BOOLEAN
MmCreateProcessAddressSpace (
IN ULONG MinimumWorkingSetSize,
IN PEPROCESS NewProcess,
OUT PULONG_PTR DirectoryTableBase
)
/*++
Routine Description:
This routine creates an address space which maps the system
portion and contains a hyper space entry.
Arguments:
MinimumWorkingSetSize - Supplies the minimum working set size for
this address space. This value is only used
to ensure that ample physical pages exist
to create this process.
NewProcess - Supplies a pointer to the process object being created.
DirectoryTableBase - Returns the value of the newly created
address space's Page Directory (PD) page and
hyper space page.
Return Value:
Returns TRUE if an address space was successfully created, FALSE
if ample physical pages do not exist.
Environment:
Kernel mode. APCs Disabled.
--*/
{
LOGICAL FlushTbNeeded;
PFN_NUMBER PageDirectoryIndex;
PFN_NUMBER HyperSpaceIndex;
PFN_NUMBER PageContainingWorkingSet;
PFN_NUMBER VadBitMapPage;
MMPTE TempPte;
MMPTE TempPte2;
PEPROCESS CurrentProcess;
KIRQL OldIrql;
PMMPFN Pfn1;
ULONG Color;
PMMPTE PointerPte;
ULONG PdeOffset;
PMMPTE MappingPte;
PMMPTE PointerFillPte;
PMMPTE CurrentAddressSpacePde;
//
// Charge commitment for the page directory pages, working set page table
// page, and working set list. If Vad bitmap lookups are enabled, then
// charge for a page or two for that as well.
//
if (MiChargeCommitment (MM_PROCESS_COMMIT_CHARGE, NULL) == FALSE) {
return FALSE;
}
FlushTbNeeded = FALSE;
CurrentProcess = PsGetCurrentProcess ();
NewProcess->NextPageColor = (USHORT) (RtlRandom (&MmProcessColorSeed));
KeInitializeSpinLock (&NewProcess->HyperSpaceLock);
//
// Get the PFN lock to get physical pages.
//
LOCK_PFN (OldIrql);
//
// Check to make sure the physical pages are available.
//
if (MI_NONPAGEABLE_MEMORY_AVAILABLE() <= (SPFN_NUMBER)MinimumWorkingSetSize){
UNLOCK_PFN (OldIrql);
MiReturnCommitment (MM_PROCESS_COMMIT_CHARGE);
//
// Indicate no directory base was allocated.
//
return FALSE;
}
MM_TRACK_COMMIT (MM_DBG_COMMIT_PROCESS_CREATE, MM_PROCESS_COMMIT_CHARGE);
MI_DECREMENT_RESIDENT_AVAILABLE (MinimumWorkingSetSize,
MM_RESAVAIL_ALLOCATE_CREATE_PROCESS);
//
// Allocate a page directory page.
//
if (MmAvailablePages < MM_HIGH_LIMIT) {
MiEnsureAvailablePageOrWait (NULL, OldIrql);
}
Color = MI_PAGE_COLOR_PTE_PROCESS (PDE_BASE,
&CurrentProcess->NextPageColor);
PageDirectoryIndex = MiRemoveZeroPageMayReleaseLocks (Color, OldIrql);
Pfn1 = MI_PFN_ELEMENT (PageDirectoryIndex);
if (Pfn1->u3.e1.CacheAttribute != MiCached) {
Pfn1->u3.e1.CacheAttribute = MiCached;
FlushTbNeeded = TRUE;
}
//
// Allocate the hyper space page table page.
//
if (MmAvailablePages < MM_HIGH_LIMIT) {
MiEnsureAvailablePageOrWait (NULL, OldIrql);
}
Color = MI_PAGE_COLOR_PTE_PROCESS (MiGetPdeAddress(HYPER_SPACE),
&CurrentProcess->NextPageColor);
HyperSpaceIndex = MiRemoveZeroPageMayReleaseLocks (Color, OldIrql);
Pfn1 = MI_PFN_ELEMENT (HyperSpaceIndex);
if (Pfn1->u3.e1.CacheAttribute != MiCached) {
Pfn1->u3.e1.CacheAttribute = MiCached;
FlushTbNeeded = TRUE;
}
//
// Remove page(s) for the VAD bitmap.
//
if (MmAvailablePages < MM_HIGH_LIMIT) {
MiEnsureAvailablePageOrWait (NULL, OldIrql);
}
Color = MI_PAGE_COLOR_VA_PROCESS (MmWorkingSetList,
&CurrentProcess->NextPageColor);
VadBitMapPage = MiRemoveZeroPageMayReleaseLocks (Color, OldIrql);
Pfn1 = MI_PFN_ELEMENT (VadBitMapPage);
if (Pfn1->u3.e1.CacheAttribute != MiCached) {
Pfn1->u3.e1.CacheAttribute = MiCached;
FlushTbNeeded = TRUE;
}
//
// Remove a page for the working set list.
//
if (MmAvailablePages < MM_HIGH_LIMIT) {
MiEnsureAvailablePageOrWait (NULL, OldIrql);
}
Color = MI_PAGE_COLOR_VA_PROCESS (MmWorkingSetList,
&CurrentProcess->NextPageColor);
PageContainingWorkingSet = MiRemoveZeroPageMayReleaseLocks (Color, OldIrql);
Pfn1 = MI_PFN_ELEMENT (PageContainingWorkingSet);
if (Pfn1->u3.e1.CacheAttribute != MiCached) {
Pfn1->u3.e1.CacheAttribute = MiCached;
FlushTbNeeded = TRUE;
}
UNLOCK_PFN (OldIrql);
if (FlushTbNeeded == TRUE) {
MI_FLUSH_TB_FOR_CACHED_ATTRIBUTE ();
}
ASSERT (NewProcess->AddressSpaceInitialized == 0);
PS_SET_BITS (&NewProcess->Flags, PS_PROCESS_FLAGS_ADDRESS_SPACE1);
ASSERT (NewProcess->AddressSpaceInitialized == 1);
NewProcess->Vm.MinimumWorkingSetSize = MinimumWorkingSetSize;
NewProcess->WorkingSetPage = PageContainingWorkingSet;
INITIALIZE_DIRECTORY_TABLE_BASE (&DirectoryTableBase[0], PageDirectoryIndex);
INITIALIZE_DIRECTORY_TABLE_BASE (&DirectoryTableBase[1], HyperSpaceIndex);
//
// Initialize the page reserved for hyper space.
//
TempPte = ValidPdePde;
MI_SET_GLOBAL_STATE (TempPte, 0);
MappingPte = MiReserveSystemPtes (1, SystemPteSpace);
if (MappingPte != NULL) {
MI_MAKE_VALID_KERNEL_PTE (TempPte2,
HyperSpaceIndex,
MM_READWRITE,
MappingPte);
MI_SET_PTE_DIRTY (TempPte2);
MI_WRITE_VALID_PTE (MappingPte, TempPte2);
PointerPte = MiGetVirtualAddressMappedByPte (MappingPte);
}
else {
PointerPte = MiMapPageInHyperSpace (CurrentProcess, HyperSpaceIndex, &OldIrql);
}
TempPte.u.Hard.PageFrameNumber = VadBitMapPage;
PointerPte[MiGetPteOffset(VAD_BITMAP_SPACE)] = TempPte;
TempPte.u.Hard.PageFrameNumber = PageContainingWorkingSet;
PointerPte[MiGetPteOffset(MmWorkingSetList)] = TempPte;
if (MappingPte != NULL) {
MiReleaseSystemPtes (MappingPte, 1, SystemPteSpace);
}
else {
MiUnmapPageInHyperSpace (CurrentProcess, PointerPte, OldIrql);
}
//
// Set the PTE address in the PFN for the page directory page.
//
Pfn1 = MI_PFN_ELEMENT (PageDirectoryIndex);
Pfn1->PteAddress = (PMMPTE)PDE_BASE;
TempPte = ValidPdePde;
TempPte.u.Hard.PageFrameNumber = HyperSpaceIndex;
MI_SET_GLOBAL_STATE (TempPte, 0);
//
// Add the new process to our internal list prior to filling any
// system PDEs so if a system PDE changes (large page map or unmap)
// it can mark this process for a subsequent update.
//
ASSERT (NewProcess->Pcb.DirectoryTableBase[0] == 0);
LOCK_EXPANSION (OldIrql);
InsertTailList (&MmProcessList, &NewProcess->MmProcessLinks);
UNLOCK_EXPANSION (OldIrql);
//
// Map the page directory page in hyperspace.
//
MappingPte = MiReserveSystemPtes (1, SystemPteSpace);
if (MappingPte != NULL) {
MI_MAKE_VALID_KERNEL_PTE (TempPte2,
PageDirectoryIndex,
MM_READWRITE,
MappingPte);
MI_SET_PTE_DIRTY (TempPte2);
MI_WRITE_VALID_PTE (MappingPte, TempPte2);
PointerPte = MiGetVirtualAddressMappedByPte (MappingPte);
}
else {
PointerPte = MiMapPageInHyperSpace (CurrentProcess, PageDirectoryIndex, &OldIrql);
}
PdeOffset = MiGetPdeOffset (MmSystemRangeStart);
PointerFillPte = &PointerPte[PdeOffset];
CurrentAddressSpacePde = MiGetPdeAddress (MmSystemRangeStart);
RtlCopyMemory (PointerFillPte,
CurrentAddressSpacePde,
PAGE_SIZE - PdeOffset * sizeof (MMPTE));
//
// Map the working set page table page.
//
PdeOffset = MiGetPdeOffset (HYPER_SPACE);
PointerPte[PdeOffset] = TempPte;
//
// Zero the remaining page directory range used to map the working
// set list and its hash.
//
PdeOffset += 1;
ASSERT (MiGetPdeOffset (MmHyperSpaceEnd) >= PdeOffset);
MiZeroMemoryPte (&PointerPte[PdeOffset],
(MiGetPdeOffset (MmHyperSpaceEnd) - PdeOffset + 1));
//
// Recursively map the page directory page so it points to itself.
//
TempPte.u.Hard.PageFrameNumber = PageDirectoryIndex;
PointerPte[MiGetPdeOffset(PTE_BASE)] = TempPte;
if (MappingPte != NULL) {
MiReleaseSystemPtes (MappingPte, 1, SystemPteSpace);
}
else {
MiUnmapPageInHyperSpace (CurrentProcess, PointerPte, OldIrql);
}
InterlockedExchangeAddSizeT (&MmProcessCommit, MM_PROCESS_COMMIT_CHARGE);
//
// Up the session space reference count.
//
MiSessionAddProcess (NewProcess);
return TRUE;
}
NTSTATUS
MiCcPutPagesInTransition (
IN PMI_READ_INFO MiReadInfo
)
/*++
Routine Description:
This routine allocates physical memory for the specified read-list and
puts all the pages in transition (so collided faults from other threads
for these same pages remain coherent). I/O for any pages not already
resident are issued here. The caller must wait for their completion.
Arguments:
MiReadInfo - Supplies a pointer to the read-list.
Return Value:
STATUS_SUCCESS - all the pages were already resident, reference counts
have been applied and no I/O needs to be waited for.
STATUS_ISSUE_PAGING_IO - the I/O has been issued and the caller must wait.
Various other failure status values indicate the operation failed.
Environment:
Kernel mode. PASSIVE_LEVEL.
--*/
{
NTSTATUS status;
PMMPTE LocalPrototypePte;
PVOID StartingVa;
PFN_NUMBER MdlPages;
KIRQL OldIrql;
MMPTE PteContents;
PFN_NUMBER PageFrameIndex;
PFN_NUMBER ResidentAvailableCharge;
PPFN_NUMBER IoPage;
PPFN_NUMBER ApiPage;
PPFN_NUMBER Page;
PPFN_NUMBER DestinationPage;
ULONG PageColor;
PMMPTE PointerPte;
PMMPTE *ProtoPteArray;
PMMPTE *EndProtoPteArray;
PFN_NUMBER DummyPage;
PMDL Mdl;
PMDL FreeMdl;
PMMPFN PfnProto;
PMMPFN Pfn1;
PMMPFN DummyPfn1;
ULONG i;
PFN_NUMBER DummyTrim;
ULONG NumberOfPagesNeedingIo;
MMPTE TempPte;
PMMPTE PointerPde;
PEPROCESS CurrentProcess;
PMMINPAGE_SUPPORT InPageSupport;
PKPRCB Prcb;
ASSERT (KeGetCurrentIrql() == PASSIVE_LEVEL);
MiReadInfo->DummyPagePfn = NULL;
FreeMdl = NULL;
CurrentProcess = PsGetCurrentProcess();
PfnProto = NULL;
PointerPde = NULL;
InPageSupport = MiReadInfo->InPageSupport;
Mdl = MI_EXTRACT_PREFETCH_MDL (InPageSupport);
ASSERT (Mdl == MiReadInfo->IoMdl);
IoPage = (PPFN_NUMBER)(Mdl + 1);
ApiPage = (PPFN_NUMBER)(MiReadInfo->ApiMdl + 1);
StartingVa = (PVOID)((PCHAR)Mdl->StartVa + Mdl->ByteOffset);
MdlPages = ADDRESS_AND_SIZE_TO_SPAN_PAGES (StartingVa,
Mdl->ByteCount);
if (MdlPages + 1 > MAXUSHORT) {
//
// The PFN ReferenceCount for the dummy page could wrap, refuse the
// request.
//
return STATUS_INSUFFICIENT_RESOURCES;
}
NumberOfPagesNeedingIo = 0;
ProtoPteArray = (PMMPTE *)InPageSupport->BasePte;
EndProtoPteArray = ProtoPteArray + MdlPages;
ASSERT (*ProtoPteArray != NULL);
LOCK_PFN (OldIrql);
//
// Ensure sufficient pages exist for the transfer plus the dummy page.
//
if (((SPFN_NUMBER)MdlPages > (SPFN_NUMBER)(MmAvailablePages - MM_HIGH_LIMIT)) ||
(MI_NONPAGEABLE_MEMORY_AVAILABLE() <= (SPFN_NUMBER)MdlPages)) {
UNLOCK_PFN (OldIrql);
return STATUS_INSUFFICIENT_RESOURCES;
}
//
// Charge resident available immediately as the PFN lock may get released
// and reacquired below before all the pages have been locked down.
// Note the dummy page is immediately charged separately.
//
MI_DECREMENT_RESIDENT_AVAILABLE (MdlPages, MM_RESAVAIL_ALLOCATE_BUILDMDL);
ResidentAvailableCharge = MdlPages;
//
// Allocate a dummy page to map discarded pages that aren't skipped.
//
DummyPage = MiRemoveAnyPage (0);
Pfn1 = MI_PFN_ELEMENT (DummyPage);
ASSERT (Pfn1->u2.ShareCount == 0);
ASSERT (Pfn1->u3.e2.ReferenceCount == 0);
MiInitializePfnForOtherProcess (DummyPage, MI_PF_DUMMY_PAGE_PTE, 0);
//
// Give the page a containing frame so MiIdentifyPfn won't crash.
//
Pfn1->u4.PteFrame = PsInitialSystemProcess->Pcb.DirectoryTableBase[0] >> PAGE_SHIFT;
//
// Always bias the reference count by 1 and charge for this locked page
// up front so the myriad increments and decrements don't get slowed
// down with needless checking.
//
Pfn1->u3.e1.PrototypePte = 0;
MI_ADD_LOCKED_PAGE_CHARGE (Pfn1);
Pfn1->u3.e1.ReadInProgress = 1;
MiReadInfo->DummyPagePfn = Pfn1;
DummyPfn1 = Pfn1;
DummyPfn1->u3.e2.ReferenceCount =
(USHORT)(DummyPfn1->u3.e2.ReferenceCount + MdlPages);
//
// Properly initialize the inpage support block fields we overloaded.
//
InPageSupport->BasePte = *ProtoPteArray;
//
// Build the proper InPageSupport and MDL to describe this run.
//
for (; ProtoPteArray < EndProtoPteArray; ProtoPteArray += 1, IoPage += 1, ApiPage += 1) {
//
// Fill the MDL entry for this RLE.
//
PointerPte = *ProtoPteArray;
ASSERT (PointerPte != NULL);
//
// The PointerPte better be inside a prototype PTE allocation
// so that subsequent page trims update the correct PTEs.
//
ASSERT (((PointerPte >= (PMMPTE)MmPagedPoolStart) &&
(PointerPte <= (PMMPTE)MmPagedPoolEnd)) ||
((PointerPte >= (PMMPTE)MmSpecialPoolStart) && (PointerPte <= (PMMPTE)MmSpecialPoolEnd)));
//
// Check the state of this prototype PTE now that the PFN lock is held.
// If the page is not resident, the PTE must be put in transition with
// read in progress before the PFN lock is released.
//
//
// Lock page containing prototype PTEs in memory by
// incrementing the reference count for the page.
// Unlock any page locked earlier containing prototype PTEs if
// the containing page is not the same for both.
//
if (PfnProto != NULL) {
if (PointerPde != MiGetPteAddress (PointerPte)) {
ASSERT (PfnProto->u3.e2.ReferenceCount > 1);
MI_REMOVE_LOCKED_PAGE_CHARGE_AND_DECREF (PfnProto);
PfnProto = NULL;
}
}
if (PfnProto == NULL) {
ASSERT (!MI_IS_PHYSICAL_ADDRESS (PointerPte));
PointerPde = MiGetPteAddress (PointerPte);
if (PointerPde->u.Hard.Valid == 0) {
MiMakeSystemAddressValidPfn (PointerPte, OldIrql);
}
PfnProto = MI_PFN_ELEMENT (PointerPde->u.Hard.PageFrameNumber);
MI_ADD_LOCKED_PAGE_CHARGE (PfnProto);
ASSERT (PfnProto->u3.e2.ReferenceCount > 1);
}
recheck:
PteContents = *PointerPte;
// LWFIX: are zero or dzero ptes possible here ?
ASSERT (PteContents.u.Long != 0);
if (PteContents.u.Hard.Valid == 1) {
PageFrameIndex = MI_GET_PAGE_FRAME_FROM_PTE (&PteContents);
Pfn1 = MI_PFN_ELEMENT (PageFrameIndex);
ASSERT (Pfn1->u3.e1.PrototypePte == 1);
MI_ADD_LOCKED_PAGE_CHARGE (Pfn1);
*ApiPage = PageFrameIndex;
*IoPage = DummyPage;
continue;
}
if ((PteContents.u.Soft.Prototype == 0) &&
(PteContents.u.Soft.Transition == 1)) {
//
// The page is in transition. If there is an inpage still in
// progress, wait for it to complete. Reference the PFN and
// then march on.
//
PageFrameIndex = MI_GET_PAGE_FRAME_FROM_TRANSITION_PTE (&PteContents);
Pfn1 = MI_PFN_ELEMENT (PageFrameIndex);
ASSERT (Pfn1->u3.e1.PrototypePte == 1);
if (Pfn1->u4.InPageError) {
//
// There was an in-page read error and there are other
// threads colliding for this page, delay to let the
// other threads complete and then retry.
//
UNLOCK_PFN (OldIrql);
KeDelayExecutionThread (KernelMode, FALSE, (PLARGE_INTEGER)&MmHalfSecond);
LOCK_PFN (OldIrql);
goto recheck;
}
if (Pfn1->u3.e1.ReadInProgress) {
// LWFIX - start with temp\aw.c
}
//
// PTE refers to a normal transition PTE.
//
ASSERT ((SPFN_NUMBER)MmAvailablePages >= 0);
if (MmAvailablePages == 0) {
//
// This can only happen if the system is utilizing a hardware
// compression cache. This ensures that only a safe amount
// of the compressed virtual cache is directly mapped so that
// if the hardware gets into trouble, we can bail it out.
//
UNLOCK_PFN (OldIrql);
KeDelayExecutionThread (KernelMode, FALSE, (PLARGE_INTEGER)&MmHalfSecond);
LOCK_PFN (OldIrql);
goto recheck;
}
//
// The PFN reference count will be 1 already here if the
// modified writer has begun a write of this page. Otherwise
// it's ordinarily 0.
//
MI_ADD_LOCKED_PAGE_CHARGE_FOR_MODIFIED_PAGE (Pfn1);
*IoPage = DummyPage;
*ApiPage = PageFrameIndex;
continue;
}
// LWFIX: need to handle protos that are now pagefile (or dzero)
// backed - prefetching it from the file here would cause us to lose
// the contents. Note this can happen for session-space images
// as we back modified (ie: for relocation fixups or IAT
// updated) portions from the pagefile. remove the assert below too.
ASSERT (PteContents.u.Soft.Prototype == 1);
if ((MmAvailablePages < MM_HIGH_LIMIT) &&
(MiEnsureAvailablePageOrWait (NULL, OldIrql))) {
//
// Had to wait so recheck all state.
//
goto recheck;
}
NumberOfPagesNeedingIo += 1;
//
// Allocate a physical page.
//
PageColor = MI_PAGE_COLOR_VA_PROCESS (
MiGetVirtualAddressMappedByPte (PointerPte),
&CurrentProcess->NextPageColor);
PageFrameIndex = MiRemoveAnyPage (PageColor);
Pfn1 = MI_PFN_ELEMENT (PageFrameIndex);
ASSERT (Pfn1->u3.e2.ReferenceCount == 0);
ASSERT (Pfn1->u2.ShareCount == 0);
ASSERT (PointerPte->u.Hard.Valid == 0);
//
// Initialize read-in-progress PFN.
//
MiInitializePfn (PageFrameIndex, PointerPte, 0);
//
// These pieces of MiInitializePfn initialization are overridden
// here as these pages are only going into prototype
// transition and not into any page tables.
//
Pfn1->u3.e1.PrototypePte = 1;
Pfn1->u2.ShareCount -= 1;
ASSERT (Pfn1->u2.ShareCount == 0);
Pfn1->u3.e1.PageLocation = ZeroedPageList;
Pfn1->u3.e2.ReferenceCount -= 1;
ASSERT (Pfn1->u3.e2.ReferenceCount == 0);
MI_ADD_LOCKED_PAGE_CHARGE_FOR_MODIFIED_PAGE (Pfn1);
//
// Initialize the I/O specific fields.
//
Pfn1->u1.Event = &InPageSupport->Event;
Pfn1->u3.e1.ReadInProgress = 1;
ASSERT (Pfn1->u4.InPageError == 0);
//
// Increment the PFN reference count in the control area for
// the subsection.
//
MiReadInfo->ControlArea->NumberOfPfnReferences += 1;
//
// Put the prototype PTE into the transition state.
//
MI_MAKE_TRANSITION_PTE (TempPte,
PageFrameIndex,
PointerPte->u.Soft.Protection,
PointerPte);
MI_WRITE_INVALID_PTE (PointerPte, TempPte);
*IoPage = PageFrameIndex;
*ApiPage = PageFrameIndex;
}
//
// If all the pages were resident, dereference the dummy page references
// now and notify our caller that I/O is not necessary.
//
if (NumberOfPagesNeedingIo == 0) {
ASSERT (DummyPfn1->u3.e2.ReferenceCount > MdlPages);
DummyPfn1->u3.e2.ReferenceCount =
(USHORT)(DummyPfn1->u3.e2.ReferenceCount - MdlPages);
//
// Unlock page containing prototype PTEs.
//
if (PfnProto != NULL) {
ASSERT (PfnProto->u3.e2.ReferenceCount > 1);
MI_REMOVE_LOCKED_PAGE_CHARGE_AND_DECREF (PfnProto);
}
UNLOCK_PFN (OldIrql);
//
// Return the upfront resident available charge as the
// individual charges have all been made at this point.
//
MI_INCREMENT_RESIDENT_AVAILABLE (ResidentAvailableCharge,
MM_RESAVAIL_FREE_BUILDMDL_EXCESS);
return STATUS_SUCCESS;
}
//
// Carefully trim leading dummy pages.
//
Page = (PPFN_NUMBER)(Mdl + 1);
DummyTrim = 0;
for (i = 0; i < MdlPages - 1; i += 1) {
if (*Page == DummyPage) {
DummyTrim += 1;
Page += 1;
}
else {
break;
}
}
if (DummyTrim != 0) {
Mdl->Size = (USHORT)(Mdl->Size - (DummyTrim * sizeof(PFN_NUMBER)));
Mdl->ByteCount -= (ULONG)(DummyTrim * PAGE_SIZE);
ASSERT (Mdl->ByteCount != 0);
InPageSupport->ReadOffset.QuadPart += (DummyTrim * PAGE_SIZE);
DummyPfn1->u3.e2.ReferenceCount =
(USHORT)(DummyPfn1->u3.e2.ReferenceCount - DummyTrim);
//
// Shuffle down the PFNs in the MDL.
// Recalculate BasePte to adjust for the shuffle.
//
Pfn1 = MI_PFN_ELEMENT (*Page);
ASSERT (Pfn1->PteAddress->u.Hard.Valid == 0);
ASSERT ((Pfn1->PteAddress->u.Soft.Prototype == 0) &&
(Pfn1->PteAddress->u.Soft.Transition == 1));
InPageSupport->BasePte = Pfn1->PteAddress;
DestinationPage = (PPFN_NUMBER)(Mdl + 1);
do {
*DestinationPage = *Page;
DestinationPage += 1;
Page += 1;
i += 1;
} while (i < MdlPages);
MdlPages -= DummyTrim;
}
//
// Carefully trim trailing dummy pages.
//
ASSERT (MdlPages != 0);
Page = (PPFN_NUMBER)(Mdl + 1) + MdlPages - 1;
if (*Page == DummyPage) {
ASSERT (MdlPages >= 2);
//
// Trim the last page specially as it may be a partial page.
//
Mdl->Size -= sizeof(PFN_NUMBER);
if (BYTE_OFFSET(Mdl->ByteCount) != 0) {
Mdl->ByteCount &= ~(PAGE_SIZE - 1);
}
else {
Mdl->ByteCount -= PAGE_SIZE;
}
ASSERT (Mdl->ByteCount != 0);
DummyPfn1->u3.e2.ReferenceCount -= 1;
//
// Now trim any other trailing pages.
//
Page -= 1;
DummyTrim = 0;
while (Page != ((PPFN_NUMBER)(Mdl + 1))) {
if (*Page != DummyPage) {
break;
}
DummyTrim += 1;
Page -= 1;
}
if (DummyTrim != 0) {
ASSERT (Mdl->Size > (USHORT)(DummyTrim * sizeof(PFN_NUMBER)));
Mdl->Size = (USHORT)(Mdl->Size - (DummyTrim * sizeof(PFN_NUMBER)));
Mdl->ByteCount -= (ULONG)(DummyTrim * PAGE_SIZE);
DummyPfn1->u3.e2.ReferenceCount =
(USHORT)(DummyPfn1->u3.e2.ReferenceCount - DummyTrim);
}
ASSERT (MdlPages > DummyTrim + 1);
MdlPages -= (DummyTrim + 1);
#if DBG
StartingVa = (PVOID)((PCHAR)Mdl->StartVa + Mdl->ByteOffset);
ASSERT (MdlPages == ADDRESS_AND_SIZE_TO_SPAN_PAGES(StartingVa,
Mdl->ByteCount));
#endif
}
//
// If the MDL is not already embedded in the inpage block, see if its
// final size qualifies it - if so, embed it now.
//
if ((Mdl != &InPageSupport->Mdl) &&
(Mdl->ByteCount <= (MM_MAXIMUM_READ_CLUSTER_SIZE + 1) * PAGE_SIZE)){
#if DBG
RtlFillMemoryUlong (&InPageSupport->Page[0],
(MM_MAXIMUM_READ_CLUSTER_SIZE+1) * sizeof (PFN_NUMBER),
0xf1f1f1f1);
#endif
RtlCopyMemory (&InPageSupport->Mdl, Mdl, Mdl->Size);
FreeMdl = Mdl;
Mdl = &InPageSupport->Mdl;
ASSERT (((ULONG_PTR)Mdl & (sizeof(QUAD) - 1)) == 0);
InPageSupport->u1.e1.PrefetchMdlHighBits = ((ULONG_PTR)Mdl >> 3);
}
LOGICAL
FASTCALL
MiCopyOnWrite (
IN PVOID FaultingAddress,
IN PMMPTE PointerPte
)
/*++
Routine Description:
This routine performs a copy on write operation for the specified
virtual address.
Arguments:
FaultingAddress - Supplies the virtual address which caused the fault.
PointerPte - Supplies the pointer to the PTE which caused the page fault.
Return Value:
Returns TRUE if the page was actually split, FALSE if not.
Environment:
Kernel mode, APCs disabled, working set mutex held.
--*/
{
MMPTE TempPte;
MMPTE TempPte2;
PMMPTE MappingPte;
PFN_NUMBER PageFrameIndex;
PFN_NUMBER NewPageIndex;
PVOID CopyTo;
PVOID CopyFrom;
KIRQL OldIrql;
PMMPFN Pfn1;
PEPROCESS CurrentProcess;
PMMCLONE_BLOCK CloneBlock;
PMMCLONE_DESCRIPTOR CloneDescriptor;
WSLE_NUMBER WorkingSetIndex;
LOGICAL FakeCopyOnWrite;
PMMWSL WorkingSetList;
PVOID SessionSpace;
PLIST_ENTRY NextEntry;
PIMAGE_ENTRY_IN_SESSION Image;
//
// This is called from MmAccessFault, the PointerPte is valid
// and the working set mutex ensures it cannot change state.
//
// Capture the PTE contents to TempPte.
//
TempPte = *PointerPte;
ASSERT (TempPte.u.Hard.Valid == 1);
PageFrameIndex = MI_GET_PAGE_FRAME_FROM_PTE (&TempPte);
Pfn1 = MI_PFN_ELEMENT (PageFrameIndex);
//
// Check to see if this is a prototype PTE with copy on write enabled.
//
FakeCopyOnWrite = FALSE;
CurrentProcess = PsGetCurrentProcess ();
CloneBlock = NULL;
if (FaultingAddress >= (PVOID) MmSessionBase) {
WorkingSetList = MmSessionSpace->Vm.VmWorkingSetList;
ASSERT (Pfn1->u3.e1.PrototypePte == 1);
SessionSpace = (PVOID) MmSessionSpace;
MM_SESSION_SPACE_WS_LOCK_ASSERT ();
if (MmSessionSpace->ImageLoadingCount != 0) {
NextEntry = MmSessionSpace->ImageList.Flink;
while (NextEntry != &MmSessionSpace->ImageList) {
Image = CONTAINING_RECORD (NextEntry, IMAGE_ENTRY_IN_SESSION, Link);
if ((FaultingAddress >= Image->Address) &&
(FaultingAddress <= Image->LastAddress)) {
if (Image->ImageLoading) {
ASSERT (Pfn1->u3.e1.PrototypePte == 1);
TempPte.u.Hard.CopyOnWrite = 0;
TempPte.u.Hard.Write = 1;
//
// The page is no longer copy on write, update the PTE
// setting both the dirty bit and the accessed bit.
//
// Even though the page's current backing is the image
// file, the modified writer will convert it to
// pagefile backing when it notices the change later.
//
MI_SET_PTE_DIRTY (TempPte);
MI_SET_ACCESSED_IN_PTE (&TempPte, 1);
MI_WRITE_VALID_PTE_NEW_PROTECTION (PointerPte, TempPte);
//
// The TB entry must be flushed as the valid PTE with
// the dirty bit clear has been fetched into the TB. If
// it isn't flushed, another fault is generated as the
// dirty bit is not set in the cached TB entry.
//
MI_FLUSH_SINGLE_TB (FaultingAddress, TRUE);
return FALSE;
}
break;
}
NextEntry = NextEntry->Flink;
}
}
}
else {
WorkingSetList = MmWorkingSetList;
SessionSpace = NULL;
//
// If a fork operation is in progress, block until the fork is
// completed, then retry the whole operation as the state of
// everything may have changed between when the mutexes were
// released and reacquired.
//
if (CurrentProcess->ForkInProgress != NULL) {
if (MiWaitForForkToComplete (CurrentProcess) == TRUE) {
return FALSE;
}
}
if (TempPte.u.Hard.CopyOnWrite == 0) {
//
// This is a fork page which is being made private in order
// to change the protection of the page.
// Do not make the page writable.
//
FakeCopyOnWrite = TRUE;
}
}
WorkingSetIndex = MiLocateWsle (FaultingAddress,
WorkingSetList,
Pfn1->u1.WsIndex,
FALSE);
//
// The page must be copied into a new page.
//
LOCK_PFN (OldIrql);
if ((MmAvailablePages < MM_HIGH_LIMIT) &&
(MiEnsureAvailablePageOrWait (SessionSpace != NULL ? HYDRA_PROCESS : CurrentProcess, OldIrql))) {
//
// A wait operation was performed to obtain an available
// page and the working set mutex and PFN lock have
// been released and various things may have changed for
// the worse. Rather than examine all the conditions again,
// return and if things are still proper, the fault will
// be taken again.
//
UNLOCK_PFN (OldIrql);
return FALSE;
}
//
// This must be a prototype PTE. Perform the copy on write.
//
ASSERT (Pfn1->u3.e1.PrototypePte == 1);
//
// A page is being copied and made private, the global state of
// the shared page needs to be updated at this point on certain
// hardware. This is done by ORing the dirty bit into the modify bit in
// the PFN element.
//
// Note that a session page cannot be dirty (no POSIX-style forking is
// supported for these drivers).
//
if (SessionSpace != NULL) {
ASSERT ((TempPte.u.Hard.Valid == 1) && (TempPte.u.Hard.Write == 0));
ASSERT (!MI_IS_PTE_DIRTY (TempPte));
NewPageIndex = MiRemoveAnyPage (MI_GET_PAGE_COLOR_FROM_SESSION(MmSessionSpace));
}
else {
MI_CAPTURE_DIRTY_BIT_TO_PFN (PointerPte, Pfn1);
CloneBlock = (PMMCLONE_BLOCK) Pfn1->PteAddress;
//
// Get a new page with the same color as this page.
//
NewPageIndex = MiRemoveAnyPage (
MI_PAGE_COLOR_PTE_PROCESS(PageFrameIndex,
&CurrentProcess->NextPageColor));
}
MiInitializeCopyOnWritePfn (NewPageIndex,
PointerPte,
WorkingSetIndex,
WorkingSetList);
UNLOCK_PFN (OldIrql);
InterlockedIncrement (&KeGetCurrentPrcb ()->MmCopyOnWriteCount);
CopyFrom = PAGE_ALIGN (FaultingAddress);
MappingPte = MiReserveSystemPtes (1, SystemPteSpace);
if (MappingPte != NULL) {
MI_MAKE_VALID_KERNEL_PTE (TempPte2,
NewPageIndex,
MM_READWRITE,
MappingPte);
MI_SET_PTE_DIRTY (TempPte2);
if (Pfn1->u3.e1.CacheAttribute == MiNonCached) {
MI_DISABLE_CACHING (TempPte2);
}
else if (Pfn1->u3.e1.CacheAttribute == MiWriteCombined) {
MI_SET_PTE_WRITE_COMBINE (TempPte2);
}
MI_WRITE_VALID_PTE (MappingPte, TempPte2);
CopyTo = MiGetVirtualAddressMappedByPte (MappingPte);
}
else {
CopyTo = MiMapPageInHyperSpace (CurrentProcess,
NewPageIndex,
&OldIrql);
}
KeCopyPage (CopyTo, CopyFrom);
if (MappingPte != NULL) {
MiReleaseSystemPtes (MappingPte, 1, SystemPteSpace);
}
else {
MiUnmapPageInHyperSpace (CurrentProcess, CopyTo, OldIrql);
}
if (!FakeCopyOnWrite) {
//
// If the page was really a copy on write page, make it
// accessed, dirty and writable. Also, clear the copy-on-write
// bit in the PTE.
//
MI_SET_PTE_DIRTY (TempPte);
TempPte.u.Hard.Write = 1;
MI_SET_ACCESSED_IN_PTE (&TempPte, 1);
TempPte.u.Hard.CopyOnWrite = 0;
}
//
// Regardless of whether the page was really a copy on write,
// the frame field of the PTE must be updated.
//
TempPte.u.Hard.PageFrameNumber = NewPageIndex;
//
// If the modify bit is set in the PFN database for the
// page, the data cache must be flushed. This is due to the
// fact that this process may have been cloned and the cache
// still contains stale data destined for the page we are
// going to remove.
//
ASSERT (TempPte.u.Hard.Valid == 1);
MI_WRITE_VALID_PTE_NEW_PAGE (PointerPte, TempPte);
//
// Flush the TB entry for this page.
//
if (SessionSpace == NULL) {
MI_FLUSH_SINGLE_TB (FaultingAddress, FALSE);
//
// Increment the number of private pages.
//
CurrentProcess->NumberOfPrivatePages += 1;
}
else {
MI_FLUSH_SINGLE_TB (FaultingAddress, TRUE);
ASSERT (Pfn1->u3.e1.PrototypePte == 1);
}
//
// Decrement the share count for the page which was copied
// as this PTE no longer refers to it.
//
LOCK_PFN (OldIrql);
MiDecrementShareCount (Pfn1, PageFrameIndex);
if (SessionSpace == NULL) {
CloneDescriptor = MiLocateCloneAddress (CurrentProcess,
(PVOID)CloneBlock);
if (CloneDescriptor != NULL) {
//
// Decrement the reference count for the clone block,
// note that this could release and reacquire the mutexes.
//
MiDecrementCloneBlockReference (CloneDescriptor,
CloneBlock,
CurrentProcess,
NULL,
OldIrql);
}
}
UNLOCK_PFN (OldIrql);
return TRUE;
}
