VirtualBox

source: vbox/trunk/src/VBox/Runtime/r0drv/nt/mp-r0drv-nt.cpp@ 55998

Last change on this file since 55998 was 54501, checked in by vboxsync, 10 years ago

Reverted r98596, only using broadcast IPI in RTMpOnSpecific for special cases.

  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 28.4 KB
Line 
1/* $Id: mp-r0drv-nt.cpp 54501 2015-02-25 15:58:43Z vboxsync $ */
2/** @file
3 * IPRT - Multiprocessor, Ring-0 Driver, NT.
4 */
5
6/*
7 * Copyright (C) 2008-2014 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*******************************************************************************/
46typedef 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 */
58typedef 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. */
74typedef RTMPNTONSPECIFICARGS *PRTMPNTONSPECIFICARGS;
75
76
77
78/* test a couple of assumption. */
79AssertCompile(MAXIMUM_PROCESSORS <= RTCPUSET_MAX_CPUS);
80AssertCompile(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
90RTDECL(RTCPUID) RTMpCpuId(void)
91{
92 /* WDK upgrade warning: PCR->Number changed from BYTE to WORD. */
93 return KeGetCurrentProcessorNumber();
94}
95
96
97RTDECL(int) RTMpCurSetIndex(void)
98{
99 /* WDK upgrade warning: PCR->Number changed from BYTE to WORD. */
100 return KeGetCurrentProcessorNumber();
101}
102
103
104RTDECL(int) RTMpCurSetIndexAndId(PRTCPUID pidCpu)
105{
106 return *pidCpu = KeGetCurrentProcessorNumber();
107}
108
109
110RTDECL(int) RTMpCpuIdToSetIndex(RTCPUID idCpu)
111{
112 return idCpu < MAXIMUM_PROCESSORS ? (int)idCpu : -1;
113}
114
115
116RTDECL(RTCPUID) RTMpCpuIdFromSetIndex(int iCpu)
117{
118 return (unsigned)iCpu < MAXIMUM_PROCESSORS ? iCpu : NIL_RTCPUID;
119}
120
121
122RTDECL(RTCPUID) RTMpGetMaxCpuId(void)
123{
124 /** @todo use KeQueryMaximumProcessorCount on vista+ */
125 return MAXIMUM_PROCESSORS - 1;
126}
127
128
129RTDECL(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
143RTDECL(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
153RTDECL(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
161RTDECL(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
169RTDECL(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
181RTDECL(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 */
193typedef struct
194{
195 LIST_ENTRY DpcListHead;
196 ULONG_PTR DpcLock;
197 volatile ULONG DpcQueueDepth;
198 ULONG DpcQueueCount;
199} KDPC_DATA, *PKDPC_DATA;
200
201RTDECL(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
220RTDECL(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 */
234static ULONG_PTR __stdcall 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 */
249static ULONG_PTR __stdcall 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 */
268static ULONG_PTR __stdcall 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 */
288static ULONG_PTR __stdcall 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 idCpu First CPU to match, ultimately specific to the
309 * pfnNativeWrapper used.
310 * @param idCpu2 Second CPU to match, ultimately specific to the
311 * pfnNativeWrapper used.
312 * @param pcHits Where to return the number of this. Optional.
313 */
314static int rtMpCallUsingBroadcastIpi(PFNRTMPWORKER pfnWorker, void *pvUser1, void *pvUser2,
315 PKIPI_BROADCAST_WORKER pfnNativeWrapper, RTCPUID idCpu, RTCPUID idCpu2,
316 uint32_t *pcHits)
317{
318 RTMPARGS Args;
319 Args.pfnWorker = pfnWorker;
320 Args.pvUser1 = pvUser1;
321 Args.pvUser2 = pvUser2;
322 Args.idCpu = idCpu;
323 Args.idCpu2 = idCpu2;
324 Args.cRefs = 0;
325 Args.cHits = 0;
326
327 AssertPtr(g_pfnrtKeIpiGenericCall);
328 g_pfnrtKeIpiGenericCall(pfnNativeWrapper, (uintptr_t)&Args);
329 if (pcHits)
330 *pcHits = Args.cHits;
331 return VINF_SUCCESS;
332}
333
334
335/**
336 * Wrapper between the native nt per-cpu callbacks and PFNRTWORKER
337 *
338 * @param Dpc DPC object
339 * @param DeferredContext Context argument specified by KeInitializeDpc
340 * @param SystemArgument1 Argument specified by KeInsertQueueDpc
341 * @param SystemArgument2 Argument specified by KeInsertQueueDpc
342 */
343static VOID __stdcall rtmpNtDPCWrapper(IN PKDPC Dpc, IN PVOID DeferredContext, IN PVOID SystemArgument1, IN PVOID SystemArgument2)
344{
345 PRTMPARGS pArgs = (PRTMPARGS)DeferredContext;
346
347 ASMAtomicIncU32(&pArgs->cHits);
348 pArgs->pfnWorker(KeGetCurrentProcessorNumber(), pArgs->pvUser1, pArgs->pvUser2);
349
350 /* Dereference the argument structure. */
351 int32_t cRefs = ASMAtomicDecS32(&pArgs->cRefs);
352 Assert(cRefs >= 0);
353 if (cRefs == 0)
354 ExFreePool(pArgs);
355}
356
357
358/**
359 * Internal worker for the RTMpOn* APIs.
360 *
361 * @returns IPRT status code.
362 * @param pfnWorker The callback.
363 * @param pvUser1 User argument 1.
364 * @param pvUser2 User argument 2.
365 * @param enmCpuid What to do / is idCpu valid.
366 * @param idCpu Used if enmCpuid is RT_NT_CPUID_SPECIFIC or
367 * RT_NT_CPUID_PAIR, otherwise ignored.
368 * @param idCpu2 Used if enmCpuid is RT_NT_CPUID_PAIR, otherwise ignored.
369 * @param pcHits Where to return the number of this. Optional.
370 */
371static int rtMpCallUsingDpcs(PFNRTMPWORKER pfnWorker, void *pvUser1, void *pvUser2,
372 RT_NT_CPUID enmCpuid, RTCPUID idCpu, RTCPUID idCpu2, uint32_t *pcHits)
373{
374 PRTMPARGS pArgs;
375 KDPC *paExecCpuDpcs;
376
377#if 0
378 /* KeFlushQueuedDpcs must be run at IRQL PASSIVE_LEVEL according to MSDN, but the
379 * driver verifier doesn't complain...
380 */
381 AssertMsg(KeGetCurrentIrql() == PASSIVE_LEVEL, ("%d != %d (PASSIVE_LEVEL)\n", KeGetCurrentIrql(), PASSIVE_LEVEL));
382#endif
383
384#ifdef IPRT_TARGET_NT4
385 KAFFINITY Mask;
386 /* g_pfnrtNt* are not present on NT anyway. */
387 return VERR_NOT_SUPPORTED;
388#else
389 KAFFINITY Mask = KeQueryActiveProcessors();
390#endif
391
392 /* KeFlushQueuedDpcs is not present in Windows 2000; import it dynamically so we can just fail this call. */
393 if (!g_pfnrtNtKeFlushQueuedDpcs)
394 return VERR_NOT_SUPPORTED;
395
396 pArgs = (PRTMPARGS)ExAllocatePoolWithTag(NonPagedPool, MAXIMUM_PROCESSORS*sizeof(KDPC) + sizeof(RTMPARGS), (ULONG)'RTMp');
397 if (!pArgs)
398 return VERR_NO_MEMORY;
399
400 pArgs->pfnWorker = pfnWorker;
401 pArgs->pvUser1 = pvUser1;
402 pArgs->pvUser2 = pvUser2;
403 pArgs->idCpu = NIL_RTCPUID;
404 pArgs->idCpu2 = NIL_RTCPUID;
405 pArgs->cHits = 0;
406 pArgs->cRefs = 1;
407
408 paExecCpuDpcs = (KDPC *)(pArgs + 1);
409
410 if (enmCpuid == RT_NT_CPUID_SPECIFIC)
411 {
412 KeInitializeDpc(&paExecCpuDpcs[0], rtmpNtDPCWrapper, pArgs);
413 KeSetImportanceDpc(&paExecCpuDpcs[0], HighImportance);
414 KeSetTargetProcessorDpc(&paExecCpuDpcs[0], (int)idCpu);
415 pArgs->idCpu = idCpu;
416 }
417 else if (enmCpuid == RT_NT_CPUID_SPECIFIC)
418 {
419 KeInitializeDpc(&paExecCpuDpcs[0], rtmpNtDPCWrapper, pArgs);
420 KeSetImportanceDpc(&paExecCpuDpcs[0], HighImportance);
421 KeSetTargetProcessorDpc(&paExecCpuDpcs[0], (int)idCpu);
422 pArgs->idCpu = idCpu;
423
424 KeInitializeDpc(&paExecCpuDpcs[1], rtmpNtDPCWrapper, pArgs);
425 KeSetImportanceDpc(&paExecCpuDpcs[1], HighImportance);
426 KeSetTargetProcessorDpc(&paExecCpuDpcs[1], (int)idCpu2);
427 pArgs->idCpu2 = idCpu2;
428 }
429 else
430 {
431 for (unsigned i = 0; i < MAXIMUM_PROCESSORS; i++)
432 {
433 KeInitializeDpc(&paExecCpuDpcs[i], rtmpNtDPCWrapper, pArgs);
434 KeSetImportanceDpc(&paExecCpuDpcs[i], HighImportance);
435 KeSetTargetProcessorDpc(&paExecCpuDpcs[i], i);
436 }
437 }
438
439 /* Raise the IRQL to DISPATCH_LEVEL so we can't be rescheduled to another cpu.
440 * KeInsertQueueDpc must also be executed at IRQL >= DISPATCH_LEVEL.
441 */
442 KIRQL oldIrql;
443 KeRaiseIrql(DISPATCH_LEVEL, &oldIrql);
444
445 /*
446 * We cannot do other than assume a 1:1 relationship between the
447 * affinity mask and the process despite the warnings in the docs.
448 * If someone knows a better way to get this done, please let bird know.
449 */
450 ASMCompilerBarrier(); /* paranoia */
451 if (enmCpuid == RT_NT_CPUID_SPECIFIC)
452 {
453 ASMAtomicIncS32(&pArgs->cRefs);
454 BOOLEAN ret = KeInsertQueueDpc(&paExecCpuDpcs[0], 0, 0);
455 Assert(ret);
456 }
457 else if (enmCpuid == RT_NT_CPUID_PAIR)
458 {
459 ASMAtomicIncS32(&pArgs->cRefs);
460 BOOLEAN ret = KeInsertQueueDpc(&paExecCpuDpcs[0], 0, 0);
461 Assert(ret);
462
463 ASMAtomicIncS32(&pArgs->cRefs);
464 ret = KeInsertQueueDpc(&paExecCpuDpcs[1], 0, 0);
465 Assert(ret);
466 }
467 else
468 {
469 unsigned iSelf = KeGetCurrentProcessorNumber();
470
471 for (unsigned i = 0; i < MAXIMUM_PROCESSORS; i++)
472 {
473 if ( (i != iSelf)
474 && (Mask & RT_BIT_64(i)))
475 {
476 ASMAtomicIncS32(&pArgs->cRefs);
477 BOOLEAN ret = KeInsertQueueDpc(&paExecCpuDpcs[i], 0, 0);
478 Assert(ret);
479 }
480 }
481 if (enmCpuid != RT_NT_CPUID_OTHERS)
482 pfnWorker(iSelf, pvUser1, pvUser2);
483 }
484
485 KeLowerIrql(oldIrql);
486
487 /* Flush all DPCs and wait for completion. (can take long!) */
488 /** @todo Consider changing this to an active wait using some atomic inc/dec
489 * stuff (and check for the current cpu above in the specific case). */
490 /** @todo Seems KeFlushQueuedDpcs doesn't wait for the DPCs to be completely
491 * executed. Seen pArgs being freed while some CPU was using it before
492 * cRefs was added. */
493 g_pfnrtNtKeFlushQueuedDpcs();
494
495 if (pcHits)
496 *pcHits = pArgs->cHits;
497
498 /* Dereference the argument structure. */
499 int32_t cRefs = ASMAtomicDecS32(&pArgs->cRefs);
500 Assert(cRefs >= 0);
501 if (cRefs == 0)
502 ExFreePool(pArgs);
503
504 return VINF_SUCCESS;
505}
506
507
508RTDECL(int) RTMpOnAll(PFNRTMPWORKER pfnWorker, void *pvUser1, void *pvUser2)
509{
510 if (g_pfnrtKeIpiGenericCall)
511 return rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnAllBroadcastIpiWrapper,
512 NIL_RTCPUID, NIL_RTCPUID, NULL);
513 return rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_ALL, NIL_RTCPUID, NIL_RTCPUID, NULL);
514}
515
516
517RTDECL(int) RTMpOnOthers(PFNRTMPWORKER pfnWorker, void *pvUser1, void *pvUser2)
518{
519 if (g_pfnrtKeIpiGenericCall)
520 return rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnOthersBroadcastIpiWrapper,
521 NIL_RTCPUID, NIL_RTCPUID, NULL);
522 return rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_OTHERS, NIL_RTCPUID, NIL_RTCPUID, NULL);
523}
524
525
526RTDECL(int) RTMpOnPair(RTCPUID idCpu1, RTCPUID idCpu2, uint32_t fFlags, PFNRTMPWORKER pfnWorker, void *pvUser1, void *pvUser2)
527{
528 int rc;
529 AssertReturn(idCpu1 != idCpu2, VERR_INVALID_PARAMETER);
530 AssertReturn(!(fFlags & RTMPON_F_VALID_MASK), VERR_INVALID_FLAGS);
531 if ((fFlags & RTMPON_F_CONCURRENT_EXEC) && !g_pfnrtKeIpiGenericCall)
532 return VERR_NOT_SUPPORTED;
533
534 /*
535 * Check that both CPUs are online before doing the broadcast call.
536 */
537 if ( RTMpIsCpuOnline(idCpu1)
538 && RTMpIsCpuOnline(idCpu2))
539 {
540 /*
541 * The broadcast IPI isn't quite as bad as it could have been, because
542 * it looks like windows doesn't synchronize CPUs on the way out, they
543 * seems to get back to normal work while the pair is still busy.
544 */
545 uint32_t cHits = 0;
546 if (g_pfnrtKeIpiGenericCall)
547 rc = rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnPairBroadcastIpiWrapper, idCpu1, idCpu2, &cHits);
548 else
549 rc = rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_PAIR, idCpu1, idCpu2, &cHits);
550 if (RT_SUCCESS(rc))
551 {
552 Assert(cHits <= 2);
553 if (cHits == 2)
554 rc = VINF_SUCCESS;
555 else if (cHits == 1)
556 rc = VERR_NOT_ALL_CPUS_SHOWED;
557 else if (cHits == 0)
558 rc = VERR_CPU_OFFLINE;
559 else
560 rc = VERR_CPU_IPE_1;
561 }
562 }
563 /*
564 * A CPU must be present to be considered just offline.
565 */
566 else if ( RTMpIsCpuPresent(idCpu1)
567 && RTMpIsCpuPresent(idCpu2))
568 rc = VERR_CPU_OFFLINE;
569 else
570 rc = VERR_CPU_NOT_FOUND;
571 return rc;
572}
573
574
575RTDECL(bool) RTMpOnPairIsConcurrentExecSupported(void)
576{
577 return g_pfnrtKeIpiGenericCall != NULL;
578}
579
580
581/**
582 * Releases a reference to a RTMPNTONSPECIFICARGS heap allocation, freeing it
583 * when the last reference is released.
584 */
585DECLINLINE(void) rtMpNtOnSpecificRelease(PRTMPNTONSPECIFICARGS pArgs)
586{
587 uint32_t cRefs = ASMAtomicDecU32(&pArgs->cRefs);
588 AssertMsg(cRefs <= 1, ("cRefs=%#x\n", cRefs));
589 if (cRefs == 0)
590 ExFreePool(pArgs);
591}
592
593
594/**
595 * Wrapper between the native nt per-cpu callbacks and PFNRTWORKER
596 *
597 * @param Dpc DPC object
598 * @param DeferredContext Context argument specified by KeInitializeDpc
599 * @param SystemArgument1 Argument specified by KeInsertQueueDpc
600 * @param SystemArgument2 Argument specified by KeInsertQueueDpc
601 */
602static VOID __stdcall rtMpNtOnSpecificDpcWrapper(IN PKDPC Dpc, IN PVOID DeferredContext,
603 IN PVOID SystemArgument1, IN PVOID SystemArgument2)
604{
605 PRTMPNTONSPECIFICARGS pArgs = (PRTMPNTONSPECIFICARGS)DeferredContext;
606 ASMAtomicWriteBool(&pArgs->fExecuting, true);
607
608 pArgs->CallbackArgs.pfnWorker(KeGetCurrentProcessorNumber(), pArgs->CallbackArgs.pvUser1, pArgs->CallbackArgs.pvUser2);
609
610 ASMAtomicWriteBool(&pArgs->fDone, true);
611 KeSetEvent(&pArgs->DoneEvt, 1 /*PriorityIncrement*/, FALSE /*Wait*/);
612
613 rtMpNtOnSpecificRelease(pArgs);
614}
615
616
617RTDECL(int) RTMpOnSpecific(RTCPUID idCpu, PFNRTMPWORKER pfnWorker, void *pvUser1, void *pvUser2)
618{
619 /*
620 * Don't try mess with an offline CPU.
621 */
622 if (!RTMpIsCpuOnline(idCpu))
623 return !RTMpIsCpuPossible(idCpu)
624 ? VERR_CPU_NOT_FOUND
625 : VERR_CPU_OFFLINE;
626
627 /*
628 * Use the broadcast IPI routine if there are no more than two CPUs online,
629 * or if the current IRQL is unsuitable for KeWaitForSingleObject.
630 */
631 int rc;
632 uint32_t cHits = 0;
633 if ( g_pfnrtKeIpiGenericCall
634 && ( RTMpGetOnlineCount() <= 2
635 || KeGetCurrentIrql() > APC_LEVEL)
636 )
637 {
638 rc = rtMpCallUsingBroadcastIpi(pfnWorker, pvUser1, pvUser2, rtmpNtOnSpecificBroadcastIpiWrapper,
639 idCpu, NIL_RTCPUID, &cHits);
640 if (RT_SUCCESS(rc))
641 {
642 if (cHits == 1)
643 return VINF_SUCCESS;
644 rc = cHits == 0 ? VERR_CPU_OFFLINE : VERR_CPU_IPE_1;
645 }
646 return rc;
647 }
648
649#if 0
650 rc = rtMpCallUsingDpcs(pfnWorker, pvUser1, pvUser2, RT_NT_CPUID_SPECIFIC, idCpu, NIL_RTCPUID, &cHits);
651 if (RT_SUCCESS(rc))
652 {
653 if (cHits == 1)
654 return VINF_SUCCESS;
655 rc = cHits == 0 ? VERR_CPU_OFFLINE : VERR_CPU_IPE_1;
656 }
657 return rc;
658
659#else
660 /*
661 * Initialize the argument package and the objects within it.
662 * The package is referenced counted to avoid unnecessary spinning to
663 * synchronize cleanup and prevent stack corruption.
664 */
665 PRTMPNTONSPECIFICARGS pArgs = (PRTMPNTONSPECIFICARGS)ExAllocatePoolWithTag(NonPagedPool, sizeof(*pArgs), (ULONG)'RTMp');
666 if (!pArgs)
667 return VERR_NO_MEMORY;
668 pArgs->cRefs = 2;
669 pArgs->fExecuting = false;
670 pArgs->fDone = false;
671 pArgs->CallbackArgs.pfnWorker = pfnWorker;
672 pArgs->CallbackArgs.pvUser1 = pvUser1;
673 pArgs->CallbackArgs.pvUser2 = pvUser2;
674 pArgs->CallbackArgs.idCpu = idCpu;
675 pArgs->CallbackArgs.cHits = 0;
676 pArgs->CallbackArgs.cRefs = 2;
677 KeInitializeEvent(&pArgs->DoneEvt, SynchronizationEvent, FALSE /* not signalled */);
678 KeInitializeDpc(&pArgs->Dpc, rtMpNtOnSpecificDpcWrapper, pArgs);
679 KeSetImportanceDpc(&pArgs->Dpc, HighImportance);
680 KeSetTargetProcessorDpc(&pArgs->Dpc, (int)idCpu);
681
682 /*
683 * Disable preemption while we check the current processor and inserts the DPC.
684 */
685 KIRQL bOldIrql;
686 KeRaiseIrql(DISPATCH_LEVEL, &bOldIrql);
687 ASMCompilerBarrier(); /* paranoia */
688
689 if (RTMpCpuId() == idCpu)
690 {
691 /* Just execute the callback on the current CPU. */
692 pfnWorker(idCpu, pvUser1, pvUser2);
693 KeLowerIrql(bOldIrql);
694
695 ExFreePool(pArgs);
696 return VINF_SUCCESS;
697 }
698
699 /* Different CPU, so queue it if the CPU is still online. */
700 if (RTMpIsCpuOnline(idCpu))
701 {
702 BOOLEAN fRc = KeInsertQueueDpc(&pArgs->Dpc, 0, 0);
703 Assert(fRc);
704 KeLowerIrql(bOldIrql);
705
706 uint64_t const nsRealWaitTS = RTTimeNanoTS();
707
708 /*
709 * Wait actively for a while in case the CPU/thread responds quickly.
710 */
711 uint32_t cLoopsLeft = 0x20000;
712 while (cLoopsLeft-- > 0)
713 {
714 if (pArgs->fDone)
715 {
716 rtMpNtOnSpecificRelease(pArgs);
717 return VINF_SUCCESS;
718 }
719 ASMNopPause();
720 }
721
722 /*
723 * It didn't respond, so wait on the event object, poking the CPU if it's slow.
724 */
725 LARGE_INTEGER Timeout;
726 Timeout.QuadPart = -10000; /* 1ms */
727 NTSTATUS rcNt = KeWaitForSingleObject(&pArgs->DoneEvt, Executive, KernelMode, FALSE /* Alertable */, &Timeout);
728 if (rcNt == STATUS_SUCCESS)
729 {
730 rtMpNtOnSpecificRelease(pArgs);
731 return VINF_SUCCESS;
732 }
733
734 /* If it hasn't respondend yet, maybe poke it and wait some more. */
735 if (rcNt == STATUS_TIMEOUT)
736 {
737#ifndef IPRT_TARGET_NT4
738 if ( !pArgs->fExecuting
739 && ( g_pfnrtMpPokeCpuWorker == rtMpPokeCpuUsingHalSendSoftwareInterrupt
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 = -30*1000*1000*10; /* 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
789static 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. */
800static ULONG_PTR __stdcall 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 */
813int rtMpPokeCpuUsingBroadcastIpi(RTCPUID idCpu)
814{
815 g_pfnrtKeIpiGenericCall(rtMpIpiGenericCall, 0);
816 return VINF_SUCCESS;
817}
818
819
820/**
821 * RTMpPokeCpu worker that uses HalSendSoftwareInterrupt to get the job done.
822 *
823 * This is only really available on AMD64, at least at the time of writing.
824 *
825 * @returns VINF_SUCCESS
826 * @param idCpu The CPU identifier.
827 */
828int rtMpPokeCpuUsingHalSendSoftwareInterrupt(RTCPUID idCpu)
829{
830 g_pfnrtNtHalSendSoftwareInterrupt(idCpu, DISPATCH_LEVEL);
831 return VINF_SUCCESS;
832}
833
834
835/**
836 * RTMpPokeCpu worker that uses the Windows 7 and later version of
837 * HalRequestIpip to get the job done.
838 *
839 * @returns VINF_SUCCESS
840 * @param idCpu The CPU identifier.
841 */
842int rtMpPokeCpuUsingHalReqestIpiW7Plus(RTCPUID idCpu)
843{
844 /*
845 * I think we'll let idCpu be an NT processor number and not a HAL processor
846 * index. KeAddProcessorAffinityEx is for HAL and uses HAL processor
847 * indexes as input from what I can tell.
848 */
849 PROCESSOR_NUMBER ProcNumber = { /*Group=*/ idCpu / 64, /*Number=*/ idCpu % 64, /* Reserved=*/ 0};
850 KAFFINITY_EX Target;
851 g_pfnrtKeInitializeAffinityEx(&Target);
852 g_pfnrtKeAddProcessorAffinityEx(&Target, g_pfnrtKeGetProcessorIndexFromNumber(&ProcNumber));
853
854 g_pfnrtHalRequestIpiW7Plus(0, &Target);
855 return VINF_SUCCESS;
856}
857
858
859/**
860 * RTMpPokeCpu worker that uses the Vista and earlier version of HalRequestIpip
861 * to get the job done.
862 *
863 * @returns VINF_SUCCESS
864 * @param idCpu The CPU identifier.
865 */
866int rtMpPokeCpuUsingHalReqestIpiPreW7(RTCPUID idCpu)
867{
868 __debugbreak(); /** @todo this code needs testing!! */
869 KAFFINITY Target = 1;
870 Target <<= idCpu;
871 g_pfnrtHalRequestIpiPreW7(Target);
872 return VINF_SUCCESS;
873}
874
875#endif /* !IPRT_TARGET_NT4 */
876
877
878int rtMpPokeCpuUsingDpc(RTCPUID idCpu)
879{
880 /*
881 * APC fallback.
882 */
883 static KDPC s_aPokeDpcs[MAXIMUM_PROCESSORS] = {0};
884 static bool s_fPokeDPCsInitialized = false;
885
886 if (!s_fPokeDPCsInitialized)
887 {
888 for (unsigned i = 0; i < RT_ELEMENTS(s_aPokeDpcs); i++)
889 {
890 KeInitializeDpc(&s_aPokeDpcs[i], rtMpNtPokeCpuDummy, NULL);
891 KeSetImportanceDpc(&s_aPokeDpcs[i], HighImportance);
892 KeSetTargetProcessorDpc(&s_aPokeDpcs[i], (int)i);
893 }
894 s_fPokeDPCsInitialized = true;
895 }
896
897 /* Raise the IRQL to DISPATCH_LEVEL so we can't be rescheduled to another cpu.
898 * KeInsertQueueDpc must also be executed at IRQL >= DISPATCH_LEVEL.
899 */
900 KIRQL oldIrql;
901 KeRaiseIrql(DISPATCH_LEVEL, &oldIrql);
902
903 KeSetImportanceDpc(&s_aPokeDpcs[idCpu], HighImportance);
904 KeSetTargetProcessorDpc(&s_aPokeDpcs[idCpu], (int)idCpu);
905
906 /* Assuming here that high importance DPCs will be delivered immediately; or at least an IPI will be sent immediately.
907 * @note: not true on at least Vista & Windows 7
908 */
909 BOOLEAN bRet = KeInsertQueueDpc(&s_aPokeDpcs[idCpu], 0, 0);
910
911 KeLowerIrql(oldIrql);
912 return (bRet == TRUE) ? VINF_SUCCESS : VERR_ACCESS_DENIED /* already queued */;
913}
914
915
916RTDECL(int) RTMpPokeCpu(RTCPUID idCpu)
917{
918 if (!RTMpIsCpuOnline(idCpu))
919 return !RTMpIsCpuPossible(idCpu)
920 ? VERR_CPU_NOT_FOUND
921 : VERR_CPU_OFFLINE;
922 /* Calls rtMpSendIpiFallback, rtMpSendIpiWin7AndLater or rtMpSendIpiVista. */
923 return g_pfnrtMpPokeCpuWorker(idCpu);
924}
925
926
927RTDECL(bool) RTMpOnAllIsConcurrentSafe(void)
928{
929 return false;
930}
931
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette