Пример #1
0
PVOID
NTAPI
MiAllocatePoolPages(IN POOL_TYPE PoolType,
                    IN SIZE_T SizeInBytes)
{
    PFN_NUMBER PageFrameNumber;
    PFN_COUNT SizeInPages, PageTableCount;
    ULONG i;
    KIRQL OldIrql;
    PLIST_ENTRY NextEntry, NextHead, LastHead;
    PMMPTE PointerPte, StartPte;
    PMMPDE PointerPde;
    ULONG EndAllocation;
    MMPTE TempPte;
    MMPDE TempPde;
    PMMPFN Pfn1;
    PVOID BaseVa, BaseVaStart;
    PMMFREE_POOL_ENTRY FreeEntry;
    PKSPIN_LOCK_QUEUE LockQueue;

    //
    // Figure out how big the allocation is in pages
    //
    SizeInPages = (PFN_COUNT)BYTES_TO_PAGES(SizeInBytes);

    //
    // Check for overflow
    //
    if (SizeInPages == 0)
    {
        //
        // Fail
        //
        return NULL;
    }

    //
    // Handle paged pool
    //
    if ((PoolType & BASE_POOL_TYPE_MASK) == PagedPool)
    {
        //
        // If only one page is being requested, try to grab it from the S-LIST
        //
        if ((SizeInPages == 1) && (ExQueryDepthSList(&MiPagedPoolSListHead)))
        {
            BaseVa = InterlockedPopEntrySList(&MiPagedPoolSListHead);
            if (BaseVa) return BaseVa;
        }

        //
        // Lock the paged pool mutex
        //
        KeAcquireGuardedMutex(&MmPagedPoolMutex);

        //
        // Find some empty allocation space
        //
        i = RtlFindClearBitsAndSet(MmPagedPoolInfo.PagedPoolAllocationMap,
                                   SizeInPages,
                                   MmPagedPoolInfo.PagedPoolHint);
        if (i == 0xFFFFFFFF)
        {
            //
            // Get the page bit count
            //
            i = ((SizeInPages - 1) / PTE_COUNT) + 1;
            DPRINT("Paged pool expansion: %lu %x\n", i, SizeInPages);

            //
            // Check if there is enougn paged pool expansion space left
            //
            if (MmPagedPoolInfo.NextPdeForPagedPoolExpansion >
                    (PMMPDE)MiAddressToPte(MmPagedPoolInfo.LastPteForPagedPool))
            {
                //
                // Out of memory!
                //
                DPRINT1("OUT OF PAGED POOL!!!\n");
                KeReleaseGuardedMutex(&MmPagedPoolMutex);
                return NULL;
            }

            //
            // Check if we'll have to expand past the last PTE we have available
            //
            if (((i - 1) + MmPagedPoolInfo.NextPdeForPagedPoolExpansion) >
                    (PMMPDE)MiAddressToPte(MmPagedPoolInfo.LastPteForPagedPool))
            {
                //
                // We can only support this much then
                //
                PointerPde = MiAddressToPte(MmPagedPoolInfo.LastPteForPagedPool);
                PageTableCount = (PFN_COUNT)(PointerPde + 1 -
                                             MmPagedPoolInfo.NextPdeForPagedPoolExpansion);
                ASSERT(PageTableCount < i);
                i = PageTableCount;
            }
            else
            {
                //
                // Otherwise, there is plenty of space left for this expansion
                //
                PageTableCount = i;
            }

            //
            // Get the template PDE we'll use to expand
            //
            TempPde = ValidKernelPde;

            //
            // Get the first PTE in expansion space
            //
            PointerPde = MmPagedPoolInfo.NextPdeForPagedPoolExpansion;
            BaseVa = MiPdeToPte(PointerPde);
            BaseVaStart = BaseVa;

            //
            // Lock the PFN database and loop pages
            //
            OldIrql = KeAcquireQueuedSpinLock(LockQueuePfnLock);
            do
            {
                //
                // It should not already be valid
                //
                ASSERT(PointerPde->u.Hard.Valid == 0);

                /* Request a page */
                MI_SET_USAGE(MI_USAGE_PAGED_POOL);
                MI_SET_PROCESS2("Kernel");
                PageFrameNumber = MiRemoveAnyPage(MI_GET_NEXT_COLOR());
                TempPde.u.Hard.PageFrameNumber = PageFrameNumber;
#if (_MI_PAGING_LEVELS >= 3)
                /* On PAE/x64 systems, there's no double-buffering */
                ASSERT(FALSE);
#else
                //
                // Save it into our double-buffered system page directory
                //
                MmSystemPagePtes[((ULONG_PTR)PointerPde & (SYSTEM_PD_SIZE - 1)) / sizeof(MMPTE)] = TempPde;

                /* Initialize the PFN */
                MiInitializePfnForOtherProcess(PageFrameNumber,
                                               (PMMPTE)PointerPde,
                                               MmSystemPageDirectory[(PointerPde - MiAddressToPde(NULL)) / PDE_COUNT]);

                /* Write the actual PDE now */
//                MI_WRITE_VALID_PDE(PointerPde, TempPde);
#endif
                //
                // Move on to the next expansion address
                //
                PointerPde++;
                BaseVa = (PVOID)((ULONG_PTR)BaseVa + PAGE_SIZE);
                i--;
            } while (i > 0);

            //
            // Release the PFN database lock
            //
            KeReleaseQueuedSpinLock(LockQueuePfnLock, OldIrql);

            //
            // These pages are now available, clear their availablity bits
            //
            EndAllocation = (ULONG)(MmPagedPoolInfo.NextPdeForPagedPoolExpansion -
                                    (PMMPDE)MiAddressToPte(MmPagedPoolInfo.FirstPteForPagedPool)) *
                            PTE_COUNT;
            RtlClearBits(MmPagedPoolInfo.PagedPoolAllocationMap,
                         EndAllocation,
                         PageTableCount * PTE_COUNT);

            //
            // Update the next expansion location
            //
            MmPagedPoolInfo.NextPdeForPagedPoolExpansion += PageTableCount;

            //
            // Zero out the newly available memory
            //
            RtlZeroMemory(BaseVaStart, PageTableCount * PAGE_SIZE);

            //
            // Now try consuming the pages again
            //
            i = RtlFindClearBitsAndSet(MmPagedPoolInfo.PagedPoolAllocationMap,
                                       SizeInPages,
                                       0);
            if (i == 0xFFFFFFFF)
            {
                //
                // Out of memory!
                //
                DPRINT1("OUT OF PAGED POOL!!!\n");
                KeReleaseGuardedMutex(&MmPagedPoolMutex);
                return NULL;
            }
        }

        //
        // Update the pool hint if the request was just one page
        //
        if (SizeInPages == 1) MmPagedPoolInfo.PagedPoolHint = i + 1;

        //
        // Update the end bitmap so we know the bounds of this allocation when
        // the time comes to free it
        //
        EndAllocation = i + SizeInPages - 1;
        RtlSetBit(MmPagedPoolInfo.EndOfPagedPoolBitmap, EndAllocation);

        //
        // Now we can release the lock (it mainly protects the bitmap)
        //
        KeReleaseGuardedMutex(&MmPagedPoolMutex);

        //
        // Now figure out where this allocation starts
        //
        BaseVa = (PVOID)((ULONG_PTR)MmPagedPoolStart + (i << PAGE_SHIFT));

        //
        // Flush the TLB
        //
        KeFlushEntireTb(TRUE, TRUE);

        /* Setup a demand-zero writable PTE */
        MI_MAKE_SOFTWARE_PTE(&TempPte, MM_READWRITE);

        //
        // Find the first and last PTE, then loop them all
        //
        PointerPte = MiAddressToPte(BaseVa);
        StartPte = PointerPte + SizeInPages;
        do
        {
            //
            // Write the demand zero PTE and keep going
            //
            MI_WRITE_INVALID_PTE(PointerPte, TempPte);
        } while (++PointerPte < StartPte);

        //
        // Return the allocation address to the caller
        //
        return BaseVa;
    }

    //
    // If only one page is being requested, try to grab it from the S-LIST
    //
    if ((SizeInPages == 1) && (ExQueryDepthSList(&MiNonPagedPoolSListHead)))
    {
        BaseVa = InterlockedPopEntrySList(&MiNonPagedPoolSListHead);
        if (BaseVa) return BaseVa;
    }

    //
    // Allocations of less than 4 pages go into their individual buckets
    //
    i = SizeInPages - 1;
    if (i >= MI_MAX_FREE_PAGE_LISTS) i = MI_MAX_FREE_PAGE_LISTS - 1;

    //
    // Loop through all the free page lists based on the page index
    //
    NextHead = &MmNonPagedPoolFreeListHead[i];
    LastHead = &MmNonPagedPoolFreeListHead[MI_MAX_FREE_PAGE_LISTS];

    //
    // Acquire the nonpaged pool lock
    //
    OldIrql = KeAcquireQueuedSpinLock(LockQueueMmNonPagedPoolLock);
    do
    {
        //
        // Now loop through all the free page entries in this given list
        //
        NextEntry = NextHead->Flink;
        while (NextEntry != NextHead)
        {
            /* Is freed non paged pool enabled */
            if (MmProtectFreedNonPagedPool)
            {
                /* We need to be able to touch this page, unprotect it */
                MiUnProtectFreeNonPagedPool(NextEntry, 0);
            }

            //
            // Grab the entry and see if it can handle our allocation
            //
            FreeEntry = CONTAINING_RECORD(NextEntry, MMFREE_POOL_ENTRY, List);
            ASSERT(FreeEntry->Signature == MM_FREE_POOL_SIGNATURE);
            if (FreeEntry->Size >= SizeInPages)
            {
                //
                // It does, so consume the pages from here
                //
                FreeEntry->Size -= SizeInPages;

                //
                // The allocation will begin in this free page area
                //
                BaseVa = (PVOID)((ULONG_PTR)FreeEntry +
                                 (FreeEntry->Size  << PAGE_SHIFT));

                /* Remove the item from the list, depending if pool is protected */
                if (MmProtectFreedNonPagedPool)
                    MiProtectedPoolRemoveEntryList(&FreeEntry->List);
                else
                    RemoveEntryList(&FreeEntry->List);

                //
                // However, check if its' still got space left
                //
                if (FreeEntry->Size != 0)
                {
                    /* Check which list to insert this entry into */
                    i = FreeEntry->Size - 1;
                    if (i >= MI_MAX_FREE_PAGE_LISTS) i = MI_MAX_FREE_PAGE_LISTS - 1;

                    /* Insert the entry into the free list head, check for prot. pool */
                    if (MmProtectFreedNonPagedPool)
                        MiProtectedPoolInsertList(&MmNonPagedPoolFreeListHead[i], &FreeEntry->List, TRUE);
                    else
                        InsertTailList(&MmNonPagedPoolFreeListHead[i], &FreeEntry->List);

                    /* Is freed non paged pool protected? */
                    if (MmProtectFreedNonPagedPool)
                    {
                        /* Protect the freed pool! */
                        MiProtectFreeNonPagedPool(FreeEntry, FreeEntry->Size);
                    }
                }

                //
                // Grab the PTE for this allocation
                //
                PointerPte = MiAddressToPte(BaseVa);
                ASSERT(PointerPte->u.Hard.Valid == 1);

                //
                // Grab the PFN NextEntry and index
                //
                Pfn1 = MiGetPfnEntry(PFN_FROM_PTE(PointerPte));

                //
                // Now mark it as the beginning of an allocation
                //
                ASSERT(Pfn1->u3.e1.StartOfAllocation == 0);
                Pfn1->u3.e1.StartOfAllocation = 1;

                /* Mark it as special pool if needed */
                ASSERT(Pfn1->u4.VerifierAllocation == 0);
                if (PoolType & VERIFIER_POOL_MASK)
                {
                    Pfn1->u4.VerifierAllocation = 1;
                }

                //
                // Check if the allocation is larger than one page
                //
                if (SizeInPages != 1)
                {
                    //
                    // Navigate to the last PFN entry and PTE
                    //
                    PointerPte += SizeInPages - 1;
                    ASSERT(PointerPte->u.Hard.Valid == 1);
                    Pfn1 = MiGetPfnEntry(PointerPte->u.Hard.PageFrameNumber);
                }

                //
                // Mark this PFN as the last (might be the same as the first)
                //
                ASSERT(Pfn1->u3.e1.EndOfAllocation == 0);
                Pfn1->u3.e1.EndOfAllocation = 1;

                //
                // Release the nonpaged pool lock, and return the allocation
                //
                KeReleaseQueuedSpinLock(LockQueueMmNonPagedPoolLock, OldIrql);
                return BaseVa;
            }

            //
            // Try the next free page entry
            //
            NextEntry = FreeEntry->List.Flink;

            /* Is freed non paged pool protected? */
            if (MmProtectFreedNonPagedPool)
            {
                /* Protect the freed pool! */
                MiProtectFreeNonPagedPool(FreeEntry, FreeEntry->Size);
            }
        }
    } while (++NextHead < LastHead);

    //
    // If we got here, we're out of space.
    // Start by releasing the lock
    //
    KeReleaseQueuedSpinLock(LockQueueMmNonPagedPoolLock, OldIrql);

    //
    // Allocate some system PTEs
    //
    StartPte = MiReserveSystemPtes(SizeInPages, NonPagedPoolExpansion);
    PointerPte = StartPte;
    if (StartPte == NULL)
    {
        //
        // Ran out of memory
        //
        DPRINT1("Out of NP Expansion Pool\n");
        return NULL;
    }

    //
    // Acquire the pool lock now
    //
    OldIrql = KeAcquireQueuedSpinLock(LockQueueMmNonPagedPoolLock);

    //
    // Lock the PFN database too
    //
    LockQueue = &KeGetCurrentPrcb()->LockQueue[LockQueuePfnLock];
    KeAcquireQueuedSpinLockAtDpcLevel(LockQueue);

    //
    // Loop the pages
    //
    TempPte = ValidKernelPte;
    do
    {
        /* Allocate a page */
        MI_SET_USAGE(MI_USAGE_PAGED_POOL);
        MI_SET_PROCESS2("Kernel");
        PageFrameNumber = MiRemoveAnyPage(MI_GET_NEXT_COLOR());

        /* Get the PFN entry for it and fill it out */
        Pfn1 = MiGetPfnEntry(PageFrameNumber);
        Pfn1->u3.e2.ReferenceCount = 1;
        Pfn1->u2.ShareCount = 1;
        Pfn1->PteAddress = PointerPte;
        Pfn1->u3.e1.PageLocation = ActiveAndValid;
        Pfn1->u4.VerifierAllocation = 0;

        /* Write the PTE for it */
        TempPte.u.Hard.PageFrameNumber = PageFrameNumber;
        MI_WRITE_VALID_PTE(PointerPte++, TempPte);
    } while (--SizeInPages > 0);

    //
    // This is the last page
    //
    Pfn1->u3.e1.EndOfAllocation = 1;

    //
    // Get the first page and mark it as such
    //
    Pfn1 = MiGetPfnEntry(StartPte->u.Hard.PageFrameNumber);
    Pfn1->u3.e1.StartOfAllocation = 1;

    /* Mark it as a verifier allocation if needed */
    ASSERT(Pfn1->u4.VerifierAllocation == 0);
    if (PoolType & VERIFIER_POOL_MASK) Pfn1->u4.VerifierAllocation = 1;

    //
    // Release the PFN and nonpaged pool lock
    //
    KeReleaseQueuedSpinLockFromDpcLevel(LockQueue);
    KeReleaseQueuedSpinLock(LockQueueMmNonPagedPoolLock, OldIrql);

    //
    // Return the address
    //
    return MiPteToAddress(StartPte);
}
Пример #2
0
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);
    }