pkgsrc/x11/kdelibs4/files/kshareddatacache.cpp
2012-03-19 11:28:48 +00:00

1496 lines
54 KiB
C++

/*
* This file is part of the KDE project.
* Copyright © 2010 Michael Pyne <mpyne@kde.org>
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Library General Public
* License version 2 as published by the Free Software Foundation.
*
* This library is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Library General Public License for more details.
*
* You should have received a copy of the GNU Library General Public License
* along with this library; see the file COPYING.LIB. If not, write to
* the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
* Boston, MA 02110-1301, USA.
*/
#include "kshareddatacache.h"
#include <kdebug.h>
#include <kglobal.h>
#include <kstandarddirs.h>
#include <krandom.h>
#include <QtCore/QDateTime>
#include <QtCore/QSharedPointer>
#include <QtCore/QByteArray>
#include <QtCore/QFile>
#include <QtCore/QAtomicInt>
#include <QtCore/QList>
#include <QtCore/QMutex>
#include <QtCore/QMutexLocker>
#include <pthread.h>
#include <sys/types.h>
#include <sys/mman.h>
#include <stdlib.h>
#include <unistd.h>
#include <time.h>
// Mac OS X, for all its POSIX compliance, does not support timeouts on its mutexes, which
// is kind of a disaster for cross-process support. So, disable support for shared memory
// and add a wrapper implementation.
#if !defined(_POSIX_TIMEOUTS) || !defined(_POSIX_THREADS) || (_POSIX_TIMEOUTS < 200112L) || (_POSIX_THREADS < 200112L)
#ifdef _POSIX_THREAD_PROCESS_SHARED
#undef _POSIX_THREAD_PROCESS_SHARED
#endif
#ifdef __GNUC__
#warning "No support for POSIX timeouts and POSIX threads -- shared memory will be ignored"
#endif
#endif
// BSD/Mac OS X compat
#if !defined(MAP_ANONYMOUS) && defined(MAP_ANON)
#define MAP_ANONYMOUS MAP_ANON
#endif
/**
* This is the hash function used for our pixmap data to hopefully make the
* hashing used to place the pixmaps as efficient as possible.
*
* The algorithm is in the public domain, by Glenn Fowler, Phong Vo, and Landon
* Curt Noll, implemented by myself (mpyne).
*/
static quint32 fnvHash32(const QByteArray &buffer)
{
static const quint32 FNVPrime = 16777619;
static const quint32 FNVOffsetBasis = 2166136261;
// uchar for correct arithmetic below
const uchar *base = reinterpret_cast<const uchar *>(buffer.constData());
quint32 result = FNVOffsetBasis;
int count = buffer.length();
while (count-- > 0) {
result *= FNVPrime;
result ^= *base;
base++;
}
return result;
}
// Alignment concerns become a big deal when we're dealing with shared memory,
// since trying to access a structure sized at, say 8 bytes at an address that
// is not evenly divisible by 8 is a crash-inducing error on some
// architectures. The compiler would normally take care of this, but with
// shared memory the compiler will not necessarily know the alignment expected,
// so make sure we account for this ourselves. To do so we need a way to find
// out the expected alignment. Enter ALIGNOF...
#ifndef ALIGNOF
#if defined(Q_CC_GNU) || defined(Q_CC_SUN)
#define ALIGNOF(x) (__alignof__ (x)) // GCC provides what we want directly
#else
#include <stddef.h> // offsetof
template<class T>
struct __alignmentHack
{
char firstEntry;
T obj;
static const size_t size = offsetof(__alignmentHack, obj);
};
#define ALIGNOF(x) (__alignmentHack<x>::size)
#endif // Non gcc
#endif // ALIGNOF undefined
// Returns a pointer properly aligned to handle size alignment.
// size should be a power of 2. start is assumed to be the lowest
// permissible address, therefore the return value will be >= start.
template<class T>
T* alignTo(const void *start, uint size = ALIGNOF(T))
{
quintptr mask = size - 1;
// Cast to int-type to handle bit-twiddling
quintptr basePointer = reinterpret_cast<quintptr>(start);
// If (and only if) we are already aligned, adding mask into basePointer
// will not increment any of the bits in ~mask and we get the right answer.
basePointer = (basePointer + mask) & ~mask;
return reinterpret_cast<T *>(basePointer);
}
/**
* Returns a pointer to a const object of type T, assumed to be @p offset
* *BYTES* greater than the base address. Note that in order to meet alignment
* requirements for T, it is possible that the returned pointer points greater
* than @p offset into @p base.
*/
template<class T>
const T *offsetAs(const void *const base, qint32 offset)
{
const char *ptr = reinterpret_cast<const char*>(base);
return alignTo<const T>(ptr + offset);
}
// Same as above, but for non-const objects
template<class T>
T *offsetAs(void *const base, qint32 offset)
{
char *ptr = reinterpret_cast<char *>(base);
return alignTo<T>(ptr + offset);
}
/**
* @return the smallest integer greater than or equal to (@p a / @p b).
* @param a Numerator, should be ≥ 0.
* @param b Denominator, should be > 0.
*/
template<class T>
T intCeil(T a, T b)
{
return (a + b - 1) / b;
}
typedef qint32 pageID;
// =========================================================================
// Description of the cache:
//
// The shared memory cache is designed to be handled as two separate objects,
// all contained in the same global memory segment. First off, there is the
// basic header data, consisting of the global header followed by the
// accounting data necessary to hold items (described in more detail
// momentarily). Following the accounting data is the start of the "page table"
// (essentially just as you'd see it in an Operating Systems text).
//
// The page table contains shared memory split into fixed-size pages, with a
// configurable page size. In the event that the data is too large to fit into
// a single logical page, it will need to occupy consecutive pages of memory.
//
// The accounting data that was referenced earlier is split into two:
//
// 1. index table, containing a fixed-size list of possible cache entries.
// Each index entry is of type IndexTableEntry (below), and holds the various
// accounting data and a pointer to the first page.
//
// 2. page table, which is used to speed up the process of searching for
// free pages of memory. There is one entry for every page in the page table,
// and it contains the index of the one entry in the index table actually
// holding the page (or <0 if the page is free).
//
// The entire segment looks like so:
// ╔════════╤═════════════╤════════════╦═══════╤═══════╤═══════╤═══════╤═══╗
// ║ Header │ Index Table │ Page Table ║ Pages │ │ │ │...║
// ╚════════╧═════════════╧════════════╩═══════╧═══════╧═══════╧═══════╧═══╝
// =========================================================================
// All elements of this struct must be "plain old data" (POD) types since it
// will be in shared memory. In addition, no pointers! To point to something
// you must use relative offsets since the pointer start addresses will be
// different in each process.
struct IndexTableEntry
{
uint fileNameHash;
uint totalItemSize; // in bytes
mutable uint useCount;
time_t addTime;
mutable time_t lastUsedTime;
pageID firstPage;
};
// Page table entry
struct PageTableEntry
{
// int so we can use values <0 for unassigned pages.
qint32 index;
};
// Each individual page contains the cached data. The first page starts off with
// the utf8-encoded key, a null '\0', and then the data follows immediately
// from the next byte, possibly crossing consecutive page boundaries to hold
// all of the data.
// There is, however, no specific struct for a page, it is simply a location in
// memory.
// This is effectively the layout of the shared memory segment. The variables
// contained within form the header, data contained afterwards is pointed to
// by using special accessor functions.
struct SharedMemory
{
enum {
PIXMAP_CACHE_VERSION = 1,
MINIMUM_CACHE_SIZE = 4096
};
// Note to those who follow me. You should not, under any circumstances, ever
// re-arrange the following two fields, even if you change the version number
// for later revisions of this code.
QAtomicInt ready; ///< DO NOT INITIALIZE
quint8 version;
mutable pthread_mutex_t lockMutex;
uint cacheSize;
uint cacheAvail;
QAtomicInt evictionPolicy;
// pageSize and cacheSize determine the number of pages. The number of
// pages determine the page table size and (indirectly) the index table
// size.
QAtomicInt pageSize;
// This variable is added to reserve space for later cache timestamping
// support. The idea is this variable will be updated when the cache is
// written to, to allow clients to detect a changed cache quickly.
QAtomicInt cacheTimestamp;
/**
* Converts the given average item size into an appropriate page size.
*/
static unsigned equivalentPageSize(unsigned itemSize)
{
if (itemSize == 0) {
return 4096; // Default average item size.
}
int log2OfSize = 0;
while ((itemSize >>= 1) != 0) {
log2OfSize++;
}
// Bound page size between 512 bytes and 256 KiB.
log2OfSize = qBound(9, log2OfSize, 18);
return (1 << log2OfSize);
}
// Returns pageSize in unsigned format.
unsigned cachePageSize() const
{
return static_cast<unsigned>(pageSize);
}
/**
* This is effectively the class ctor. But since we're in shared memory,
* there's a few rules:
*
* 1. To allow for some form of locking in the initial-setup case, we
* use an atomic int, which will be initialized to 0 by mmap(). Then to
* take the lock we atomically increment the 0 to 1. If we end up calling
* the QAtomicInt constructor we can mess that up, so we can't use a
* constructor for this class either.
* 2. Any member variable you add takes up space in shared memory as well,
* so make sure you need it.
*/
bool performInitialSetup(bool processShared, uint _cacheSize, uint _pageSize)
{
if (_cacheSize < MINIMUM_CACHE_SIZE) {
kError(264) << "Internal error: Attempted to create a cache sized < "
<< MINIMUM_CACHE_SIZE;
return false;
}
if (_pageSize == 0) {
kError(264) << "Internal error: Attempted to create a cache with 0-sized pages.";
return false;
}
if (processShared) {
// Perform initialization. We effectively hold a mini-lock right
// now as long as all clients cooperate...
pthread_mutexattr_t mutexAttr;
// Initialize attributes, enable process-shared primitives, and setup
// the mutex.
if (pthread_mutexattr_init(&mutexAttr) == 0) {
if (
#ifdef _POSIX_THREAD_PROCESS_SHARED
pthread_mutexattr_setpshared(&mutexAttr, PTHREAD_PROCESS_SHARED) != 0 ||
#endif
pthread_mutex_init(&lockMutex, &mutexAttr) != 0)
{
pthread_mutexattr_destroy(&mutexAttr);
return false; // No process sharing
}
pthread_mutexattr_destroy(&mutexAttr);
}
else {
return false; // No mutex attrs
}
}
else {
// Only thread-shared
if (pthread_mutex_init(&lockMutex, 0) != 0) {
return false;
}
}
// These must be updated to make some of our auxiliary functions
// work right since their values will be based on the cache size.
cacheSize = _cacheSize;
pageSize = _pageSize;
version = PIXMAP_CACHE_VERSION;
cacheTimestamp = static_cast<unsigned>(::time(0));
clearInternalTables();
// Unlock the mini-lock, and introduce a total memory barrier to make
// sure all changes have propagated even without a mutex.
ready.ref();
return true;
}
void clearInternalTables()
{
// Assumes we're already locked somehow.
cacheAvail = pageTableSize();
// Setup page tables to point nowhere
PageTableEntry *table = pageTable();
for (uint i = 0; i < pageTableSize(); ++i) {
table[i].index = -1;
}
// Setup index tables to be accurate.
IndexTableEntry *indices = indexTable();
for (uint i = 0; i < indexTableSize(); ++i) {
indices[i].firstPage = -1;
}
}
const IndexTableEntry *indexTable() const
{
// Index Table goes immediately after this struct, at the first byte
// where alignment constraints are met (accounted for by offsetAs).
return offsetAs<IndexTableEntry>(this, sizeof(*this));
}
const PageTableEntry *pageTable() const
{
const IndexTableEntry *base = indexTable();
base += indexTableSize();
// Let's call wherever we end up the start of the page table...
return alignTo<PageTableEntry>(base);
}
const void *cachePages() const
{
const PageTableEntry *tableStart = pageTable();
tableStart += pageTableSize();
// Let's call wherever we end up the start of the data...
return alignTo<void>(tableStart, cachePageSize());
}
const void *page(pageID at) const
{
if (static_cast<int>(at) >= static_cast<int>(pageTableSize())) {
return 0;
}
// We must manually calculate this one since pageSize varies.
const char *pageStart = reinterpret_cast<const char *>(cachePages());
pageStart += (at * cachePageSize());
return reinterpret_cast<const void *>(pageStart);
}
// The following are non-const versions of some of the methods defined
// above. They use const_cast<> because I feel that is better than
// duplicating the code. I suppose template member functions (?)
// may work, may investigate later.
IndexTableEntry *indexTable()
{
const SharedMemory *that = const_cast<const SharedMemory*>(this);
return const_cast<IndexTableEntry *>(that->indexTable());
}
PageTableEntry *pageTable()
{
const SharedMemory *that = const_cast<const SharedMemory*>(this);
return const_cast<PageTableEntry *>(that->pageTable());
}
void *cachePages()
{
const SharedMemory *that = const_cast<const SharedMemory*>(this);
return const_cast<void *>(that->cachePages());
}
void *page(pageID at)
{
const SharedMemory *that = const_cast<const SharedMemory*>(this);
return const_cast<void *>(that->page(at));
}
uint pageTableSize() const
{
return cacheSize / cachePageSize();
}
uint indexTableSize() const
{
// Assume 2 pages on average are needed -> the number of entries
// would be half of the number of pages.
return pageTableSize() / 2;
}
/**
* @return the index of the first page, for the set of contiguous
* pages that can hold @p pagesNeeded PAGES.
*/
pageID findEmptyPages(uint pagesNeeded) const
{
// Loop through the page table, find the first empty page, and just
// makes sure that there are enough free pages.
const PageTableEntry *table = pageTable();
uint contiguousPagesFound = 0;
pageID base = 0;
for (pageID i = 0; i < static_cast<int>(pageTableSize() - pagesNeeded + 1); ++i) {
if (table[i].index < 0) {
if (contiguousPagesFound == 0) {
base = i;
}
contiguousPagesFound++;
}
else {
contiguousPagesFound = 0;
}
if (contiguousPagesFound == pagesNeeded) {
return base;
}
}
return pageTableSize();
}
// left < right?
static bool lruCompare(const IndexTableEntry &l, const IndexTableEntry &r)
{
// Ensure invalid entries migrate to the end
if (l.firstPage == -1 && r.firstPage >= 0) {
return false;
}
if (l.firstPage >= 0 && r.firstPage == -1) {
return true;
}
// Most recently used will have the highest absolute time =>
// least recently used (lowest) should go first => use left < right
return l.lastUsedTime < r.lastUsedTime;
}
// left < right?
static bool seldomUsedCompare(const IndexTableEntry &l, const IndexTableEntry &r)
{
// Ensure invalid entries migrate to the end
if (l.firstPage == -1 && r.firstPage >= 0) {
return false;
}
if (l.firstPage >= 0 && r.firstPage == -1) {
return true;
}
// Put lowest use count at start by using left < right
return l.useCount < r.useCount;
}
// left < right?
static bool ageCompare(const IndexTableEntry &l, const IndexTableEntry &r)
{
// Ensure invalid entries migrate to the end
if (l.firstPage == -1 && r.firstPage >= 0) {
return false;
}
if (l.firstPage >= 0 && r.firstPage == -1) {
return true;
}
// Oldest entries die first -- they have the lowest absolute add time,
// so just like the others use left < right
return l.addTime < r.addTime;
}
void defragment()
{
if (cacheAvail * cachePageSize() == cacheSize) {
return; // That was easy
}
kDebug(264) << "Defragmenting the shared cache";
// Just do a linear scan, and anytime there is free space, swap it
// with the pages to its right. In order to meet the precondition
// we need to skip any used pages first.
pageID currentPage = 0;
pageID idLimit = static_cast<pageID>(pageTableSize());
PageTableEntry *pages = pageTable();
// Skip used pages
while (currentPage < idLimit && pages[currentPage].index >= 0) {
++currentPage;
}
pageID freeSpot = currentPage;
// Main loop, starting from a free page, skip to the used pages and
// move them back.
while (currentPage < idLimit) {
// Find the next used page
while (currentPage < idLimit && pages[currentPage].index < 0) {
++currentPage;
}
if (currentPage >= idLimit) {
break;
}
// Found an entry, move it.
qint32 affectedIndex = pages[currentPage].index;
Q_ASSERT(affectedIndex >= 0);
Q_ASSERT(indexTable()[affectedIndex].firstPage == currentPage);
indexTable()[affectedIndex].firstPage = freeSpot;
// Moving one page at a time guarantees we can use memcpy safely
// (in other words, the source and destination will not overlap).
while (currentPage < idLimit && pages[currentPage].index >= 0) {
::memcpy(page(freeSpot), page(currentPage), cachePageSize());
pages[freeSpot].index = affectedIndex;
pages[currentPage].index = -1;
++currentPage;
++freeSpot;
// If we've just moved the very last page and it happened to
// be at the very end of the cache then we're done.
if (currentPage >= idLimit) {
break;
}
// We're moving consecutive used pages whether they belong to
// our affected entry or not, so detect if we've started moving
// the data for a different entry and adjust if necessary.
if (affectedIndex != pages[currentPage].index) {
indexTable()[pages[currentPage].index].firstPage = freeSpot;
}
affectedIndex = pages[currentPage].index;
}
// At this point currentPage is on a page that is unused, and the
// cycle repeats. However, currentPage is not the first unused
// page, freeSpot is, so leave it alone.
}
}
qint32 findNamedEntry(const QByteArray &key) const
{
uint keyHash = fnvHash32(key);
uint position = keyHash % indexTableSize();
int probeNumber = 1; // See insert() for description
while (indexTable()[position].fileNameHash != keyHash &&
indexTable()[position].useCount > 0 &&
probeNumber < 6)
{
position = (keyHash + (probeNumber + probeNumber * probeNumber) / 2)
% indexTableSize();
probeNumber++;
}
if (indexTable()[position].fileNameHash == keyHash) {
pageID firstPage = indexTable()[position].firstPage;
if (firstPage < 0) {
return -1;
}
const void *resultPage = page(firstPage);
const char *utf8FileName = reinterpret_cast<const char *>(resultPage);
if (qstrcmp(utf8FileName, key.constData()) == 0) {
return position;
}
}
return -1; // Not found, or a different one found.
}
// Function to use with QSharedPointer in removeUsedPages below...
static void deleteTable(IndexTableEntry *table) {
delete [] table;
}
uint removeUsedPages(uint numberNeeded)
{
if (numberNeeded > pageTableSize()) {
kError(264) << "Internal error: Requested more space than exists in the cache.";
kError(264) << numberNeeded << "requested, " << pageTableSize() << "is the total possible.";
return pageTableSize();
}
// To avoid calling findEmptyPages() all the time we will figure out
// the minimum number of pages required to fulfill the request if the
// page table were perfectly defragmented, and remove at least that
// amount first. If the cache free space is large enough we will
// defragment first instead since it's likely we're highly fragmented.
uint freedPagesRequired = 0;
if (numberNeeded > cacheAvail) {
freedPagesRequired = numberNeeded - cacheAvail;
}
kDebug(264) << "Removing old entries to free up" << numberNeeded << "pages,"
<< cacheAvail << "are already theoretically available.";
if (cacheAvail > 3 * numberNeeded) {
defragment();
uint result = findEmptyPages(numberNeeded);
if (result < pageTableSize()) {
return result;
}
else {
kError(264) << "Just defragmented a locked cache, but still there"
<< "isn't enough room for the current request.";
}
}
// At this point we know we'll have to free some space up, so sort our
// list of entries by whatever the current criteria are and start
// killing expired entries.
QSharedPointer<IndexTableEntry> tablePtr(new IndexTableEntry[indexTableSize()], deleteTable);
// We use tablePtr to ensure the data is destroyed, but do the access
// via a helper pointer to allow for array ops.
IndexTableEntry *table = tablePtr.data();
::memcpy(table, indexTable(), sizeof(IndexTableEntry) * indexTableSize());
// Our entry ID is simply its index into the
// index table, which qSort will rearrange all willy-nilly, so first
// we'll save the *real* entry ID into firstPage (which is useless in
// our copy of the index table. On the other hand if the entry is not
// used then we note that with -1.
for (uint i = 0; i < indexTableSize(); ++i) {
table[i].firstPage = table[i].useCount > 0 ? static_cast<pageID>(i)
: -1;
}
// Declare the comparison function that we'll use to pass to qSort,
// based on our cache eviction policy.
bool (*compareFunction)(const IndexTableEntry &, const IndexTableEntry &);
switch((int) evictionPolicy) {
case (int) KSharedDataCache::EvictLeastOftenUsed:
case (int) KSharedDataCache::NoEvictionPreference:
default:
compareFunction = seldomUsedCompare;
break;
case (int) KSharedDataCache::EvictLeastRecentlyUsed:
compareFunction = lruCompare;
break;
case (int) KSharedDataCache::EvictOldest:
compareFunction = ageCompare;
break;
}
qSort(table, table + indexTableSize(), compareFunction);
// Least recently used entries will be in the front.
// Start killing until we have room.
// Note on removeEntry: It expects an index into the index table,
// but our sorted list is all jumbled. But we stored the real index
// in the firstPage member.
// Remove entries until we've removed at least the required number
// of pages.
uint i = 0;
while (i < indexTableSize() && numberNeeded > cacheAvail) {
int curIndex = table[i++].firstPage;
// Removed everything, still no luck. At *this* point,
// pagesRemoved < numberNeeded or in other words we can't fulfill
// the request even if we defragment. This is really a logic error.
if (curIndex < 0) {
kError(264) << "Unable to remove enough used pages to allocate"
<< numberNeeded << "pages. In theory the cache is empty, weird.";
return pageTableSize();
}
kDebug(264) << "Removing entry of" << indexTable()[curIndex].totalItemSize
<< "size";
removeEntry(curIndex);
}
// At this point let's see if we have freed up enough data by
// defragmenting first and seeing if we can find that free space.
defragment();
pageID result = pageTableSize();
while (i < indexTableSize() &&
(result = findEmptyPages(numberNeeded)) >= static_cast<int>(pageTableSize()))
{
int curIndex = table[i++].firstPage;
if (curIndex < 0) {
// One last shot.
defragment();
return findEmptyPages(numberNeeded);
}
removeEntry(curIndex);
}
// Whew.
return result;
}
// Returns the total size required for a given cache size.
static uint totalSize(uint cacheSize, uint effectivePageSize)
{
uint numberPages = intCeil(cacheSize, effectivePageSize);
uint indexTableSize = numberPages / 2;
// Knowing the number of pages, we can determine what addresses we'd be
// using (properly aligned), and from there determine how much memory
// we'd use.
IndexTableEntry *indexTableStart =
offsetAs<IndexTableEntry>(static_cast<void*>(0), sizeof (SharedMemory));
indexTableStart += indexTableSize;
PageTableEntry *pageTableStart = reinterpret_cast<PageTableEntry *>(indexTableStart);
pageTableStart = alignTo<PageTableEntry>(pageTableStart);
pageTableStart += numberPages;
// The weird part, we must manually adjust the pointer based on the page size.
char *cacheStart = reinterpret_cast<char *>(pageTableStart);
cacheStart += (numberPages * effectivePageSize);
// ALIGNOF gives pointer alignment
cacheStart = alignTo<char>(cacheStart, ALIGNOF(void*));
// We've traversed the header, index, page table, and cache.
// Wherever we're at now is the size of the enchilada.
return static_cast<uint>(reinterpret_cast<quintptr>(cacheStart));
}
uint fileNameHash(const QByteArray &utf8FileName) const
{
return fnvHash32(utf8FileName) % indexTableSize();
}
bool lock() const
{
#ifdef _POSIX_THREAD_PROCESS_SHARED
struct timespec timeout;
// Long timeout, but if we fail to meet this timeout it's probably a cache
// corruption (and if we take 8 seconds then it should be much much quicker
// the next time anyways since we'd be paged back in from disk)
timeout.tv_sec = 10 + ::time(NULL); // Absolute time, so 10 seconds from now
timeout.tv_nsec = 0;
return pthread_mutex_timedlock(&lockMutex, &timeout) >= 0;
#else
// Some POSIX platforms don't have full support for pthreads. On these typically
// there will be no timedlock, so just don't bother and accept hangs on weird
// platforms.
return pthread_mutex_lock(&lockMutex) == 0;
#endif
}
void unlock() const
{
pthread_mutex_unlock(&lockMutex);
}
void clear()
{
lock();
clearInternalTables();
unlock();
}
void removeEntry(uint index);
};
// The per-instance private data, such as map size, whether
// attached or not, pointer to shared memory, etc.
class KSharedDataCache::Private
{
public:
Private(const QString &name,
unsigned defaultCacheSize,
unsigned expectedItemSize
)
: shm(0)
, m_cacheName(name)
, m_attached(false)
, m_mapSize(0)
, m_defaultCacheSize(defaultCacheSize)
, m_expectedItemSize(expectedItemSize)
{
mapSharedMemory();
}
// This function does a lot of the important work, attempting to connect to shared
// memory, a private anonymous mapping if that fails, and failing that, nothing (but
// the cache remains "valid", we just don't actually do anything).
void mapSharedMemory()
{
// 0-sized caches are fairly useless.
unsigned cacheSize = qMax(m_defaultCacheSize, uint(SharedMemory::MINIMUM_CACHE_SIZE));
unsigned pageSize = SharedMemory::equivalentPageSize(m_expectedItemSize);
// Ensure that the cache is sized such that there is a minimum number of
// pages available. (i.e. a cache consisting of only 1 page is fairly
// useless and probably crash-prone).
cacheSize = qMax(pageSize * 256, cacheSize);
// The m_cacheName is used to find the file to store the cache in.
QString cacheName = KGlobal::dirs()->locateLocal("cache", m_cacheName + QLatin1String(".kcache"));
QFile file(cacheName);
kDebug(264) << "Opening cache" << cacheName << "page size is" << pageSize;
// The basic idea is to open the file that we want to map into shared
// memory, and then actually establish the mapping. Once we have mapped the
// file into shared memory we can close the file handle, the mapping will
// still be maintained (unless the file is resized to be shorter than
// expected, which we don't handle yet :-( )
// size accounts for the overhead over the desired cacheSize
uint size = SharedMemory::totalSize(cacheSize, pageSize);
void *mapAddress = MAP_FAILED;
if (size < cacheSize) {
kError(264) << "Asked for a cache size less than requested size somehow -- Logic Error :(";
return;
}
bool systemSupportsProcessSharing = false;
// Compile time check ensures the required constants from the .h are available to be
// queried at all.
#ifdef _POSIX_THREAD_PROCESS_SHARED
systemSupportsProcessSharing = ::sysconf(_SC_THREAD_PROCESS_SHARED) > 0;
#endif
// We establish the shared memory mapping here, only if we will have appropriate
// mutex support (systemSupportsProcessSharing), then we:
// Open the file and resize to some sane value if the file is too small.
if (systemSupportsProcessSharing &&
file.open(QIODevice::ReadWrite) &&
(file.size() >= size || file.resize(size)))
{
// Use mmap directly instead of QFile::map since the QFile (and its
// shared mapping) will disappear unless we hang onto the QFile for no
// reason (see the note below, we don't care about the file per se...)
mapAddress = ::mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, file.handle(), 0);
// So... it is possible that someone else has mapped this cache already
// with a larger size. If that's the case we need to at least match
// the size to be able to access every entry, so fixup the mapping.
if (mapAddress != MAP_FAILED) {
SharedMemory *mapped = reinterpret_cast<SharedMemory *>(mapAddress);
// First make sure that the version of the cache on disk is
// valid. We also need to check that version != 0 to
// disambiguate against an uninitialized cache.
if (mapped->version != SharedMemory::PIXMAP_CACHE_VERSION &&
mapped->version > 0)
{
kWarning(264) << "Deleting wrong version of cache" << cacheName;
// CAUTION: Potentially recursive since the recovery
// involves calling this function again.
m_attached = true;
m_mapSize = size;
shm = mapped;
recoverCorruptedCache();
return;
}
else if (mapped->cacheSize > cacheSize) {
// This order is very important. We must save the cache size
// before we remove the mapping, but unmap before overwriting
// the previous mapping size...
cacheSize = mapped->cacheSize;
unsigned actualPageSize = mapped->cachePageSize();
::munmap(mapAddress, size);
size = SharedMemory::totalSize(cacheSize, actualPageSize);
mapAddress = ::mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, file.handle(), 0);
}
}
}
// We could be here without the mapping established if:
// 1) Process-shared synchronization is not supported, either at compile or run time,
// 2) Unable to open the required file.
// 3) Unable to resize the file to be large enough.
// 4) Establishing the mapping failed.
// 5) The mapping succeeded, but the size was wrong and we were unable to map when
// we tried again.
// 6) The incorrect version of the cache was detected.
// In any of these cases, attempt to fallback to the
// better-supported anonymous private page style of mmap. This memory won't
// be shared, but our code will still work the same.
// NOTE: We never use the on-disk representation independently of the
// shared memory. If we don't get shared memory the disk info is ignored,
// if we do get shared memory we never look at disk again.
bool usingSharedMapping = true;
if (mapAddress == MAP_FAILED) {
kWarning(264) << "Failed to establish shared memory mapping, will fallback"
<< "to private memory -- memory usage will increase";
mapAddress = ::mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
usingSharedMapping = false;
}
// Well now we're really hosed. We can still work, but we can't even cache
// data.
if (mapAddress == MAP_FAILED) {
kError(264) << "Unable to allocate shared memory segment for shared data cache"
<< cacheName << "of size" << cacheSize;
return;
}
kDebug(264) << "Attached to cache, determining if it must be initialized";
m_mapSize = size;
// We never actually construct shm, but we assign it the same address as the
// shared memory we just mapped, so effectively shm is now a SharedMemory that
// happens to be located at mapAddress.
shm = reinterpret_cast<SharedMemory *>(mapAddress);
// If we were first to create this memory map, all data will be 0.
// Therefore if ready == 0 we're not initialized. A fully initialized
// header will have ready == 2. Why?
// Because 0 means "safe to initialize"
// 1 means "in progress of initing"
// 2 means "ready"
uint usecSleepTime = 8; // Start by sleeping for 8 microseconds
while (shm->ready != 2) {
if (usecSleepTime >= (1 << 21)) {
// Didn't acquire within ~8 seconds? Assume an issue exists
kError(264) << "Unable to acquire shared lock, is the cache corrupt?";
::munmap(mapAddress, size);
file.remove(); // Unlink the cache in case it's corrupt.
return; // Fallback to QCache (later)
}
if (shm->ready.testAndSetAcquire(0, 1)) {
if (!shm->performInitialSetup(usingSharedMapping, cacheSize, pageSize)) {
kError(264) << "Unable to perform initial setup, this system probably "
"does not really support process-shared pthreads, even "
"though it claims otherwise.";
::munmap(mapAddress, size);
file.remove();
return;
}
}
else {
usleep(usecSleepTime); // spin
// Exponential fallback as in Ethernet and similar collision resolution methods
usecSleepTime *= 2;
}
}
// We are "attached" if we have a valid memory mapping, whether it is
// shared or private.
kDebug(264) << "Cache fully initialized -- attached to memory mapping";
kDebug(264) << shm->cacheAvail * shm->cachePageSize() << "bytes available out of"
<< shm->cacheSize;
m_attached = true;
}
// Called whenever the cache is apparently corrupt (for instance, a timeout trying to
// lock the cache). In this situation it is safer just to destroy it all and try again.
void recoverCorruptedCache()
{
KSharedDataCache::deleteCache(m_cacheName);
if (m_attached) {
::munmap(shm, m_mapSize);
m_attached = false;
shm = 0;
m_mapSize = 0;
}
// Do this even if we weren't previously cached -- it might work now.
mapSharedMemory();
}
class CacheLocker
{
mutable Private * d;
bool cautiousLock()
{
int lockCount = 0;
// Locking can fail due to a timeout. If it happens too often even though
// we're taking corrective action assume there's some disastrous problem
// and give up.
while (!d->shm->lock()) {
d->recoverCorruptedCache();
if (!d->m_attached) {
kWarning(264) << "Lost the connection to shared memory for cache"
<< d->m_cacheName;
return false;
}
if (lockCount++ > 4) {
kError(264) << "There is a very serious problem with the KDE data cache"
<< d->m_cacheName << "giving up trying to access cache.";
::munmap(d->shm, d->m_mapSize);
d->shm = 0;
d->m_attached = false;
return false;
}
}
return true;
}
public:
CacheLocker(const Private *_d) : d(const_cast<Private *>(_d))
{
if (d->m_attached) {
// A separate mutex for the shm lock? What gives?
// The reason is that we have to check to see if the cache
// was made bigger by a different process. If so, we need to
// re-map the cache to accomodate that. In that event, we will
// need to un-map the cache first.
//
// Now imagine what happens if two threads in this same process
// tried to do this concurrently. Since we wouldn't be attached
// to shm we couldn't use d->shm->lock to be safe during that
// critical section (in between unlock and the subsequent
// lock), so one thread could set shm = 0 after unmapping and
// cause the other thread to crash. So we need a separate
// single-process/multiple-thread lock to be super-safe.
QMutexLocker d_locker(&d->m_threadLock);
if (!cautiousLock()) {
return;
}
uint testSize = SharedMemory::totalSize(d->shm->cacheSize, d->shm->cachePageSize());
// A while loop? Indeed, think what happens if this happens
// twice -- hard to debug race conditions.
while (testSize > d->m_mapSize) {
kDebug(264) << "Someone enlarged the cache on us,"
<< "attempting to match new configuration.";
// Linux supports mremap, but it's not portable. So,
// drop the map and (try to) re-establish.
d->shm->unlock();
::munmap(d->shm, d->m_mapSize);
d->m_attached = false;
d->m_mapSize = 0;
d->shm = 0;
QFile f(d->m_cacheName);
if (!f.open(QFile::ReadWrite)) {
kError(264) << "Unable to re-open cache, unfortunately"
<< "the connection had to be dropped for"
<< "crash safety -- things will be much"
<< "slower now.";
return;
}
void *newMap = ::mmap(0, testSize, PROT_READ | PROT_WRITE,
MAP_SHARED, f.handle(), 0);
if (newMap == MAP_FAILED) {
kError(264) << "Unopen to re-map the cache into memory"
<< "things will be much slower now";
return;
}
d->shm = reinterpret_cast<SharedMemory *>(newMap);
d->m_attached = true;
d->m_mapSize = testSize;
if (!cautiousLock()) {
return;
}
testSize = SharedMemory::totalSize(d->shm->cacheSize, d->shm->cachePageSize());
}
}
}
~CacheLocker()
{
if (d->m_attached) {
d->shm->unlock();
}
}
bool failed() const
{
return d->shm == 0;
}
};
SharedMemory *shm;
QString m_cacheName;
QMutex m_threadLock;
bool m_attached;
uint m_mapSize;
uint m_defaultCacheSize;
uint m_expectedItemSize;
};
// Must be called while the lock is already held!
void SharedMemory::removeEntry(uint index)
{
Q_ASSERT(index < indexTableSize());
Q_ASSERT(cacheAvail <= pageTableSize());
PageTableEntry *pageTableEntries = pageTable();
IndexTableEntry *entriesIndex = indexTable();
// Update page table first
pageID firstPage = entriesIndex[index].firstPage;
if (firstPage < 0 || firstPage >= pageTableSize()) {
kError(264) << "Removing" << index << "which is already marked as empty!";
return;
}
if (index != static_cast<uint>(pageTableEntries[firstPage].index)) {
kError(264) << "Removing" << index << "will not work as it is assigned"
<< "to page" << firstPage << "which is itself assigned to"
<< "entry" << pageTableEntries[firstPage].index << "instead!";
return;
}
uint entriesToRemove = intCeil(entriesIndex[index].totalItemSize, cachePageSize());
uint savedCacheSize = cacheAvail;
for (uint i = firstPage; i < pageTableSize() &&
(uint) pageTableEntries[i].index == index; ++i)
{
pageTableEntries[i].index = -1;
cacheAvail++;
}
if ((cacheAvail - savedCacheSize) != entriesToRemove) {
kError(264) << "We somehow did not remove" << entriesToRemove
<< "when removing entry" << index << ", instead we removed"
<< (cacheAvail - savedCacheSize);
}
// For debugging
#ifdef NDEBUG
QByteArray str((const char *)page(firstPage));
str.prepend(" REMOVED: ");
str.prepend(QByteArray::number(index));
str.prepend("ENTRY ");
::memcpy(page(firstPage), str.constData(), str.size() + 1);
#endif
// Update the index
entriesIndex[index].fileNameHash = 0;
entriesIndex[index].totalItemSize = 0;
entriesIndex[index].useCount = 0;
entriesIndex[index].lastUsedTime = 0;
entriesIndex[index].addTime = 0;
entriesIndex[index].firstPage = -1;
}
KSharedDataCache::KSharedDataCache(const QString &cacheName,
unsigned defaultCacheSize,
unsigned expectedItemSize)
: d(new Private(cacheName, defaultCacheSize, expectedItemSize))
{
}
KSharedDataCache::~KSharedDataCache()
{
// Note that there is no other actions required to separate from the
// shared memory segment, simply unmapping is enough. This makes things
// *much* easier so I'd recommend maintaining this ideal.
if (d->m_attached) {
::munmap(d->shm, d->m_mapSize);
}
// Do not delete d->shm, it was never constructed, it's just an alias.
d->shm = 0;
delete d;
}
bool KSharedDataCache::insert(const QString &key, const QByteArray &data)
{
Private::CacheLocker lock(d);
if (lock.failed()) {
return false;
}
QByteArray encodedKey = key.toUtf8();
uint keyHash = fnvHash32(encodedKey);
uint position = keyHash % d->shm->indexTableSize();
// See if we're overwriting an existing entry.
IndexTableEntry *indices = d->shm->indexTable();
// In order to avoid the issue of a very long-lived cache having items
// with a use count of 1 near-permanently, we attempt to artifically
// reduce the use count of long-lived items when there is high load on
// the cache. We do this randomly, with a weighting that makes the event
// impossible if load < 0.5, and guaranteed if load >= 0.96.
static double startCullPoint = 0.5l;
static double mustCullPoint = 0.96l;
// cacheAvail is in pages, cacheSize is in bytes.
double loadFactor = (1.0l * d->shm->cacheAvail * d->shm->cachePageSize()
/ d->shm->cacheSize);
bool cullCollisions = false;
if (KDE_ISUNLIKELY(loadFactor >= mustCullPoint)) {
cullCollisions = true;
}
else {
int tripWireValue = RAND_MAX * (loadFactor - startCullPoint) / (mustCullPoint - startCullPoint);
if (KRandom::random() >= tripWireValue) {
cullCollisions = true;
}
}
// In case of collisions, use quadratic chaining to attempt to find an
// empty slot. The equation we use is
// position = (hash + (i + i*i) / 2) % size, where i is the probe number.
int probeNumber = 1;
while (indices[position].useCount > 0 && probeNumber < 6) {
// If we are "culling" old entries, see if this one is old and if so
// reduce its use count. If it reduces to zero then eliminate it and
// use its old spot.
if (cullCollisions && (indices[position].lastUsedTime - ::time(0)) > 60) {
indices[position].useCount >>= 1;
if (indices[position].useCount == 0) {
kDebug(264) << "Overwriting existing old cached entry due to collision.";
d->shm->removeEntry(position); // Remove it first
break;
}
}
position = (keyHash + (probeNumber + probeNumber * probeNumber) / 2)
% d->shm->indexTableSize();
probeNumber++;
}
if (indices[position].useCount > 0 && indices[position].firstPage >= 0) {
kDebug(264) << "Overwriting existing cached entry due to collision.";
d->shm->removeEntry(position); // Remove it first
}
// Data will be stored as fileNamefoo\0PNGimagedata.....
// So total size required is the length of the encoded file name + 1
// for the trailing null, and then the length of the image data.
uint fileNameLength = 1 + encodedKey.length();
uint requiredSize = fileNameLength + data.size();
uint pagesNeeded = intCeil(requiredSize, d->shm->cachePageSize());
uint firstPage = (uint) -1;
if (pagesNeeded >= d->shm->pageTableSize()) {
kWarning(264) << key << "is too large to be cached.";
return false;
}
// If the cache has no room, or the fragmentation is too great to find
// the required number of consecutive free pages, take action.
if (pagesNeeded > d->shm->cacheAvail ||
(firstPage = d->shm->findEmptyPages(pagesNeeded)) >= d->shm->pageTableSize())
{
// If we have enough free space just defragment
uint freePagesDesired = 3 * qMax(1u, pagesNeeded / 2);
if (d->shm->cacheAvail > freePagesDesired) {
// TODO: How the hell long does this actually take on real
// caches?
d->shm->defragment();
firstPage = d->shm->findEmptyPages(pagesNeeded);
}
else {
// If we already have free pages we don't want to remove a ton
// extra. However we can't rely on the return value of
// removeUsedPages giving us a good location since we're not
// passing in the actual number of pages that we need.
d->shm->removeUsedPages(qMin(2 * freePagesDesired, d->shm->pageTableSize())
- d->shm->cacheAvail);
firstPage = d->shm->findEmptyPages(pagesNeeded);
}
if (firstPage >= d->shm->pageTableSize() ||
d->shm->cacheAvail < pagesNeeded)
{
kError(264) << "Unable to free up memory for" << key;
return false;
}
}
// Update page table
PageTableEntry *table = d->shm->pageTable();
for (uint i = 0; i < pagesNeeded; ++i) {
table[firstPage + i].index = position;
}
// Update index
indices[position].fileNameHash = keyHash;
indices[position].totalItemSize = requiredSize;
indices[position].useCount = 1;
indices[position].addTime = ::time(0);
indices[position].lastUsedTime = indices[position].addTime;
indices[position].firstPage = firstPage;
// Update cache
d->shm->cacheAvail -= pagesNeeded;
// Actually move the data in place
void *dataPage = d->shm->page(firstPage);
// Cast for byte-sized pointer arithmetic
uchar *startOfPageData = reinterpret_cast<uchar *>(dataPage);
::memcpy(startOfPageData, encodedKey.constData(), fileNameLength);
::memcpy(startOfPageData + fileNameLength, data.constData(), data.size());
return true;
}
bool KSharedDataCache::find(const QString &key, QByteArray *destination) const
{
if (!d->m_attached) {
return false;
}
Private::CacheLocker lock(d);
if (lock.failed()) {
return false;
}
// Search in the index for our data, hashed by key;
QByteArray encodedKey = key.toUtf8();
qint32 entry = d->shm->findNamedEntry(encodedKey);
if (entry >= 0) {
const IndexTableEntry *header = &d->shm->indexTable()[entry];
const void *resultPage = d->shm->page(header->firstPage);
header->useCount++;
header->lastUsedTime = ::time(0);
// Our item is the key followed immediately by the data, so skip
// past the key.
const char *cacheData = reinterpret_cast<const char *>(resultPage);
cacheData += encodedKey.size();
cacheData++; // Skip trailing null -- now we're pointing to start of data
if (destination) {
*destination = QByteArray(cacheData, header->totalItemSize - encodedKey.size() - 1);
}
return true;
}
return false;
}
void KSharedDataCache::clear()
{
d->shm->clear();
}
bool KSharedDataCache::contains(const QString &key) const
{
Private::CacheLocker lock(d);
if (lock.failed()) {
return false;
}
return d->shm->findNamedEntry(key.toUtf8()) >= 0;
}
void KSharedDataCache::deleteCache(const QString &cacheName)
{
QString cachePath = KGlobal::dirs()->locateLocal("cache", cacheName + QLatin1String(".kcache"));
// Note that it is important to simply unlink the file, and not truncate it
// smaller first to avoid SIGBUS errors and similar with shared memory
// attached to the underlying inode.
kDebug(264) << "Removing cache at" << cachePath;
QFile::remove(cachePath);
}
unsigned KSharedDataCache::totalSize() const
{
Private::CacheLocker lock(d);
if (lock.failed()) {
return 0u;
}
return d->shm->cacheSize;
}
unsigned KSharedDataCache::freeSize() const
{
Private::CacheLocker lock(d);
if (lock.failed()) {
return 0u;
}
return d->shm->cacheAvail * d->shm->cachePageSize();
}
KSharedDataCache::EvictionPolicy KSharedDataCache::evictionPolicy() const
{
if (d->m_attached) {
int policy(d->shm->evictionPolicy);
return static_cast<EvictionPolicy>(policy);
}
return NoEvictionPreference;
}
void KSharedDataCache::setEvictionPolicy(EvictionPolicy newPolicy)
{
if (d->m_attached) {
d->shm->evictionPolicy.fetchAndStoreRelease(static_cast<int>(newPolicy));
}
}
unsigned KSharedDataCache::timestamp() const
{
if (d->m_attached) {
return static_cast<unsigned>(d->shm->cacheTimestamp.fetchAndAddAcquire(0));
}
return 0;
}
void KSharedDataCache::setTimestamp(unsigned newTimestamp)
{
if (d->m_attached) {
d->shm->cacheTimestamp.fetchAndStoreRelease(static_cast<int>(newTimestamp));
}
}