VOID
CcPrepareMdlWrite (
IN PFILE_OBJECT FileObject,
IN PLARGE_INTEGER FileOffset,
IN ULONG Length,
OUT PMDL *MdlChain,
OUT PIO_STATUS_BLOCK IoStatus
)
/*++
Routine Description:
This routine attempts to lock the specified file data in the cache
and return a description of it in an Mdl along with the correct
I/O status. Pages to be completely overwritten may be satisfied
with emtpy pages. It is *not* safe to call this routine from Dpc level.
This call is synchronous and raises on error.
When this call returns, the caller may immediately begin
to transfer data into the buffers via the Mdl.
When the call returns with TRUE, the pages described by the Mdl are
locked in memory, but not mapped in system space. If the caller
needs the pages mapped in system space, then it must map them.
On the subsequent call to CcMdlWriteComplete the pages will be
unmapped if they were mapped, and in any case unlocked and the Mdl
deallocated.
Arguments:
FileObject - Pointer to the file object for a file which was
opened with NO_INTERMEDIATE_BUFFERING clear, i.e., for
which CcInitializeCacheMap was called by the file system.
FileOffset - Byte offset in file for desired data.
Length - Length of desired data in bytes.
MdlChain - On output it returns a pointer to an Mdl chain describing
the desired data. Note that even if FALSE is returned,
one or more Mdls may have been allocated, as may be ascertained
by the IoStatus.Information field (see below).
IoStatus - Pointer to standard I/O status block to receive the status
for the in-transfer of the data. (STATUS_SUCCESS guaranteed
for cache hits, otherwise the actual I/O status is returned.)
The I/O Information Field indicates how many bytes have been
successfully locked down in the Mdl Chain.
Return Value:
None
--*/
{
PSHARED_CACHE_MAP SharedCacheMap;
PVOID CacheBuffer;
LARGE_INTEGER FOffset;
PMDL Mdl = NULL;
PMDL MdlTemp;
LARGE_INTEGER Temp;
ULONG SavedState = 0;
ULONG ZeroFlags = 0;
ULONG Information = 0;
KIRQL OldIrql;
ULONG ActivePage;
ULONG PageIsDirty;
PVACB Vacb = NULL;
DebugTrace(+1, me, "CcPrepareMdlWrite\n", 0 );
DebugTrace( 0, me, " FileObject = %08lx\n", FileObject );
DebugTrace2(0, me, " FileOffset = %08lx, %08lx\n", FileOffset->LowPart,
FileOffset->HighPart );
DebugTrace( 0, me, " Length = %08lx\n", Length );
//
// Get pointer to SharedCacheMap.
//
SharedCacheMap = FileObject->SectionObjectPointer->SharedCacheMap;
//
// See if we have an active Vacb, that we need to free.
//
GetActiveVacb( SharedCacheMap, OldIrql, Vacb, ActivePage, PageIsDirty );
//
// If there is an end of a page to be zeroed, then free that page now,
// so it does not cause our data to get zeroed. If there is an active
// page, free it so we have the correct ValidDataGoal.
//
if ((Vacb != NULL) || (SharedCacheMap->NeedToZero != NULL)) {
CcFreeActiveVacb( SharedCacheMap, Vacb, ActivePage, PageIsDirty );
Vacb = NULL;
}
FOffset = *FileOffset;
//
// Put try-finally around the loop to deal with exceptions
//
try {
//
// Not all of the transfer will come back at once, so we have to loop
// until the entire transfer is complete.
//
while (Length != 0) {
ULONG ReceivedLength;
LARGE_INTEGER BeyondLastByte;
//
// Map and see how much we could potentially access at this
// FileOffset, then cut it down if it is more than we need.
//
CacheBuffer = CcGetVirtualAddress( SharedCacheMap,
FOffset,
&Vacb,
&ReceivedLength );
if (ReceivedLength > Length) {
ReceivedLength = Length;
}
BeyondLastByte.QuadPart = FOffset.QuadPart + (LONGLONG)ReceivedLength;
//
// At this point we can calculate the ZeroFlags.
//
//
// We can always zero middle pages, if any.
//
ZeroFlags = ZERO_MIDDLE_PAGES;
//
// See if we are completely overwriting the first or last page.
//
if (((FOffset.LowPart & (PAGE_SIZE - 1)) == 0) &&
(ReceivedLength >= PAGE_SIZE)) {
ZeroFlags |= ZERO_FIRST_PAGE;
}
if ((BeyondLastByte.LowPart & (PAGE_SIZE - 1)) == 0) {
ZeroFlags |= ZERO_LAST_PAGE;
}
//
// See if the entire transfer is beyond valid data length,
// or at least starting from the second page.
//
Temp = FOffset;
Temp.LowPart &= ~(PAGE_SIZE -1);
ExAcquireFastLock( &SharedCacheMap->BcbSpinLock, &OldIrql );
Temp.QuadPart = SharedCacheMap->ValidDataGoal.QuadPart - Temp.QuadPart;
ExReleaseFastLock( &SharedCacheMap->BcbSpinLock, OldIrql );
if (Temp.QuadPart <= 0) {
ZeroFlags |= ZERO_FIRST_PAGE | ZERO_MIDDLE_PAGES | ZERO_LAST_PAGE;
} else if ((Temp.HighPart == 0) && (Temp.LowPart <= PAGE_SIZE)) {
ZeroFlags |= ZERO_MIDDLE_PAGES | ZERO_LAST_PAGE;
}
(VOID)CcMapAndRead( SharedCacheMap,
&FOffset,
ReceivedLength,
ZeroFlags,
TRUE,
CacheBuffer );
//
// Now attempt to allocate an Mdl to describe the mapped data.
//
DebugTrace( 0, mm, "IoAllocateMdl:\n", 0 );
DebugTrace( 0, mm, " BaseAddress = %08lx\n", CacheBuffer );
DebugTrace( 0, mm, " Length = %08lx\n", ReceivedLength );
Mdl = IoAllocateMdl( CacheBuffer,
ReceivedLength,
FALSE,
FALSE,
NULL );
DebugTrace( 0, mm, " <Mdl = %08lx\n", Mdl );
if (Mdl == NULL) {
DebugTrace( 0, 0, "Failed to allocate Mdl\n", 0 );
ExRaiseStatus( STATUS_INSUFFICIENT_RESOURCES );
}
DebugTrace( 0, mm, "MmProbeAndLockPages:\n", 0 );
DebugTrace( 0, mm, " Mdl = %08lx\n", Mdl );
MmDisablePageFaultClustering(&SavedState);
MmProbeAndLockPages( Mdl, KernelMode, IoWriteAccess );
MmEnablePageFaultClustering(SavedState);
SavedState = 0;
//
// Now that some data (maybe zeros) is locked in memory and
// set dirty, it is safe, and necessary for us to advance
// valid data goal, so that we will not subsequently ask
// for a zero page. Note if we are extending valid data,
// our caller has the file exclusive.
//
ExAcquireFastLock( &SharedCacheMap->BcbSpinLock, &OldIrql );
if (BeyondLastByte.QuadPart > SharedCacheMap->ValidDataGoal.QuadPart) {
SharedCacheMap->ValidDataGoal = BeyondLastByte;
}
ExReleaseFastLock( &SharedCacheMap->BcbSpinLock, OldIrql );
//
// Unmap the data now, now that the pages are locked down.
//
CcFreeVirtualAddress( Vacb );
Vacb = NULL;
//
// Now link the Mdl into the caller's chain
//
if ( *MdlChain == NULL ) {
*MdlChain = Mdl;
} else {
MdlTemp = CONTAINING_RECORD( *MdlChain, MDL, Next );
while (MdlTemp->Next != NULL) {
MdlTemp = MdlTemp->Next;
}
MdlTemp->Next = Mdl;
}
Mdl = NULL;
//
// Assume we did not get all the data we wanted, and set FOffset
// to the end of the returned data.
//
FOffset = BeyondLastByte;
//
// Update number of bytes transferred.
//
Information += ReceivedLength;
//
// Calculate length left to transfer.
//
Length -= ReceivedLength;
}
}
finally {
if (AbnormalTermination()) {
if (SavedState != 0) {
MmEnablePageFaultClustering(SavedState);
}
if (Vacb != NULL) {
CcFreeVirtualAddress( Vacb );
}
if (Mdl != NULL) {
IoFreeMdl( Mdl );
}
//
// Otherwise loop to deallocate the Mdls
//
FOffset = *FileOffset;
while (*MdlChain != NULL) {
MdlTemp = (*MdlChain)->Next;
DebugTrace( 0, mm, "MmUnlockPages/IoFreeMdl:\n", 0 );
DebugTrace( 0, mm, " Mdl = %08lx\n", *MdlChain );
MmUnlockPages( *MdlChain );
//
// Extract the File Offset for this part of the transfer, and
// tell the lazy writer to write these pages, since we have
// marked them dirty. Ignore the only exception (allocation
// error), and console ourselves for having tried.
//
CcSetDirtyInMask( SharedCacheMap, &FOffset, (*MdlChain)->ByteCount );
FOffset.QuadPart = FOffset.QuadPart + (LONGLONG)((*MdlChain)->ByteCount);
IoFreeMdl( *MdlChain );
*MdlChain = MdlTemp;
}
DebugTrace(-1, me, "CcPrepareMdlWrite -> Unwinding\n", 0 );
}
else {
IoStatus->Status = STATUS_SUCCESS;
IoStatus->Information = Information;
//
// Make sure the SharedCacheMap does not go away while
// the Mdl write is in progress. We decrment below.
//
CcAcquireMasterLock( &OldIrql );
CcIncrementOpenCount( SharedCacheMap, 'ldmP' );
CcReleaseMasterLock( OldIrql );
}
}
DebugTrace( 0, me, " <MdlChain = %08lx\n", *MdlChain );
DebugTrace(-1, me, "CcPrepareMdlWrite -> VOID\n", 0 );
return;
}
NTKERNELAPI
BOOLEAN
CcSetPrivateWriteFile(
PFILE_OBJECT FileObject
)
/*++
Routine Description:
This routine will instruct the cache manager to treat the file as
a private-write stream, so that a caller can implement a private
logging mechanism for it. We will turn on both Mm's modify-no-write
and our disable-write-behind, and disallow non-aware flush/purge for
the file.
Caching must already be initiated on the file.
This routine is only exported to the kernel.
Arguments:
FileObject - File to make private-write.
Return Value:
None.
--*/
{
PSHARED_CACHE_MAP SharedCacheMap;
BOOLEAN Disabled;
KIRQL OldIrql;
PVACB Vacb;
ULONG ActivePage;
ULONG PageIsDirty;
//
// Pick up the file exclusive to synchronize against readahead and
// other purge/map activity.
//
FsRtlAcquireFileExclusive( FileObject );
//
// Get a pointer to the SharedCacheMap. Be sure to release the FileObject
// in case an error condition forces a premature exit.
//
if ((FileObject->SectionObjectPointer == NULL) ||
((SharedCacheMap = FileObject->SectionObjectPointer->SharedCacheMap) == NULL)){
FsRtlReleaseFile( FileObject );
return FALSE;
}
//
// Unmap all the views in preparation for making the disable mw call.
//
//
// We still need to wait for any dangling cache read or writes.
//
// In fact we have to loop and wait because the lazy writer can
// sneak in and do an CcGetVirtualAddressIfMapped, and we are not
// synchronized.
//
// This is the same bit of code that our purge will do. We assume
// that a private writer has succesfully blocked out other activity.
//
//
// If there is an active Vacb, then delete it now (before waiting!).
//
CcAcquireMasterLock( &OldIrql );
GetActiveVacbAtDpcLevel( SharedCacheMap, Vacb, ActivePage, PageIsDirty );
CcReleaseMasterLock( OldIrql );
if (Vacb != NULL) {
CcFreeActiveVacb( SharedCacheMap, Vacb, ActivePage, PageIsDirty );
}
while ((SharedCacheMap->Vacbs != NULL) &&
!CcUnmapVacbArray( SharedCacheMap, NULL, 0, FALSE )) {
CcWaitOnActiveCount( SharedCacheMap );
}
//
// Knock the file down.
//
CcFlushCache( FileObject->SectionObjectPointer, NULL, 0, NULL );
//
// Now the file is clean and unmapped. We can still have a racing
// lazy writer, though.
//
// We just wait for the lazy writer queue to drain before disabling
// modified write. There may be a better way to do this by having
// an event for the WRITE_QUEUED flag. ? This would also let us
// dispense with the pagingio pick/drop in the FS cache coherency
// paths, but there could be reasons why CcFlushCache shouldn't
// always do such a block. Investigate this.
//
// This wait takes on the order of ~.5s avg. case.
//
CcAcquireMasterLock( &OldIrql );
if (FlagOn( SharedCacheMap->Flags, WRITE_QUEUED ) ||
FlagOn( SharedCacheMap->Flags, READ_AHEAD_QUEUED )) {
CcReleaseMasterLock( OldIrql );
FsRtlReleaseFile( FileObject );
CcWaitForCurrentLazyWriterActivity();
FsRtlAcquireFileExclusive( FileObject );
} else {
CcReleaseMasterLock( OldIrql );
}
//
// Now set the flags and return. We do not set our MODIFIED_WRITE_DISABLED
// since we don't want to fully promote this cache map. Future?
//
Disabled = MmDisableModifiedWriteOfSection( FileObject->SectionObjectPointer );
if (Disabled) {
CcAcquireMasterLock( &OldIrql );
SetFlag(SharedCacheMap->Flags, DISABLE_WRITE_BEHIND | PRIVATE_WRITE);
CcReleaseMasterLock( OldIrql );
}
//
// Now release the file for regular operation.
//
FsRtlReleaseFile( FileObject );
return Disabled;
}
VOID
CcMdlRead (
IN PFILE_OBJECT FileObject,
IN PLARGE_INTEGER FileOffset,
IN ULONG Length,
OUT PMDL *MdlChain,
OUT PIO_STATUS_BLOCK IoStatus
)
/*++
Routine Description:
This routine attempts to lock the specified file data in the cache
and return a description of it in an Mdl along with the correct
I/O status. It is *not* safe to call this routine from Dpc level.
This routine is synchronous, and raises on errors.
As each call returns, the pages described by the Mdl are
locked in memory, but not mapped in system space. If the caller
needs the pages mapped in system space, then it must map them.
Note that each call is a "single shot" which should be followed by
a call to CcMdlReadComplete. To resume an Mdl-based transfer, the
caller must form one or more subsequent calls to CcMdlRead with
appropriately adjusted parameters.
Arguments:
FileObject - Pointer to the file object for a file which was
opened with NO_INTERMEDIATE_BUFFERING clear, i.e., for
which CcInitializeCacheMap was called by the file system.
FileOffset - Byte offset in file for desired data.
Length - Length of desired data in bytes.
MdlChain - On output it returns a pointer to an Mdl chain describing
the desired data. Note that even if FALSE is returned,
one or more Mdls may have been allocated, as may be ascertained
by the IoStatus.Information field (see below).
IoStatus - Pointer to standard I/O status block to receive the status
for the transfer. (STATUS_SUCCESS guaranteed for cache
hits, otherwise the actual I/O status is returned.) The
I/O Information Field indicates how many bytes have been
successfully locked down in the Mdl Chain.
Return Value:
None
Raises:
STATUS_INSUFFICIENT_RESOURCES - If a pool allocation failure occurs.
--*/
{
PSHARED_CACHE_MAP SharedCacheMap;
PPRIVATE_CACHE_MAP PrivateCacheMap;
PVOID CacheBuffer;
LARGE_INTEGER FOffset;
PMDL Mdl = NULL;
PMDL MdlTemp;
PETHREAD Thread = PsGetCurrentThread();
ULONG SavedState = 0;
ULONG OriginalLength = Length;
ULONG Information = 0;
PVACB Vacb = NULL;
ULONG SavedMissCounter = 0;
KIRQL OldIrql;
ULONG ActivePage;
ULONG PageIsDirty;
PVACB ActiveVacb = NULL;
DebugTrace(+1, me, "CcMdlRead\n", 0 );
DebugTrace( 0, me, " FileObject = %08lx\n", FileObject );
DebugTrace2(0, me, " FileOffset = %08lx, %08lx\n", FileOffset->LowPart,
FileOffset->HighPart );
DebugTrace( 0, me, " Length = %08lx\n", Length );
//
// Save the current readahead hints.
//
MmSavePageFaultReadAhead( Thread, &SavedState );
//
// Get pointer to SharedCacheMap.
//
SharedCacheMap = FileObject->SectionObjectPointer->SharedCacheMap;
PrivateCacheMap = FileObject->PrivateCacheMap;
//
// See if we have an active Vacb, that we need to free.
//
GetActiveVacb( SharedCacheMap, OldIrql, ActiveVacb, ActivePage, PageIsDirty );
//
// If there is an end of a page to be zeroed, then free that page now,
// so we don't send Greg the uninitialized data...
//
if ((ActiveVacb != NULL) || (SharedCacheMap->NeedToZero != NULL)) {
CcFreeActiveVacb( SharedCacheMap, ActiveVacb, ActivePage, PageIsDirty );
}
//
// If read ahead is enabled, then do the read ahead here so it
// overlaps with the copy (otherwise we will do it below).
// Note that we are assuming that we will not get ahead of our
// current transfer - if read ahead is working it should either
// already be in memory or else underway.
//
if (PrivateCacheMap->ReadAheadEnabled && (PrivateCacheMap->ReadAheadLength[1] == 0)) {
CcScheduleReadAhead( FileObject, FileOffset, Length );
}
//
// Increment performance counters
//
CcMdlReadWait += 1;
//
// This is not an exact solution, but when IoPageRead gets a miss,
// it cannot tell whether it was CcCopyRead or CcMdlRead, but since
// the miss should occur very soon, by loading the pointer here
// probably the right counter will get incremented, and in any case,
// we hope the errrors average out!
//
CcMissCounter = &CcMdlReadWaitMiss;
FOffset = *FileOffset;
//
// Check for read past file size, the caller must filter this case out.
//
ASSERT( ( FOffset.QuadPart + (LONGLONG)Length ) <= SharedCacheMap->FileSize.QuadPart );
//
// Put try-finally around the loop to deal with any exceptions
//
try {
//
// Not all of the transfer will come back at once, so we have to loop
// until the entire transfer is complete.
//
while (Length != 0) {
ULONG ReceivedLength;
LARGE_INTEGER BeyondLastByte;
//
// Map the data and read it in (if necessary) with the
// MmProbeAndLockPages call below.
//
CacheBuffer = CcGetVirtualAddress( SharedCacheMap,
FOffset,
&Vacb,
&ReceivedLength );
if (ReceivedLength > Length) {
ReceivedLength = Length;
}
BeyondLastByte.QuadPart = FOffset.QuadPart + (LONGLONG)ReceivedLength;
//
// Now attempt to allocate an Mdl to describe the mapped data.
//
DebugTrace( 0, mm, "IoAllocateMdl:\n", 0 );
DebugTrace( 0, mm, " BaseAddress = %08lx\n", CacheBuffer );
DebugTrace( 0, mm, " Length = %08lx\n", ReceivedLength );
Mdl = IoAllocateMdl( CacheBuffer,
ReceivedLength,
FALSE,
FALSE,
NULL );
DebugTrace( 0, mm, " <Mdl = %08lx\n", Mdl );
if (Mdl == NULL) {
DebugTrace( 0, 0, "Failed to allocate Mdl\n", 0 );
ExRaiseStatus( STATUS_INSUFFICIENT_RESOURCES );
}
DebugTrace( 0, mm, "MmProbeAndLockPages:\n", 0 );
DebugTrace( 0, mm, " Mdl = %08lx\n", Mdl );
//
// Set to see if the miss counter changes in order to
// detect when we should turn on read ahead.
//
SavedMissCounter += CcMdlReadWaitMiss;
MmSetPageFaultReadAhead( Thread, COMPUTE_PAGES_SPANNED( CacheBuffer, ReceivedLength ) - 1);
MmProbeAndLockPages( Mdl, KernelMode, IoReadAccess );
SavedMissCounter -= CcMdlReadWaitMiss;
//
// Unmap the data now, now that the pages are locked down.
//
CcFreeVirtualAddress( Vacb );
Vacb = NULL;
//
// Now link the Mdl into the caller's chain
//
if ( *MdlChain == NULL ) {
*MdlChain = Mdl;
} else {
MdlTemp = CONTAINING_RECORD( *MdlChain, MDL, Next );
while (MdlTemp->Next != NULL) {
MdlTemp = MdlTemp->Next;
}
MdlTemp->Next = Mdl;
}
Mdl = NULL;
//
// Assume we did not get all the data we wanted, and set FOffset
// to the end of the returned data.
//
FOffset = BeyondLastByte;
//
// Update number of bytes transferred.
//
Information += ReceivedLength;
//
// Calculate length left to transfer.
//
Length -= ReceivedLength;
}
}
finally {
CcMissCounter = &CcThrowAway;
//
// Restore the readahead hints.
//
MmResetPageFaultReadAhead( Thread, SavedState );
if (AbnormalTermination()) {
//
// We may have failed to allocate an Mdl while still having
// data mapped.
//
if (Vacb != NULL) {
CcFreeVirtualAddress( Vacb );
}
if (Mdl != NULL) {
IoFreeMdl( Mdl );
}
//
// Otherwise loop to deallocate the Mdls
//
while (*MdlChain != NULL) {
MdlTemp = (*MdlChain)->Next;
DebugTrace( 0, mm, "MmUnlockPages/IoFreeMdl:\n", 0 );
DebugTrace( 0, mm, " Mdl = %08lx\n", *MdlChain );
MmUnlockPages( *MdlChain );
IoFreeMdl( *MdlChain );
*MdlChain = MdlTemp;
}
DebugTrace(-1, me, "CcMdlRead -> Unwinding\n", 0 );
}
else {
//
// Now enable read ahead if it looks like we got any misses, and do
// the first one.
//
if (!FlagOn( FileObject->Flags, FO_RANDOM_ACCESS ) &&
!PrivateCacheMap->ReadAheadEnabled &&
(SavedMissCounter != 0)) {
PrivateCacheMap->ReadAheadEnabled = TRUE;
CcScheduleReadAhead( FileObject, FileOffset, OriginalLength );
}
//
// Now that we have described our desired read ahead, let's
// shift the read history down.
//
PrivateCacheMap->FileOffset1 = PrivateCacheMap->FileOffset2;
PrivateCacheMap->BeyondLastByte1 = PrivateCacheMap->BeyondLastByte2;
PrivateCacheMap->FileOffset2 = *FileOffset;
PrivateCacheMap->BeyondLastByte2.QuadPart =
FileOffset->QuadPart + (LONGLONG)OriginalLength;
IoStatus->Status = STATUS_SUCCESS;
IoStatus->Information = Information;
}
}
DebugTrace( 0, me, " <MdlChain = %08lx\n", *MdlChain );
DebugTrace2(0, me, " <IoStatus = %08lx, %08lx\n", IoStatus->Status,
IoStatus->Information );
DebugTrace(-1, me, "CcMdlRead -> VOID\n", 0 );
return;
}
