You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
218 lines
6.7 KiB
218 lines
6.7 KiB
/*++
|
|
|
|
Copyright (c) 1998-2000 Microsoft Corporation
|
|
|
|
Module Name :
|
|
readmost.hxx
|
|
|
|
Abstract:
|
|
Read-mostly Data Cache
|
|
|
|
Author:
|
|
George V. Reilly (GeorgeRe) 14-Sep-1998
|
|
(from an idea by Neel Jain)
|
|
|
|
Environment:
|
|
Win32 - User Mode
|
|
|
|
Project:
|
|
Internet Information Server RunTime Library
|
|
|
|
Revision History:
|
|
|
|
--*/
|
|
|
|
#ifndef __READMOST_HXX__
|
|
#define __READMOST_HXX__
|
|
|
|
#include "irtlmisc.h"
|
|
|
|
//=====================================================================
|
|
// Locks are expensive and they kill concurrency on multiprocessor
|
|
// systems. CDataCache<_T> is a lock-free cache that is suitable for
|
|
// "read-mostly" data structures; i.e., data structures that are hardly
|
|
// ever updated. We use a monotonically increasing sequence number to
|
|
// version stamp the data in the cache. Whenever the data is altered
|
|
// (which can only happen through the Write() method), the version number
|
|
// is updated. For a Read(), if the version number is the same both
|
|
// before and after the data itself is copied into an out parameter, then
|
|
// the Read() obtained a valid copy of the data.
|
|
//=====================================================================
|
|
|
|
|
|
// Use a portable implementation with interlocked routines that doesn't
|
|
// rely on processor-specific memory barrier intrinsics?
|
|
#undef READMOST_INTERLOCKED
|
|
|
|
|
|
#ifndef READMOST_INTERLOCKED
|
|
#if defined(_M_IA64)
|
|
extern "C" void __mf(void);
|
|
#pragma intrinsic(__mf)
|
|
#endif // _M_IA64
|
|
#endif // !READMOST_INTERLOCKED
|
|
|
|
|
|
template <class _T>
|
|
class IRTL_DLLEXP CDataCache
|
|
{
|
|
protected:
|
|
// Place the cached data first to preserve its alignment constraints.
|
|
volatile _T m_tData;
|
|
|
|
// Mark the sequence number (version stamp) as volatile to ensure that
|
|
// the compiler doesn't cache its value in a register. Mark it as mutable
|
|
// so that we can use the Interlocked operations on the sequence
|
|
// number in const member functions.
|
|
mutable volatile LONG m_nSequence;
|
|
|
|
enum {
|
|
UPDATING = 0xffffffff, // out-of-band odd value => cache is invalid
|
|
INITIAL = UPDATING + 1, // even value
|
|
STEP = 2, // ensures m_nSequence will never == UPDATING
|
|
BOGUS = UPDATING + STEP,// impossible value, never used
|
|
};
|
|
|
|
|
|
#ifdef READMOST_INTERLOCKED
|
|
|
|
LONG
|
|
_ReadSequence() const
|
|
{
|
|
// Since m_nSequence will never be equal to BOGUS, this
|
|
// will atomically read the value of m_nSequence, but not
|
|
// modify it. On architectures that need such things, it
|
|
// will have the side effect of erecting a read memory
|
|
// barrier both before and after reading the value of m_nSequence.
|
|
return InterlockedCompareExchange((LONG*) &m_nSequence, BOGUS, BOGUS);
|
|
}
|
|
|
|
#else // !READMOST_INTERLOCKED
|
|
|
|
// On some systems, such as Alphas and Itaniums, the compiler or
|
|
// processor can issue out-of-order (speculative) reads and writes.
|
|
// _ReadMemoryBarrier() and _WriteMemoryBarrier() force serialization
|
|
// of memory accesses.
|
|
|
|
static void
|
|
_ReadMemoryBarrier()
|
|
{
|
|
#if defined(_M_IA64)
|
|
__mf();
|
|
#endif // _M_IA64
|
|
}
|
|
|
|
// Read the value of m_nSequence, imposing memory barriers
|
|
// both before and after reading m_nSequence.
|
|
LONG
|
|
_ReadSequence() const
|
|
{
|
|
_ReadMemoryBarrier();
|
|
const LONG nSequence = m_nSequence;
|
|
_ReadMemoryBarrier();
|
|
return nSequence;
|
|
}
|
|
|
|
// Not currently used, as we rely on InterlockedExchange in
|
|
// _SetSequence to do the right thing with write memory barriers.
|
|
static void
|
|
_WriteMemoryBarrier()
|
|
{
|
|
#if defined(_M_IA64)
|
|
__mf();
|
|
#endif // _M_IA64
|
|
}
|
|
|
|
#endif // !READMOST_INTERLOCKED
|
|
|
|
// Update m_nSequence, returning its old value. InterlockedExchange
|
|
// has the side effect of erecting a write memory barrier both
|
|
// before and after updating m_nSequence.
|
|
LONG
|
|
_SetSequence(
|
|
LONG nNewValue)
|
|
{
|
|
return InterlockedExchange((LONG*) &m_nSequence, nNewValue);
|
|
}
|
|
|
|
|
|
public:
|
|
// Default ctor. Rely on _T::_T() to do something useful.
|
|
CDataCache()
|
|
: m_nSequence(INITIAL)
|
|
{}
|
|
|
|
|
|
// Ctor.
|
|
CDataCache(const _T& t)
|
|
: m_tData(t), m_nSequence(INITIAL)
|
|
{}
|
|
|
|
|
|
// Read the contents of the cache into rtOut. Returns `true' if
|
|
// successful, `false' otherwise (in which case rtOut is garbage).
|
|
// You should retry if Read() returns `false'.
|
|
bool
|
|
Read(
|
|
_T& rtOut) const
|
|
{
|
|
const LONG nSequence1 = _ReadSequence();
|
|
|
|
// Is the data being updated on another thread?
|
|
if (nSequence1 != UPDATING)
|
|
{
|
|
// No, so read the data into rtOut.
|
|
|
|
// The weird const_cast syntax is necessitated by the volatile
|
|
// attribute on m_tData.
|
|
rtOut = * const_cast<_T*>(&m_tData);
|
|
|
|
// If the sequence number is unchanged, the read was valid.
|
|
const LONG nSequence2 = _ReadSequence();
|
|
|
|
return (nSequence1 == nSequence2);
|
|
}
|
|
|
|
// Another thread was updating the cache, so Read failed.
|
|
// The caller should probably retry.
|
|
return false;
|
|
}
|
|
|
|
|
|
// Updates the contents of the cache. Returns `true' if the cache was
|
|
// successfully updated, `false' otherwise (because the cache is already
|
|
// being updated on some other thread).
|
|
bool
|
|
Write(
|
|
const _T& rtIn)
|
|
{
|
|
// Atomically set m_nSequence to UPDATING.
|
|
const LONG nSequence = _SetSequence(UPDATING);
|
|
|
|
// If the old value of m_nSequence was not UPDATING,
|
|
// then we now "own" the cache.
|
|
if (nSequence != UPDATING)
|
|
{
|
|
// Update the cached data. The weird const_cast syntax is
|
|
// necessitated by the volatile attribute on m_tData.
|
|
* const_cast<_T*>(&m_tData) = rtIn;
|
|
|
|
// Finally, update the sequence number. The implicit
|
|
// memory barriers in InterlockedExchange will force
|
|
// the write of m_tData to complete before m_nSequence
|
|
// acquires its new value, and will force the write
|
|
// of m_nSequence to complete before Write() returns.
|
|
_SetSequence(nSequence + STEP);
|
|
|
|
return true;
|
|
}
|
|
|
|
// Another thread already owned the cache, so Write failed.
|
|
// This is probably fine, but that determination must be
|
|
// made by the routine that called Write(), since it
|
|
// understands the semantics of its caching and Write() doesn't.
|
|
return false;
|
|
}
|
|
};
|
|
|
|
#endif // __READMOST_HXX__
|