VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR0/GVMMR0.cpp@ 30756

Last change on this file since 30756 was 30334, checked in by vboxsync, 14 years ago

Must initialize idHostCpu properly (breaks SMP guest restore)
  • Property svn:eol-style set to native
  • Property svn:keywords set to Id Revision
File size: 68.2 KB
Line 
1/* $Id: GVMMR0.cpp 30334 2010-06-21 14:19:33Z vboxsync $ */
2/** @file
3 * GVMM - Global VM Manager.
4 */
5
6/*
7 * Copyright (C) 2007 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. It's main function at the moment
22 * is to manage a list of all running VMs, keep a ring-0 only structure (GVM)
23 * for each of them, and assign them unique identifiers (so GMM can track
24 * page owners). The idea for the future is to add an idle priority kernel
25 * thread that can take care of tasks like page sharing.
26 *
27 * The GVMM will create a ring-0 object for each VM when it's registered,
28 * this is both for session cleanup purposes and for having a point where
29 * it's possible to implement usage polices later (in SUPR0ObjRegister).
30 */
31
32
33/*******************************************************************************
34* Header Files *
35*******************************************************************************/
36#define LOG_GROUP LOG_GROUP_GVMM
37#include <VBox/gvmm.h>
38#include <VBox/gmm.h>
39#include "GVMMR0Internal.h"
40#include <VBox/gvm.h>
41#include <VBox/vm.h>
42#include <VBox/vmm.h>
43#include <VBox/param.h>
44#include <VBox/err.h>
45#include <iprt/asm.h>
46#include <iprt/asm-amd64-x86.h>
47#include <iprt/mem.h>
48#include <iprt/semaphore.h>
49#include <iprt/time.h>
50#include <VBox/log.h>
51#include <iprt/thread.h>
52#include <iprt/process.h>
53#include <iprt/param.h>
54#include <iprt/string.h>
55#include <iprt/assert.h>
56#include <iprt/mem.h>
57#include <iprt/memobj.h>
58#include <iprt/mp.h>
59
60
61/*******************************************************************************
62* Structures and Typedefs *
63*******************************************************************************/
64
65/**
66 * Global VM handle.
67 */
68typedef struct GVMHANDLE
69{
70 /** The index of the next handle in the list (free or used). (0 is nil.) */
71 uint16_t volatile iNext;
72 /** Our own index / handle value. */
73 uint16_t iSelf;
74 /** The pointer to the ring-0 only (aka global) VM structure. */
75 PGVM pGVM;
76 /** The ring-0 mapping of the shared VM instance data. */
77 PVM pVM;
78 /** The virtual machine object. */
79 void *pvObj;
80 /** The session this VM is associated with. */
81 PSUPDRVSESSION pSession;
82 /** The ring-0 handle of the EMT0 thread.
83 * This is used for ownership checks as well as looking up a VM handle by thread
84 * at times like assertions. */
85 RTNATIVETHREAD hEMT0;
86 /** The process ID of the handle owner.
87 * This is used for access checks. */
88 RTPROCESS ProcId;
89} GVMHANDLE;
90/** Pointer to a global VM handle. */
91typedef GVMHANDLE *PGVMHANDLE;
92
93/** Number of GVM handles (including the NIL handle). */
94#if HC_ARCH_BITS == 64
95# define GVMM_MAX_HANDLES 1024
96#else
97# define GVMM_MAX_HANDLES 128
98#endif
99
100/**
101 * The GVMM instance data.
102 */
103typedef struct GVMM
104{
105 /** Eyecatcher / magic. */
106 uint32_t u32Magic;
107 /** The index of the head of the free handle chain. (0 is nil.) */
108 uint16_t volatile iFreeHead;
109 /** The index of the head of the active handle chain. (0 is nil.) */
110 uint16_t volatile iUsedHead;
111 /** The number of VMs. */
112 uint16_t volatile cVMs;
113// /** The number of halted EMT threads. */
114// uint16_t volatile cHaltedEMTs;
115 /** The number of EMTs. */
116 uint32_t volatile cEMTs;
117 /** The lock used to serialize VM creation, destruction and associated events that
118 * isn't performance critical. Owners may acquire the list lock. */
119 RTSEMFASTMUTEX CreateDestroyLock;
120 /** The lock used to serialize used list updates and accesses.
121 * This indirectly includes scheduling since the scheduler will have to walk the
122 * used list to examin running VMs. Owners may not acquire any other locks. */
123 RTSEMFASTMUTEX UsedLock;
124 /** The handle array.
125 * The size of this array defines the maximum number of currently running VMs.
126 * The first entry is unused as it represents the NIL handle. */
127 GVMHANDLE aHandles[GVMM_MAX_HANDLES];
128
129 /** @gcfgm{/GVMM/cEMTsMeansCompany, 32-bit, 0, UINT32_MAX, 1}
130 * The number of EMTs that means we no longer consider ourselves alone on a
131 * CPU/Core.
132 */
133 uint32_t cEMTsMeansCompany;
134 /** @gcfgm{/GVMM/MinSleepAlone,32-bit, 0, 100000000, 750000, ns}
135 * The minimum sleep time for when we're alone, in nano seconds.
136 */
137 uint32_t nsMinSleepAlone;
138 /** @gcfgm{/GVMM/MinSleepCompany,32-bit,0, 100000000, 15000, ns}
139 * The minimum sleep time for when we've got company, in nano seconds.
140 */
141 uint32_t nsMinSleepCompany;
142 /** @gcfgm{/GVMM/EarlyWakeUp1, 32-bit, 0, 100000000, 25000, ns}
143 * The limit for the first round of early wakeups, given in nano seconds.
144 */
145 uint32_t nsEarlyWakeUp1;
146 /** @gcfgm{/GVMM/EarlyWakeUp2, 32-bit, 0, 100000000, 50000, ns}
147 * The limit for the second round of early wakeups, given in nano seconds.
148 */
149 uint32_t nsEarlyWakeUp2;
150} GVMM;
151/** Pointer to the GVMM instance data. */
152typedef GVMM *PGVMM;
153
154/** The GVMM::u32Magic value (Charlie Haden). */
155#define GVMM_MAGIC 0x19370806
156
157
158
159/*******************************************************************************
160* Global Variables *
161*******************************************************************************/
162/** Pointer to the GVMM instance data.
163 * (Just my general dislike for global variables.) */
164static PGVMM g_pGVMM = NULL;
165
166/** Macro for obtaining and validating the g_pGVMM pointer.
167 * On failure it will return from the invoking function with the specified return value.
168 *
169 * @param pGVMM The name of the pGVMM variable.
170 * @param rc The return value on failure. Use VERR_INTERNAL_ERROR for
171 * VBox status codes.
172 */
173#define GVMM_GET_VALID_INSTANCE(pGVMM, rc) \
174 do { \
175 (pGVMM) = g_pGVMM;\
176 AssertPtrReturn((pGVMM), (rc)); \
177 AssertMsgReturn((pGVMM)->u32Magic == GVMM_MAGIC, ("%p - %#x\n", (pGVMM), (pGVMM)->u32Magic), (rc)); \
178 } while (0)
179
180/** Macro for obtaining and validating the g_pGVMM pointer, void function variant.
181 * On failure it will return from the invoking function.
182 *
183 * @param pGVMM The name of the pGVMM variable.
184 */
185#define GVMM_GET_VALID_INSTANCE_VOID(pGVMM) \
186 do { \
187 (pGVMM) = g_pGVMM;\
188 AssertPtrReturnVoid((pGVMM)); \
189 AssertMsgReturnVoid((pGVMM)->u32Magic == GVMM_MAGIC, ("%p - %#x\n", (pGVMM), (pGVMM)->u32Magic)); \
190 } while (0)
191
192
193/*******************************************************************************
194* Internal Functions *
195*******************************************************************************/
196static void gvmmR0InitPerVMData(PGVM pGVM);
197static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void *pvObj, void *pvGVMM, void *pvHandle);
198static int gvmmR0ByVM(PVM pVM, PGVM *ppGVM, PGVMM *ppGVMM, bool fTakeUsedLock);
199static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM, PGVMM *ppGVMM);
200
201
202/**
203 * Initializes the GVMM.
204 *
205 * This is called while owninng the loader sempahore (see supdrvIOCtl_LdrLoad()).
206 *
207 * @returns VBox status code.
208 */
209GVMMR0DECL(int) GVMMR0Init(void)
210{
211 LogFlow(("GVMMR0Init:\n"));
212
213 /*
214 * Allocate and initialize the instance data.
215 */
216 PGVMM pGVMM = (PGVMM)RTMemAllocZ(sizeof(*pGVMM));
217 if (!pGVMM)
218 return VERR_NO_MEMORY;
219 int rc = RTSemFastMutexCreate(&pGVMM->CreateDestroyLock);
220 if (RT_SUCCESS(rc))
221 {
222 rc = RTSemFastMutexCreate(&pGVMM->UsedLock);
223 if (RT_SUCCESS(rc))
224 {
225 pGVMM->u32Magic = GVMM_MAGIC;
226 pGVMM->iUsedHead = 0;
227 pGVMM->iFreeHead = 1;
228
229 /* the nil handle */
230 pGVMM->aHandles[0].iSelf = 0;
231 pGVMM->aHandles[0].iNext = 0;
232
233 /* the tail */
234 unsigned i = RT_ELEMENTS(pGVMM->aHandles) - 1;
235 pGVMM->aHandles[i].iSelf = i;
236 pGVMM->aHandles[i].iNext = 0; /* nil */
237
238 /* the rest */
239 while (i-- > 1)
240 {
241 pGVMM->aHandles[i].iSelf = i;
242 pGVMM->aHandles[i].iNext = i + 1;
243 }
244
245 /* The default configuration values. */
246 pGVMM->cEMTsMeansCompany = 1; /** @todo should be adjusted to relative to the cpu count or something... */
247 pGVMM->nsMinSleepAlone = 750000 /* ns (0.750 ms) */; /** @todo this should be adjusted to be 75% (or something) of the scheduler granularity... */
248 pGVMM->nsMinSleepCompany = 15000 /* ns (0.015 ms) */;
249 pGVMM->nsEarlyWakeUp1 = 25000 /* ns (0.025 ms) */;
250 pGVMM->nsEarlyWakeUp2 = 50000 /* ns (0.050 ms) */;
251
252 g_pGVMM = pGVMM;
253 LogFlow(("GVMMR0Init: pGVMM=%p\n", pGVMM));
254 return VINF_SUCCESS;
255 }
256
257 RTSemFastMutexDestroy(pGVMM->CreateDestroyLock);
258 }
259
260 RTMemFree(pGVMM);
261 return rc;
262}
263
264
265/**
266 * Terminates the GVM.
267 *
268 * This is called while owning the loader semaphore (see supdrvLdrFree()).
269 * And unless something is wrong, there should be absolutely no VMs
270 * registered at this point.
271 */
272GVMMR0DECL(void) GVMMR0Term(void)
273{
274 LogFlow(("GVMMR0Term:\n"));
275
276 PGVMM pGVMM = g_pGVMM;
277 g_pGVMM = NULL;
278 if (RT_UNLIKELY(!VALID_PTR(pGVMM)))
279 {
280 SUPR0Printf("GVMMR0Term: pGVMM=%p\n", pGVMM);
281 return;
282 }
283
284 pGVMM->u32Magic++;
285
286 RTSemFastMutexDestroy(pGVMM->UsedLock);
287 pGVMM->UsedLock = NIL_RTSEMFASTMUTEX;
288 RTSemFastMutexDestroy(pGVMM->CreateDestroyLock);
289 pGVMM->CreateDestroyLock = NIL_RTSEMFASTMUTEX;
290
291 pGVMM->iFreeHead = 0;
292 if (pGVMM->iUsedHead)
293 {
294 SUPR0Printf("GVMMR0Term: iUsedHead=%#x! (cVMs=%#x cEMTs=%#x)\n", pGVMM->iUsedHead, pGVMM->cVMs, pGVMM->cEMTs);
295 pGVMM->iUsedHead = 0;
296 }
297
298 RTMemFree(pGVMM);
299}
300
301
302/**
303 * A quick hack for setting global config values.
304 *
305 * @returns VBox status code.
306 *
307 * @param pSession The session handle. Used for authentication.
308 * @param pszName The variable name.
309 * @param u64Value The new value.
310 */
311GVMMR0DECL(int) GVMMR0SetConfig(PSUPDRVSESSION pSession, const char *pszName, uint64_t u64Value)
312{
313 /*
314 * Validate input.
315 */
316 PGVMM pGVMM;
317 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
318 AssertPtrReturn(pSession, VERR_INVALID_HANDLE);
319 AssertPtrReturn(pszName, VERR_INVALID_POINTER);
320
321 /*
322 * String switch time!
323 */
324 if (strncmp(pszName, "/GVMM/", sizeof("/GVMM/") - 1))
325 return VERR_CFGM_VALUE_NOT_FOUND; /* borrow status codes from CFGM... */
326 int rc = VINF_SUCCESS;
327 pszName += sizeof("/GVMM/") - 1;
328 if (!strcmp(pszName, "cEMTsMeansCompany"))
329 {
330 if (u64Value <= UINT32_MAX)
331 pGVMM->cEMTsMeansCompany = u64Value;
332 else
333 rc = VERR_OUT_OF_RANGE;
334 }
335 else if (!strcmp(pszName, "MinSleepAlone"))
336 {
337 if (u64Value <= 100000000)
338 pGVMM->nsMinSleepAlone = u64Value;
339 else
340 rc = VERR_OUT_OF_RANGE;
341 }
342 else if (!strcmp(pszName, "MinSleepCompany"))
343 {
344 if (u64Value <= 100000000)
345 pGVMM->nsMinSleepCompany = u64Value;
346 else
347 rc = VERR_OUT_OF_RANGE;
348 }
349 else if (!strcmp(pszName, "EarlyWakeUp1"))
350 {
351 if (u64Value <= 100000000)
352 pGVMM->nsEarlyWakeUp1 = u64Value;
353 else
354 rc = VERR_OUT_OF_RANGE;
355 }
356 else if (!strcmp(pszName, "EarlyWakeUp2"))
357 {
358 if (u64Value <= 100000000)
359 pGVMM->nsEarlyWakeUp2 = u64Value;
360 else
361 rc = VERR_OUT_OF_RANGE;
362 }
363 else
364 rc = VERR_CFGM_VALUE_NOT_FOUND;
365 return rc;
366}
367
368
369/**
370 * A quick hack for getting global config values.
371 *
372 * @returns VBox status code.
373 *
374 * @param pSession The session handle. Used for authentication.
375 * @param pszName The variable name.
376 * @param u64Value The new value.
377 */
378GVMMR0DECL(int) GVMMR0QueryConfig(PSUPDRVSESSION pSession, const char *pszName, uint64_t *pu64Value)
379{
380 /*
381 * Validate input.
382 */
383 PGVMM pGVMM;
384 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
385 AssertPtrReturn(pSession, VERR_INVALID_HANDLE);
386 AssertPtrReturn(pszName, VERR_INVALID_POINTER);
387 AssertPtrReturn(pu64Value, VERR_INVALID_POINTER);
388
389 /*
390 * String switch time!
391 */
392 if (strncmp(pszName, "/GVMM/", sizeof("/GVMM/") - 1))
393 return VERR_CFGM_VALUE_NOT_FOUND; /* borrow status codes from CFGM... */
394 int rc = VINF_SUCCESS;
395 pszName += sizeof("/GVMM/") - 1;
396 if (!strcmp(pszName, "cEMTsMeansCompany"))
397 *pu64Value = pGVMM->cEMTsMeansCompany;
398 else if (!strcmp(pszName, "MinSleepAlone"))
399 *pu64Value = pGVMM->nsMinSleepAlone;
400 else if (!strcmp(pszName, "MinSleepCompany"))
401 *pu64Value = pGVMM->nsMinSleepCompany;
402 else if (!strcmp(pszName, "EarlyWakeUp1"))
403 *pu64Value = pGVMM->nsEarlyWakeUp1;
404 else if (!strcmp(pszName, "EarlyWakeUp2"))
405 *pu64Value = pGVMM->nsEarlyWakeUp2;
406 else
407 rc = VERR_CFGM_VALUE_NOT_FOUND;
408 return rc;
409}
410
411
412/**
413 * Try acquire the 'used' lock.
414 *
415 * @returns IPRT status code, see RTSemFastMutexRequest.
416 * @param pGVMM The GVMM instance data.
417 */
418DECLINLINE(int) gvmmR0UsedLock(PGVMM pGVMM)
419{
420 LogFlow(("++gvmmR0UsedLock(%p)\n", pGVMM));
421 int rc = RTSemFastMutexRequest(pGVMM->UsedLock);
422 LogFlow(("gvmmR0UsedLock(%p)->%Rrc\n", pGVMM, rc));
423 return rc;
424}
425
426
427/**
428 * Release the 'used' lock.
429 *
430 * @returns IPRT status code, see RTSemFastMutexRelease.
431 * @param pGVMM The GVMM instance data.
432 */
433DECLINLINE(int) gvmmR0UsedUnlock(PGVMM pGVMM)
434{
435 LogFlow(("--gvmmR0UsedUnlock(%p)\n", pGVMM));
436 int rc = RTSemFastMutexRelease(pGVMM->UsedLock);
437 AssertRC(rc);
438 return rc;
439}
440
441
442/**
443 * Try acquire the 'create & destroy' lock.
444 *
445 * @returns IPRT status code, see RTSemFastMutexRequest.
446 * @param pGVMM The GVMM instance data.
447 */
448DECLINLINE(int) gvmmR0CreateDestroyLock(PGVMM pGVMM)
449{
450 LogFlow(("++gvmmR0CreateDestroyLock(%p)\n", pGVMM));
451 int rc = RTSemFastMutexRequest(pGVMM->CreateDestroyLock);
452 LogFlow(("gvmmR0CreateDestroyLock(%p)->%Rrc\n", pGVMM, rc));
453 return rc;
454}
455
456
457/**
458 * Release the 'create & destroy' lock.
459 *
460 * @returns IPRT status code, see RTSemFastMutexRequest.
461 * @param pGVMM The GVMM instance data.
462 */
463DECLINLINE(int) gvmmR0CreateDestroyUnlock(PGVMM pGVMM)
464{
465 LogFlow(("--gvmmR0CreateDestroyUnlock(%p)\n", pGVMM));
466 int rc = RTSemFastMutexRelease(pGVMM->CreateDestroyLock);
467 AssertRC(rc);
468 return rc;
469}
470
471
472/**
473 * Request wrapper for the GVMMR0CreateVM API.
474 *
475 * @returns VBox status code.
476 * @param pReq The request buffer.
477 */
478GVMMR0DECL(int) GVMMR0CreateVMReq(PGVMMCREATEVMREQ pReq)
479{
480 /*
481 * Validate the request.
482 */
483 if (!VALID_PTR(pReq))
484 return VERR_INVALID_POINTER;
485 if (pReq->Hdr.cbReq != sizeof(*pReq))
486 return VERR_INVALID_PARAMETER;
487 if (!VALID_PTR(pReq->pSession))
488 return VERR_INVALID_POINTER;
489
490 /*
491 * Execute it.
492 */
493 PVM pVM;
494 pReq->pVMR0 = NULL;
495 pReq->pVMR3 = NIL_RTR3PTR;
496 int rc = GVMMR0CreateVM(pReq->pSession, pReq->cCpus, &pVM);
497 if (RT_SUCCESS(rc))
498 {
499 pReq->pVMR0 = pVM;
500 pReq->pVMR3 = pVM->pVMR3;
501 }
502 return rc;
503}
504
505
506/**
507 * Allocates the VM structure and registers it with GVM.
508 *
509 * The caller will become the VM owner and there by the EMT.
510 *
511 * @returns VBox status code.
512 * @param pSession The support driver session.
513 * @param cCpus Number of virtual CPUs for the new VM.
514 * @param ppVM Where to store the pointer to the VM structure.
515 *
516 * @thread EMT.
517 */
518GVMMR0DECL(int) GVMMR0CreateVM(PSUPDRVSESSION pSession, uint32_t cCpus, PVM *ppVM)
519{
520 LogFlow(("GVMMR0CreateVM: pSession=%p\n", pSession));
521 PGVMM pGVMM;
522 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
523
524 AssertPtrReturn(ppVM, VERR_INVALID_POINTER);
525 *ppVM = NULL;
526
527 if ( cCpus == 0
528 || cCpus > VMM_MAX_CPU_COUNT)
529 return VERR_INVALID_PARAMETER;
530
531 RTNATIVETHREAD hEMT0 = RTThreadNativeSelf();
532 AssertReturn(hEMT0 != NIL_RTNATIVETHREAD, VERR_INTERNAL_ERROR);
533 RTNATIVETHREAD ProcId = RTProcSelf();
534 AssertReturn(ProcId != NIL_RTPROCESS, VERR_INTERNAL_ERROR);
535
536 /*
537 * The whole allocation process is protected by the lock.
538 */
539 int rc = gvmmR0CreateDestroyLock(pGVMM);
540 AssertRCReturn(rc, rc);
541
542 /*
543 * Allocate a handle first so we don't waste resources unnecessarily.
544 */
545 uint16_t iHandle = pGVMM->iFreeHead;
546 if (iHandle)
547 {
548 PGVMHANDLE pHandle = &pGVMM->aHandles[iHandle];
549
550 /* consistency checks, a bit paranoid as always. */
551 if ( !pHandle->pVM
552 && !pHandle->pGVM
553 && !pHandle->pvObj
554 && pHandle->iSelf == iHandle)
555 {
556 pHandle->pvObj = SUPR0ObjRegister(pSession, SUPDRVOBJTYPE_VM, gvmmR0HandleObjDestructor, pGVMM, pHandle);
557 if (pHandle->pvObj)
558 {
559 /*
560 * Move the handle from the free to used list and perform permission checks.
561 */
562 rc = gvmmR0UsedLock(pGVMM);
563 AssertRC(rc);
564
565 pGVMM->iFreeHead = pHandle->iNext;
566 pHandle->iNext = pGVMM->iUsedHead;
567 pGVMM->iUsedHead = iHandle;
568 pGVMM->cVMs++;
569
570 pHandle->pVM = NULL;
571 pHandle->pGVM = NULL;
572 pHandle->pSession = pSession;
573 pHandle->hEMT0 = NIL_RTNATIVETHREAD;
574 pHandle->ProcId = NIL_RTPROCESS;
575
576 gvmmR0UsedUnlock(pGVMM);
577
578 rc = SUPR0ObjVerifyAccess(pHandle->pvObj, pSession, NULL);
579 if (RT_SUCCESS(rc))
580 {
581 /*
582 * Allocate the global VM structure (GVM) and initialize it.
583 */
584 PGVM pGVM = (PGVM)RTMemAllocZ(RT_UOFFSETOF(GVM, aCpus[cCpus]));
585 if (pGVM)
586 {
587 pGVM->u32Magic = GVM_MAGIC;
588 pGVM->hSelf = iHandle;
589 pGVM->pVM = NULL;
590 pGVM->cCpus = cCpus;
591
592 gvmmR0InitPerVMData(pGVM);
593 GMMR0InitPerVMData(pGVM);
594
595 /*
596 * Allocate the shared VM structure and associated page array.
597 */
598 const uint32_t cbVM = RT_UOFFSETOF(VM, aCpus[cCpus]);
599 const uint32_t cPages = RT_ALIGN_32(cbVM, PAGE_SIZE) >> PAGE_SHIFT;
600#ifdef RT_OS_DARWIN /** @todo Figure out why this is broken. Is it only on snow leopard? */
601 rc = RTR0MemObjAllocLow(&pGVM->gvmm.s.VMMemObj, (cPages + 1) << PAGE_SHIFT, false /* fExecutable */);
602#else
603 rc = RTR0MemObjAllocLow(&pGVM->gvmm.s.VMMemObj, cPages << PAGE_SHIFT, false /* fExecutable */);
604#endif
605 if (RT_SUCCESS(rc))
606 {
607 PVM pVM = (PVM)RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj); AssertPtr(pVM);
608 memset(pVM, 0, cPages << PAGE_SHIFT);
609 pVM->enmVMState = VMSTATE_CREATING;
610 pVM->pVMR0 = pVM;
611 pVM->pSession = pSession;
612 pVM->hSelf = iHandle;
613 pVM->cbSelf = cbVM;
614 pVM->cCpus = cCpus;
615 pVM->offVMCPU = RT_UOFFSETOF(VM, aCpus);
616
617 rc = RTR0MemObjAllocPage(&pGVM->gvmm.s.VMPagesMemObj, cPages * sizeof(SUPPAGE), false /* fExecutable */);
618 if (RT_SUCCESS(rc))
619 {
620 PSUPPAGE paPages = (PSUPPAGE)RTR0MemObjAddress(pGVM->gvmm.s.VMPagesMemObj); AssertPtr(paPages);
621 for (uint32_t iPage = 0; iPage < cPages; iPage++)
622 {
623 paPages[iPage].uReserved = 0;
624 paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pGVM->gvmm.s.VMMemObj, iPage);
625 Assert(paPages[iPage].Phys != NIL_RTHCPHYS);
626 }
627
628 /*
629 * Map them into ring-3.
630 */
631 rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMMapObj, pGVM->gvmm.s.VMMemObj, (RTR3PTR)-1, 0,
632 RTMEM_PROT_READ | RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
633 if (RT_SUCCESS(rc))
634 {
635 pVM->pVMR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMMapObj);
636 AssertPtr((void *)pVM->pVMR3);
637
638 /* Initialize all the VM pointers. */
639 for (uint32_t i = 0; i < cCpus; i++)
640 {
641 pVM->aCpus[i].pVMR0 = pVM;
642 pVM->aCpus[i].pVMR3 = pVM->pVMR3;
643 pVM->aCpus[i].idHostCpu = NIL_RTCPUID;
644 }
645
646 rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMPagesMapObj, pGVM->gvmm.s.VMPagesMemObj, (RTR3PTR)-1, 0,
647 RTMEM_PROT_READ | RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
648 if (RT_SUCCESS(rc))
649 {
650 pVM->paVMPagesR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMPagesMapObj);
651 AssertPtr((void *)pVM->paVMPagesR3);
652
653 /* complete the handle - take the UsedLock sem just to be careful. */
654 rc = gvmmR0UsedLock(pGVMM);
655 AssertRC(rc);
656
657 pHandle->pVM = pVM;
658 pHandle->pGVM = pGVM;
659 pHandle->hEMT0 = hEMT0;
660 pHandle->ProcId = ProcId;
661 pGVM->pVM = pVM;
662 pGVM->aCpus[0].hEMT = hEMT0;
663 pGVMM->cEMTs += cCpus;
664
665 gvmmR0UsedUnlock(pGVMM);
666 gvmmR0CreateDestroyUnlock(pGVMM);
667
668 *ppVM = pVM;
669 Log(("GVMMR0CreateVM: pVM=%p pVMR3=%p pGVM=%p hGVM=%d\n", pVM, pVM->pVMR3, pGVM, iHandle));
670 return VINF_SUCCESS;
671 }
672
673 RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */);
674 pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
675 }
676 RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */);
677 pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
678 }
679 RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */);
680 pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
681 }
682 }
683 }
684 /* else: The user wasn't permitted to create this VM. */
685
686 /*
687 * The handle will be freed by gvmmR0HandleObjDestructor as we release the
688 * object reference here. A little extra mess because of non-recursive lock.
689 */
690 void *pvObj = pHandle->pvObj;
691 pHandle->pvObj = NULL;
692 gvmmR0CreateDestroyUnlock(pGVMM);
693
694 SUPR0ObjRelease(pvObj, pSession);
695
696 SUPR0Printf("GVMMR0CreateVM: failed, rc=%d\n", rc);
697 return rc;
698 }
699
700 rc = VERR_NO_MEMORY;
701 }
702 else
703 rc = VERR_INTERNAL_ERROR;
704 }
705 else
706 rc = VERR_GVM_TOO_MANY_VMS;
707
708 gvmmR0CreateDestroyUnlock(pGVMM);
709 return rc;
710}
711
712
713/**
714 * Initializes the per VM data belonging to GVMM.
715 *
716 * @param pGVM Pointer to the global VM structure.
717 */
718static void gvmmR0InitPerVMData(PGVM pGVM)
719{
720 AssertCompile(RT_SIZEOFMEMB(GVM,gvmm.s) <= RT_SIZEOFMEMB(GVM,gvmm.padding));
721 AssertCompile(RT_SIZEOFMEMB(GVMCPU,gvmm.s) <= RT_SIZEOFMEMB(GVMCPU,gvmm.padding));
722 pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
723 pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
724 pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
725 pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
726 pGVM->gvmm.s.fDoneVMMR0Init = false;
727 pGVM->gvmm.s.fDoneVMMR0Term = false;
728
729 for (VMCPUID i = 0; i < pGVM->cCpus; i++)
730 {
731 pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
732 pGVM->aCpus[i].hEMT = NIL_RTNATIVETHREAD;
733 }
734}
735
736
737/**
738 * Does the VM initialization.
739 *
740 * @returns VBox status code.
741 * @param pVM Pointer to the shared VM structure.
742 */
743GVMMR0DECL(int) GVMMR0InitVM(PVM pVM)
744{
745 LogFlow(("GVMMR0InitVM: pVM=%p\n", pVM));
746
747 /*
748 * Validate the VM structure, state and handle.
749 */
750 PGVM pGVM;
751 PGVMM pGVMM;
752 int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
753 if (RT_SUCCESS(rc))
754 {
755 if ( !pGVM->gvmm.s.fDoneVMMR0Init
756 && pGVM->aCpus[0].gvmm.s.HaltEventMulti == NIL_RTSEMEVENTMULTI)
757 {
758 for (VMCPUID i = 0; i < pGVM->cCpus; i++)
759 {
760 rc = RTSemEventMultiCreate(&pGVM->aCpus[i].gvmm.s.HaltEventMulti);
761 if (RT_FAILURE(rc))
762 {
763 pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
764 break;
765 }
766 }
767 }
768 else
769 rc = VERR_WRONG_ORDER;
770 }
771
772 LogFlow(("GVMMR0InitVM: returns %Rrc\n", rc));
773 return rc;
774}
775
776
777/**
778 * Indicates that we're done with the ring-0 initialization
779 * of the VM.
780 *
781 * @param pVM Pointer to the shared VM structure.
782 * @thread EMT(0)
783 */
784GVMMR0DECL(void) GVMMR0DoneInitVM(PVM pVM)
785{
786 /* Validate the VM structure, state and handle. */
787 PGVM pGVM;
788 PGVMM pGVMM;
789 int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
790 AssertRCReturnVoid(rc);
791
792 /* Set the indicator. */
793 pGVM->gvmm.s.fDoneVMMR0Init = true;
794}
795
796
797/**
798 * Indicates that we're doing the ring-0 termination of the VM.
799 *
800 * @returns true if termination hasn't been done already, false if it has.
801 * @param pVM Pointer to the shared VM structure.
802 * @param pGVM Pointer to the global VM structure. Optional.
803 * @thread EMT(0)
804 */
805GVMMR0DECL(bool) GVMMR0DoingTermVM(PVM pVM, PGVM pGVM)
806{
807 /* Validate the VM structure, state and handle. */
808 AssertPtrNullReturn(pGVM, false);
809 AssertReturn(!pGVM || pGVM->u32Magic == GVM_MAGIC, false);
810 if (!pGVM)
811 {
812 PGVMM pGVMM;
813 int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
814 AssertRCReturn(rc, false);
815 }
816
817 /* Set the indicator. */
818 if (pGVM->gvmm.s.fDoneVMMR0Term)
819 return false;
820 pGVM->gvmm.s.fDoneVMMR0Term = true;
821 return true;
822}
823
824
825/**
826 * Destroys the VM, freeing all associated resources (the ring-0 ones anyway).
827 *
828 * This is call from the vmR3DestroyFinalBit and from a error path in VMR3Create,
829 * and the caller is not the EMT thread, unfortunately. For security reasons, it
830 * would've been nice if the caller was actually the EMT thread or that we somehow
831 * could've associated the calling thread with the VM up front.
832 *
833 * @returns VBox status code.
834 * @param pVM Where to store the pointer to the VM structure.
835 *
836 * @thread EMT(0) if it's associated with the VM, otherwise any thread.
837 */
838GVMMR0DECL(int) GVMMR0DestroyVM(PVM pVM)
839{
840 LogFlow(("GVMMR0DestroyVM: pVM=%p\n", pVM));
841 PGVMM pGVMM;
842 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
843
844
845 /*
846 * Validate the VM structure, state and caller.
847 */
848 AssertPtrReturn(pVM, VERR_INVALID_POINTER);
849 AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
850 AssertMsgReturn(pVM->enmVMState >= VMSTATE_CREATING && pVM->enmVMState <= VMSTATE_TERMINATED, ("%d\n", pVM->enmVMState), VERR_WRONG_ORDER);
851
852 uint32_t hGVM = pVM->hSelf;
853 AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
854 AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
855
856 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
857 AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
858
859 RTPROCESS ProcId = RTProcSelf();
860 RTNATIVETHREAD hSelf = RTThreadNativeSelf();
861 AssertReturn( ( pHandle->hEMT0 == hSelf
862 && pHandle->ProcId == ProcId)
863 || pHandle->hEMT0 == NIL_RTNATIVETHREAD, VERR_NOT_OWNER);
864
865 /*
866 * Lookup the handle and destroy the object.
867 * Since the lock isn't recursive and we'll have to leave it before dereferencing the
868 * object, we take some precautions against racing callers just in case...
869 */
870 int rc = gvmmR0CreateDestroyLock(pGVMM);
871 AssertRC(rc);
872
873 /* be careful here because we might theoretically be racing someone else cleaning up. */
874 if ( pHandle->pVM == pVM
875 && ( ( pHandle->hEMT0 == hSelf
876 && pHandle->ProcId == ProcId)
877 || pHandle->hEMT0 == NIL_RTNATIVETHREAD)
878 && VALID_PTR(pHandle->pvObj)
879 && VALID_PTR(pHandle->pSession)
880 && VALID_PTR(pHandle->pGVM)
881 && pHandle->pGVM->u32Magic == GVM_MAGIC)
882 {
883 void *pvObj = pHandle->pvObj;
884 pHandle->pvObj = NULL;
885 gvmmR0CreateDestroyUnlock(pGVMM);
886
887 SUPR0ObjRelease(pvObj, pHandle->pSession);
888 }
889 else
890 {
891 SUPR0Printf("GVMMR0DestroyVM: pHandle=%p:{.pVM=%p, .hEMT0=%p, .ProcId=%u, .pvObj=%p} pVM=%p hSelf=%p\n",
892 pHandle, pHandle->pVM, pHandle->hEMT0, pHandle->ProcId, pHandle->pvObj, pVM, hSelf);
893 gvmmR0CreateDestroyUnlock(pGVMM);
894 rc = VERR_INTERNAL_ERROR;
895 }
896
897 return rc;
898}
899
900
901/**
902 * Performs VM cleanup task as part of object destruction.
903 *
904 * @param pGVM The GVM pointer.
905 */
906static void gvmmR0CleanupVM(PGVM pGVM)
907{
908 if ( pGVM->gvmm.s.fDoneVMMR0Init
909 && !pGVM->gvmm.s.fDoneVMMR0Term)
910 {
911 if ( pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ
912 && RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj) == pGVM->pVM)
913 {
914 LogFlow(("gvmmR0CleanupVM: Calling VMMR0TermVM\n"));
915 VMMR0TermVM(pGVM->pVM, pGVM);
916 }
917 else
918 AssertMsgFailed(("gvmmR0CleanupVM: VMMemObj=%p pVM=%p\n", pGVM->gvmm.s.VMMemObj, pGVM->pVM));
919 }
920
921 GMMR0CleanupVM(pGVM);
922}
923
924
925/**
926 * Handle destructor.
927 *
928 * @param pvGVMM The GVM instance pointer.
929 * @param pvHandle The handle pointer.
930 */
931static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void *pvObj, void *pvGVMM, void *pvHandle)
932{
933 LogFlow(("gvmmR0HandleObjDestructor: %p %p %p\n", pvObj, pvGVMM, pvHandle));
934
935 /*
936 * Some quick, paranoid, input validation.
937 */
938 PGVMHANDLE pHandle = (PGVMHANDLE)pvHandle;
939 AssertPtr(pHandle);
940 PGVMM pGVMM = (PGVMM)pvGVMM;
941 Assert(pGVMM == g_pGVMM);
942 const uint16_t iHandle = pHandle - &pGVMM->aHandles[0];
943 if ( !iHandle
944 || iHandle >= RT_ELEMENTS(pGVMM->aHandles)
945 || iHandle != pHandle->iSelf)
946 {
947 SUPR0Printf("GVM: handle %d is out of range or corrupt (iSelf=%d)!\n", iHandle, pHandle->iSelf);
948 return;
949 }
950
951 int rc = gvmmR0CreateDestroyLock(pGVMM);
952 AssertRC(rc);
953 rc = gvmmR0UsedLock(pGVMM);
954 AssertRC(rc);
955
956 /*
957 * This is a tad slow but a doubly linked list is too much hazzle.
958 */
959 if (RT_UNLIKELY(pHandle->iNext >= RT_ELEMENTS(pGVMM->aHandles)))
960 {
961 SUPR0Printf("GVM: used list index %d is out of range!\n", pHandle->iNext);
962 gvmmR0UsedUnlock(pGVMM);
963 gvmmR0CreateDestroyUnlock(pGVMM);
964 return;
965 }
966
967 if (pGVMM->iUsedHead == iHandle)
968 pGVMM->iUsedHead = pHandle->iNext;
969 else
970 {
971 uint16_t iPrev = pGVMM->iUsedHead;
972 int c = RT_ELEMENTS(pGVMM->aHandles) + 2;
973 while (iPrev)
974 {
975 if (RT_UNLIKELY(iPrev >= RT_ELEMENTS(pGVMM->aHandles)))
976 {
977 SUPR0Printf("GVM: used list index %d is out of range!\n", iPrev);
978 gvmmR0UsedUnlock(pGVMM);
979 gvmmR0CreateDestroyUnlock(pGVMM);
980 return;
981 }
982 if (RT_UNLIKELY(c-- <= 0))
983 {
984 iPrev = 0;
985 break;
986 }
987
988 if (pGVMM->aHandles[iPrev].iNext == iHandle)
989 break;
990 iPrev = pGVMM->aHandles[iPrev].iNext;
991 }
992 if (!iPrev)
993 {
994 SUPR0Printf("GVM: can't find the handle previous previous of %d!\n", pHandle->iSelf);
995 gvmmR0UsedUnlock(pGVMM);
996 gvmmR0CreateDestroyUnlock(pGVMM);
997 return;
998 }
999
1000 Assert(pGVMM->aHandles[iPrev].iNext == iHandle);
1001 pGVMM->aHandles[iPrev].iNext = pHandle->iNext;
1002 }
1003 pHandle->iNext = 0;
1004 pGVMM->cVMs--;
1005
1006 /*
1007 * Do the global cleanup round.
1008 */
1009 PGVM pGVM = pHandle->pGVM;
1010 if ( VALID_PTR(pGVM)
1011 && pGVM->u32Magic == GVM_MAGIC)
1012 {
1013 pGVMM->cEMTs -= pGVM->cCpus;
1014 gvmmR0UsedUnlock(pGVMM);
1015
1016 gvmmR0CleanupVM(pGVM);
1017
1018 /*
1019 * Do the GVMM cleanup - must be done last.
1020 */
1021 /* The VM and VM pages mappings/allocations. */
1022 if (pGVM->gvmm.s.VMPagesMapObj != NIL_RTR0MEMOBJ)
1023 {
1024 rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMapObj, false /* fFreeMappings */); AssertRC(rc);
1025 pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
1026 }
1027
1028 if (pGVM->gvmm.s.VMMapObj != NIL_RTR0MEMOBJ)
1029 {
1030 rc = RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */); AssertRC(rc);
1031 pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
1032 }
1033
1034 if (pGVM->gvmm.s.VMPagesMemObj != NIL_RTR0MEMOBJ)
1035 {
1036 rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */); AssertRC(rc);
1037 pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
1038 }
1039
1040 if (pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ)
1041 {
1042 rc = RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */); AssertRC(rc);
1043 pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
1044 }
1045
1046 for (VMCPUID i = 0; i < pGVM->cCpus; i++)
1047 {
1048 if (pGVM->aCpus[i].gvmm.s.HaltEventMulti != NIL_RTSEMEVENTMULTI)
1049 {
1050 rc = RTSemEventMultiDestroy(pGVM->aCpus[i].gvmm.s.HaltEventMulti); AssertRC(rc);
1051 pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
1052 }
1053 }
1054
1055 /* the GVM structure itself. */
1056 pGVM->u32Magic |= UINT32_C(0x80000000);
1057 RTMemFree(pGVM);
1058
1059 /* Re-acquire the UsedLock before freeing the handle since we're updating handle fields. */
1060 rc = gvmmR0UsedLock(pGVMM);
1061 AssertRC(rc);
1062 }
1063 /* else: GVMMR0CreateVM cleanup. */
1064
1065 /*
1066 * Free the handle.
1067 */
1068 pHandle->iNext = pGVMM->iFreeHead;
1069 pGVMM->iFreeHead = iHandle;
1070 ASMAtomicWriteNullPtr(&pHandle->pGVM);
1071 ASMAtomicWriteNullPtr(&pHandle->pVM);
1072 ASMAtomicWriteNullPtr(&pHandle->pvObj);
1073 ASMAtomicWriteNullPtr(&pHandle->pSession);
1074 ASMAtomicWriteSize(&pHandle->hEMT0, NIL_RTNATIVETHREAD);
1075 ASMAtomicWriteSize(&pHandle->ProcId, NIL_RTPROCESS);
1076
1077 gvmmR0UsedUnlock(pGVMM);
1078 gvmmR0CreateDestroyUnlock(pGVMM);
1079 LogFlow(("gvmmR0HandleObjDestructor: returns\n"));
1080}
1081
1082
1083/**
1084 * Registers the calling thread as the EMT of a Virtual CPU.
1085 *
1086 * Note that VCPU 0 is automatically registered during VM creation.
1087 *
1088 * @returns VBox status code
1089 * @param pVM The shared VM structure (the ring-0 mapping).
1090 * @param idCpu VCPU id.
1091 */
1092GVMMR0DECL(int) GVMMR0RegisterVCpu(PVM pVM, VMCPUID idCpu)
1093{
1094 AssertReturn(idCpu != 0, VERR_NOT_OWNER);
1095
1096 /*
1097 * Validate the VM structure, state and handle.
1098 */
1099 PGVM pGVM;
1100 PGVMM pGVMM;
1101 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, false /* fTakeUsedLock */);
1102 if (RT_FAILURE(rc))
1103 return rc;
1104
1105 AssertReturn(idCpu < pVM->cCpus, VERR_INVALID_CPU_ID);
1106 AssertReturn(pGVM->aCpus[idCpu].hEMT == NIL_RTNATIVETHREAD, VERR_ACCESS_DENIED);
1107
1108 pGVM->aCpus[idCpu].hEMT = RTThreadNativeSelf();
1109 return VINF_SUCCESS;
1110}
1111
1112
1113/**
1114 * Lookup a GVM structure by its handle.
1115 *
1116 * @returns The GVM pointer on success, NULL on failure.
1117 * @param hGVM The global VM handle. Asserts on bad handle.
1118 */
1119GVMMR0DECL(PGVM) GVMMR0ByHandle(uint32_t hGVM)
1120{
1121 PGVMM pGVMM;
1122 GVMM_GET_VALID_INSTANCE(pGVMM, NULL);
1123
1124 /*
1125 * Validate.
1126 */
1127 AssertReturn(hGVM != NIL_GVM_HANDLE, NULL);
1128 AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), NULL);
1129
1130 /*
1131 * Look it up.
1132 */
1133 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1134 AssertPtrReturn(pHandle->pVM, NULL);
1135 AssertPtrReturn(pHandle->pvObj, NULL);
1136 PGVM pGVM = pHandle->pGVM;
1137 AssertPtrReturn(pGVM, NULL);
1138 AssertReturn(pGVM->pVM == pHandle->pVM, NULL);
1139
1140 return pHandle->pGVM;
1141}
1142
1143
1144/**
1145 * Lookup a GVM structure by the shared VM structure.
1146 *
1147 * The calling thread must be in the same process as the VM. All current lookups
1148 * are by threads inside the same process, so this will not be an issue.
1149 *
1150 * @returns VBox status code.
1151 * @param pVM The shared VM structure (the ring-0 mapping).
1152 * @param ppGVM Where to store the GVM pointer.
1153 * @param ppGVMM Where to store the pointer to the GVMM instance data.
1154 * @param fTakeUsedLock Whether to take the used lock or not.
1155 * Be very careful if not taking the lock as it's possible that
1156 * the VM will disappear then.
1157 *
1158 * @remark This will not assert on an invalid pVM but try return sliently.
1159 */
1160static int gvmmR0ByVM(PVM pVM, PGVM *ppGVM, PGVMM *ppGVMM, bool fTakeUsedLock)
1161{
1162 RTPROCESS ProcId = RTProcSelf();
1163 PGVMM pGVMM;
1164 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
1165
1166 /*
1167 * Validate.
1168 */
1169 if (RT_UNLIKELY( !VALID_PTR(pVM)
1170 || ((uintptr_t)pVM & PAGE_OFFSET_MASK)))
1171 return VERR_INVALID_POINTER;
1172 if (RT_UNLIKELY( pVM->enmVMState < VMSTATE_CREATING
1173 || pVM->enmVMState >= VMSTATE_TERMINATED))
1174 return VERR_INVALID_POINTER;
1175
1176 uint16_t hGVM = pVM->hSelf;
1177 if (RT_UNLIKELY( hGVM == NIL_GVM_HANDLE
1178 || hGVM >= RT_ELEMENTS(pGVMM->aHandles)))
1179 return VERR_INVALID_HANDLE;
1180
1181 /*
1182 * Look it up.
1183 */
1184 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1185 PGVM pGVM;
1186 if (fTakeUsedLock)
1187 {
1188 int rc = gvmmR0UsedLock(pGVMM);
1189 AssertRCReturn(rc, rc);
1190
1191 pGVM = pHandle->pGVM;
1192 if (RT_UNLIKELY( pHandle->pVM != pVM
1193 || pHandle->ProcId != ProcId
1194 || !VALID_PTR(pHandle->pvObj)
1195 || !VALID_PTR(pGVM)
1196 || pGVM->pVM != pVM))
1197 {
1198 gvmmR0UsedUnlock(pGVMM);
1199 return VERR_INVALID_HANDLE;
1200 }
1201 }
1202 else
1203 {
1204 if (RT_UNLIKELY(pHandle->pVM != pVM))
1205 return VERR_INVALID_HANDLE;
1206 if (RT_UNLIKELY(pHandle->ProcId != ProcId))
1207 return VERR_INVALID_HANDLE;
1208 if (RT_UNLIKELY(!VALID_PTR(pHandle->pvObj)))
1209 return VERR_INVALID_HANDLE;
1210
1211 pGVM = pHandle->pGVM;
1212 if (RT_UNLIKELY(!VALID_PTR(pGVM)))
1213 return VERR_INVALID_HANDLE;
1214 if (RT_UNLIKELY(pGVM->pVM != pVM))
1215 return VERR_INVALID_HANDLE;
1216 }
1217
1218 *ppGVM = pGVM;
1219 *ppGVMM = pGVMM;
1220 return VINF_SUCCESS;
1221}
1222
1223
1224/**
1225 * Lookup a GVM structure by the shared VM structure.
1226 *
1227 * @returns VBox status code.
1228 * @param pVM The shared VM structure (the ring-0 mapping).
1229 * @param ppGVM Where to store the GVM pointer.
1230 *
1231 * @remark This will not take the 'used'-lock because it doesn't do
1232 * nesting and this function will be used from under the lock.
1233 */
1234GVMMR0DECL(int) GVMMR0ByVM(PVM pVM, PGVM *ppGVM)
1235{
1236 PGVMM pGVMM;
1237 return gvmmR0ByVM(pVM, ppGVM, &pGVMM, false /* fTakeUsedLock */);
1238}
1239
1240
1241/**
1242 * Lookup a GVM structure by the shared VM structure and ensuring that the
1243 * caller is an EMT thread.
1244 *
1245 * @returns VBox status code.
1246 * @param pVM The shared VM structure (the ring-0 mapping).
1247 * @param idCpu The Virtual CPU ID of the calling EMT.
1248 * @param ppGVM Where to store the GVM pointer.
1249 * @param ppGVMM Where to store the pointer to the GVMM instance data.
1250 * @thread EMT
1251 *
1252 * @remark This will assert in all failure paths.
1253 */
1254static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM, PGVMM *ppGVMM)
1255{
1256 PGVMM pGVMM;
1257 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
1258
1259 /*
1260 * Validate.
1261 */
1262 AssertPtrReturn(pVM, VERR_INVALID_POINTER);
1263 AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
1264
1265 uint16_t hGVM = pVM->hSelf;
1266 AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
1267 AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
1268
1269 /*
1270 * Look it up.
1271 */
1272 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1273 AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
1274 RTPROCESS ProcId = RTProcSelf();
1275 AssertReturn(pHandle->ProcId == ProcId, VERR_NOT_OWNER);
1276 AssertPtrReturn(pHandle->pvObj, VERR_INTERNAL_ERROR);
1277
1278 PGVM pGVM = pHandle->pGVM;
1279 AssertPtrReturn(pGVM, VERR_INTERNAL_ERROR);
1280 AssertReturn(pGVM->pVM == pVM, VERR_INTERNAL_ERROR);
1281 RTNATIVETHREAD hAllegedEMT = RTThreadNativeSelf();
1282 AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
1283 AssertReturn(pGVM->aCpus[idCpu].hEMT == hAllegedEMT, VERR_INTERNAL_ERROR);
1284
1285 *ppGVM = pGVM;
1286 *ppGVMM = pGVMM;
1287 return VINF_SUCCESS;
1288}
1289
1290
1291/**
1292 * Lookup a GVM structure by the shared VM structure
1293 * and ensuring that the caller is the EMT thread.
1294 *
1295 * @returns VBox status code.
1296 * @param pVM The shared VM structure (the ring-0 mapping).
1297 * @param idCpu The Virtual CPU ID of the calling EMT.
1298 * @param ppGVM Where to store the GVM pointer.
1299 * @thread EMT
1300 */
1301GVMMR0DECL(int) GVMMR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM)
1302{
1303 AssertPtrReturn(ppGVM, VERR_INVALID_POINTER);
1304 PGVMM pGVMM;
1305 return gvmmR0ByVMAndEMT(pVM, idCpu, ppGVM, &pGVMM);
1306}
1307
1308
1309/**
1310 * Lookup a VM by its global handle.
1311 *
1312 * @returns The VM handle on success, NULL on failure.
1313 * @param hGVM The global VM handle. Asserts on bad handle.
1314 */
1315GVMMR0DECL(PVM) GVMMR0GetVMByHandle(uint32_t hGVM)
1316{
1317 PGVM pGVM = GVMMR0ByHandle(hGVM);
1318 return pGVM ? pGVM->pVM : NULL;
1319}
1320
1321
1322/**
1323 * Looks up the VM belonging to the specified EMT thread.
1324 *
1325 * This is used by the assertion machinery in VMMR0.cpp to avoid causing
1326 * unnecessary kernel panics when the EMT thread hits an assertion. The
1327 * call may or not be an EMT thread.
1328 *
1329 * @returns The VM handle on success, NULL on failure.
1330 * @param hEMT The native thread handle of the EMT.
1331 * NIL_RTNATIVETHREAD means the current thread
1332 */
1333GVMMR0DECL(PVM) GVMMR0GetVMByEMT(RTNATIVETHREAD hEMT)
1334{
1335 /*
1336 * No Assertions here as we're usually called in a AssertMsgN or
1337 * RTAssert* context.
1338 */
1339 PGVMM pGVMM = g_pGVMM;
1340 if ( !VALID_PTR(pGVMM)
1341 || pGVMM->u32Magic != GVMM_MAGIC)
1342 return NULL;
1343
1344 if (hEMT == NIL_RTNATIVETHREAD)
1345 hEMT = RTThreadNativeSelf();
1346 RTPROCESS ProcId = RTProcSelf();
1347
1348 /*
1349 * Search the handles in a linear fashion as we don't dare to take the lock (assert).
1350 */
1351 for (unsigned i = 1; i < RT_ELEMENTS(pGVMM->aHandles); i++)
1352 {
1353 if ( pGVMM->aHandles[i].iSelf == i
1354 && pGVMM->aHandles[i].ProcId == ProcId
1355 && VALID_PTR(pGVMM->aHandles[i].pvObj)
1356 && VALID_PTR(pGVMM->aHandles[i].pVM)
1357 && VALID_PTR(pGVMM->aHandles[i].pGVM))
1358 {
1359 if (pGVMM->aHandles[i].hEMT0 == hEMT)
1360 return pGVMM->aHandles[i].pVM;
1361
1362 /* This is fearly safe with the current process per VM approach. */
1363 PGVM pGVM = pGVMM->aHandles[i].pGVM;
1364 VMCPUID const cCpus = pGVM->cCpus;
1365 if ( cCpus < 1
1366 || cCpus > VMM_MAX_CPU_COUNT)
1367 continue;
1368 for (VMCPUID idCpu = 1; idCpu < cCpus; idCpu++)
1369 if (pGVM->aCpus[idCpu].hEMT == hEMT)
1370 return pGVMM->aHandles[i].pVM;
1371 }
1372 }
1373 return NULL;
1374}
1375
1376
1377/**
1378 * This is will wake up expired and soon-to-be expired VMs.
1379 *
1380 * @returns Number of VMs that has been woken up.
1381 * @param pGVMM Pointer to the GVMM instance data.
1382 * @param u64Now The current time.
1383 */
1384static unsigned gvmmR0SchedDoWakeUps(PGVMM pGVMM, uint64_t u64Now)
1385{
1386/** @todo Rewrite this algorithm. See performance defect XYZ. */
1387
1388 /*
1389 * The first pass will wake up VMs which have actually expired
1390 * and look for VMs that should be woken up in the 2nd and 3rd passes.
1391 */
1392 unsigned cWoken = 0;
1393 unsigned cHalted = 0;
1394 unsigned cTodo2nd = 0;
1395 unsigned cTodo3rd = 0;
1396 for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1397 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1398 i = pGVMM->aHandles[i].iNext)
1399 {
1400 PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1401 if ( VALID_PTR(pCurGVM)
1402 && pCurGVM->u32Magic == GVM_MAGIC)
1403 {
1404 for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1405 {
1406 PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1407
1408 uint64_t u64 = pCurGVCpu->gvmm.s.u64HaltExpire;
1409 if (u64)
1410 {
1411 if (u64 <= u64Now)
1412 {
1413 if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1414 {
1415 int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1416 AssertRC(rc);
1417 cWoken++;
1418 }
1419 }
1420 else
1421 {
1422 cHalted++;
1423 if (u64 <= u64Now + pGVMM->nsEarlyWakeUp1)
1424 cTodo2nd++;
1425 else if (u64 <= u64Now + pGVMM->nsEarlyWakeUp2)
1426 cTodo3rd++;
1427 }
1428 }
1429 }
1430 }
1431 AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1432 }
1433
1434 if (cTodo2nd)
1435 {
1436 for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1437 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1438 i = pGVMM->aHandles[i].iNext)
1439 {
1440 PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1441 if ( VALID_PTR(pCurGVM)
1442 && pCurGVM->u32Magic == GVM_MAGIC)
1443 {
1444 for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1445 {
1446 PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1447
1448 if ( pCurGVCpu->gvmm.s.u64HaltExpire
1449 && pCurGVCpu->gvmm.s.u64HaltExpire <= u64Now + pGVMM->nsEarlyWakeUp1)
1450 {
1451 if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1452 {
1453 int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1454 AssertRC(rc);
1455 cWoken++;
1456 }
1457 }
1458 }
1459 }
1460 AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1461 }
1462 }
1463
1464 if (cTodo3rd)
1465 {
1466 for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1467 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1468 i = pGVMM->aHandles[i].iNext)
1469 {
1470 PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1471 if ( VALID_PTR(pCurGVM)
1472 && pCurGVM->u32Magic == GVM_MAGIC)
1473 {
1474 for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1475 {
1476 PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1477
1478 if ( pCurGVCpu->gvmm.s.u64HaltExpire
1479 && pCurGVCpu->gvmm.s.u64HaltExpire <= u64Now + pGVMM->nsEarlyWakeUp2)
1480 {
1481 if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1482 {
1483 int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1484 AssertRC(rc);
1485 cWoken++;
1486 }
1487 }
1488 }
1489 }
1490 AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1491 }
1492 }
1493
1494 return cWoken;
1495}
1496
1497
1498/**
1499 * Halt the EMT thread.
1500 *
1501 * @returns VINF_SUCCESS normal wakeup (timeout or kicked by other thread).
1502 * VERR_INTERRUPTED if a signal was scheduled for the thread.
1503 * @param pVM Pointer to the shared VM structure.
1504 * @param idCpu The Virtual CPU ID of the calling EMT.
1505 * @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
1506 * @thread EMT(idCpu).
1507 */
1508GVMMR0DECL(int) GVMMR0SchedHalt(PVM pVM, VMCPUID idCpu, uint64_t u64ExpireGipTime)
1509{
1510 LogFlow(("GVMMR0SchedHalt: pVM=%p\n", pVM));
1511
1512 /*
1513 * Validate the VM structure, state and handle.
1514 */
1515 PGVM pGVM;
1516 PGVMM pGVMM;
1517 int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
1518 if (RT_FAILURE(rc))
1519 return rc;
1520 pGVM->gvmm.s.StatsSched.cHaltCalls++;
1521
1522 PGVMCPU pCurGVCpu = &pGVM->aCpus[idCpu];
1523 Assert(!pCurGVCpu->gvmm.s.u64HaltExpire);
1524
1525 /*
1526 * Take the UsedList semaphore, get the current time
1527 * and check if anyone needs waking up.
1528 * Interrupts must NOT be disabled at this point because we ask for GIP time!
1529 */
1530 rc = gvmmR0UsedLock(pGVMM);
1531 AssertRC(rc);
1532
1533 pCurGVCpu->gvmm.s.iCpuEmt = ASMGetApicId();
1534
1535 Assert(ASMGetFlags() & X86_EFL_IF);
1536 const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1537 pGVM->gvmm.s.StatsSched.cHaltWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1538
1539 /*
1540 * Go to sleep if we must...
1541 */
1542 if ( u64Now < u64ExpireGipTime
1543 && u64ExpireGipTime - u64Now > (pGVMM->cEMTs > pGVMM->cEMTsMeansCompany
1544 ? pGVMM->nsMinSleepCompany
1545 : pGVMM->nsMinSleepAlone))
1546 {
1547 pGVM->gvmm.s.StatsSched.cHaltBlocking++;
1548 ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, u64ExpireGipTime);
1549 gvmmR0UsedUnlock(pGVMM);
1550
1551 uint32_t cMillies = (u64ExpireGipTime - u64Now) / 1000000;
1552 rc = RTSemEventMultiWaitNoResume(pCurGVCpu->gvmm.s.HaltEventMulti, cMillies ? cMillies : 1);
1553 ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0);
1554 if (rc == VERR_TIMEOUT)
1555 {
1556 pGVM->gvmm.s.StatsSched.cHaltTimeouts++;
1557 rc = VINF_SUCCESS;
1558 }
1559 }
1560 else
1561 {
1562 pGVM->gvmm.s.StatsSched.cHaltNotBlocking++;
1563 gvmmR0UsedUnlock(pGVMM);
1564 }
1565
1566 /* Make sure false wake up calls (gvmmR0SchedDoWakeUps) cause us to spin. */
1567 RTSemEventMultiReset(pCurGVCpu->gvmm.s.HaltEventMulti);
1568
1569 return rc;
1570}
1571
1572
1573/**
1574 * Worker for GVMMR0SchedWakeUp and GVMMR0SchedWakeUpAndPokeCpus that wakes up
1575 * the a sleeping EMT.
1576 *
1577 * @retval VINF_SUCCESS if successfully woken up.
1578 * @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1579 *
1580 * @param pGVM The global (ring-0) VM structure.
1581 * @param pGVCpu The global (ring-0) VCPU structure.
1582 */
1583DECLINLINE(int) gvmmR0SchedWakeUpOne(PGVM pGVM, PGVMCPU pGVCpu)
1584{
1585 pGVM->gvmm.s.StatsSched.cWakeUpCalls++;
1586
1587 /*
1588 * Signal the semaphore regardless of whether it's current blocked on it.
1589 *
1590 * The reason for this is that there is absolutely no way we can be 100%
1591 * certain that it isn't *about* go to go to sleep on it and just got
1592 * delayed a bit en route. So, we will always signal the semaphore when
1593 * the it is flagged as halted in the VMM.
1594 */
1595/** @todo we can optimize some of that by means of the pVCpu->enmState now. */
1596 int rc;
1597 if (pGVCpu->gvmm.s.u64HaltExpire)
1598 {
1599 rc = VINF_SUCCESS;
1600 ASMAtomicXchgU64(&pGVCpu->gvmm.s.u64HaltExpire, 0);
1601 }
1602 else
1603 {
1604 rc = VINF_GVM_NOT_BLOCKED;
1605 pGVM->gvmm.s.StatsSched.cWakeUpNotHalted++;
1606 }
1607
1608 int rc2 = RTSemEventMultiSignal(pGVCpu->gvmm.s.HaltEventMulti);
1609 AssertRC(rc2);
1610
1611 return rc;
1612}
1613
1614
1615/**
1616 * Wakes up the halted EMT thread so it can service a pending request.
1617 *
1618 * @returns VBox status code.
1619 * @retval VINF_SUCCESS if successfully woken up.
1620 * @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1621 *
1622 * @param pVM Pointer to the shared VM structure.
1623 * @param idCpu The Virtual CPU ID of the EMT to wake up.
1624 * @param fTakeUsedLock Take the used lock or not
1625 * @thread Any but EMT.
1626 */
1627GVMMR0DECL(int) GVMMR0SchedWakeUpEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
1628{
1629 /*
1630 * Validate input and take the UsedLock.
1631 */
1632 PGVM pGVM;
1633 PGVMM pGVMM;
1634 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
1635 if (RT_SUCCESS(rc))
1636 {
1637 if (idCpu < pGVM->cCpus)
1638 {
1639 /*
1640 * Do the actual job.
1641 */
1642 rc = gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
1643
1644 if (fTakeUsedLock)
1645 {
1646 /*
1647 * While we're here, do a round of scheduling.
1648 */
1649 Assert(ASMGetFlags() & X86_EFL_IF);
1650 const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1651 pGVM->gvmm.s.StatsSched.cWakeUpWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1652 }
1653 }
1654 else
1655 rc = VERR_INVALID_CPU_ID;
1656
1657 if (fTakeUsedLock)
1658 {
1659 int rc2 = gvmmR0UsedUnlock(pGVMM);
1660 AssertRC(rc2);
1661 }
1662 }
1663
1664 LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
1665 return rc;
1666}
1667
1668
1669/**
1670 * Wakes up the halted EMT thread so it can service a pending request.
1671 *
1672 * @returns VBox status code.
1673 * @retval VINF_SUCCESS if successfully woken up.
1674 * @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1675 *
1676 * @param pVM Pointer to the shared VM structure.
1677 * @param idCpu The Virtual CPU ID of the EMT to wake up.
1678 * @thread Any but EMT.
1679 */
1680GVMMR0DECL(int) GVMMR0SchedWakeUp(PVM pVM, VMCPUID idCpu)
1681{
1682 return GVMMR0SchedWakeUpEx(pVM, idCpu, true /* fTakeUsedLock */);
1683}
1684
1685/**
1686 * Worker common to GVMMR0SchedPoke and GVMMR0SchedWakeUpAndPokeCpus that pokes
1687 * the Virtual CPU if it's still busy executing guest code.
1688 *
1689 * @returns VBox status code.
1690 * @retval VINF_SUCCESS if poked successfully.
1691 * @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1692 *
1693 * @param pGVM The global (ring-0) VM structure.
1694 * @param pVCpu The Virtual CPU handle.
1695 */
1696DECLINLINE(int) gvmmR0SchedPokeOne(PGVM pGVM, PVMCPU pVCpu)
1697{
1698 pGVM->gvmm.s.StatsSched.cPokeCalls++;
1699
1700 RTCPUID idHostCpu = pVCpu->idHostCpu;
1701 if ( idHostCpu == NIL_RTCPUID
1702 || VMCPU_GET_STATE(pVCpu) != VMCPUSTATE_STARTED_EXEC)
1703 {
1704 pGVM->gvmm.s.StatsSched.cPokeNotBusy++;
1705 return VINF_GVM_NOT_BUSY_IN_GC;
1706 }
1707
1708 /* Note: this function is not implemented on Darwin and Linux (kernel < 2.6.19) */
1709 RTMpPokeCpu(idHostCpu);
1710 return VINF_SUCCESS;
1711}
1712
1713/**
1714 * Pokes an EMT if it's still busy running guest code.
1715 *
1716 * @returns VBox status code.
1717 * @retval VINF_SUCCESS if poked successfully.
1718 * @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1719 *
1720 * @param pVM Pointer to the shared VM structure.
1721 * @param idCpu The ID of the virtual CPU to poke.
1722 * @param fTakeUsedLock Take the used lock or not
1723 */
1724GVMMR0DECL(int) GVMMR0SchedPokeEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
1725{
1726 /*
1727 * Validate input and take the UsedLock.
1728 */
1729 PGVM pGVM;
1730 PGVMM pGVMM;
1731 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
1732 if (RT_SUCCESS(rc))
1733 {
1734 if (idCpu < pGVM->cCpus)
1735 rc = gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
1736 else
1737 rc = VERR_INVALID_CPU_ID;
1738
1739 if (fTakeUsedLock)
1740 {
1741 int rc2 = gvmmR0UsedUnlock(pGVMM);
1742 AssertRC(rc2);
1743 }
1744 }
1745
1746 LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
1747 return rc;
1748}
1749
1750
1751/**
1752 * Pokes an EMT if it's still busy running guest code.
1753 *
1754 * @returns VBox status code.
1755 * @retval VINF_SUCCESS if poked successfully.
1756 * @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1757 *
1758 * @param pVM Pointer to the shared VM structure.
1759 * @param idCpu The ID of the virtual CPU to poke.
1760 */
1761GVMMR0DECL(int) GVMMR0SchedPoke(PVM pVM, VMCPUID idCpu)
1762{
1763 return GVMMR0SchedPokeEx(pVM, idCpu, true /* fTakeUsedLock */);
1764}
1765
1766
1767/**
1768 * Wakes up a set of halted EMT threads so they can service pending request.
1769 *
1770 * @returns VBox status code, no informational stuff.
1771 *
1772 * @param pVM Pointer to the shared VM structure.
1773 * @param pSleepSet The set of sleepers to wake up.
1774 * @param pPokeSet The set of CPUs to poke.
1775 */
1776GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpus(PVM pVM, PCVMCPUSET pSleepSet, PCVMCPUSET pPokeSet)
1777{
1778 AssertPtrReturn(pSleepSet, VERR_INVALID_POINTER);
1779 AssertPtrReturn(pPokeSet, VERR_INVALID_POINTER);
1780 RTNATIVETHREAD hSelf = RTThreadNativeSelf();
1781
1782 /*
1783 * Validate input and take the UsedLock.
1784 */
1785 PGVM pGVM;
1786 PGVMM pGVMM;
1787 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /* fTakeUsedLock */);
1788 if (RT_SUCCESS(rc))
1789 {
1790 rc = VINF_SUCCESS;
1791 VMCPUID idCpu = pGVM->cCpus;
1792 while (idCpu-- > 0)
1793 {
1794 /* Don't try poke or wake up ourselves. */
1795 if (pGVM->aCpus[idCpu].hEMT == hSelf)
1796 continue;
1797
1798 /* just ignore errors for now. */
1799 if (VMCPUSET_IS_PRESENT(pSleepSet, idCpu))
1800 gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
1801 else if (VMCPUSET_IS_PRESENT(pPokeSet, idCpu))
1802 gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
1803 }
1804
1805 int rc2 = gvmmR0UsedUnlock(pGVMM);
1806 AssertRC(rc2);
1807 }
1808
1809 LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
1810 return rc;
1811}
1812
1813
1814/**
1815 * VMMR0 request wrapper for GVMMR0SchedWakeUpAndPokeCpus.
1816 *
1817 * @returns see GVMMR0SchedWakeUpAndPokeCpus.
1818 * @param pVM Pointer to the shared VM structure.
1819 * @param pReq The request packet.
1820 */
1821GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpusReq(PVM pVM, PGVMMSCHEDWAKEUPANDPOKECPUSREQ pReq)
1822{
1823 /*
1824 * Validate input and pass it on.
1825 */
1826 AssertPtrReturn(pReq, VERR_INVALID_POINTER);
1827 AssertMsgReturn(pReq->Hdr.cbReq == sizeof(*pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(*pReq)), VERR_INVALID_PARAMETER);
1828
1829 return GVMMR0SchedWakeUpAndPokeCpus(pVM, &pReq->SleepSet, &pReq->PokeSet);
1830}
1831
1832
1833
1834/**
1835 * Poll the schedule to see if someone else should get a chance to run.
1836 *
1837 * This is a bit hackish and will not work too well if the machine is
1838 * under heavy load from non-VM processes.
1839 *
1840 * @returns VINF_SUCCESS if not yielded.
1841 * VINF_GVM_YIELDED if an attempt to switch to a different VM task was made.
1842 * @param pVM Pointer to the shared VM structure.
1843 * @param idCpu The Virtual CPU ID of the calling EMT.
1844 * @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
1845 * @param fYield Whether to yield or not.
1846 * This is for when we're spinning in the halt loop.
1847 * @thread EMT(idCpu).
1848 */
1849GVMMR0DECL(int) GVMMR0SchedPoll(PVM pVM, VMCPUID idCpu, bool fYield)
1850{
1851 /*
1852 * Validate input.
1853 */
1854 PGVM pGVM;
1855 PGVMM pGVMM;
1856 int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
1857 if (RT_SUCCESS(rc))
1858 {
1859 rc = gvmmR0UsedLock(pGVMM);
1860 AssertRC(rc);
1861 pGVM->gvmm.s.StatsSched.cPollCalls++;
1862
1863 Assert(ASMGetFlags() & X86_EFL_IF);
1864 const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1865
1866 if (!fYield)
1867 pGVM->gvmm.s.StatsSched.cPollWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1868 else
1869 {
1870 /** @todo implement this... */
1871 rc = VERR_NOT_IMPLEMENTED;
1872 }
1873
1874 gvmmR0UsedUnlock(pGVMM);
1875 }
1876
1877 LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
1878 return rc;
1879}
1880
1881
1882
1883/**
1884 * Retrieves the GVMM statistics visible to the caller.
1885 *
1886 * @returns VBox status code.
1887 *
1888 * @param pStats Where to put the statistics.
1889 * @param pSession The current session.
1890 * @param pVM The VM to obtain statistics for. Optional.
1891 */
1892GVMMR0DECL(int) GVMMR0QueryStatistics(PGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
1893{
1894 LogFlow(("GVMMR0QueryStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
1895
1896 /*
1897 * Validate input.
1898 */
1899 AssertPtrReturn(pSession, VERR_INVALID_POINTER);
1900 AssertPtrReturn(pStats, VERR_INVALID_POINTER);
1901 pStats->cVMs = 0; /* (crash before taking the sem...) */
1902
1903 /*
1904 * Take the lock and get the VM statistics.
1905 */
1906 PGVMM pGVMM;
1907 if (pVM)
1908 {
1909 PGVM pGVM;
1910 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /*fTakeUsedLock*/);
1911 if (RT_FAILURE(rc))
1912 return rc;
1913 pStats->SchedVM = pGVM->gvmm.s.StatsSched;
1914 }
1915 else
1916 {
1917 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
1918 memset(&pStats->SchedVM, 0, sizeof(pStats->SchedVM));
1919
1920 int rc = gvmmR0UsedLock(pGVMM);
1921 AssertRCReturn(rc, rc);
1922 }
1923
1924 /*
1925 * Enumerate the VMs and add the ones visibile to the statistics.
1926 */
1927 pStats->cVMs = 0;
1928 pStats->cEMTs = 0;
1929 memset(&pStats->SchedSum, 0, sizeof(pStats->SchedSum));
1930
1931 for (unsigned i = pGVMM->iUsedHead;
1932 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1933 i = pGVMM->aHandles[i].iNext)
1934 {
1935 PGVM pGVM = pGVMM->aHandles[i].pGVM;
1936 void *pvObj = pGVMM->aHandles[i].pvObj;
1937 if ( VALID_PTR(pvObj)
1938 && VALID_PTR(pGVM)
1939 && pGVM->u32Magic == GVM_MAGIC
1940 && RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
1941 {
1942 pStats->cVMs++;
1943 pStats->cEMTs += pGVM->cCpus;
1944
1945 pStats->SchedSum.cHaltCalls += pGVM->gvmm.s.StatsSched.cHaltCalls;
1946 pStats->SchedSum.cHaltBlocking += pGVM->gvmm.s.StatsSched.cHaltBlocking;
1947 pStats->SchedSum.cHaltTimeouts += pGVM->gvmm.s.StatsSched.cHaltTimeouts;
1948 pStats->SchedSum.cHaltNotBlocking += pGVM->gvmm.s.StatsSched.cHaltNotBlocking;
1949 pStats->SchedSum.cHaltWakeUps += pGVM->gvmm.s.StatsSched.cHaltWakeUps;
1950
1951 pStats->SchedSum.cWakeUpCalls += pGVM->gvmm.s.StatsSched.cWakeUpCalls;
1952 pStats->SchedSum.cWakeUpNotHalted += pGVM->gvmm.s.StatsSched.cWakeUpNotHalted;
1953 pStats->SchedSum.cWakeUpWakeUps += pGVM->gvmm.s.StatsSched.cWakeUpWakeUps;
1954
1955 pStats->SchedSum.cPokeCalls += pGVM->gvmm.s.StatsSched.cPokeCalls;
1956 pStats->SchedSum.cPokeNotBusy += pGVM->gvmm.s.StatsSched.cPokeNotBusy;
1957
1958 pStats->SchedSum.cPollCalls += pGVM->gvmm.s.StatsSched.cPollCalls;
1959 pStats->SchedSum.cPollHalts += pGVM->gvmm.s.StatsSched.cPollHalts;
1960 pStats->SchedSum.cPollWakeUps += pGVM->gvmm.s.StatsSched.cPollWakeUps;
1961 }
1962 }
1963
1964 gvmmR0UsedUnlock(pGVMM);
1965
1966 return VINF_SUCCESS;
1967}
1968
1969
1970/**
1971 * VMMR0 request wrapper for GVMMR0QueryStatistics.
1972 *
1973 * @returns see GVMMR0QueryStatistics.
1974 * @param pVM Pointer to the shared VM structure. Optional.
1975 * @param pReq The request packet.
1976 */
1977GVMMR0DECL(int) GVMMR0QueryStatisticsReq(PVM pVM, PGVMMQUERYSTATISTICSSREQ pReq)
1978{
1979 /*
1980 * Validate input and pass it on.
1981 */
1982 AssertPtrReturn(pReq, VERR_INVALID_POINTER);
1983 AssertMsgReturn(pReq->Hdr.cbReq == sizeof(*pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(*pReq)), VERR_INVALID_PARAMETER);
1984
1985 return GVMMR0QueryStatistics(&pReq->Stats, pReq->pSession, pVM);
1986}
1987
1988
1989/**
1990 * Resets the specified GVMM statistics.
1991 *
1992 * @returns VBox status code.
1993 *
1994 * @param pStats Which statistics to reset, that is, non-zero fields indicates which to reset.
1995 * @param pSession The current session.
1996 * @param pVM The VM to reset statistics for. Optional.
1997 */
1998GVMMR0DECL(int) GVMMR0ResetStatistics(PCGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
1999{
2000 LogFlow(("GVMMR0ResetStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
2001
2002 /*
2003 * Validate input.
2004 */
2005 AssertPtrReturn(pSession, VERR_INVALID_POINTER);
2006 AssertPtrReturn(pStats, VERR_INVALID_POINTER);
2007
2008 /*
2009 * Take the lock and get the VM statistics.
2010 */
2011 PGVMM pGVMM;
2012 if (pVM)
2013 {
2014 PGVM pGVM;
2015 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /*fTakeUsedLock*/);
2016 if (RT_FAILURE(rc))
2017 return rc;
2018# define MAYBE_RESET_FIELD(field) \
2019 do { if (pStats->SchedVM. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2020 MAYBE_RESET_FIELD(cHaltCalls);
2021 MAYBE_RESET_FIELD(cHaltBlocking);
2022 MAYBE_RESET_FIELD(cHaltTimeouts);
2023 MAYBE_RESET_FIELD(cHaltNotBlocking);
2024 MAYBE_RESET_FIELD(cHaltWakeUps);
2025 MAYBE_RESET_FIELD(cWakeUpCalls);
2026 MAYBE_RESET_FIELD(cWakeUpNotHalted);
2027 MAYBE_RESET_FIELD(cWakeUpWakeUps);
2028 MAYBE_RESET_FIELD(cPokeCalls);
2029 MAYBE_RESET_FIELD(cPokeNotBusy);
2030 MAYBE_RESET_FIELD(cPollCalls);
2031 MAYBE_RESET_FIELD(cPollHalts);
2032 MAYBE_RESET_FIELD(cPollWakeUps);
2033# undef MAYBE_RESET_FIELD
2034 }
2035 else
2036 {
2037 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
2038
2039 int rc = gvmmR0UsedLock(pGVMM);
2040 AssertRCReturn(rc, rc);
2041 }
2042
2043 /*
2044 * Enumerate the VMs and add the ones visibile to the statistics.
2045 */
2046 if (ASMMemIsAll8(&pStats->SchedSum, sizeof(pStats->SchedSum), 0))
2047 {
2048 for (unsigned i = pGVMM->iUsedHead;
2049 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
2050 i = pGVMM->aHandles[i].iNext)
2051 {
2052 PGVM pGVM = pGVMM->aHandles[i].pGVM;
2053 void *pvObj = pGVMM->aHandles[i].pvObj;
2054 if ( VALID_PTR(pvObj)
2055 && VALID_PTR(pGVM)
2056 && pGVM->u32Magic == GVM_MAGIC
2057 && RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
2058 {
2059# define MAYBE_RESET_FIELD(field) \
2060 do { if (pStats->SchedSum. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2061 MAYBE_RESET_FIELD(cHaltCalls);
2062 MAYBE_RESET_FIELD(cHaltBlocking);
2063 MAYBE_RESET_FIELD(cHaltTimeouts);
2064 MAYBE_RESET_FIELD(cHaltNotBlocking);
2065 MAYBE_RESET_FIELD(cHaltWakeUps);
2066 MAYBE_RESET_FIELD(cWakeUpCalls);
2067 MAYBE_RESET_FIELD(cWakeUpNotHalted);
2068 MAYBE_RESET_FIELD(cWakeUpWakeUps);
2069 MAYBE_RESET_FIELD(cPokeCalls);
2070 MAYBE_RESET_FIELD(cPokeNotBusy);
2071 MAYBE_RESET_FIELD(cPollCalls);
2072 MAYBE_RESET_FIELD(cPollHalts);
2073 MAYBE_RESET_FIELD(cPollWakeUps);
2074# undef MAYBE_RESET_FIELD
2075 }
2076 }
2077 }
2078
2079 gvmmR0UsedUnlock(pGVMM);
2080
2081 return VINF_SUCCESS;
2082}
2083
2084
2085/**
2086 * VMMR0 request wrapper for GVMMR0ResetStatistics.
2087 *
2088 * @returns see GVMMR0ResetStatistics.
2089 * @param pVM Pointer to the shared VM structure. Optional.
2090 * @param pReq The request packet.
2091 */
2092GVMMR0DECL(int) GVMMR0ResetStatisticsReq(PVM pVM, PGVMMRESETSTATISTICSSREQ pReq)
2093{
2094 /*
2095 * Validate input and pass it on.
2096 */
2097 AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2098 AssertMsgReturn(pReq->Hdr.cbReq == sizeof(*pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(*pReq)), VERR_INVALID_PARAMETER);
2099
2100 return GVMMR0ResetStatistics(&pReq->Stats, pReq->pSession, pVM);
2101}
2102
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