mp-r0drv-nt.cpp@ 63558

Last change on this file since 63558 was 63063, checked in by vboxsync, 9 years ago
IPRT: warnings
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 28.5 KB

Line
1	/* $Id: mp-r0drv-nt.cpp 63063 2016-08-05 21:14:33Z vboxsync $ */
2	/** @file
3	* IPRT - Multiprocessor, Ring-0 Driver, NT.
4	*/
5
6	/*
7	* Copyright (C) 2008-2016 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*/
26
27
28	/*********************************************************************************************************************************
29	* Header Files *
30	*********************************************************************************************************************************/
31	#include "the-nt-kernel.h"
32
33	#include <iprt/mp.h>
34	#include <iprt/cpuset.h>
35	#include <iprt/err.h>
36	#include <iprt/asm.h>
37	#include <iprt/log.h>
38	#include <iprt/time.h>
39	#include "r0drv/mp-r0drv.h"
40	#include "internal-r0drv-nt.h"
41
42
43	/*********************************************************************************************************************************
44	* Structures and Typedefs *
45	*********************************************************************************************************************************/
46	typedef enum
47	{
48	RT_NT_CPUID_SPECIFIC,
49	RT_NT_CPUID_PAIR,
50	RT_NT_CPUID_OTHERS,
51	RT_NT_CPUID_ALL
52	} RT_NT_CPUID;
53
54
55	/**
56	* Used by the RTMpOnSpecific.
57	*/
58	typedef struct RTMPNTONSPECIFICARGS
59	{
60	/** Set if we're executing. */
61	bool volatile fExecuting;
62	/** Set when done executing. */
63	bool volatile fDone;
64	/** Number of references to this heap block. */
65	uint32_t volatile cRefs;
66	/** Event that the calling thread is waiting on. */
67	KEVENT DoneEvt;
68	/** The deferred procedure call object. */
69	KDPC Dpc;
70	/** The callback argument package. */
71	RTMPARGS CallbackArgs;
72	} RTMPNTONSPECIFICARGS;
73	/** Pointer to an argument/state structure for RTMpOnSpecific on NT. */
74	typedef RTMPNTONSPECIFICARGS *PRTMPNTONSPECIFICARGS;
75
76
77
78	/* test a couple of assumption. */
79	AssertCompile(MAXIMUM_PROCESSORS <= RTCPUSET_MAX_CPUS);
80	AssertCompile(NIL_RTCPUID >= MAXIMUM_PROCESSORS);
81
82	/** @todo
83	* We cannot do other than assume a 1:1 relationship between the
84	* affinity mask and the process despite the vagueness/warnings in
85	* the docs. If someone knows a better way to get this done, please
86	* let bird know.
87	*/
88
89
90	RTDECL(RTCPUID) RTMpCpuId(void)
91	{
92	/* WDK upgrade warning: PCR->Number changed from BYTE to WORD. */
93	return KeGetCurrentProcessorNumber();
94	}
95
96
97	RTDECL(int) RTMpCurSetIndex(void)
98	{
99	/* WDK upgrade warning: PCR->Number changed from BYTE to WORD. */
100	return KeGetCurrentProcessorNumber();
101	}
102
103
104	RTDECL(int) RTMpCurSetIndexAndId(PRTCPUID pidCpu)
105	{
106	return *pidCpu = KeGetCurrentProcessorNumber();
107	}
108
109
110	RTDECL(int) RTMpCpuIdToSetIndex(RTCPUID idCpu)
111	{
112	return idCpu < MAXIMUM_PROCESSORS ? (int)idCpu : -1;
113	}
114
115
116	RTDECL(RTCPUID) RTMpCpuIdFromSetIndex(int iCpu)
117	{
118	return (unsigned)iCpu < MAXIMUM_PROCESSORS ? iCpu : NIL_RTCPUID;
119	}
120
121
122	RTDECL(RTCPUID) RTMpGetMaxCpuId(void)
123	{
124	/** @todo use KeQueryMaximumProcessorCount on vista+ */
125	return MAXIMUM_PROCESSORS - 1;
126	}
127
128
129	RTDECL(bool) RTMpIsCpuOnline(RTCPUID idCpu)
130	{
131	if (idCpu >= MAXIMUM_PROCESSORS)
132	return false;
133
134	#if 0 /* this isn't safe at all IRQLs (great work guys) */
135	KAFFINITY Mask = KeQueryActiveProcessors();
136	return !!(Mask & RT_BIT_64(idCpu));
137	#else
138	return RTCpuSetIsMember(&g_rtMpNtCpuSet, idCpu);
139	#endif
140	}
141
142
143	RTDECL(bool) RTMpIsCpuPossible(RTCPUID idCpu)
144	{
145	/* Cannot easily distinguish between online and offline cpus. */
146	/** @todo online/present cpu stuff must be corrected for proper W2K8 support
147	* (KeQueryMaximumProcessorCount). */
148	return RTMpIsCpuOnline(idCpu);
149	}
150
151
152
153	RTDECL(PRTCPUSET) RTMpGetSet(PRTCPUSET pSet)
154	{
155	/** @todo online/present cpu stuff must be corrected for proper W2K8 support
156	* (KeQueryMaximumProcessorCount). */
157	return RTMpGetOnlineSet(pSet);
158	}
159
160
161	RTDECL(RTCPUID) RTMpGetCount(void)
162	{
163	/** @todo online/present cpu stuff must be corrected for proper W2K8 support
164	* (KeQueryMaximumProcessorCount). */
165	return RTMpGetOnlineCount();
166	}
167
168
169	RTDECL(PRTCPUSET) RTMpGetOnlineSet(PRTCPUSET pSet)
170	{
171	#if 0 /* this isn't safe at all IRQLs (great work guys) */
172	KAFFINITY Mask = KeQueryActiveProcessors();
173	return RTCpuSetFromU64(pSet, Mask);
174	#else
175	*pSet = g_rtMpNtCpuSet;
176	return pSet;
177	#endif
178	}
179
180
181	RTDECL(RTCPUID) RTMpGetOnlineCount(void)
182	{
183	RTCPUSET Set;
184	RTMpGetOnlineSet(&Set);
185	return RTCpuSetCount(&Set);
186	}
187
188
189	#if 0
190	/* Experiment with checking the undocumented KPRCB structure
191	* 'dt nt!_kprcb 0xaddress' shows the layout
192	*/
193	typedef struct
194	{
195	LIST_ENTRY DpcListHead;
196	ULONG_PTR DpcLock;
197	volatile ULONG DpcQueueDepth;
198	ULONG DpcQueueCount;
199	} KDPC_DATA, *PKDPC_DATA;
200
201	RTDECL(bool) RTMpIsCpuWorkPending(void)
202	{
203	uint8_t *pkprcb;
204	PKDPC_DATA pDpcData;
205
206	_asm {
207	mov eax, fs:0x20
208	mov pkprcb, eax
209	}
210	pDpcData = (PKDPC_DATA)(pkprcb + 0x19e0);
211	if (pDpcData->DpcQueueDepth)
212	return true;
213
214	pDpcData++;
215	if (pDpcData->DpcQueueDepth)
216	return true;
217	return false;
218	}
219	#else
220	RTDECL(bool) RTMpIsCpuWorkPending(void)
221	{
222	/** @todo not implemented */
223	return false;
224	}
225	#endif
226
227
228	/**
229	* Wrapper between the native KIPI_BROADCAST_WORKER and IPRT's PFNRTMPWORKER for
230	* the RTMpOnAll case.
231	*
232	* @param uUserCtx The user context argument (PRTMPARGS).
233	*/
234	static ULONG_PTR rtmpNtOnAllBroadcastIpiWrapper(ULONG_PTR uUserCtx)
235	{
236	PRTMPARGS pArgs = (PRTMPARGS)uUserCtx;
237	/ASMAtomicIncU32(&pArgs->cHits); - not needed /
238	pArgs->pfnWorker(KeGetCurrentProcessorNumber(), pArgs->pvUser1, pArgs->pvUser2);
239	return 0;
240	}
241
242
243	/**
244	* Wrapper between the native KIPI_BROADCAST_WORKER and IPRT's PFNRTMPWORKER for
245	* the RTMpOnOthers case.
246	*
247	* @param uUserCtx The user context argument (PRTMPARGS).
248	*/
249	static ULONG_PTR rtmpNtOnOthersBroadcastIpiWrapper(ULONG_PTR uUserCtx)
250	{
251	PRTMPARGS pArgs = (PRTMPARGS)uUserCtx;
252	RTCPUID idCpu = KeGetCurrentProcessorNumber();
253	if (pArgs->idCpu != idCpu)
254	{
255	/ASMAtomicIncU32(&pArgs->cHits); - not needed /
256	pArgs->pfnWorker(idCpu, pArgs->pvUser1, pArgs->pvUser2);
257	}
258	return 0;
259	}
260
261
262	/**
263	* Wrapper between the native KIPI_BROADCAST_WORKER and IPRT's PFNRTMPWORKER for
264	* the RTMpOnPair case.
265	*
266	* @param uUserCtx The user context argument (PRTMPARGS).
267	*/
268	static ULONG_PTR rtmpNtOnPairBroadcastIpiWrapper(ULONG_PTR uUserCtx)
269	{
270	PRTMPARGS pArgs = (PRTMPARGS)uUserCtx;
271	RTCPUID idCpu = KeGetCurrentProcessorNumber();
272	if ( pArgs->idCpu == idCpu
273	\|\| pArgs->idCpu2 == idCpu)
274	{
275	ASMAtomicIncU32(&pArgs->cHits);
276	pArgs->pfnWorker(idCpu, pArgs->pvUser1, pArgs->pvUser2);
277	}
278	return 0;
279	}
280
281
282	/**
283	* Wrapper between the native KIPI_BROADCAST_WORKER and IPRT's PFNRTMPWORKER for
284	* the RTMpOnSpecific case.
285	*
286	* @param uUserCtx The user context argument (PRTMPARGS).
287	*/
288	static ULONG_PTR rtmpNtOnSpecificBroadcastIpiWrapper(ULONG_PTR uUserCtx)
289	{
290	PRTMPARGS pArgs = (PRTMPARGS)uUserCtx;
291	RTCPUID idCpu = KeGetCurrentProcessorNumber();
292	if (pArgs->idCpu == idCpu)
293	{
294	ASMAtomicIncU32(&pArgs->cHits);
295	pArgs->pfnWorker(idCpu, pArgs->pvUser1, pArgs->pvUser2);
296	}
297	return 0;
298	}
299
300
301	/**
302	* Internal worker for the RTMpOn* APIs using KeIpiGenericCall.
303	*
304	* @returns VINF_SUCCESS.
305	* @param pfnWorker The callback.
306	* @param pvUser1 User argument 1.
307	* @param pvUser2 User argument 2.
308	* @param pfnNativeWrapper The wrapper between the NT and IPRT callbacks.
309	* @param idCpu First CPU to match, ultimately specific to the
310	* pfnNativeWrapper used.
311	* @param idCpu2 Second CPU to match, ultimately specific to the
312	* pfnNativeWrapper used.
313	* @param pcHits Where to return the number of this. Optional.
314	*/
315	static int rtMpCallUsingBroadcastIpi(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2,
316	PKIPI_BROADCAST_WORKER pfnNativeWrapper, RTCPUID idCpu, RTCPUID idCpu2,
317	uint32_t *pcHits)
318	{
319	RTMPARGS Args;
320	Args.pfnWorker = pfnWorker;
321	Args.pvUser1 = pvUser1;
322	Args.pvUser2 = pvUser2;
323	Args.idCpu = idCpu;
324	Args.idCpu2 = idCpu2;
325	Args.cRefs = 0;
326	Args.cHits = 0;
327
328	AssertPtr(g_pfnrtKeIpiGenericCall);
329	g_pfnrtKeIpiGenericCall(pfnNativeWrapper, (uintptr_t)&Args);
330	if (pcHits)
331	*pcHits = Args.cHits;
332	return VINF_SUCCESS;
333	}
334
335
336	/**
337	* Wrapper between the native nt per-cpu callbacks and PFNRTWORKER
338	*
339	* @param Dpc DPC object
340	* @param DeferredContext Context argument specified by KeInitializeDpc
341	* @param SystemArgument1 Argument specified by KeInsertQueueDpc
342	* @param SystemArgument2 Argument specified by KeInsertQueueDpc
343	*/
344	static VOID rtmpNtDPCWrapper(IN PKDPC Dpc, IN PVOID DeferredContext, IN PVOID SystemArgument1, IN PVOID SystemArgument2)
345	{
346	PRTMPARGS pArgs = (PRTMPARGS)DeferredContext;
347	RT_NOREF3(Dpc, SystemArgument1, SystemArgument2);
348
349	ASMAtomicIncU32(&pArgs->cHits);
350	pArgs->pfnWorker(KeGetCurrentProcessorNumber(), pArgs->pvUser1, pArgs->pvUser2);
351
352	/* Dereference the argument structure. */
353	int32_t cRefs = ASMAtomicDecS32(&pArgs->cRefs);
354	Assert(cRefs >= 0);
355	if (cRefs == 0)
356	ExFreePool(pArgs);
357	}
358
359
360	/**
361	* Internal worker for the RTMpOn* APIs.
362	*
363	* @returns IPRT status code.
364	* @param pfnWorker The callback.
365	* @param pvUser1 User argument 1.
366	* @param pvUser2 User argument 2.
367	* @param enmCpuid What to do / is idCpu valid.
368	* @param idCpu Used if enmCpuid is RT_NT_CPUID_SPECIFIC or
369	* RT_NT_CPUID_PAIR, otherwise ignored.
370	* @param idCpu2 Used if enmCpuid is RT_NT_CPUID_PAIR, otherwise ignored.
371	* @param pcHits Where to return the number of this. Optional.
372	*/
373	static int rtMpCallUsingDpcs(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2,
374	RT_NT_CPUID enmCpuid, RTCPUID idCpu, RTCPUID idCpu2, uint32_t *pcHits)
375	{
376	#ifdef IPRT_TARGET_NT4
377	RT_NOREF(pfnWorker, pvUser1, pvUser2, enmCpuid, idCpu, idCpu2, pcHits);
378	/* g_pfnrtNt* are not present on NT anyway. */
379	return VERR_NOT_SUPPORTED;
380
381	#else /* !IPRT_TARGET_NT4 */
382	PRTMPARGS pArgs;
383	KDPC *paExecCpuDpcs;
384
385	# if 0
386	/* KeFlushQueuedDpcs must be run at IRQL PASSIVE_LEVEL according to MSDN, but the
387	* driver verifier doesn't complain...
388	*/
389	AssertMsg(KeGetCurrentIrql() == PASSIVE_LEVEL, ("%d != %d (PASSIVE_LEVEL)\n", KeGetCurrentIrql(), PASSIVE_LEVEL));
390	# endif
391
392	KAFFINITY Mask = KeQueryActiveProcessors();
393
394	/* KeFlushQueuedDpcs is not present in Windows 2000; import it dynamically so we can just fail this call. */
395	if (!g_pfnrtNtKeFlushQueuedDpcs)
396	return VERR_NOT_SUPPORTED;
397
398	pArgs = (PRTMPARGS)ExAllocatePoolWithTag(NonPagedPool, MAXIMUM_PROCESSORS*sizeof(KDPC) + sizeof(RTMPARGS), (ULONG)'RTMp');
399	if (!pArgs)
400	return VERR_NO_MEMORY;
401
402	pArgs->pfnWorker = pfnWorker;
403	pArgs->pvUser1 = pvUser1;
404	pArgs->pvUser2 = pvUser2;
405	pArgs->idCpu = NIL_RTCPUID;
406	pArgs->idCpu2 = NIL_RTCPUID;
407	pArgs->cHits = 0;
408	pArgs->cRefs = 1;
409
410	paExecCpuDpcs = (KDPC *)(pArgs + 1);
411
412	if (enmCpuid == RT_NT_CPUID_SPECIFIC)
413	{
414	KeInitializeDpc(&paExecCpuDpcs[0], rtmpNtDPCWrapper, pArgs);
415	KeSetImportanceDpc(&paExecCpuDpcs[0], HighImportance);
416	KeSetTargetProcessorDpc(&paExecCpuDpcs[0], (int)idCpu);
417	pArgs->idCpu = idCpu;
418	}
419	else if (enmCpuid == RT_NT_CPUID_PAIR)
420	{
421	KeInitializeDpc(&paExecCpuDpcs[0], rtmpNtDPCWrapper, pArgs);
422	KeSetImportanceDpc(&paExecCpuDpcs[0], HighImportance);
423	KeSetTargetProcessorDpc(&paExecCpuDpcs[0], (int)idCpu);
424	pArgs->idCpu = idCpu;
425
426	KeInitializeDpc(&paExecCpuDpcs[1], rtmpNtDPCWrapper, pArgs);
427	KeSetImportanceDpc(&paExecCpuDpcs[1], HighImportance);
428	KeSetTargetProcessorDpc(&paExecCpuDpcs[1], (int)idCpu2);
429	pArgs->idCpu2 = idCpu2;
430	}
431	else
432	{
433	for (unsigned i = 0; i < MAXIMUM_PROCESSORS; i++)
434	{
435	KeInitializeDpc(&paExecCpuDpcs[i], rtmpNtDPCWrapper, pArgs);
436	KeSetImportanceDpc(&paExecCpuDpcs[i], HighImportance);
437	KeSetTargetProcessorDpc(&paExecCpuDpcs[i], i);
438	}
439	}
440
441	/* Raise the IRQL to DISPATCH_LEVEL so we can't be rescheduled to another cpu.
442	* KeInsertQueueDpc must also be executed at IRQL >= DISPATCH_LEVEL.
443	*/
444	KIRQL oldIrql;
445	KeRaiseIrql(DISPATCH_LEVEL, &oldIrql);
446
447	/*
448	* We cannot do other than assume a 1:1 relationship between the
449	* affinity mask and the process despite the warnings in the docs.
450	* If someone knows a better way to get this done, please let bird know.
451	*/
452	ASMCompilerBarrier(); /* paranoia */
453	if (enmCpuid == RT_NT_CPUID_SPECIFIC)
454	{
455	ASMAtomicIncS32(&pArgs->cRefs);
456	BOOLEAN fRc = KeInsertQueueDpc(&paExecCpuDpcs[0], 0, 0);
457	Assert(fRc); NOREF(fRc);
458	}
459	else if (enmCpuid == RT_NT_CPUID_PAIR)
460	{
461	ASMAtomicIncS32(&pArgs->cRefs);
462	BOOLEAN fRc = KeInsertQueueDpc(&paExecCpuDpcs[0], 0, 0);
463	Assert(fRc); NOREF(fRc);
464
465	ASMAtomicIncS32(&pArgs->cRefs);
466	fRc = KeInsertQueueDpc(&paExecCpuDpcs[1], 0, 0);
467	Assert(fRc); NOREF(fRc);
468	}
469	else
470	{
471	unsigned iSelf = KeGetCurrentProcessorNumber();
472
473	for (unsigned i = 0; i < MAXIMUM_PROCESSORS; i++)
474	{
475	if ( (i != iSelf)
476	&& (Mask & RT_BIT_64(i)))
477	{
478	ASMAtomicIncS32(&pArgs->cRefs);
479	BOOLEAN fRc = KeInsertQueueDpc(&paExecCpuDpcs[i], 0, 0);
480	Assert(fRc); NOREF(fRc);
481	}
482	}
483	if (enmCpuid != RT_NT_CPUID_OTHERS)
484	pfnWorker(iSelf, pvUser1, pvUser2);
485	}
486
487	KeLowerIrql(oldIrql);
488
489	/* Flush all DPCs and wait for completion. (can take long!) */
490	/** @todo Consider changing this to an active wait using some atomic inc/dec
491	* stuff (and check for the current cpu above in the specific case). */
492	/** @todo Seems KeFlushQueuedDpcs doesn't wait for the DPCs to be completely
493	* executed. Seen pArgs being freed while some CPU was using it before
494	* cRefs was added. */
495	g_pfnrtNtKeFlushQueuedDpcs();
496
497	if (pcHits)
498	*pcHits = pArgs->cHits;
499
500	/* Dereference the argument structure. */
501	int32_t cRefs = ASMAtomicDecS32(&pArgs->cRefs);
502	Assert(cRefs >= 0);
503	if (cRefs == 0)
504	ExFreePool(pArgs);
505
506	return VINF_SUCCESS;
507	#endif /* */
508	}
509
510
511	RTDECL(int) RTMpOnAll(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
512	{
513	if (g_pfnrtKeIpiGenericCall)
514	return rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnAllBroadcastIpiWrapper,
515	NIL_RTCPUID, NIL_RTCPUID, NULL);
516	return rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_ALL, NIL_RTCPUID, NIL_RTCPUID, NULL);
517	}
518
519
520	RTDECL(int) RTMpOnOthers(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
521	{
522	if (g_pfnrtKeIpiGenericCall)
523	return rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnOthersBroadcastIpiWrapper,
524	NIL_RTCPUID, NIL_RTCPUID, NULL);
525	return rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_OTHERS, NIL_RTCPUID, NIL_RTCPUID, NULL);
526	}
527
528
529	RTDECL(int) RTMpOnPair(RTCPUID idCpu1, RTCPUID idCpu2, uint32_t fFlags, PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
530	{
531	int rc;
532	AssertReturn(idCpu1 != idCpu2, VERR_INVALID_PARAMETER);
533	AssertReturn(!(fFlags & RTMPON_F_VALID_MASK), VERR_INVALID_FLAGS);
534	if ((fFlags & RTMPON_F_CONCURRENT_EXEC) && !g_pfnrtKeIpiGenericCall)
535	return VERR_NOT_SUPPORTED;
536
537	/*
538	* Check that both CPUs are online before doing the broadcast call.
539	*/
540	if ( RTMpIsCpuOnline(idCpu1)
541	&& RTMpIsCpuOnline(idCpu2))
542	{
543	/*
544	* The broadcast IPI isn't quite as bad as it could have been, because
545	* it looks like windows doesn't synchronize CPUs on the way out, they
546	* seems to get back to normal work while the pair is still busy.
547	*/
548	uint32_t cHits = 0;
549	if (g_pfnrtKeIpiGenericCall)
550	rc = rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnPairBroadcastIpiWrapper, idCpu1, idCpu2, &cHits);
551	else
552	rc = rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_PAIR, idCpu1, idCpu2, &cHits);
553	if (RT_SUCCESS(rc))
554	{
555	Assert(cHits <= 2);
556	if (cHits == 2)
557	rc = VINF_SUCCESS;
558	else if (cHits == 1)
559	rc = VERR_NOT_ALL_CPUS_SHOWED;
560	else if (cHits == 0)
561	rc = VERR_CPU_OFFLINE;
562	else
563	rc = VERR_CPU_IPE_1;
564	}
565	}
566	/*
567	* A CPU must be present to be considered just offline.
568	*/
569	else if ( RTMpIsCpuPresent(idCpu1)
570	&& RTMpIsCpuPresent(idCpu2))
571	rc = VERR_CPU_OFFLINE;
572	else
573	rc = VERR_CPU_NOT_FOUND;
574	return rc;
575	}
576
577
578	RTDECL(bool) RTMpOnPairIsConcurrentExecSupported(void)
579	{
580	return g_pfnrtKeIpiGenericCall != NULL;
581	}
582
583
584	/**
585	* Releases a reference to a RTMPNTONSPECIFICARGS heap allocation, freeing it
586	* when the last reference is released.
587	*/
588	DECLINLINE(void) rtMpNtOnSpecificRelease(PRTMPNTONSPECIFICARGS pArgs)
589	{
590	uint32_t cRefs = ASMAtomicDecU32(&pArgs->cRefs);
591	AssertMsg(cRefs <= 1, ("cRefs=%#x\n", cRefs));
592	if (cRefs == 0)
593	ExFreePool(pArgs);
594	}
595
596
597	/**
598	* Wrapper between the native nt per-cpu callbacks and PFNRTWORKER
599	*
600	* @param Dpc DPC object
601	* @param DeferredContext Context argument specified by KeInitializeDpc
602	* @param SystemArgument1 Argument specified by KeInsertQueueDpc
603	* @param SystemArgument2 Argument specified by KeInsertQueueDpc
604	*/
605	static VOID rtMpNtOnSpecificDpcWrapper(IN PKDPC Dpc, IN PVOID DeferredContext,
606	IN PVOID SystemArgument1, IN PVOID SystemArgument2)
607	{
608	PRTMPNTONSPECIFICARGS pArgs = (PRTMPNTONSPECIFICARGS)DeferredContext;
609	RT_NOREF3(Dpc, SystemArgument1, SystemArgument2);
610
611	ASMAtomicWriteBool(&pArgs->fExecuting, true);
612
613	pArgs->CallbackArgs.pfnWorker(KeGetCurrentProcessorNumber(), pArgs->CallbackArgs.pvUser1, pArgs->CallbackArgs.pvUser2);
614
615	ASMAtomicWriteBool(&pArgs->fDone, true);
616	KeSetEvent(&pArgs->DoneEvt, 1 /PriorityIncrement/, FALSE /Wait/);
617
618	rtMpNtOnSpecificRelease(pArgs);
619	}
620
621
622	RTDECL(int) RTMpOnSpecific(RTCPUID idCpu, PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
623	{
624	/*
625	* Don't try mess with an offline CPU.
626	*/
627	if (!RTMpIsCpuOnline(idCpu))
628	return !RTMpIsCpuPossible(idCpu)
629	? VERR_CPU_NOT_FOUND
630	: VERR_CPU_OFFLINE;
631
632	/*
633	* Use the broadcast IPI routine if there are no more than two CPUs online,
634	* or if the current IRQL is unsuitable for KeWaitForSingleObject.
635	*/
636	int rc;
637	uint32_t cHits = 0;
638	if ( g_pfnrtKeIpiGenericCall
639	&& ( RTMpGetOnlineCount() <= 2
640	\|\| KeGetCurrentIrql() > APC_LEVEL)
641	)
642	{
643	rc = rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnSpecificBroadcastIpiWrapper,
644	idCpu, NIL_RTCPUID, &cHits);
645	if (RT_SUCCESS(rc))
646	{
647	if (cHits == 1)
648	return VINF_SUCCESS;
649	rc = cHits == 0 ? VERR_CPU_OFFLINE : VERR_CPU_IPE_1;
650	}
651	return rc;
652	}
653
654	#if 0
655	rc = rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_SPECIFIC, idCpu, NIL_RTCPUID, &cHits);
656	if (RT_SUCCESS(rc))
657	{
658	if (cHits == 1)
659	return VINF_SUCCESS;
660	rc = cHits == 0 ? VERR_CPU_OFFLINE : VERR_CPU_IPE_1;
661	}
662	return rc;
663
664	#else
665	/*
666	* Initialize the argument package and the objects within it.
667	* The package is referenced counted to avoid unnecessary spinning to
668	* synchronize cleanup and prevent stack corruption.
669	*/
670	PRTMPNTONSPECIFICARGS pArgs = (PRTMPNTONSPECIFICARGS)ExAllocatePoolWithTag(NonPagedPool, sizeof(*pArgs), (ULONG)'RTMp');
671	if (!pArgs)
672	return VERR_NO_MEMORY;
673	pArgs->cRefs = 2;
674	pArgs->fExecuting = false;
675	pArgs->fDone = false;
676	pArgs->CallbackArgs.pfnWorker = pfnWorker;
677	pArgs->CallbackArgs.pvUser1 = pvUser1;
678	pArgs->CallbackArgs.pvUser2 = pvUser2;
679	pArgs->CallbackArgs.idCpu = idCpu;
680	pArgs->CallbackArgs.cHits = 0;
681	pArgs->CallbackArgs.cRefs = 2;
682	KeInitializeEvent(&pArgs->DoneEvt, SynchronizationEvent, FALSE /* not signalled */);
683	KeInitializeDpc(&pArgs->Dpc, rtMpNtOnSpecificDpcWrapper, pArgs);
684	KeSetImportanceDpc(&pArgs->Dpc, HighImportance);
685	KeSetTargetProcessorDpc(&pArgs->Dpc, (int)idCpu);
686
687	/*
688	* Disable preemption while we check the current processor and inserts the DPC.
689	*/
690	KIRQL bOldIrql;
691	KeRaiseIrql(DISPATCH_LEVEL, &bOldIrql);
692	ASMCompilerBarrier(); /* paranoia */
693
694	if (RTMpCpuId() == idCpu)
695	{
696	/* Just execute the callback on the current CPU. */
697	pfnWorker(idCpu, pvUser1, pvUser2);
698	KeLowerIrql(bOldIrql);
699
700	ExFreePool(pArgs);
701	return VINF_SUCCESS;
702	}
703
704	/* Different CPU, so queue it if the CPU is still online. */
705	if (RTMpIsCpuOnline(idCpu))
706	{
707	BOOLEAN fRc = KeInsertQueueDpc(&pArgs->Dpc, 0, 0);
708	Assert(fRc); NOREF(fRc);
709	KeLowerIrql(bOldIrql);
710
711	uint64_t const nsRealWaitTS = RTTimeNanoTS();
712
713	/*
714	* Wait actively for a while in case the CPU/thread responds quickly.
715	*/
716	uint32_t cLoopsLeft = 0x20000;
717	while (cLoopsLeft-- > 0)
718	{
719	if (pArgs->fDone)
720	{
721	rtMpNtOnSpecificRelease(pArgs);
722	return VINF_SUCCESS;
723	}
724	ASMNopPause();
725	}
726
727	/*
728	* It didn't respond, so wait on the event object, poking the CPU if it's slow.
729	*/
730	LARGE_INTEGER Timeout;
731	Timeout.QuadPart = -10000; /* 1ms */
732	NTSTATUS rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
733	if (rcNt == STATUS_SUCCESS)
734	{
735	rtMpNtOnSpecificRelease(pArgs);
736	return VINF_SUCCESS;
737	}
738
739	/* If it hasn't respondend yet, maybe poke it and wait some more. */
740	if (rcNt == STATUS_TIMEOUT)
741	{
742	#ifndef IPRT_TARGET_NT4
743	if ( !pArgs->fExecuting
744	&& ( g_pfnrtMpPokeCpuWorker == rtMpPokeCpuUsingHalReqestIpiW7Plus
745	\|\| g_pfnrtMpPokeCpuWorker == rtMpPokeCpuUsingHalReqestIpiPreW7))
746	RTMpPokeCpu(idCpu);
747	#endif
748
749	Timeout.QuadPart = -1280000; /* 128ms */
750	rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
751	if (rcNt == STATUS_SUCCESS)
752	{
753	rtMpNtOnSpecificRelease(pArgs);
754	return VINF_SUCCESS;
755	}
756	}
757
758	/*
759	* Something weird is happening, try bail out.
760	*/
761	if (KeRemoveQueueDpc(&pArgs->Dpc))
762	{
763	ExFreePool(pArgs); /* DPC was still queued, so we can return without further ado. */
764	LogRel(("RTMpOnSpecific(%#x): Not processed after %llu ns: rcNt=%#x\n", idCpu, RTTimeNanoTS() - nsRealWaitTS, rcNt));
765	}
766	else
767	{
768	/* DPC is running, wait a good while for it to complete. */
769	LogRel(("RTMpOnSpecific(%#x): Still running after %llu ns: rcNt=%#x\n", idCpu, RTTimeNanoTS() - nsRealWaitTS, rcNt));
770
771	Timeout.QuadPart = -301000100010; / 30 seconds */
772	rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
773	if (rcNt != STATUS_SUCCESS)
774	LogRel(("RTMpOnSpecific(%#x): Giving up on running worker after %llu ns: rcNt=%#x\n", idCpu, RTTimeNanoTS() - nsRealWaitTS, rcNt));
775	}
776	rc = RTErrConvertFromNtStatus(rcNt);
777	}
778	else
779	{
780	/* CPU is offline.*/
781	KeLowerIrql(bOldIrql);
782	rc = !RTMpIsCpuPossible(idCpu) ? VERR_CPU_NOT_FOUND : VERR_CPU_OFFLINE;
783	}
784
785	rtMpNtOnSpecificRelease(pArgs);
786	return rc;
787	#endif
788	}
789
790
791
792
793	static VOID rtMpNtPokeCpuDummy(IN PKDPC Dpc, IN PVOID DeferredContext, IN PVOID SystemArgument1, IN PVOID SystemArgument2)
794	{
795	NOREF(Dpc);
796	NOREF(DeferredContext);
797	NOREF(SystemArgument1);
798	NOREF(SystemArgument2);
799	}
800
801	#ifndef IPRT_TARGET_NT4
802
803	/** Callback used by rtMpPokeCpuUsingBroadcastIpi. */
804	static ULONG_PTR rtMpIpiGenericCall(ULONG_PTR Argument)
805	{
806	NOREF(Argument);
807	return 0;
808	}
809
810
811	/**
812	* RTMpPokeCpu worker that uses broadcast IPIs for doing the work.
813	*
814	* @returns VINF_SUCCESS
815	* @param idCpu The CPU identifier.
816	*/
817	int rtMpPokeCpuUsingBroadcastIpi(RTCPUID idCpu)
818	{
819	NOREF(idCpu);
820	g_pfnrtKeIpiGenericCall(rtMpIpiGenericCall, 0);
821	return VINF_SUCCESS;
822	}
823
824
825	/**
826	* RTMpPokeCpu worker that uses the Windows 7 and later version of
827	* HalRequestIpip to get the job done.
828	*
829	* @returns VINF_SUCCESS
830	* @param idCpu The CPU identifier.
831	*/
832	int rtMpPokeCpuUsingHalReqestIpiW7Plus(RTCPUID idCpu)
833	{
834	/*
835	* I think we'll let idCpu be an NT processor number and not a HAL processor
836	* index. KeAddProcessorAffinityEx is for HAL and uses HAL processor
837	* indexes as input from what I can tell.
838	*/
839	PROCESSOR_NUMBER ProcNumber = { /Group=/ idCpu / 64, /Number=/ idCpu % 64, /* Reserved=*/ 0};
840	KAFFINITY_EX Target;
841	g_pfnrtKeInitializeAffinityEx(&Target);
842	g_pfnrtKeAddProcessorAffinityEx(&Target, g_pfnrtKeGetProcessorIndexFromNumber(&ProcNumber));
843
844	g_pfnrtHalRequestIpiW7Plus(0, &Target);
845	return VINF_SUCCESS;
846	}
847
848
849	/**
850	* RTMpPokeCpu worker that uses the Vista and earlier version of HalRequestIpip
851	* to get the job done.
852	*
853	* @returns VINF_SUCCESS
854	* @param idCpu The CPU identifier.
855	*/
856	int rtMpPokeCpuUsingHalReqestIpiPreW7(RTCPUID idCpu)
857	{
858	__debugbreak(); /** @todo this code needs testing!! */
859	KAFFINITY Target = 1;
860	Target <<= idCpu;
861	g_pfnrtHalRequestIpiPreW7(Target);
862	return VINF_SUCCESS;
863	}
864
865	#endif /* !IPRT_TARGET_NT4 */
866
867
868	int rtMpPokeCpuUsingDpc(RTCPUID idCpu)
869	{
870	/*
871	* APC fallback.
872	*/
873	static KDPC s_aPokeDpcs[MAXIMUM_PROCESSORS] = {0};
874	static bool s_fPokeDPCsInitialized = false;
875
876	if (!s_fPokeDPCsInitialized)
877	{
878	for (unsigned i = 0; i < RT_ELEMENTS(s_aPokeDpcs); i++)
879	{
880	KeInitializeDpc(&s_aPokeDpcs[i], rtMpNtPokeCpuDummy, NULL);
881	KeSetImportanceDpc(&s_aPokeDpcs[i], HighImportance);
882	KeSetTargetProcessorDpc(&s_aPokeDpcs[i], (int)i);
883	}
884	s_fPokeDPCsInitialized = true;
885	}
886
887	/* Raise the IRQL to DISPATCH_LEVEL so we can't be rescheduled to another cpu.
888	* KeInsertQueueDpc must also be executed at IRQL >= DISPATCH_LEVEL.
889	*/
890	KIRQL oldIrql;
891	KeRaiseIrql(DISPATCH_LEVEL, &oldIrql);
892
893	KeSetImportanceDpc(&s_aPokeDpcs[idCpu], HighImportance);
894	KeSetTargetProcessorDpc(&s_aPokeDpcs[idCpu], (int)idCpu);
895
896	/* Assuming here that high importance DPCs will be delivered immediately; or at least an IPI will be sent immediately.
897	* @note: not true on at least Vista & Windows 7
898	*/
899	BOOLEAN bRet = KeInsertQueueDpc(&s_aPokeDpcs[idCpu], 0, 0);
900
901	KeLowerIrql(oldIrql);
902	return (bRet == TRUE) ? VINF_SUCCESS : VERR_ACCESS_DENIED /* already queued */;
903	}
904
905
906	RTDECL(int) RTMpPokeCpu(RTCPUID idCpu)
907	{
908	if (!RTMpIsCpuOnline(idCpu))
909	return !RTMpIsCpuPossible(idCpu)
910	? VERR_CPU_NOT_FOUND
911	: VERR_CPU_OFFLINE;
912	/* Calls rtMpSendIpiFallback, rtMpSendIpiWin7AndLater or rtMpSendIpiVista. */
913	return g_pfnrtMpPokeCpuWorker(idCpu);
914	}
915
916
917	RTDECL(bool) RTMpOnAllIsConcurrentSafe(void)
918	{
919	return false;
920	}
921

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

source: vbox/trunk/src/VBox/Runtime/r0drv/nt/mp-r0drv-nt.cpp@ 63558

Download in other formats: