHz);/
2378	RTTimerStop(pTimer);
2379	}
2380	}
2381	else
2382	RTSpinlockRelease(pCpu->Ppt.hSpinlock);
2383	}
2384	#endif /* GVMM_SCHED_WITH_PPT */
2385
2386
2387	/**
2388	* Updates the periodic preemption timer for the calling CPU.
2389	*
2390	* The caller must have disabled preemption!
2391	* The caller must check that the host can do high resolution timers.
2392	*
2393	* @param pVM The cross context VM structure.
2394	* @param idHostCpu The current host CPU id.
2395	* @param uHz The desired frequency.
2396	*/
2397	GVMMR0DECL(void) GVMMR0SchedUpdatePeriodicPreemptionTimer(PVM pVM, RTCPUID idHostCpu, uint32_t uHz)
2398	{
2399	NOREF(pVM);
2400	#ifdef GVMM_SCHED_WITH_PPT
2401	Assert(!RTThreadPreemptIsEnabled(NIL_RTTHREAD));
2402	Assert(RTTimerCanDoHighResolution());
2403
2404	/*
2405	* Resolve the per CPU data.
2406	*/
2407	uint32_t iCpu = RTMpCpuIdToSetIndex(idHostCpu);
2408	PGVMM pGVMM = g_pGVMM;
2409	if ( !VALID_PTR(pGVMM)
2410	\|\| pGVMM->u32Magic != GVMM_MAGIC)
2411	return;
2412	AssertMsgReturnVoid(iCpu < pGVMM->cHostCpus, ("iCpu=%d cHostCpus=%d\n", iCpu, pGVMM->cHostCpus));
2413	PGVMMHOSTCPU pCpu = &pGVMM->aHostCpus[iCpu];
2414	AssertMsgReturnVoid( pCpu->u32Magic == GVMMHOSTCPU_MAGIC
2415	&& pCpu->idCpu == idHostCpu,
2416	("u32Magic=%#x idCpu=% idHostCpu=%d\n", pCpu->u32Magic, pCpu->idCpu, idHostCpu));
2417
2418	/*
2419	* Check whether we need to do anything about the timer.
2420	* We have to be a little bit careful since we might be race the timer
2421	* callback here.
2422	*/
2423	if (uHz > 16384)
2424	uHz = 16384; /** @todo add a query method for this! */
2425	if (RT_UNLIKELY( uHz > ASMAtomicReadU32(&pCpu->Ppt.uDesiredHz)
2426	&& uHz >= pCpu->Ppt.uMinHz
2427	&& !pCpu->Ppt.fStarting /* solaris paranoia */))
2428	{
2429	RTSpinlockAcquire(pCpu->Ppt.hSpinlock);
2430
2431	pCpu->Ppt.uDesiredHz = uHz;
2432	uint32_t cNsInterval = 0;
2433	if (!pCpu->Ppt.fStarted)
2434	{
2435	pCpu->Ppt.cStarts++;
2436	pCpu->Ppt.fStarted = true;
2437	pCpu->Ppt.fStarting = true;
2438	pCpu->Ppt.iTickHistorization = 0;
2439	pCpu->Ppt.uTimerHz = uHz;
2440	pCpu->Ppt.cNsInterval = cNsInterval = RT_NS_1SEC / uHz;
2441	if (cNsInterval < GVMMHOSTCPU_PPT_HIST_INTERVAL_NS)
2442	pCpu->Ppt.cTicksHistoriziationInterval = ( GVMMHOSTCPU_PPT_HIST_INTERVAL_NS
2443	+ GVMMHOSTCPU_PPT_HIST_INTERVAL_NS / 2 - 1)
2444	/ cNsInterval;
2445	else
2446	pCpu->Ppt.cTicksHistoriziationInterval = 1;
2447	}
2448
2449	RTSpinlockRelease(pCpu->Ppt.hSpinlock);
2450
2451	if (cNsInterval)
2452	{
2453	RTTimerChangeInterval(pCpu->Ppt.pTimer, cNsInterval);
2454	int rc = RTTimerStart(pCpu->Ppt.pTimer, cNsInterval);
2455	AssertRC(rc);
2456
2457	RTSpinlockAcquire(pCpu->Ppt.hSpinlock);
2458	if (RT_FAILURE(rc))
2459	pCpu->Ppt.fStarted = false;
2460	pCpu->Ppt.fStarting = false;
2461	RTSpinlockRelease(pCpu->Ppt.hSpinlock);
2462	}
2463	}
2464	#else /* !GVMM_SCHED_WITH_PPT */
2465	NOREF(idHostCpu); NOREF(uHz);
2466	#endif /* !GVMM_SCHED_WITH_PPT */
2467	}
2468
2469
2470	/**
2471	* Retrieves the GVMM statistics visible to the caller.
2472	*
2473	* @returns VBox status code.
2474	*
2475	* @param pStats Where to put the statistics.
2476	* @param pSession The current session.
2477	* @param pVM The VM to obtain statistics for. Optional.
2478	*/
2479	GVMMR0DECL(int) GVMMR0QueryStatistics(PGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
2480	{
2481	LogFlow(("GVMMR0QueryStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
2482
2483	/*
2484	* Validate input.
2485	*/
2486	AssertPtrReturn(pSession, VERR_INVALID_POINTER);
2487	AssertPtrReturn(pStats, VERR_INVALID_POINTER);
2488	pStats->cVMs = 0; /* (crash before taking the sem...) */
2489
2490	/*
2491	* Take the lock and get the VM statistics.
2492	*/
2493	PGVMM pGVMM;
2494	if (pVM)
2495	{
2496	PGVM pGVM;
2497	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /fTakeUsedLock/);
2498	if (RT_FAILURE(rc))
2499	return rc;
2500	pStats->SchedVM = pGVM->gvmm.s.StatsSched;
2501	}
2502	else
2503	{
2504	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
2505	memset(&pStats->SchedVM, 0, sizeof(pStats->SchedVM));
2506
2507	int rc = GVMMR0_USED_SHARED_LOCK(pGVMM);
2508	AssertRCReturn(rc, rc);
2509	}
2510
2511	/*
2512	* Enumerate the VMs and add the ones visible to the statistics.
2513	*/
2514	pStats->cVMs = 0;
2515	pStats->cEMTs = 0;
2516	memset(&pStats->SchedSum, 0, sizeof(pStats->SchedSum));
2517
2518	for (unsigned i = pGVMM->iUsedHead;
2519	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
2520	i = pGVMM->aHandles[i].iNext)
2521	{
2522	PGVM pGVM = pGVMM->aHandles[i].pGVM;
2523	void *pvObj = pGVMM->aHandles[i].pvObj;
2524	if ( VALID_PTR(pvObj)
2525	&& VALID_PTR(pGVM)
2526	&& pGVM->u32Magic == GVM_MAGIC
2527	&& RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
2528	{
2529	pStats->cVMs++;
2530	pStats->cEMTs += pGVM->cCpus;
2531
2532	pStats->SchedSum.cHaltCalls += pGVM->gvmm.s.StatsSched.cHaltCalls;
2533	pStats->SchedSum.cHaltBlocking += pGVM->gvmm.s.StatsSched.cHaltBlocking;
2534	pStats->SchedSum.cHaltTimeouts += pGVM->gvmm.s.StatsSched.cHaltTimeouts;
2535	pStats->SchedSum.cHaltNotBlocking += pGVM->gvmm.s.StatsSched.cHaltNotBlocking;
2536	pStats->SchedSum.cHaltWakeUps += pGVM->gvmm.s.StatsSched.cHaltWakeUps;
2537
2538	pStats->SchedSum.cWakeUpCalls += pGVM->gvmm.s.StatsSched.cWakeUpCalls;
2539	pStats->SchedSum.cWakeUpNotHalted += pGVM->gvmm.s.StatsSched.cWakeUpNotHalted;
2540	pStats->SchedSum.cWakeUpWakeUps += pGVM->gvmm.s.StatsSched.cWakeUpWakeUps;
2541
2542	pStats->SchedSum.cPokeCalls += pGVM->gvmm.s.StatsSched.cPokeCalls;
2543	pStats->SchedSum.cPokeNotBusy += pGVM->gvmm.s.StatsSched.cPokeNotBusy;
2544
2545	pStats->SchedSum.cPollCalls += pGVM->gvmm.s.StatsSched.cPollCalls;
2546	pStats->SchedSum.cPollHalts += pGVM->gvmm.s.StatsSched.cPollHalts;
2547	pStats->SchedSum.cPollWakeUps += pGVM->gvmm.s.StatsSched.cPollWakeUps;
2548	}
2549	}
2550
2551	/*
2552	* Copy out the per host CPU statistics.
2553	*/
2554	uint32_t iDstCpu = 0;
2555	uint32_t cSrcCpus = pGVMM->cHostCpus;
2556	for (uint32_t iSrcCpu = 0; iSrcCpu < cSrcCpus; iSrcCpu++)
2557	{
2558	if (pGVMM->aHostCpus[iSrcCpu].idCpu != NIL_RTCPUID)
2559	{
2560	pStats->aHostCpus[iDstCpu].idCpu = pGVMM->aHostCpus[iSrcCpu].idCpu;
2561	pStats->aHostCpus[iDstCpu].idxCpuSet = pGVMM->aHostCpus[iSrcCpu].idxCpuSet;
2562	#ifdef GVMM_SCHED_WITH_PPT
2563	pStats->aHostCpus[iDstCpu].uDesiredHz = pGVMM->aHostCpus[iSrcCpu].Ppt.uDesiredHz;
2564	pStats->aHostCpus[iDstCpu].uTimerHz = pGVMM->aHostCpus[iSrcCpu].Ppt.uTimerHz;
2565	pStats->aHostCpus[iDstCpu].cChanges = pGVMM->aHostCpus[iSrcCpu].Ppt.cChanges;
2566	pStats->aHostCpus[iDstCpu].cStarts = pGVMM->aHostCpus[iSrcCpu].Ppt.cStarts;
2567	#else
2568	pStats->aHostCpus[iDstCpu].uDesiredHz = 0;
2569	pStats->aHostCpus[iDstCpu].uTimerHz = 0;
2570	pStats->aHostCpus[iDstCpu].cChanges = 0;
2571	pStats->aHostCpus[iDstCpu].cStarts = 0;
2572	#endif
2573	iDstCpu++;
2574	if (iDstCpu >= RT_ELEMENTS(pStats->aHostCpus))
2575	break;
2576	}
2577	}
2578	pStats->cHostCpus = iDstCpu;
2579
2580	GVMMR0_USED_SHARED_UNLOCK(pGVMM);
2581
2582	return VINF_SUCCESS;
2583	}
2584
2585
2586	/**
2587	* VMMR0 request wrapper for GVMMR0QueryStatistics.
2588	*
2589	* @returns see GVMMR0QueryStatistics.
2590	* @param pVM The cross context VM structure. Optional.
2591	* @param pReq Pointer to the request packet.
2592	*/
2593	GVMMR0DECL(int) GVMMR0QueryStatisticsReq(PVM pVM, PGVMMQUERYSTATISTICSSREQ pReq)
2594	{
2595	/*
2596	* Validate input and pass it on.
2597	*/
2598	AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2599	AssertMsgReturn(pReq->Hdr.cbReq == sizeof(pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(pReq)), VERR_INVALID_PARAMETER);
2600
2601	return GVMMR0QueryStatistics(&pReq->Stats, pReq->pSession, pVM);
2602	}
2603
2604
2605	/**
2606	* Resets the specified GVMM statistics.
2607	*
2608	* @returns VBox status code.
2609	*
2610	* @param pStats Which statistics to reset, that is, non-zero fields indicates which to reset.
2611	* @param pSession The current session.
2612	* @param pVM The VM to reset statistics for. Optional.
2613	*/
2614	GVMMR0DECL(int) GVMMR0ResetStatistics(PCGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
2615	{
2616	LogFlow(("GVMMR0ResetStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
2617
2618	/*
2619	* Validate input.
2620	*/
2621	AssertPtrReturn(pSession, VERR_INVALID_POINTER);
2622	AssertPtrReturn(pStats, VERR_INVALID_POINTER);
2623
2624	/*
2625	* Take the lock and get the VM statistics.
2626	*/
2627	PGVMM pGVMM;
2628	if (pVM)
2629	{
2630	PGVM pGVM;
2631	int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /fTakeUsedLock/);
2632	if (RT_FAILURE(rc))
2633	return rc;
2634	# define MAYBE_RESET_FIELD(field) \
2635	do { if (pStats->SchedVM. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2636	MAYBE_RESET_FIELD(cHaltCalls);
2637	MAYBE_RESET_FIELD(cHaltBlocking);
2638	MAYBE_RESET_FIELD(cHaltTimeouts);
2639	MAYBE_RESET_FIELD(cHaltNotBlocking);
2640	MAYBE_RESET_FIELD(cHaltWakeUps);
2641	MAYBE_RESET_FIELD(cWakeUpCalls);
2642	MAYBE_RESET_FIELD(cWakeUpNotHalted);
2643	MAYBE_RESET_FIELD(cWakeUpWakeUps);
2644	MAYBE_RESET_FIELD(cPokeCalls);
2645	MAYBE_RESET_FIELD(cPokeNotBusy);
2646	MAYBE_RESET_FIELD(cPollCalls);
2647	MAYBE_RESET_FIELD(cPollHalts);
2648	MAYBE_RESET_FIELD(cPollWakeUps);
2649	# undef MAYBE_RESET_FIELD
2650	}
2651	else
2652	{
2653	GVMM_GET_VALID_INSTANCE(pGVMM, VERR_GVMM_INSTANCE);
2654
2655	int rc = GVMMR0_USED_SHARED_LOCK(pGVMM);
2656	AssertRCReturn(rc, rc);
2657	}
2658
2659	/*
2660	* Enumerate the VMs and add the ones visible to the statistics.
2661	*/
2662	if (!ASMMemIsZero(&pStats->SchedSum, sizeof(pStats->SchedSum)))
2663	{
2664	for (unsigned i = pGVMM->iUsedHead;
2665	i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
2666	i = pGVMM->aHandles[i].iNext)
2667	{
2668	PGVM pGVM = pGVMM->aHandles[i].pGVM;
2669	void *pvObj = pGVMM->aHandles[i].pvObj;
2670	if ( VALID_PTR(pvObj)
2671	&& VALID_PTR(pGVM)
2672	&& pGVM->u32Magic == GVM_MAGIC
2673	&& RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
2674	{
2675	# define MAYBE_RESET_FIELD(field) \
2676	do { if (pStats->SchedSum. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2677	MAYBE_RESET_FIELD(cHaltCalls);
2678	MAYBE_RESET_FIELD(cHaltBlocking);
2679	MAYBE_RESET_FIELD(cHaltTimeouts);
2680	MAYBE_RESET_FIELD(cHaltNotBlocking);
2681	MAYBE_RESET_FIELD(cHaltWakeUps);
2682	MAYBE_RESET_FIELD(cWakeUpCalls);
2683	MAYBE_RESET_FIELD(cWakeUpNotHalted);
2684	MAYBE_RESET_FIELD(cWakeUpWakeUps);
2685	MAYBE_RESET_FIELD(cPokeCalls);
2686	MAYBE_RESET_FIELD(cPokeNotBusy);
2687	MAYBE_RESET_FIELD(cPollCalls);
2688	MAYBE_RESET_FIELD(cPollHalts);
2689	MAYBE_RESET_FIELD(cPollWakeUps);
2690	# undef MAYBE_RESET_FIELD
2691	}
2692	}
2693	}
2694
2695	GVMMR0_USED_SHARED_UNLOCK(pGVMM);
2696
2697	return VINF_SUCCESS;
2698	}
2699
2700
2701	/**
2702	* VMMR0 request wrapper for GVMMR0ResetStatistics.
2703	*
2704	* @returns see GVMMR0ResetStatistics.
2705	* @param pVM The cross context VM structure. Optional.
2706	* @param pReq Pointer to the request packet.
2707	*/
2708	GVMMR0DECL(int) GVMMR0ResetStatisticsReq(PVM pVM, PGVMMRESETSTATISTICSSREQ pReq)
2709	{
2710	/*
2711	* Validate input and pass it on.
2712	*/
2713	AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2714	AssertMsgReturn(pReq->Hdr.cbReq == sizeof(pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(pReq)), VERR_INVALID_PARAMETER);
2715
2716	return GVMMR0ResetStatistics(&pReq->Stats, pReq->pSession, pVM);
2717	}
2718

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

source: vbox/trunk/src/VBox/VMM/VMMR0/GVMMR0.cpp@ 61632

Download in other formats: