VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR0/GVMMR0.cpp@ 31326

Last change on this file since 31326 was 31326, checked in by vboxsync, 15 years ago

Pass on cpu priority property.
  • Property svn:eol-style set to native
  • Property svn:keywords set to Id Revision
File size: 68.3 KB
Line 
1/* $Id: GVMMR0.cpp 31326 2010-08-03 09:56:22Z 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->uCpuPriority = 100; /* default is maximum priority. */
616 pVM->offVMCPU = RT_UOFFSETOF(VM, aCpus);
617
618 rc = RTR0MemObjAllocPage(&pGVM->gvmm.s.VMPagesMemObj, cPages * sizeof(SUPPAGE), false /* fExecutable */);
619 if (RT_SUCCESS(rc))
620 {
621 PSUPPAGE paPages = (PSUPPAGE)RTR0MemObjAddress(pGVM->gvmm.s.VMPagesMemObj); AssertPtr(paPages);
622 for (uint32_t iPage = 0; iPage < cPages; iPage++)
623 {
624 paPages[iPage].uReserved = 0;
625 paPages[iPage].Phys = RTR0MemObjGetPagePhysAddr(pGVM->gvmm.s.VMMemObj, iPage);
626 Assert(paPages[iPage].Phys != NIL_RTHCPHYS);
627 }
628
629 /*
630 * Map them into ring-3.
631 */
632 rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMMapObj, pGVM->gvmm.s.VMMemObj, (RTR3PTR)-1, 0,
633 RTMEM_PROT_READ | RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
634 if (RT_SUCCESS(rc))
635 {
636 pVM->pVMR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMMapObj);
637 AssertPtr((void *)pVM->pVMR3);
638
639 /* Initialize all the VM pointers. */
640 for (uint32_t i = 0; i < cCpus; i++)
641 {
642 pVM->aCpus[i].pVMR0 = pVM;
643 pVM->aCpus[i].pVMR3 = pVM->pVMR3;
644 pVM->aCpus[i].idHostCpu = NIL_RTCPUID;
645 }
646
647 rc = RTR0MemObjMapUser(&pGVM->gvmm.s.VMPagesMapObj, pGVM->gvmm.s.VMPagesMemObj, (RTR3PTR)-1, 0,
648 RTMEM_PROT_READ | RTMEM_PROT_WRITE, NIL_RTR0PROCESS);
649 if (RT_SUCCESS(rc))
650 {
651 pVM->paVMPagesR3 = RTR0MemObjAddressR3(pGVM->gvmm.s.VMPagesMapObj);
652 AssertPtr((void *)pVM->paVMPagesR3);
653
654 /* complete the handle - take the UsedLock sem just to be careful. */
655 rc = gvmmR0UsedLock(pGVMM);
656 AssertRC(rc);
657
658 pHandle->pVM = pVM;
659 pHandle->pGVM = pGVM;
660 pHandle->hEMT0 = hEMT0;
661 pHandle->ProcId = ProcId;
662 pGVM->pVM = pVM;
663 pGVM->aCpus[0].hEMT = hEMT0;
664 pGVMM->cEMTs += cCpus;
665
666 gvmmR0UsedUnlock(pGVMM);
667 gvmmR0CreateDestroyUnlock(pGVMM);
668
669 *ppVM = pVM;
670 Log(("GVMMR0CreateVM: pVM=%p pVMR3=%p pGVM=%p hGVM=%d\n", pVM, pVM->pVMR3, pGVM, iHandle));
671 return VINF_SUCCESS;
672 }
673
674 RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */);
675 pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
676 }
677 RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */);
678 pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
679 }
680 RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */);
681 pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
682 }
683 }
684 }
685 /* else: The user wasn't permitted to create this VM. */
686
687 /*
688 * The handle will be freed by gvmmR0HandleObjDestructor as we release the
689 * object reference here. A little extra mess because of non-recursive lock.
690 */
691 void *pvObj = pHandle->pvObj;
692 pHandle->pvObj = NULL;
693 gvmmR0CreateDestroyUnlock(pGVMM);
694
695 SUPR0ObjRelease(pvObj, pSession);
696
697 SUPR0Printf("GVMMR0CreateVM: failed, rc=%d\n", rc);
698 return rc;
699 }
700
701 rc = VERR_NO_MEMORY;
702 }
703 else
704 rc = VERR_INTERNAL_ERROR;
705 }
706 else
707 rc = VERR_GVM_TOO_MANY_VMS;
708
709 gvmmR0CreateDestroyUnlock(pGVMM);
710 return rc;
711}
712
713
714/**
715 * Initializes the per VM data belonging to GVMM.
716 *
717 * @param pGVM Pointer to the global VM structure.
718 */
719static void gvmmR0InitPerVMData(PGVM pGVM)
720{
721 AssertCompile(RT_SIZEOFMEMB(GVM,gvmm.s) <= RT_SIZEOFMEMB(GVM,gvmm.padding));
722 AssertCompile(RT_SIZEOFMEMB(GVMCPU,gvmm.s) <= RT_SIZEOFMEMB(GVMCPU,gvmm.padding));
723 pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
724 pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
725 pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
726 pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
727 pGVM->gvmm.s.fDoneVMMR0Init = false;
728 pGVM->gvmm.s.fDoneVMMR0Term = false;
729
730 for (VMCPUID i = 0; i < pGVM->cCpus; i++)
731 {
732 pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
733 pGVM->aCpus[i].hEMT = NIL_RTNATIVETHREAD;
734 }
735}
736
737
738/**
739 * Does the VM initialization.
740 *
741 * @returns VBox status code.
742 * @param pVM Pointer to the shared VM structure.
743 */
744GVMMR0DECL(int) GVMMR0InitVM(PVM pVM)
745{
746 LogFlow(("GVMMR0InitVM: pVM=%p\n", pVM));
747
748 /*
749 * Validate the VM structure, state and handle.
750 */
751 PGVM pGVM;
752 PGVMM pGVMM;
753 int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
754 if (RT_SUCCESS(rc))
755 {
756 if ( !pGVM->gvmm.s.fDoneVMMR0Init
757 && pGVM->aCpus[0].gvmm.s.HaltEventMulti == NIL_RTSEMEVENTMULTI)
758 {
759 for (VMCPUID i = 0; i < pGVM->cCpus; i++)
760 {
761 rc = RTSemEventMultiCreate(&pGVM->aCpus[i].gvmm.s.HaltEventMulti);
762 if (RT_FAILURE(rc))
763 {
764 pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
765 break;
766 }
767 }
768 }
769 else
770 rc = VERR_WRONG_ORDER;
771 }
772
773 LogFlow(("GVMMR0InitVM: returns %Rrc\n", rc));
774 return rc;
775}
776
777
778/**
779 * Indicates that we're done with the ring-0 initialization
780 * of the VM.
781 *
782 * @param pVM Pointer to the shared VM structure.
783 * @thread EMT(0)
784 */
785GVMMR0DECL(void) GVMMR0DoneInitVM(PVM pVM)
786{
787 /* Validate the VM structure, state and handle. */
788 PGVM pGVM;
789 PGVMM pGVMM;
790 int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
791 AssertRCReturnVoid(rc);
792
793 /* Set the indicator. */
794 pGVM->gvmm.s.fDoneVMMR0Init = true;
795}
796
797
798/**
799 * Indicates that we're doing the ring-0 termination of the VM.
800 *
801 * @returns true if termination hasn't been done already, false if it has.
802 * @param pVM Pointer to the shared VM structure.
803 * @param pGVM Pointer to the global VM structure. Optional.
804 * @thread EMT(0)
805 */
806GVMMR0DECL(bool) GVMMR0DoingTermVM(PVM pVM, PGVM pGVM)
807{
808 /* Validate the VM structure, state and handle. */
809 AssertPtrNullReturn(pGVM, false);
810 AssertReturn(!pGVM || pGVM->u32Magic == GVM_MAGIC, false);
811 if (!pGVM)
812 {
813 PGVMM pGVMM;
814 int rc = gvmmR0ByVMAndEMT(pVM, 0 /* idCpu */, &pGVM, &pGVMM);
815 AssertRCReturn(rc, false);
816 }
817
818 /* Set the indicator. */
819 if (pGVM->gvmm.s.fDoneVMMR0Term)
820 return false;
821 pGVM->gvmm.s.fDoneVMMR0Term = true;
822 return true;
823}
824
825
826/**
827 * Destroys the VM, freeing all associated resources (the ring-0 ones anyway).
828 *
829 * This is call from the vmR3DestroyFinalBit and from a error path in VMR3Create,
830 * and the caller is not the EMT thread, unfortunately. For security reasons, it
831 * would've been nice if the caller was actually the EMT thread or that we somehow
832 * could've associated the calling thread with the VM up front.
833 *
834 * @returns VBox status code.
835 * @param pVM Where to store the pointer to the VM structure.
836 *
837 * @thread EMT(0) if it's associated with the VM, otherwise any thread.
838 */
839GVMMR0DECL(int) GVMMR0DestroyVM(PVM pVM)
840{
841 LogFlow(("GVMMR0DestroyVM: pVM=%p\n", pVM));
842 PGVMM pGVMM;
843 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
844
845
846 /*
847 * Validate the VM structure, state and caller.
848 */
849 AssertPtrReturn(pVM, VERR_INVALID_POINTER);
850 AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
851 AssertMsgReturn(pVM->enmVMState >= VMSTATE_CREATING && pVM->enmVMState <= VMSTATE_TERMINATED, ("%d\n", pVM->enmVMState), VERR_WRONG_ORDER);
852
853 uint32_t hGVM = pVM->hSelf;
854 AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
855 AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
856
857 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
858 AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
859
860 RTPROCESS ProcId = RTProcSelf();
861 RTNATIVETHREAD hSelf = RTThreadNativeSelf();
862 AssertReturn( ( pHandle->hEMT0 == hSelf
863 && pHandle->ProcId == ProcId)
864 || pHandle->hEMT0 == NIL_RTNATIVETHREAD, VERR_NOT_OWNER);
865
866 /*
867 * Lookup the handle and destroy the object.
868 * Since the lock isn't recursive and we'll have to leave it before dereferencing the
869 * object, we take some precautions against racing callers just in case...
870 */
871 int rc = gvmmR0CreateDestroyLock(pGVMM);
872 AssertRC(rc);
873
874 /* be careful here because we might theoretically be racing someone else cleaning up. */
875 if ( pHandle->pVM == pVM
876 && ( ( pHandle->hEMT0 == hSelf
877 && pHandle->ProcId == ProcId)
878 || pHandle->hEMT0 == NIL_RTNATIVETHREAD)
879 && VALID_PTR(pHandle->pvObj)
880 && VALID_PTR(pHandle->pSession)
881 && VALID_PTR(pHandle->pGVM)
882 && pHandle->pGVM->u32Magic == GVM_MAGIC)
883 {
884 void *pvObj = pHandle->pvObj;
885 pHandle->pvObj = NULL;
886 gvmmR0CreateDestroyUnlock(pGVMM);
887
888 SUPR0ObjRelease(pvObj, pHandle->pSession);
889 }
890 else
891 {
892 SUPR0Printf("GVMMR0DestroyVM: pHandle=%p:{.pVM=%p, .hEMT0=%p, .ProcId=%u, .pvObj=%p} pVM=%p hSelf=%p\n",
893 pHandle, pHandle->pVM, pHandle->hEMT0, pHandle->ProcId, pHandle->pvObj, pVM, hSelf);
894 gvmmR0CreateDestroyUnlock(pGVMM);
895 rc = VERR_INTERNAL_ERROR;
896 }
897
898 return rc;
899}
900
901
902/**
903 * Performs VM cleanup task as part of object destruction.
904 *
905 * @param pGVM The GVM pointer.
906 */
907static void gvmmR0CleanupVM(PGVM pGVM)
908{
909 if ( pGVM->gvmm.s.fDoneVMMR0Init
910 && !pGVM->gvmm.s.fDoneVMMR0Term)
911 {
912 if ( pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ
913 && RTR0MemObjAddress(pGVM->gvmm.s.VMMemObj) == pGVM->pVM)
914 {
915 LogFlow(("gvmmR0CleanupVM: Calling VMMR0TermVM\n"));
916 VMMR0TermVM(pGVM->pVM, pGVM);
917 }
918 else
919 AssertMsgFailed(("gvmmR0CleanupVM: VMMemObj=%p pVM=%p\n", pGVM->gvmm.s.VMMemObj, pGVM->pVM));
920 }
921
922 GMMR0CleanupVM(pGVM);
923}
924
925
926/**
927 * Handle destructor.
928 *
929 * @param pvGVMM The GVM instance pointer.
930 * @param pvHandle The handle pointer.
931 */
932static DECLCALLBACK(void) gvmmR0HandleObjDestructor(void *pvObj, void *pvGVMM, void *pvHandle)
933{
934 LogFlow(("gvmmR0HandleObjDestructor: %p %p %p\n", pvObj, pvGVMM, pvHandle));
935
936 /*
937 * Some quick, paranoid, input validation.
938 */
939 PGVMHANDLE pHandle = (PGVMHANDLE)pvHandle;
940 AssertPtr(pHandle);
941 PGVMM pGVMM = (PGVMM)pvGVMM;
942 Assert(pGVMM == g_pGVMM);
943 const uint16_t iHandle = pHandle - &pGVMM->aHandles[0];
944 if ( !iHandle
945 || iHandle >= RT_ELEMENTS(pGVMM->aHandles)
946 || iHandle != pHandle->iSelf)
947 {
948 SUPR0Printf("GVM: handle %d is out of range or corrupt (iSelf=%d)!\n", iHandle, pHandle->iSelf);
949 return;
950 }
951
952 int rc = gvmmR0CreateDestroyLock(pGVMM);
953 AssertRC(rc);
954 rc = gvmmR0UsedLock(pGVMM);
955 AssertRC(rc);
956
957 /*
958 * This is a tad slow but a doubly linked list is too much hazzle.
959 */
960 if (RT_UNLIKELY(pHandle->iNext >= RT_ELEMENTS(pGVMM->aHandles)))
961 {
962 SUPR0Printf("GVM: used list index %d is out of range!\n", pHandle->iNext);
963 gvmmR0UsedUnlock(pGVMM);
964 gvmmR0CreateDestroyUnlock(pGVMM);
965 return;
966 }
967
968 if (pGVMM->iUsedHead == iHandle)
969 pGVMM->iUsedHead = pHandle->iNext;
970 else
971 {
972 uint16_t iPrev = pGVMM->iUsedHead;
973 int c = RT_ELEMENTS(pGVMM->aHandles) + 2;
974 while (iPrev)
975 {
976 if (RT_UNLIKELY(iPrev >= RT_ELEMENTS(pGVMM->aHandles)))
977 {
978 SUPR0Printf("GVM: used list index %d is out of range!\n", iPrev);
979 gvmmR0UsedUnlock(pGVMM);
980 gvmmR0CreateDestroyUnlock(pGVMM);
981 return;
982 }
983 if (RT_UNLIKELY(c-- <= 0))
984 {
985 iPrev = 0;
986 break;
987 }
988
989 if (pGVMM->aHandles[iPrev].iNext == iHandle)
990 break;
991 iPrev = pGVMM->aHandles[iPrev].iNext;
992 }
993 if (!iPrev)
994 {
995 SUPR0Printf("GVM: can't find the handle previous previous of %d!\n", pHandle->iSelf);
996 gvmmR0UsedUnlock(pGVMM);
997 gvmmR0CreateDestroyUnlock(pGVMM);
998 return;
999 }
1000
1001 Assert(pGVMM->aHandles[iPrev].iNext == iHandle);
1002 pGVMM->aHandles[iPrev].iNext = pHandle->iNext;
1003 }
1004 pHandle->iNext = 0;
1005 pGVMM->cVMs--;
1006
1007 /*
1008 * Do the global cleanup round.
1009 */
1010 PGVM pGVM = pHandle->pGVM;
1011 if ( VALID_PTR(pGVM)
1012 && pGVM->u32Magic == GVM_MAGIC)
1013 {
1014 pGVMM->cEMTs -= pGVM->cCpus;
1015 gvmmR0UsedUnlock(pGVMM);
1016
1017 gvmmR0CleanupVM(pGVM);
1018
1019 /*
1020 * Do the GVMM cleanup - must be done last.
1021 */
1022 /* The VM and VM pages mappings/allocations. */
1023 if (pGVM->gvmm.s.VMPagesMapObj != NIL_RTR0MEMOBJ)
1024 {
1025 rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMapObj, false /* fFreeMappings */); AssertRC(rc);
1026 pGVM->gvmm.s.VMPagesMapObj = NIL_RTR0MEMOBJ;
1027 }
1028
1029 if (pGVM->gvmm.s.VMMapObj != NIL_RTR0MEMOBJ)
1030 {
1031 rc = RTR0MemObjFree(pGVM->gvmm.s.VMMapObj, false /* fFreeMappings */); AssertRC(rc);
1032 pGVM->gvmm.s.VMMapObj = NIL_RTR0MEMOBJ;
1033 }
1034
1035 if (pGVM->gvmm.s.VMPagesMemObj != NIL_RTR0MEMOBJ)
1036 {
1037 rc = RTR0MemObjFree(pGVM->gvmm.s.VMPagesMemObj, false /* fFreeMappings */); AssertRC(rc);
1038 pGVM->gvmm.s.VMPagesMemObj = NIL_RTR0MEMOBJ;
1039 }
1040
1041 if (pGVM->gvmm.s.VMMemObj != NIL_RTR0MEMOBJ)
1042 {
1043 rc = RTR0MemObjFree(pGVM->gvmm.s.VMMemObj, false /* fFreeMappings */); AssertRC(rc);
1044 pGVM->gvmm.s.VMMemObj = NIL_RTR0MEMOBJ;
1045 }
1046
1047 for (VMCPUID i = 0; i < pGVM->cCpus; i++)
1048 {
1049 if (pGVM->aCpus[i].gvmm.s.HaltEventMulti != NIL_RTSEMEVENTMULTI)
1050 {
1051 rc = RTSemEventMultiDestroy(pGVM->aCpus[i].gvmm.s.HaltEventMulti); AssertRC(rc);
1052 pGVM->aCpus[i].gvmm.s.HaltEventMulti = NIL_RTSEMEVENTMULTI;
1053 }
1054 }
1055
1056 /* the GVM structure itself. */
1057 pGVM->u32Magic |= UINT32_C(0x80000000);
1058 RTMemFree(pGVM);
1059
1060 /* Re-acquire the UsedLock before freeing the handle since we're updating handle fields. */
1061 rc = gvmmR0UsedLock(pGVMM);
1062 AssertRC(rc);
1063 }
1064 /* else: GVMMR0CreateVM cleanup. */
1065
1066 /*
1067 * Free the handle.
1068 */
1069 pHandle->iNext = pGVMM->iFreeHead;
1070 pGVMM->iFreeHead = iHandle;
1071 ASMAtomicWriteNullPtr(&pHandle->pGVM);
1072 ASMAtomicWriteNullPtr(&pHandle->pVM);
1073 ASMAtomicWriteNullPtr(&pHandle->pvObj);
1074 ASMAtomicWriteNullPtr(&pHandle->pSession);
1075 ASMAtomicWriteSize(&pHandle->hEMT0, NIL_RTNATIVETHREAD);
1076 ASMAtomicWriteSize(&pHandle->ProcId, NIL_RTPROCESS);
1077
1078 gvmmR0UsedUnlock(pGVMM);
1079 gvmmR0CreateDestroyUnlock(pGVMM);
1080 LogFlow(("gvmmR0HandleObjDestructor: returns\n"));
1081}
1082
1083
1084/**
1085 * Registers the calling thread as the EMT of a Virtual CPU.
1086 *
1087 * Note that VCPU 0 is automatically registered during VM creation.
1088 *
1089 * @returns VBox status code
1090 * @param pVM The shared VM structure (the ring-0 mapping).
1091 * @param idCpu VCPU id.
1092 */
1093GVMMR0DECL(int) GVMMR0RegisterVCpu(PVM pVM, VMCPUID idCpu)
1094{
1095 AssertReturn(idCpu != 0, VERR_NOT_OWNER);
1096
1097 /*
1098 * Validate the VM structure, state and handle.
1099 */
1100 PGVM pGVM;
1101 PGVMM pGVMM;
1102 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, false /* fTakeUsedLock */);
1103 if (RT_FAILURE(rc))
1104 return rc;
1105
1106 AssertReturn(idCpu < pVM->cCpus, VERR_INVALID_CPU_ID);
1107 AssertReturn(pGVM->aCpus[idCpu].hEMT == NIL_RTNATIVETHREAD, VERR_ACCESS_DENIED);
1108
1109 pGVM->aCpus[idCpu].hEMT = RTThreadNativeSelf();
1110 return VINF_SUCCESS;
1111}
1112
1113
1114/**
1115 * Lookup a GVM structure by its handle.
1116 *
1117 * @returns The GVM pointer on success, NULL on failure.
1118 * @param hGVM The global VM handle. Asserts on bad handle.
1119 */
1120GVMMR0DECL(PGVM) GVMMR0ByHandle(uint32_t hGVM)
1121{
1122 PGVMM pGVMM;
1123 GVMM_GET_VALID_INSTANCE(pGVMM, NULL);
1124
1125 /*
1126 * Validate.
1127 */
1128 AssertReturn(hGVM != NIL_GVM_HANDLE, NULL);
1129 AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), NULL);
1130
1131 /*
1132 * Look it up.
1133 */
1134 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1135 AssertPtrReturn(pHandle->pVM, NULL);
1136 AssertPtrReturn(pHandle->pvObj, NULL);
1137 PGVM pGVM = pHandle->pGVM;
1138 AssertPtrReturn(pGVM, NULL);
1139 AssertReturn(pGVM->pVM == pHandle->pVM, NULL);
1140
1141 return pHandle->pGVM;
1142}
1143
1144
1145/**
1146 * Lookup a GVM structure by the shared VM structure.
1147 *
1148 * The calling thread must be in the same process as the VM. All current lookups
1149 * are by threads inside the same process, so this will not be an issue.
1150 *
1151 * @returns VBox status code.
1152 * @param pVM The shared VM structure (the ring-0 mapping).
1153 * @param ppGVM Where to store the GVM pointer.
1154 * @param ppGVMM Where to store the pointer to the GVMM instance data.
1155 * @param fTakeUsedLock Whether to take the used lock or not.
1156 * Be very careful if not taking the lock as it's possible that
1157 * the VM will disappear then.
1158 *
1159 * @remark This will not assert on an invalid pVM but try return sliently.
1160 */
1161static int gvmmR0ByVM(PVM pVM, PGVM *ppGVM, PGVMM *ppGVMM, bool fTakeUsedLock)
1162{
1163 RTPROCESS ProcId = RTProcSelf();
1164 PGVMM pGVMM;
1165 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
1166
1167 /*
1168 * Validate.
1169 */
1170 if (RT_UNLIKELY( !VALID_PTR(pVM)
1171 || ((uintptr_t)pVM & PAGE_OFFSET_MASK)))
1172 return VERR_INVALID_POINTER;
1173 if (RT_UNLIKELY( pVM->enmVMState < VMSTATE_CREATING
1174 || pVM->enmVMState >= VMSTATE_TERMINATED))
1175 return VERR_INVALID_POINTER;
1176
1177 uint16_t hGVM = pVM->hSelf;
1178 if (RT_UNLIKELY( hGVM == NIL_GVM_HANDLE
1179 || hGVM >= RT_ELEMENTS(pGVMM->aHandles)))
1180 return VERR_INVALID_HANDLE;
1181
1182 /*
1183 * Look it up.
1184 */
1185 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1186 PGVM pGVM;
1187 if (fTakeUsedLock)
1188 {
1189 int rc = gvmmR0UsedLock(pGVMM);
1190 AssertRCReturn(rc, rc);
1191
1192 pGVM = pHandle->pGVM;
1193 if (RT_UNLIKELY( pHandle->pVM != pVM
1194 || pHandle->ProcId != ProcId
1195 || !VALID_PTR(pHandle->pvObj)
1196 || !VALID_PTR(pGVM)
1197 || pGVM->pVM != pVM))
1198 {
1199 gvmmR0UsedUnlock(pGVMM);
1200 return VERR_INVALID_HANDLE;
1201 }
1202 }
1203 else
1204 {
1205 if (RT_UNLIKELY(pHandle->pVM != pVM))
1206 return VERR_INVALID_HANDLE;
1207 if (RT_UNLIKELY(pHandle->ProcId != ProcId))
1208 return VERR_INVALID_HANDLE;
1209 if (RT_UNLIKELY(!VALID_PTR(pHandle->pvObj)))
1210 return VERR_INVALID_HANDLE;
1211
1212 pGVM = pHandle->pGVM;
1213 if (RT_UNLIKELY(!VALID_PTR(pGVM)))
1214 return VERR_INVALID_HANDLE;
1215 if (RT_UNLIKELY(pGVM->pVM != pVM))
1216 return VERR_INVALID_HANDLE;
1217 }
1218
1219 *ppGVM = pGVM;
1220 *ppGVMM = pGVMM;
1221 return VINF_SUCCESS;
1222}
1223
1224
1225/**
1226 * Lookup a GVM structure by the shared VM structure.
1227 *
1228 * @returns VBox status code.
1229 * @param pVM The shared VM structure (the ring-0 mapping).
1230 * @param ppGVM Where to store the GVM pointer.
1231 *
1232 * @remark This will not take the 'used'-lock because it doesn't do
1233 * nesting and this function will be used from under the lock.
1234 */
1235GVMMR0DECL(int) GVMMR0ByVM(PVM pVM, PGVM *ppGVM)
1236{
1237 PGVMM pGVMM;
1238 return gvmmR0ByVM(pVM, ppGVM, &pGVMM, false /* fTakeUsedLock */);
1239}
1240
1241
1242/**
1243 * Lookup a GVM structure by the shared VM structure and ensuring that the
1244 * caller is an EMT thread.
1245 *
1246 * @returns VBox status code.
1247 * @param pVM The shared VM structure (the ring-0 mapping).
1248 * @param idCpu The Virtual CPU ID of the calling EMT.
1249 * @param ppGVM Where to store the GVM pointer.
1250 * @param ppGVMM Where to store the pointer to the GVMM instance data.
1251 * @thread EMT
1252 *
1253 * @remark This will assert in all failure paths.
1254 */
1255static int gvmmR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM, PGVMM *ppGVMM)
1256{
1257 PGVMM pGVMM;
1258 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
1259
1260 /*
1261 * Validate.
1262 */
1263 AssertPtrReturn(pVM, VERR_INVALID_POINTER);
1264 AssertReturn(!((uintptr_t)pVM & PAGE_OFFSET_MASK), VERR_INVALID_POINTER);
1265
1266 uint16_t hGVM = pVM->hSelf;
1267 AssertReturn(hGVM != NIL_GVM_HANDLE, VERR_INVALID_HANDLE);
1268 AssertReturn(hGVM < RT_ELEMENTS(pGVMM->aHandles), VERR_INVALID_HANDLE);
1269
1270 /*
1271 * Look it up.
1272 */
1273 PGVMHANDLE pHandle = &pGVMM->aHandles[hGVM];
1274 AssertReturn(pHandle->pVM == pVM, VERR_NOT_OWNER);
1275 RTPROCESS ProcId = RTProcSelf();
1276 AssertReturn(pHandle->ProcId == ProcId, VERR_NOT_OWNER);
1277 AssertPtrReturn(pHandle->pvObj, VERR_INTERNAL_ERROR);
1278
1279 PGVM pGVM = pHandle->pGVM;
1280 AssertPtrReturn(pGVM, VERR_INTERNAL_ERROR);
1281 AssertReturn(pGVM->pVM == pVM, VERR_INTERNAL_ERROR);
1282 RTNATIVETHREAD hAllegedEMT = RTThreadNativeSelf();
1283 AssertReturn(idCpu < pGVM->cCpus, VERR_INVALID_CPU_ID);
1284 AssertReturn(pGVM->aCpus[idCpu].hEMT == hAllegedEMT, VERR_INTERNAL_ERROR);
1285
1286 *ppGVM = pGVM;
1287 *ppGVMM = pGVMM;
1288 return VINF_SUCCESS;
1289}
1290
1291
1292/**
1293 * Lookup a GVM structure by the shared VM structure
1294 * and ensuring that the caller is the EMT thread.
1295 *
1296 * @returns VBox status code.
1297 * @param pVM The shared VM structure (the ring-0 mapping).
1298 * @param idCpu The Virtual CPU ID of the calling EMT.
1299 * @param ppGVM Where to store the GVM pointer.
1300 * @thread EMT
1301 */
1302GVMMR0DECL(int) GVMMR0ByVMAndEMT(PVM pVM, VMCPUID idCpu, PGVM *ppGVM)
1303{
1304 AssertPtrReturn(ppGVM, VERR_INVALID_POINTER);
1305 PGVMM pGVMM;
1306 return gvmmR0ByVMAndEMT(pVM, idCpu, ppGVM, &pGVMM);
1307}
1308
1309
1310/**
1311 * Lookup a VM by its global handle.
1312 *
1313 * @returns The VM handle on success, NULL on failure.
1314 * @param hGVM The global VM handle. Asserts on bad handle.
1315 */
1316GVMMR0DECL(PVM) GVMMR0GetVMByHandle(uint32_t hGVM)
1317{
1318 PGVM pGVM = GVMMR0ByHandle(hGVM);
1319 return pGVM ? pGVM->pVM : NULL;
1320}
1321
1322
1323/**
1324 * Looks up the VM belonging to the specified EMT thread.
1325 *
1326 * This is used by the assertion machinery in VMMR0.cpp to avoid causing
1327 * unnecessary kernel panics when the EMT thread hits an assertion. The
1328 * call may or not be an EMT thread.
1329 *
1330 * @returns The VM handle on success, NULL on failure.
1331 * @param hEMT The native thread handle of the EMT.
1332 * NIL_RTNATIVETHREAD means the current thread
1333 */
1334GVMMR0DECL(PVM) GVMMR0GetVMByEMT(RTNATIVETHREAD hEMT)
1335{
1336 /*
1337 * No Assertions here as we're usually called in a AssertMsgN or
1338 * RTAssert* context.
1339 */
1340 PGVMM pGVMM = g_pGVMM;
1341 if ( !VALID_PTR(pGVMM)
1342 || pGVMM->u32Magic != GVMM_MAGIC)
1343 return NULL;
1344
1345 if (hEMT == NIL_RTNATIVETHREAD)
1346 hEMT = RTThreadNativeSelf();
1347 RTPROCESS ProcId = RTProcSelf();
1348
1349 /*
1350 * Search the handles in a linear fashion as we don't dare to take the lock (assert).
1351 */
1352 for (unsigned i = 1; i < RT_ELEMENTS(pGVMM->aHandles); i++)
1353 {
1354 if ( pGVMM->aHandles[i].iSelf == i
1355 && pGVMM->aHandles[i].ProcId == ProcId
1356 && VALID_PTR(pGVMM->aHandles[i].pvObj)
1357 && VALID_PTR(pGVMM->aHandles[i].pVM)
1358 && VALID_PTR(pGVMM->aHandles[i].pGVM))
1359 {
1360 if (pGVMM->aHandles[i].hEMT0 == hEMT)
1361 return pGVMM->aHandles[i].pVM;
1362
1363 /* This is fearly safe with the current process per VM approach. */
1364 PGVM pGVM = pGVMM->aHandles[i].pGVM;
1365 VMCPUID const cCpus = pGVM->cCpus;
1366 if ( cCpus < 1
1367 || cCpus > VMM_MAX_CPU_COUNT)
1368 continue;
1369 for (VMCPUID idCpu = 1; idCpu < cCpus; idCpu++)
1370 if (pGVM->aCpus[idCpu].hEMT == hEMT)
1371 return pGVMM->aHandles[i].pVM;
1372 }
1373 }
1374 return NULL;
1375}
1376
1377
1378/**
1379 * This is will wake up expired and soon-to-be expired VMs.
1380 *
1381 * @returns Number of VMs that has been woken up.
1382 * @param pGVMM Pointer to the GVMM instance data.
1383 * @param u64Now The current time.
1384 */
1385static unsigned gvmmR0SchedDoWakeUps(PGVMM pGVMM, uint64_t u64Now)
1386{
1387/** @todo Rewrite this algorithm. See performance defect XYZ. */
1388
1389 /*
1390 * The first pass will wake up VMs which have actually expired
1391 * and look for VMs that should be woken up in the 2nd and 3rd passes.
1392 */
1393 unsigned cWoken = 0;
1394 unsigned cHalted = 0;
1395 unsigned cTodo2nd = 0;
1396 unsigned cTodo3rd = 0;
1397 for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1398 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1399 i = pGVMM->aHandles[i].iNext)
1400 {
1401 PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1402 if ( VALID_PTR(pCurGVM)
1403 && pCurGVM->u32Magic == GVM_MAGIC)
1404 {
1405 for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1406 {
1407 PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1408
1409 uint64_t u64 = pCurGVCpu->gvmm.s.u64HaltExpire;
1410 if (u64)
1411 {
1412 if (u64 <= u64Now)
1413 {
1414 if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1415 {
1416 int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1417 AssertRC(rc);
1418 cWoken++;
1419 }
1420 }
1421 else
1422 {
1423 cHalted++;
1424 if (u64 <= u64Now + pGVMM->nsEarlyWakeUp1)
1425 cTodo2nd++;
1426 else if (u64 <= u64Now + pGVMM->nsEarlyWakeUp2)
1427 cTodo3rd++;
1428 }
1429 }
1430 }
1431 }
1432 AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1433 }
1434
1435 if (cTodo2nd)
1436 {
1437 for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1438 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1439 i = pGVMM->aHandles[i].iNext)
1440 {
1441 PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1442 if ( VALID_PTR(pCurGVM)
1443 && pCurGVM->u32Magic == GVM_MAGIC)
1444 {
1445 for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1446 {
1447 PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1448
1449 if ( pCurGVCpu->gvmm.s.u64HaltExpire
1450 && pCurGVCpu->gvmm.s.u64HaltExpire <= u64Now + pGVMM->nsEarlyWakeUp1)
1451 {
1452 if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1453 {
1454 int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1455 AssertRC(rc);
1456 cWoken++;
1457 }
1458 }
1459 }
1460 }
1461 AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1462 }
1463 }
1464
1465 if (cTodo3rd)
1466 {
1467 for (unsigned i = pGVMM->iUsedHead, cGuard = 0;
1468 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1469 i = pGVMM->aHandles[i].iNext)
1470 {
1471 PGVM pCurGVM = pGVMM->aHandles[i].pGVM;
1472 if ( VALID_PTR(pCurGVM)
1473 && pCurGVM->u32Magic == GVM_MAGIC)
1474 {
1475 for (VMCPUID idCpu = 0; idCpu < pCurGVM->cCpus; idCpu++)
1476 {
1477 PGVMCPU pCurGVCpu = &pCurGVM->aCpus[idCpu];
1478
1479 if ( pCurGVCpu->gvmm.s.u64HaltExpire
1480 && pCurGVCpu->gvmm.s.u64HaltExpire <= u64Now + pGVMM->nsEarlyWakeUp2)
1481 {
1482 if (ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0))
1483 {
1484 int rc = RTSemEventMultiSignal(pCurGVCpu->gvmm.s.HaltEventMulti);
1485 AssertRC(rc);
1486 cWoken++;
1487 }
1488 }
1489 }
1490 }
1491 AssertLogRelBreak(cGuard++ < RT_ELEMENTS(pGVMM->aHandles));
1492 }
1493 }
1494
1495 return cWoken;
1496}
1497
1498
1499/**
1500 * Halt the EMT thread.
1501 *
1502 * @returns VINF_SUCCESS normal wakeup (timeout or kicked by other thread).
1503 * VERR_INTERRUPTED if a signal was scheduled for the thread.
1504 * @param pVM Pointer to the shared VM structure.
1505 * @param idCpu The Virtual CPU ID of the calling EMT.
1506 * @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
1507 * @thread EMT(idCpu).
1508 */
1509GVMMR0DECL(int) GVMMR0SchedHalt(PVM pVM, VMCPUID idCpu, uint64_t u64ExpireGipTime)
1510{
1511 LogFlow(("GVMMR0SchedHalt: pVM=%p\n", pVM));
1512
1513 /*
1514 * Validate the VM structure, state and handle.
1515 */
1516 PGVM pGVM;
1517 PGVMM pGVMM;
1518 int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
1519 if (RT_FAILURE(rc))
1520 return rc;
1521 pGVM->gvmm.s.StatsSched.cHaltCalls++;
1522
1523 PGVMCPU pCurGVCpu = &pGVM->aCpus[idCpu];
1524 Assert(!pCurGVCpu->gvmm.s.u64HaltExpire);
1525
1526 /*
1527 * Take the UsedList semaphore, get the current time
1528 * and check if anyone needs waking up.
1529 * Interrupts must NOT be disabled at this point because we ask for GIP time!
1530 */
1531 rc = gvmmR0UsedLock(pGVMM);
1532 AssertRC(rc);
1533
1534 pCurGVCpu->gvmm.s.iCpuEmt = ASMGetApicId();
1535
1536 Assert(ASMGetFlags() & X86_EFL_IF);
1537 const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1538 pGVM->gvmm.s.StatsSched.cHaltWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1539
1540 /*
1541 * Go to sleep if we must...
1542 */
1543 if ( u64Now < u64ExpireGipTime
1544 && u64ExpireGipTime - u64Now > (pGVMM->cEMTs > pGVMM->cEMTsMeansCompany
1545 ? pGVMM->nsMinSleepCompany
1546 : pGVMM->nsMinSleepAlone))
1547 {
1548 pGVM->gvmm.s.StatsSched.cHaltBlocking++;
1549 ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, u64ExpireGipTime);
1550 gvmmR0UsedUnlock(pGVMM);
1551
1552 uint32_t cMillies = (u64ExpireGipTime - u64Now) / 1000000;
1553 rc = RTSemEventMultiWaitNoResume(pCurGVCpu->gvmm.s.HaltEventMulti, cMillies ? cMillies : 1);
1554 ASMAtomicXchgU64(&pCurGVCpu->gvmm.s.u64HaltExpire, 0);
1555 if (rc == VERR_TIMEOUT)
1556 {
1557 pGVM->gvmm.s.StatsSched.cHaltTimeouts++;
1558 rc = VINF_SUCCESS;
1559 }
1560 }
1561 else
1562 {
1563 pGVM->gvmm.s.StatsSched.cHaltNotBlocking++;
1564 gvmmR0UsedUnlock(pGVMM);
1565 }
1566
1567 /* Make sure false wake up calls (gvmmR0SchedDoWakeUps) cause us to spin. */
1568 RTSemEventMultiReset(pCurGVCpu->gvmm.s.HaltEventMulti);
1569
1570 return rc;
1571}
1572
1573
1574/**
1575 * Worker for GVMMR0SchedWakeUp and GVMMR0SchedWakeUpAndPokeCpus that wakes up
1576 * the a sleeping EMT.
1577 *
1578 * @retval VINF_SUCCESS if successfully woken up.
1579 * @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1580 *
1581 * @param pGVM The global (ring-0) VM structure.
1582 * @param pGVCpu The global (ring-0) VCPU structure.
1583 */
1584DECLINLINE(int) gvmmR0SchedWakeUpOne(PGVM pGVM, PGVMCPU pGVCpu)
1585{
1586 pGVM->gvmm.s.StatsSched.cWakeUpCalls++;
1587
1588 /*
1589 * Signal the semaphore regardless of whether it's current blocked on it.
1590 *
1591 * The reason for this is that there is absolutely no way we can be 100%
1592 * certain that it isn't *about* go to go to sleep on it and just got
1593 * delayed a bit en route. So, we will always signal the semaphore when
1594 * the it is flagged as halted in the VMM.
1595 */
1596/** @todo we can optimize some of that by means of the pVCpu->enmState now. */
1597 int rc;
1598 if (pGVCpu->gvmm.s.u64HaltExpire)
1599 {
1600 rc = VINF_SUCCESS;
1601 ASMAtomicXchgU64(&pGVCpu->gvmm.s.u64HaltExpire, 0);
1602 }
1603 else
1604 {
1605 rc = VINF_GVM_NOT_BLOCKED;
1606 pGVM->gvmm.s.StatsSched.cWakeUpNotHalted++;
1607 }
1608
1609 int rc2 = RTSemEventMultiSignal(pGVCpu->gvmm.s.HaltEventMulti);
1610 AssertRC(rc2);
1611
1612 return rc;
1613}
1614
1615
1616/**
1617 * Wakes up the halted EMT thread so it can service a pending request.
1618 *
1619 * @returns VBox status code.
1620 * @retval VINF_SUCCESS if successfully woken up.
1621 * @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1622 *
1623 * @param pVM Pointer to the shared VM structure.
1624 * @param idCpu The Virtual CPU ID of the EMT to wake up.
1625 * @param fTakeUsedLock Take the used lock or not
1626 * @thread Any but EMT.
1627 */
1628GVMMR0DECL(int) GVMMR0SchedWakeUpEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
1629{
1630 /*
1631 * Validate input and take the UsedLock.
1632 */
1633 PGVM pGVM;
1634 PGVMM pGVMM;
1635 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
1636 if (RT_SUCCESS(rc))
1637 {
1638 if (idCpu < pGVM->cCpus)
1639 {
1640 /*
1641 * Do the actual job.
1642 */
1643 rc = gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
1644
1645 if (fTakeUsedLock)
1646 {
1647 /*
1648 * While we're here, do a round of scheduling.
1649 */
1650 Assert(ASMGetFlags() & X86_EFL_IF);
1651 const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1652 pGVM->gvmm.s.StatsSched.cWakeUpWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1653 }
1654 }
1655 else
1656 rc = VERR_INVALID_CPU_ID;
1657
1658 if (fTakeUsedLock)
1659 {
1660 int rc2 = gvmmR0UsedUnlock(pGVMM);
1661 AssertRC(rc2);
1662 }
1663 }
1664
1665 LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
1666 return rc;
1667}
1668
1669
1670/**
1671 * Wakes up the halted EMT thread so it can service a pending request.
1672 *
1673 * @returns VBox status code.
1674 * @retval VINF_SUCCESS if successfully woken up.
1675 * @retval VINF_GVM_NOT_BLOCKED if the EMT wasn't blocked.
1676 *
1677 * @param pVM Pointer to the shared VM structure.
1678 * @param idCpu The Virtual CPU ID of the EMT to wake up.
1679 * @thread Any but EMT.
1680 */
1681GVMMR0DECL(int) GVMMR0SchedWakeUp(PVM pVM, VMCPUID idCpu)
1682{
1683 return GVMMR0SchedWakeUpEx(pVM, idCpu, true /* fTakeUsedLock */);
1684}
1685
1686/**
1687 * Worker common to GVMMR0SchedPoke and GVMMR0SchedWakeUpAndPokeCpus that pokes
1688 * the Virtual CPU if it's still busy executing guest code.
1689 *
1690 * @returns VBox status code.
1691 * @retval VINF_SUCCESS if poked successfully.
1692 * @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1693 *
1694 * @param pGVM The global (ring-0) VM structure.
1695 * @param pVCpu The Virtual CPU handle.
1696 */
1697DECLINLINE(int) gvmmR0SchedPokeOne(PGVM pGVM, PVMCPU pVCpu)
1698{
1699 pGVM->gvmm.s.StatsSched.cPokeCalls++;
1700
1701 RTCPUID idHostCpu = pVCpu->idHostCpu;
1702 if ( idHostCpu == NIL_RTCPUID
1703 || VMCPU_GET_STATE(pVCpu) != VMCPUSTATE_STARTED_EXEC)
1704 {
1705 pGVM->gvmm.s.StatsSched.cPokeNotBusy++;
1706 return VINF_GVM_NOT_BUSY_IN_GC;
1707 }
1708
1709 /* Note: this function is not implemented on Darwin and Linux (kernel < 2.6.19) */
1710 RTMpPokeCpu(idHostCpu);
1711 return VINF_SUCCESS;
1712}
1713
1714/**
1715 * Pokes an EMT if it's still busy running guest code.
1716 *
1717 * @returns VBox status code.
1718 * @retval VINF_SUCCESS if poked successfully.
1719 * @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1720 *
1721 * @param pVM Pointer to the shared VM structure.
1722 * @param idCpu The ID of the virtual CPU to poke.
1723 * @param fTakeUsedLock Take the used lock or not
1724 */
1725GVMMR0DECL(int) GVMMR0SchedPokeEx(PVM pVM, VMCPUID idCpu, bool fTakeUsedLock)
1726{
1727 /*
1728 * Validate input and take the UsedLock.
1729 */
1730 PGVM pGVM;
1731 PGVMM pGVMM;
1732 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, fTakeUsedLock);
1733 if (RT_SUCCESS(rc))
1734 {
1735 if (idCpu < pGVM->cCpus)
1736 rc = gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
1737 else
1738 rc = VERR_INVALID_CPU_ID;
1739
1740 if (fTakeUsedLock)
1741 {
1742 int rc2 = gvmmR0UsedUnlock(pGVMM);
1743 AssertRC(rc2);
1744 }
1745 }
1746
1747 LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
1748 return rc;
1749}
1750
1751
1752/**
1753 * Pokes an EMT if it's still busy running guest code.
1754 *
1755 * @returns VBox status code.
1756 * @retval VINF_SUCCESS if poked successfully.
1757 * @retval VINF_GVM_NOT_BUSY_IN_GC if the EMT wasn't busy in GC.
1758 *
1759 * @param pVM Pointer to the shared VM structure.
1760 * @param idCpu The ID of the virtual CPU to poke.
1761 */
1762GVMMR0DECL(int) GVMMR0SchedPoke(PVM pVM, VMCPUID idCpu)
1763{
1764 return GVMMR0SchedPokeEx(pVM, idCpu, true /* fTakeUsedLock */);
1765}
1766
1767
1768/**
1769 * Wakes up a set of halted EMT threads so they can service pending request.
1770 *
1771 * @returns VBox status code, no informational stuff.
1772 *
1773 * @param pVM Pointer to the shared VM structure.
1774 * @param pSleepSet The set of sleepers to wake up.
1775 * @param pPokeSet The set of CPUs to poke.
1776 */
1777GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpus(PVM pVM, PCVMCPUSET pSleepSet, PCVMCPUSET pPokeSet)
1778{
1779 AssertPtrReturn(pSleepSet, VERR_INVALID_POINTER);
1780 AssertPtrReturn(pPokeSet, VERR_INVALID_POINTER);
1781 RTNATIVETHREAD hSelf = RTThreadNativeSelf();
1782
1783 /*
1784 * Validate input and take the UsedLock.
1785 */
1786 PGVM pGVM;
1787 PGVMM pGVMM;
1788 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /* fTakeUsedLock */);
1789 if (RT_SUCCESS(rc))
1790 {
1791 rc = VINF_SUCCESS;
1792 VMCPUID idCpu = pGVM->cCpus;
1793 while (idCpu-- > 0)
1794 {
1795 /* Don't try poke or wake up ourselves. */
1796 if (pGVM->aCpus[idCpu].hEMT == hSelf)
1797 continue;
1798
1799 /* just ignore errors for now. */
1800 if (VMCPUSET_IS_PRESENT(pSleepSet, idCpu))
1801 gvmmR0SchedWakeUpOne(pGVM, &pGVM->aCpus[idCpu]);
1802 else if (VMCPUSET_IS_PRESENT(pPokeSet, idCpu))
1803 gvmmR0SchedPokeOne(pGVM, &pVM->aCpus[idCpu]);
1804 }
1805
1806 int rc2 = gvmmR0UsedUnlock(pGVMM);
1807 AssertRC(rc2);
1808 }
1809
1810 LogFlow(("GVMMR0SchedWakeUpAndPokeCpus: returns %Rrc\n", rc));
1811 return rc;
1812}
1813
1814
1815/**
1816 * VMMR0 request wrapper for GVMMR0SchedWakeUpAndPokeCpus.
1817 *
1818 * @returns see GVMMR0SchedWakeUpAndPokeCpus.
1819 * @param pVM Pointer to the shared VM structure.
1820 * @param pReq The request packet.
1821 */
1822GVMMR0DECL(int) GVMMR0SchedWakeUpAndPokeCpusReq(PVM pVM, PGVMMSCHEDWAKEUPANDPOKECPUSREQ pReq)
1823{
1824 /*
1825 * Validate input and pass it on.
1826 */
1827 AssertPtrReturn(pReq, VERR_INVALID_POINTER);
1828 AssertMsgReturn(pReq->Hdr.cbReq == sizeof(*pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(*pReq)), VERR_INVALID_PARAMETER);
1829
1830 return GVMMR0SchedWakeUpAndPokeCpus(pVM, &pReq->SleepSet, &pReq->PokeSet);
1831}
1832
1833
1834
1835/**
1836 * Poll the schedule to see if someone else should get a chance to run.
1837 *
1838 * This is a bit hackish and will not work too well if the machine is
1839 * under heavy load from non-VM processes.
1840 *
1841 * @returns VINF_SUCCESS if not yielded.
1842 * VINF_GVM_YIELDED if an attempt to switch to a different VM task was made.
1843 * @param pVM Pointer to the shared VM structure.
1844 * @param idCpu The Virtual CPU ID of the calling EMT.
1845 * @param u64ExpireGipTime The time for the sleep to expire expressed as GIP time.
1846 * @param fYield Whether to yield or not.
1847 * This is for when we're spinning in the halt loop.
1848 * @thread EMT(idCpu).
1849 */
1850GVMMR0DECL(int) GVMMR0SchedPoll(PVM pVM, VMCPUID idCpu, bool fYield)
1851{
1852 /*
1853 * Validate input.
1854 */
1855 PGVM pGVM;
1856 PGVMM pGVMM;
1857 int rc = gvmmR0ByVMAndEMT(pVM, idCpu, &pGVM, &pGVMM);
1858 if (RT_SUCCESS(rc))
1859 {
1860 rc = gvmmR0UsedLock(pGVMM);
1861 AssertRC(rc);
1862 pGVM->gvmm.s.StatsSched.cPollCalls++;
1863
1864 Assert(ASMGetFlags() & X86_EFL_IF);
1865 const uint64_t u64Now = RTTimeNanoTS(); /* (GIP time) */
1866
1867 if (!fYield)
1868 pGVM->gvmm.s.StatsSched.cPollWakeUps += gvmmR0SchedDoWakeUps(pGVMM, u64Now);
1869 else
1870 {
1871 /** @todo implement this... */
1872 rc = VERR_NOT_IMPLEMENTED;
1873 }
1874
1875 gvmmR0UsedUnlock(pGVMM);
1876 }
1877
1878 LogFlow(("GVMMR0SchedWakeUp: returns %Rrc\n", rc));
1879 return rc;
1880}
1881
1882
1883
1884/**
1885 * Retrieves the GVMM statistics visible to the caller.
1886 *
1887 * @returns VBox status code.
1888 *
1889 * @param pStats Where to put the statistics.
1890 * @param pSession The current session.
1891 * @param pVM The VM to obtain statistics for. Optional.
1892 */
1893GVMMR0DECL(int) GVMMR0QueryStatistics(PGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
1894{
1895 LogFlow(("GVMMR0QueryStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
1896
1897 /*
1898 * Validate input.
1899 */
1900 AssertPtrReturn(pSession, VERR_INVALID_POINTER);
1901 AssertPtrReturn(pStats, VERR_INVALID_POINTER);
1902 pStats->cVMs = 0; /* (crash before taking the sem...) */
1903
1904 /*
1905 * Take the lock and get the VM statistics.
1906 */
1907 PGVMM pGVMM;
1908 if (pVM)
1909 {
1910 PGVM pGVM;
1911 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /*fTakeUsedLock*/);
1912 if (RT_FAILURE(rc))
1913 return rc;
1914 pStats->SchedVM = pGVM->gvmm.s.StatsSched;
1915 }
1916 else
1917 {
1918 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
1919 memset(&pStats->SchedVM, 0, sizeof(pStats->SchedVM));
1920
1921 int rc = gvmmR0UsedLock(pGVMM);
1922 AssertRCReturn(rc, rc);
1923 }
1924
1925 /*
1926 * Enumerate the VMs and add the ones visibile to the statistics.
1927 */
1928 pStats->cVMs = 0;
1929 pStats->cEMTs = 0;
1930 memset(&pStats->SchedSum, 0, sizeof(pStats->SchedSum));
1931
1932 for (unsigned i = pGVMM->iUsedHead;
1933 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
1934 i = pGVMM->aHandles[i].iNext)
1935 {
1936 PGVM pGVM = pGVMM->aHandles[i].pGVM;
1937 void *pvObj = pGVMM->aHandles[i].pvObj;
1938 if ( VALID_PTR(pvObj)
1939 && VALID_PTR(pGVM)
1940 && pGVM->u32Magic == GVM_MAGIC
1941 && RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
1942 {
1943 pStats->cVMs++;
1944 pStats->cEMTs += pGVM->cCpus;
1945
1946 pStats->SchedSum.cHaltCalls += pGVM->gvmm.s.StatsSched.cHaltCalls;
1947 pStats->SchedSum.cHaltBlocking += pGVM->gvmm.s.StatsSched.cHaltBlocking;
1948 pStats->SchedSum.cHaltTimeouts += pGVM->gvmm.s.StatsSched.cHaltTimeouts;
1949 pStats->SchedSum.cHaltNotBlocking += pGVM->gvmm.s.StatsSched.cHaltNotBlocking;
1950 pStats->SchedSum.cHaltWakeUps += pGVM->gvmm.s.StatsSched.cHaltWakeUps;
1951
1952 pStats->SchedSum.cWakeUpCalls += pGVM->gvmm.s.StatsSched.cWakeUpCalls;
1953 pStats->SchedSum.cWakeUpNotHalted += pGVM->gvmm.s.StatsSched.cWakeUpNotHalted;
1954 pStats->SchedSum.cWakeUpWakeUps += pGVM->gvmm.s.StatsSched.cWakeUpWakeUps;
1955
1956 pStats->SchedSum.cPokeCalls += pGVM->gvmm.s.StatsSched.cPokeCalls;
1957 pStats->SchedSum.cPokeNotBusy += pGVM->gvmm.s.StatsSched.cPokeNotBusy;
1958
1959 pStats->SchedSum.cPollCalls += pGVM->gvmm.s.StatsSched.cPollCalls;
1960 pStats->SchedSum.cPollHalts += pGVM->gvmm.s.StatsSched.cPollHalts;
1961 pStats->SchedSum.cPollWakeUps += pGVM->gvmm.s.StatsSched.cPollWakeUps;
1962 }
1963 }
1964
1965 gvmmR0UsedUnlock(pGVMM);
1966
1967 return VINF_SUCCESS;
1968}
1969
1970
1971/**
1972 * VMMR0 request wrapper for GVMMR0QueryStatistics.
1973 *
1974 * @returns see GVMMR0QueryStatistics.
1975 * @param pVM Pointer to the shared VM structure. Optional.
1976 * @param pReq The request packet.
1977 */
1978GVMMR0DECL(int) GVMMR0QueryStatisticsReq(PVM pVM, PGVMMQUERYSTATISTICSSREQ pReq)
1979{
1980 /*
1981 * Validate input and pass it on.
1982 */
1983 AssertPtrReturn(pReq, VERR_INVALID_POINTER);
1984 AssertMsgReturn(pReq->Hdr.cbReq == sizeof(*pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(*pReq)), VERR_INVALID_PARAMETER);
1985
1986 return GVMMR0QueryStatistics(&pReq->Stats, pReq->pSession, pVM);
1987}
1988
1989
1990/**
1991 * Resets the specified GVMM statistics.
1992 *
1993 * @returns VBox status code.
1994 *
1995 * @param pStats Which statistics to reset, that is, non-zero fields indicates which to reset.
1996 * @param pSession The current session.
1997 * @param pVM The VM to reset statistics for. Optional.
1998 */
1999GVMMR0DECL(int) GVMMR0ResetStatistics(PCGVMMSTATS pStats, PSUPDRVSESSION pSession, PVM pVM)
2000{
2001 LogFlow(("GVMMR0ResetStatistics: pStats=%p pSession=%p pVM=%p\n", pStats, pSession, pVM));
2002
2003 /*
2004 * Validate input.
2005 */
2006 AssertPtrReturn(pSession, VERR_INVALID_POINTER);
2007 AssertPtrReturn(pStats, VERR_INVALID_POINTER);
2008
2009 /*
2010 * Take the lock and get the VM statistics.
2011 */
2012 PGVMM pGVMM;
2013 if (pVM)
2014 {
2015 PGVM pGVM;
2016 int rc = gvmmR0ByVM(pVM, &pGVM, &pGVMM, true /*fTakeUsedLock*/);
2017 if (RT_FAILURE(rc))
2018 return rc;
2019# define MAYBE_RESET_FIELD(field) \
2020 do { if (pStats->SchedVM. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2021 MAYBE_RESET_FIELD(cHaltCalls);
2022 MAYBE_RESET_FIELD(cHaltBlocking);
2023 MAYBE_RESET_FIELD(cHaltTimeouts);
2024 MAYBE_RESET_FIELD(cHaltNotBlocking);
2025 MAYBE_RESET_FIELD(cHaltWakeUps);
2026 MAYBE_RESET_FIELD(cWakeUpCalls);
2027 MAYBE_RESET_FIELD(cWakeUpNotHalted);
2028 MAYBE_RESET_FIELD(cWakeUpWakeUps);
2029 MAYBE_RESET_FIELD(cPokeCalls);
2030 MAYBE_RESET_FIELD(cPokeNotBusy);
2031 MAYBE_RESET_FIELD(cPollCalls);
2032 MAYBE_RESET_FIELD(cPollHalts);
2033 MAYBE_RESET_FIELD(cPollWakeUps);
2034# undef MAYBE_RESET_FIELD
2035 }
2036 else
2037 {
2038 GVMM_GET_VALID_INSTANCE(pGVMM, VERR_INTERNAL_ERROR);
2039
2040 int rc = gvmmR0UsedLock(pGVMM);
2041 AssertRCReturn(rc, rc);
2042 }
2043
2044 /*
2045 * Enumerate the VMs and add the ones visibile to the statistics.
2046 */
2047 if (ASMMemIsAll8(&pStats->SchedSum, sizeof(pStats->SchedSum), 0))
2048 {
2049 for (unsigned i = pGVMM->iUsedHead;
2050 i != NIL_GVM_HANDLE && i < RT_ELEMENTS(pGVMM->aHandles);
2051 i = pGVMM->aHandles[i].iNext)
2052 {
2053 PGVM pGVM = pGVMM->aHandles[i].pGVM;
2054 void *pvObj = pGVMM->aHandles[i].pvObj;
2055 if ( VALID_PTR(pvObj)
2056 && VALID_PTR(pGVM)
2057 && pGVM->u32Magic == GVM_MAGIC
2058 && RT_SUCCESS(SUPR0ObjVerifyAccess(pvObj, pSession, NULL)))
2059 {
2060# define MAYBE_RESET_FIELD(field) \
2061 do { if (pStats->SchedSum. field ) { pGVM->gvmm.s.StatsSched. field = 0; } } while (0)
2062 MAYBE_RESET_FIELD(cHaltCalls);
2063 MAYBE_RESET_FIELD(cHaltBlocking);
2064 MAYBE_RESET_FIELD(cHaltTimeouts);
2065 MAYBE_RESET_FIELD(cHaltNotBlocking);
2066 MAYBE_RESET_FIELD(cHaltWakeUps);
2067 MAYBE_RESET_FIELD(cWakeUpCalls);
2068 MAYBE_RESET_FIELD(cWakeUpNotHalted);
2069 MAYBE_RESET_FIELD(cWakeUpWakeUps);
2070 MAYBE_RESET_FIELD(cPokeCalls);
2071 MAYBE_RESET_FIELD(cPokeNotBusy);
2072 MAYBE_RESET_FIELD(cPollCalls);
2073 MAYBE_RESET_FIELD(cPollHalts);
2074 MAYBE_RESET_FIELD(cPollWakeUps);
2075# undef MAYBE_RESET_FIELD
2076 }
2077 }
2078 }
2079
2080 gvmmR0UsedUnlock(pGVMM);
2081
2082 return VINF_SUCCESS;
2083}
2084
2085
2086/**
2087 * VMMR0 request wrapper for GVMMR0ResetStatistics.
2088 *
2089 * @returns see GVMMR0ResetStatistics.
2090 * @param pVM Pointer to the shared VM structure. Optional.
2091 * @param pReq The request packet.
2092 */
2093GVMMR0DECL(int) GVMMR0ResetStatisticsReq(PVM pVM, PGVMMRESETSTATISTICSSREQ pReq)
2094{
2095 /*
2096 * Validate input and pass it on.
2097 */
2098 AssertPtrReturn(pReq, VERR_INVALID_POINTER);
2099 AssertMsgReturn(pReq->Hdr.cbReq == sizeof(*pReq), ("%#x != %#x\n", pReq->Hdr.cbReq, sizeof(*pReq)), VERR_INVALID_PARAMETER);
2100
2101 return GVMMR0ResetStatistics(&pReq->Stats, pReq->pSession, pVM);
2102}
2103
Note: See TracBrowser for help on using the repository browser.

© 2025 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette