mp-r0drv-nt.cpp@ 61572

Last change on this file since 61572 was 60771, checked in by vboxsync, 9 years ago
RTMpPoke/nt: Removed all the HalSendSoftwareInterrupt code, don't want to spend time trying to figure out why it doesn't work. Just use HalRequestIpi, then no 32/64 differences either.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 28.2 KB

Line
1	/* $Id: mp-r0drv-nt.cpp 60771 2016-04-29 20:49:59Z vboxsync $ */
2	/** @file
3	* IPRT - Multiprocessor, Ring-0 Driver, NT.
4	*/
5
6	/*
7	* Copyright (C) 2008-2015 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	* 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
348	ASMAtomicIncU32(&pArgs->cHits);
349	pArgs->pfnWorker(KeGetCurrentProcessorNumber(), pArgs->pvUser1, pArgs->pvUser2);
350
351	/* Dereference the argument structure. */
352	int32_t cRefs = ASMAtomicDecS32(&pArgs->cRefs);
353	Assert(cRefs >= 0);
354	if (cRefs == 0)
355	ExFreePool(pArgs);
356	}
357
358
359	/**
360	* Internal worker for the RTMpOn* APIs.
361	*
362	* @returns IPRT status code.
363	* @param pfnWorker The callback.
364	* @param pvUser1 User argument 1.
365	* @param pvUser2 User argument 2.
366	* @param enmCpuid What to do / is idCpu valid.
367	* @param idCpu Used if enmCpuid is RT_NT_CPUID_SPECIFIC or
368	* RT_NT_CPUID_PAIR, otherwise ignored.
369	* @param idCpu2 Used if enmCpuid is RT_NT_CPUID_PAIR, otherwise ignored.
370	* @param pcHits Where to return the number of this. Optional.
371	*/
372	static int rtMpCallUsingDpcs(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2,
373	RT_NT_CPUID enmCpuid, RTCPUID idCpu, RTCPUID idCpu2, uint32_t *pcHits)
374	{
375	PRTMPARGS pArgs;
376	KDPC *paExecCpuDpcs;
377
378	#if 0
379	/* KeFlushQueuedDpcs must be run at IRQL PASSIVE_LEVEL according to MSDN, but the
380	* driver verifier doesn't complain...
381	*/
382	AssertMsg(KeGetCurrentIrql() == PASSIVE_LEVEL, ("%d != %d (PASSIVE_LEVEL)\n", KeGetCurrentIrql(), PASSIVE_LEVEL));
383	#endif
384
385	#ifdef IPRT_TARGET_NT4
386	KAFFINITY Mask;
387	/* g_pfnrtNt* are not present on NT anyway. */
388	return VERR_NOT_SUPPORTED;
389	#else
390	KAFFINITY Mask = KeQueryActiveProcessors();
391	#endif
392
393	/* KeFlushQueuedDpcs is not present in Windows 2000; import it dynamically so we can just fail this call. */
394	if (!g_pfnrtNtKeFlushQueuedDpcs)
395	return VERR_NOT_SUPPORTED;
396
397	pArgs = (PRTMPARGS)ExAllocatePoolWithTag(NonPagedPool, MAXIMUM_PROCESSORS*sizeof(KDPC) + sizeof(RTMPARGS), (ULONG)'RTMp');
398	if (!pArgs)
399	return VERR_NO_MEMORY;
400
401	pArgs->pfnWorker = pfnWorker;
402	pArgs->pvUser1 = pvUser1;
403	pArgs->pvUser2 = pvUser2;
404	pArgs->idCpu = NIL_RTCPUID;
405	pArgs->idCpu2 = NIL_RTCPUID;
406	pArgs->cHits = 0;
407	pArgs->cRefs = 1;
408
409	paExecCpuDpcs = (KDPC *)(pArgs + 1);
410
411	if (enmCpuid == RT_NT_CPUID_SPECIFIC)
412	{
413	KeInitializeDpc(&paExecCpuDpcs[0], rtmpNtDPCWrapper, pArgs);
414	KeSetImportanceDpc(&paExecCpuDpcs[0], HighImportance);
415	KeSetTargetProcessorDpc(&paExecCpuDpcs[0], (int)idCpu);
416	pArgs->idCpu = idCpu;
417	}
418	else if (enmCpuid == RT_NT_CPUID_PAIR)
419	{
420	KeInitializeDpc(&paExecCpuDpcs[0], rtmpNtDPCWrapper, pArgs);
421	KeSetImportanceDpc(&paExecCpuDpcs[0], HighImportance);
422	KeSetTargetProcessorDpc(&paExecCpuDpcs[0], (int)idCpu);
423	pArgs->idCpu = idCpu;
424
425	KeInitializeDpc(&paExecCpuDpcs[1], rtmpNtDPCWrapper, pArgs);
426	KeSetImportanceDpc(&paExecCpuDpcs[1], HighImportance);
427	KeSetTargetProcessorDpc(&paExecCpuDpcs[1], (int)idCpu2);
428	pArgs->idCpu2 = idCpu2;
429	}
430	else
431	{
432	for (unsigned i = 0; i < MAXIMUM_PROCESSORS; i++)
433	{
434	KeInitializeDpc(&paExecCpuDpcs[i], rtmpNtDPCWrapper, pArgs);
435	KeSetImportanceDpc(&paExecCpuDpcs[i], HighImportance);
436	KeSetTargetProcessorDpc(&paExecCpuDpcs[i], i);
437	}
438	}
439
440	/* Raise the IRQL to DISPATCH_LEVEL so we can't be rescheduled to another cpu.
441	* KeInsertQueueDpc must also be executed at IRQL >= DISPATCH_LEVEL.
442	*/
443	KIRQL oldIrql;
444	KeRaiseIrql(DISPATCH_LEVEL, &oldIrql);
445
446	/*
447	* We cannot do other than assume a 1:1 relationship between the
448	* affinity mask and the process despite the warnings in the docs.
449	* If someone knows a better way to get this done, please let bird know.
450	*/
451	ASMCompilerBarrier(); /* paranoia */
452	if (enmCpuid == RT_NT_CPUID_SPECIFIC)
453	{
454	ASMAtomicIncS32(&pArgs->cRefs);
455	BOOLEAN ret = KeInsertQueueDpc(&paExecCpuDpcs[0], 0, 0);
456	Assert(ret);
457	}
458	else if (enmCpuid == RT_NT_CPUID_PAIR)
459	{
460	ASMAtomicIncS32(&pArgs->cRefs);
461	BOOLEAN ret = KeInsertQueueDpc(&paExecCpuDpcs[0], 0, 0);
462	Assert(ret);
463
464	ASMAtomicIncS32(&pArgs->cRefs);
465	ret = KeInsertQueueDpc(&paExecCpuDpcs[1], 0, 0);
466	Assert(ret);
467	}
468	else
469	{
470	unsigned iSelf = KeGetCurrentProcessorNumber();
471
472	for (unsigned i = 0; i < MAXIMUM_PROCESSORS; i++)
473	{
474	if ( (i != iSelf)
475	&& (Mask & RT_BIT_64(i)))
476	{
477	ASMAtomicIncS32(&pArgs->cRefs);
478	BOOLEAN ret = KeInsertQueueDpc(&paExecCpuDpcs[i], 0, 0);
479	Assert(ret);
480	}
481	}
482	if (enmCpuid != RT_NT_CPUID_OTHERS)
483	pfnWorker(iSelf, pvUser1, pvUser2);
484	}
485
486	KeLowerIrql(oldIrql);
487
488	/* Flush all DPCs and wait for completion. (can take long!) */
489	/** @todo Consider changing this to an active wait using some atomic inc/dec
490	* stuff (and check for the current cpu above in the specific case). */
491	/** @todo Seems KeFlushQueuedDpcs doesn't wait for the DPCs to be completely
492	* executed. Seen pArgs being freed while some CPU was using it before
493	* cRefs was added. */
494	g_pfnrtNtKeFlushQueuedDpcs();
495
496	if (pcHits)
497	*pcHits = pArgs->cHits;
498
499	/* Dereference the argument structure. */
500	int32_t cRefs = ASMAtomicDecS32(&pArgs->cRefs);
501	Assert(cRefs >= 0);
502	if (cRefs == 0)
503	ExFreePool(pArgs);
504
505	return VINF_SUCCESS;
506	}
507
508
509	RTDECL(int) RTMpOnAll(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
510	{
511	if (g_pfnrtKeIpiGenericCall)
512	return rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnAllBroadcastIpiWrapper,
513	NIL_RTCPUID, NIL_RTCPUID, NULL);
514	return rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_ALL, NIL_RTCPUID, NIL_RTCPUID, NULL);
515	}
516
517
518	RTDECL(int) RTMpOnOthers(PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
519	{
520	if (g_pfnrtKeIpiGenericCall)
521	return rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnOthersBroadcastIpiWrapper,
522	NIL_RTCPUID, NIL_RTCPUID, NULL);
523	return rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_OTHERS, NIL_RTCPUID, NIL_RTCPUID, NULL);
524	}
525
526
527	RTDECL(int) RTMpOnPair(RTCPUID idCpu1, RTCPUID idCpu2, uint32_t fFlags, PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
528	{
529	int rc;
530	AssertReturn(idCpu1 != idCpu2, VERR_INVALID_PARAMETER);
531	AssertReturn(!(fFlags & RTMPON_F_VALID_MASK), VERR_INVALID_FLAGS);
532	if ((fFlags & RTMPON_F_CONCURRENT_EXEC) && !g_pfnrtKeIpiGenericCall)
533	return VERR_NOT_SUPPORTED;
534
535	/*
536	* Check that both CPUs are online before doing the broadcast call.
537	*/
538	if ( RTMpIsCpuOnline(idCpu1)
539	&& RTMpIsCpuOnline(idCpu2))
540	{
541	/*
542	* The broadcast IPI isn't quite as bad as it could have been, because
543	* it looks like windows doesn't synchronize CPUs on the way out, they
544	* seems to get back to normal work while the pair is still busy.
545	*/
546	uint32_t cHits = 0;
547	if (g_pfnrtKeIpiGenericCall)
548	rc = rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnPairBroadcastIpiWrapper, idCpu1, idCpu2, &cHits);
549	else
550	rc = rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_PAIR, idCpu1, idCpu2, &cHits);
551	if (RT_SUCCESS(rc))
552	{
553	Assert(cHits <= 2);
554	if (cHits == 2)
555	rc = VINF_SUCCESS;
556	else if (cHits == 1)
557	rc = VERR_NOT_ALL_CPUS_SHOWED;
558	else if (cHits == 0)
559	rc = VERR_CPU_OFFLINE;
560	else
561	rc = VERR_CPU_IPE_1;
562	}
563	}
564	/*
565	* A CPU must be present to be considered just offline.
566	*/
567	else if ( RTMpIsCpuPresent(idCpu1)
568	&& RTMpIsCpuPresent(idCpu2))
569	rc = VERR_CPU_OFFLINE;
570	else
571	rc = VERR_CPU_NOT_FOUND;
572	return rc;
573	}
574
575
576	RTDECL(bool) RTMpOnPairIsConcurrentExecSupported(void)
577	{
578	return g_pfnrtKeIpiGenericCall != NULL;
579	}
580
581
582	/**
583	* Releases a reference to a RTMPNTONSPECIFICARGS heap allocation, freeing it
584	* when the last reference is released.
585	*/
586	DECLINLINE(void) rtMpNtOnSpecificRelease(PRTMPNTONSPECIFICARGS pArgs)
587	{
588	uint32_t cRefs = ASMAtomicDecU32(&pArgs->cRefs);
589	AssertMsg(cRefs <= 1, ("cRefs=%#x\n", cRefs));
590	if (cRefs == 0)
591	ExFreePool(pArgs);
592	}
593
594
595	/**
596	* Wrapper between the native nt per-cpu callbacks and PFNRTWORKER
597	*
598	* @param Dpc DPC object
599	* @param DeferredContext Context argument specified by KeInitializeDpc
600	* @param SystemArgument1 Argument specified by KeInsertQueueDpc
601	* @param SystemArgument2 Argument specified by KeInsertQueueDpc
602	*/
603	static VOID rtMpNtOnSpecificDpcWrapper(IN PKDPC Dpc, IN PVOID DeferredContext,
604	IN PVOID SystemArgument1, IN PVOID SystemArgument2)
605	{
606	PRTMPNTONSPECIFICARGS pArgs = (PRTMPNTONSPECIFICARGS)DeferredContext;
607	ASMAtomicWriteBool(&pArgs->fExecuting, true);
608
609	pArgs->CallbackArgs.pfnWorker(KeGetCurrentProcessorNumber(), pArgs->CallbackArgs.pvUser1, pArgs->CallbackArgs.pvUser2);
610
611	ASMAtomicWriteBool(&pArgs->fDone, true);
612	KeSetEvent(&pArgs->DoneEvt, 1 /PriorityIncrement/, FALSE /Wait/);
613
614	rtMpNtOnSpecificRelease(pArgs);
615	}
616
617
618	RTDECL(int) RTMpOnSpecific(RTCPUID idCpu, PFNRTMPWORKER pfnWorker, void pvUser1, void pvUser2)
619	{
620	/*
621	* Don't try mess with an offline CPU.
622	*/
623	if (!RTMpIsCpuOnline(idCpu))
624	return !RTMpIsCpuPossible(idCpu)
625	? VERR_CPU_NOT_FOUND
626	: VERR_CPU_OFFLINE;
627
628	/*
629	* Use the broadcast IPI routine if there are no more than two CPUs online,
630	* or if the current IRQL is unsuitable for KeWaitForSingleObject.
631	*/
632	int rc;
633	uint32_t cHits = 0;
634	if ( g_pfnrtKeIpiGenericCall
635	&& ( RTMpGetOnlineCount() <= 2
636	\|\| KeGetCurrentIrql() > APC_LEVEL)
637	)
638	{
639	rc = rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnSpecificBroadcastIpiWrapper,
640	idCpu, NIL_RTCPUID, &cHits);
641	if (RT_SUCCESS(rc))
642	{
643	if (cHits == 1)
644	return VINF_SUCCESS;
645	rc = cHits == 0 ? VERR_CPU_OFFLINE : VERR_CPU_IPE_1;
646	}
647	return rc;
648	}
649
650	#if 0
651	rc = rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_SPECIFIC, idCpu, NIL_RTCPUID, &cHits);
652	if (RT_SUCCESS(rc))
653	{
654	if (cHits == 1)
655	return VINF_SUCCESS;
656	rc = cHits == 0 ? VERR_CPU_OFFLINE : VERR_CPU_IPE_1;
657	}
658	return rc;
659
660	#else
661	/*
662	* Initialize the argument package and the objects within it.
663	* The package is referenced counted to avoid unnecessary spinning to
664	* synchronize cleanup and prevent stack corruption.
665	*/
666	PRTMPNTONSPECIFICARGS pArgs = (PRTMPNTONSPECIFICARGS)ExAllocatePoolWithTag(NonPagedPool, sizeof(*pArgs), (ULONG)'RTMp');
667	if (!pArgs)
668	return VERR_NO_MEMORY;
669	pArgs->cRefs = 2;
670	pArgs->fExecuting = false;
671	pArgs->fDone = false;
672	pArgs->CallbackArgs.pfnWorker = pfnWorker;
673	pArgs->CallbackArgs.pvUser1 = pvUser1;
674	pArgs->CallbackArgs.pvUser2 = pvUser2;
675	pArgs->CallbackArgs.idCpu = idCpu;
676	pArgs->CallbackArgs.cHits = 0;
677	pArgs->CallbackArgs.cRefs = 2;
678	KeInitializeEvent(&pArgs->DoneEvt, SynchronizationEvent, FALSE /* not signalled */);
679	KeInitializeDpc(&pArgs->Dpc, rtMpNtOnSpecificDpcWrapper, pArgs);
680	KeSetImportanceDpc(&pArgs->Dpc, HighImportance);
681	KeSetTargetProcessorDpc(&pArgs->Dpc, (int)idCpu);
682
683	/*
684	* Disable preemption while we check the current processor and inserts the DPC.
685	*/
686	KIRQL bOldIrql;
687	KeRaiseIrql(DISPATCH_LEVEL, &bOldIrql);
688	ASMCompilerBarrier(); /* paranoia */
689
690	if (RTMpCpuId() == idCpu)
691	{
692	/* Just execute the callback on the current CPU. */
693	pfnWorker(idCpu, pvUser1, pvUser2);
694	KeLowerIrql(bOldIrql);
695
696	ExFreePool(pArgs);
697	return VINF_SUCCESS;
698	}
699
700	/* Different CPU, so queue it if the CPU is still online. */
701	if (RTMpIsCpuOnline(idCpu))
702	{
703	BOOLEAN fRc = KeInsertQueueDpc(&pArgs->Dpc, 0, 0);
704	Assert(fRc);
705	KeLowerIrql(bOldIrql);
706
707	uint64_t const nsRealWaitTS = RTTimeNanoTS();
708
709	/*
710	* Wait actively for a while in case the CPU/thread responds quickly.
711	*/
712	uint32_t cLoopsLeft = 0x20000;
713	while (cLoopsLeft-- > 0)
714	{
715	if (pArgs->fDone)
716	{
717	rtMpNtOnSpecificRelease(pArgs);
718	return VINF_SUCCESS;
719	}
720	ASMNopPause();
721	}
722
723	/*
724	* It didn't respond, so wait on the event object, poking the CPU if it's slow.
725	*/
726	LARGE_INTEGER Timeout;
727	Timeout.QuadPart = -10000; /* 1ms */
728	NTSTATUS rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
729	if (rcNt == STATUS_SUCCESS)
730	{
731	rtMpNtOnSpecificRelease(pArgs);
732	return VINF_SUCCESS;
733	}
734
735	/* If it hasn't respondend yet, maybe poke it and wait some more. */
736	if (rcNt == STATUS_TIMEOUT)
737	{
738	#ifndef IPRT_TARGET_NT4
739	if ( !pArgs->fExecuting
740	&& ( g_pfnrtMpPokeCpuWorker == rtMpPokeCpuUsingHalReqestIpiW7Plus
741	\|\| g_pfnrtMpPokeCpuWorker == rtMpPokeCpuUsingHalReqestIpiPreW7))
742	RTMpPokeCpu(idCpu);
743	#endif
744
745	Timeout.QuadPart = -1280000; /* 128ms */
746	rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
747	if (rcNt == STATUS_SUCCESS)
748	{
749	rtMpNtOnSpecificRelease(pArgs);
750	return VINF_SUCCESS;
751	}
752	}
753
754	/*
755	* Something weird is happening, try bail out.
756	*/
757	if (KeRemoveQueueDpc(&pArgs->Dpc))
758	{
759	ExFreePool(pArgs); /* DPC was still queued, so we can return without further ado. */
760	LogRel(("RTMpOnSpecific(%#x): Not processed after %llu ns: rcNt=%#x\n", idCpu, RTTimeNanoTS() - nsRealWaitTS, rcNt));
761	}
762	else
763	{
764	/* DPC is running, wait a good while for it to complete. */
765	LogRel(("RTMpOnSpecific(%#x): Still running after %llu ns: rcNt=%#x\n", idCpu, RTTimeNanoTS() - nsRealWaitTS, rcNt));
766
767	Timeout.QuadPart = -301000100010; / 30 seconds */
768	rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
769	if (rcNt != STATUS_SUCCESS)
770	LogRel(("RTMpOnSpecific(%#x): Giving up on running worker after %llu ns: rcNt=%#x\n", idCpu, RTTimeNanoTS() - nsRealWaitTS, rcNt));
771	}
772	rc = RTErrConvertFromNtStatus(rcNt);
773	}
774	else
775	{
776	/* CPU is offline.*/
777	KeLowerIrql(bOldIrql);
778	rc = !RTMpIsCpuPossible(idCpu) ? VERR_CPU_NOT_FOUND : VERR_CPU_OFFLINE;
779	}
780
781	rtMpNtOnSpecificRelease(pArgs);
782	return rc;
783	#endif
784	}
785
786
787
788
789	static VOID rtMpNtPokeCpuDummy(IN PKDPC Dpc, IN PVOID DeferredContext, IN PVOID SystemArgument1, IN PVOID SystemArgument2)
790	{
791	NOREF(Dpc);
792	NOREF(DeferredContext);
793	NOREF(SystemArgument1);
794	NOREF(SystemArgument2);
795	}
796
797	#ifndef IPRT_TARGET_NT4
798
799	/** Callback used by rtMpPokeCpuUsingBroadcastIpi. */
800	static ULONG_PTR rtMpIpiGenericCall(ULONG_PTR Argument)
801	{
802	NOREF(Argument);
803	return 0;
804	}
805
806
807	/**
808	* RTMpPokeCpu worker that uses broadcast IPIs for doing the work.
809	*
810	* @returns VINF_SUCCESS
811	* @param idCpu The CPU identifier.
812	*/
813	int rtMpPokeCpuUsingBroadcastIpi(RTCPUID idCpu)
814	{
815	g_pfnrtKeIpiGenericCall(rtMpIpiGenericCall, 0);
816	return VINF_SUCCESS;
817	}
818
819
820	/**
821	* RTMpPokeCpu worker that uses the Windows 7 and later version of
822	* HalRequestIpip to get the job done.
823	*
824	* @returns VINF_SUCCESS
825	* @param idCpu The CPU identifier.
826	*/
827	int rtMpPokeCpuUsingHalReqestIpiW7Plus(RTCPUID idCpu)
828	{
829	/*
830	* I think we'll let idCpu be an NT processor number and not a HAL processor
831	* index. KeAddProcessorAffinityEx is for HAL and uses HAL processor
832	* indexes as input from what I can tell.
833	*/
834	PROCESSOR_NUMBER ProcNumber = { /Group=/ idCpu / 64, /Number=/ idCpu % 64, /* Reserved=*/ 0};
835	KAFFINITY_EX Target;
836	g_pfnrtKeInitializeAffinityEx(&Target);
837	g_pfnrtKeAddProcessorAffinityEx(&Target, g_pfnrtKeGetProcessorIndexFromNumber(&ProcNumber));
838
839	g_pfnrtHalRequestIpiW7Plus(0, &Target);
840	return VINF_SUCCESS;
841	}
842
843
844	/**
845	* RTMpPokeCpu worker that uses the Vista and earlier version of HalRequestIpip
846	* to get the job done.
847	*
848	* @returns VINF_SUCCESS
849	* @param idCpu The CPU identifier.
850	*/
851	int rtMpPokeCpuUsingHalReqestIpiPreW7(RTCPUID idCpu)
852	{
853	__debugbreak(); /** @todo this code needs testing!! */
854	KAFFINITY Target = 1;
855	Target <<= idCpu;
856	g_pfnrtHalRequestIpiPreW7(Target);
857	return VINF_SUCCESS;
858	}
859
860	#endif /* !IPRT_TARGET_NT4 */
861
862
863	int rtMpPokeCpuUsingDpc(RTCPUID idCpu)
864	{
865	/*
866	* APC fallback.
867	*/
868	static KDPC s_aPokeDpcs[MAXIMUM_PROCESSORS] = {0};
869	static bool s_fPokeDPCsInitialized = false;
870
871	if (!s_fPokeDPCsInitialized)
872	{
873	for (unsigned i = 0; i < RT_ELEMENTS(s_aPokeDpcs); i++)
874	{
875	KeInitializeDpc(&s_aPokeDpcs[i], rtMpNtPokeCpuDummy, NULL);
876	KeSetImportanceDpc(&s_aPokeDpcs[i], HighImportance);
877	KeSetTargetProcessorDpc(&s_aPokeDpcs[i], (int)i);
878	}
879	s_fPokeDPCsInitialized = true;
880	}
881
882	/* Raise the IRQL to DISPATCH_LEVEL so we can't be rescheduled to another cpu.
883	* KeInsertQueueDpc must also be executed at IRQL >= DISPATCH_LEVEL.
884	*/
885	KIRQL oldIrql;
886	KeRaiseIrql(DISPATCH_LEVEL, &oldIrql);
887
888	KeSetImportanceDpc(&s_aPokeDpcs[idCpu], HighImportance);
889	KeSetTargetProcessorDpc(&s_aPokeDpcs[idCpu], (int)idCpu);
890
891	/* Assuming here that high importance DPCs will be delivered immediately; or at least an IPI will be sent immediately.
892	* @note: not true on at least Vista & Windows 7
893	*/
894	BOOLEAN bRet = KeInsertQueueDpc(&s_aPokeDpcs[idCpu], 0, 0);
895
896	KeLowerIrql(oldIrql);
897	return (bRet == TRUE) ? VINF_SUCCESS : VERR_ACCESS_DENIED /* already queued */;
898	}
899
900
901	RTDECL(int) RTMpPokeCpu(RTCPUID idCpu)
902	{
903	if (!RTMpIsCpuOnline(idCpu))
904	return !RTMpIsCpuPossible(idCpu)
905	? VERR_CPU_NOT_FOUND
906	: VERR_CPU_OFFLINE;
907	/* Calls rtMpSendIpiFallback, rtMpSendIpiWin7AndLater or rtMpSendIpiVista. */
908	return g_pfnrtMpPokeCpuWorker(idCpu);
909	}
910
911
912	RTDECL(bool) RTMpOnAllIsConcurrentSafe(void)
913	{
914	return false;
915	}
916

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

source: vbox/trunk/src/VBox/Runtime/r0drv/nt/mp-r0drv-nt.cpp@ 61572

Download in other formats: