/**
* Flush a specified ledger
*
* @param ledger Ledger Id
* @throws IOException
*/
private void flushSpecificLedger(long ledger) throws IOException {
LinkedList<Long> firstEntryList = pageMapAndList.getFirstEntryListToBeFlushed(ledger);
// flush ledger index file header if necessary
indexPersistenceManager.flushLedgerHeader(ledger);
if (null == firstEntryList || firstEntryList.size() == 0) {
LOG.debug("Nothing to flush for ledger {}.", ledger);
// nothing to do
return;
}
// Now flush all the pages of a ledger
List<LedgerEntryPage> entries = new ArrayList<LedgerEntryPage>(firstEntryList.size());
try {
for (Long firstEntry : firstEntryList) {
LedgerEntryPage lep = getLedgerEntryPage(ledger, firstEntry, true);
if (lep != null) {
entries.add(lep);
}
}
indexPersistenceManager.flushLedgerEntries(ledger, entries);
} finally {
for (LedgerEntryPage lep : entries) {
lep.releasePage();
}
}
}
/**
* Grab ledger entry page whose first entry is <code>pageEntry</code>.
*
* <p>If the page doesn't existed before, we allocate a memory page. Otherwise, we grab a clean
* page and read it from disk.
*
* @param ledger Ledger Id
* @param pageEntry Start entry of this entry page.
*/
private LedgerEntryPage grabLedgerEntryPage(long ledger, long pageEntry) throws IOException {
LedgerEntryPage lep = grabCleanPage(ledger, pageEntry);
try {
// should get the up to date page from the persistence manager
// before we put it into table otherwise we would put
// an empty page in it
indexPersistenceManager.updatePage(lep);
LedgerEntryPage oldLep;
if (lep != (oldLep = pageMapAndList.putPage(lep))) {
lep.releasePage();
// Decrement the page count because we couldn't put this lep in the page cache.
pageCount.decrementAndGet();
// Increment the use count of the old lep because this is unexpected
oldLep.usePage();
lep = oldLep;
}
} catch (IOException ie) {
// if we grab a clean page, but failed to update the page
// we are exhausting the count of ledger entry pages.
// since this page will be never used, so we need to decrement
// page count of ledger cache.
lep.releasePage();
pageCount.decrementAndGet();
throw ie;
}
return lep;
}
private LedgerEntryPage grabCleanPage(long ledger, long entry) throws IOException {
if (entry % entriesPerPage != 0) {
throw new IllegalArgumentException(entry + " is not a multiple of " + entriesPerPage);
}
while (true) {
boolean canAllocate = false;
if (pageCount.incrementAndGet() <= pageLimit) {
canAllocate = true;
} else {
pageCount.decrementAndGet();
}
if (canAllocate) {
LedgerEntryPage lep = new LedgerEntryPage(pageSize, entriesPerPage, pageMapAndList);
lep.setLedgerAndFirstEntry(ledger, entry);
lep.usePage();
return lep;
}
LedgerEntryPage lep = pageMapAndList.grabCleanPage(ledger, entry);
if (null != lep) {
return lep;
}
LOG.info(
"Could not grab a clean page for ledger {}, entry {}, force flushing dirty ledgers.",
ledger,
entry);
flushOneOrMoreLedgers(false);
}
}
/**
* Add the LedgerEntryPage to the clean page LRU map
*
* @param lep Ledger Entry Page object
*/
private void addToCleanPagesList(LedgerEntryPage lep) {
synchronized (lruCleanPageMap) {
if (lep.isClean() && !lep.inUse()) {
lruCleanPageMap.put(lep.getEntryKey(), lep);
}
}
}
/**
* Remove the LedgerEntryPage from the clean page LRU map
*
* @param lep Ledger Entry Page object
*/
private void removeFromCleanPageList(LedgerEntryPage lep) {
synchronized (lruCleanPageMap) {
if (!lep.isClean() || lep.inUse()) {
lruCleanPageMap.remove(lep.getEntryKey());
}
}
}
LedgerEntryPage getLedgerEntryPage(Long ledger, Long firstEntry, boolean onlyDirty) {
LedgerEntryPage lep = pageMapAndList.getPage(ledger, firstEntry);
if (onlyDirty && null != lep && lep.isClean()) {
return null;
}
if (null != lep) {
lep.usePage();
}
return lep;
}
void flushLedgerEntries(long l, List<LedgerEntryPage> entries) throws IOException {
FileInfo fi = null;
try {
Collections.sort(
entries,
new Comparator<LedgerEntryPage>() {
@Override
public int compare(LedgerEntryPage o1, LedgerEntryPage o2) {
return (int) (o1.getFirstEntry() - o2.getFirstEntry());
}
});
int[] versions = new int[entries.size()];
try {
fi = getFileInfo(l, null);
} catch (Bookie.NoLedgerException nle) {
// ledger has been deleted
LOG.info("No ledger {} found when flushing entries.", l);
return;
}
// flush the header if necessary
relocateIndexFileAndFlushHeader(l, fi);
int start = 0;
long lastOffset = -1;
for (int i = 0; i < entries.size(); i++) {
versions[i] = entries.get(i).getVersion();
if (lastOffset != -1 && (entries.get(i).getFirstEntry() - lastOffset) != entriesPerPage) {
// send up a sequential list
int count = i - start;
if (count == 0) {
LOG.warn("Count cannot possibly be zero!");
}
writeBuffers(l, entries, fi, start, count);
start = i;
}
lastOffset = entries.get(i).getFirstEntry();
}
if (entries.size() - start == 0 && entries.size() != 0) {
LOG.warn("Nothing to write, but there were entries!");
}
writeBuffers(l, entries, fi, start, entries.size() - start);
for (int i = 0; i < entries.size(); i++) {
LedgerEntryPage lep = entries.get(i);
lep.setClean(versions[i]);
}
if (LOG.isDebugEnabled()) {
LOG.debug("Flushed ledger {} with {} pages.", l, entries.size());
}
} finally {
if (fi != null) {
fi.release();
}
}
}
void putEntryOffset(long ledger, long entry, long offset) throws IOException {
int offsetInPage = (int) (entry % entriesPerPage);
// find the id of the first entry of the page that has the entry
// we are looking for
long pageEntry = entry - offsetInPage;
LedgerEntryPage lep = getLedgerEntryPage(ledger, pageEntry, false);
if (lep == null) {
lep = grabLedgerEntryPage(ledger, pageEntry);
}
assert lep != null;
lep.setOffset(offset, offsetInPage * LedgerEntryPage.getIndexEntrySize());
lep.releasePage();
}
long getEntryOffset(long ledger, long entry) throws IOException {
int offsetInPage = (int) (entry % entriesPerPage);
// find the id of the first entry of the page that has the entry
// we are looking for
long pageEntry = entry - offsetInPage;
LedgerEntryPage lep = getLedgerEntryPage(ledger, pageEntry, false);
try {
if (lep == null) {
lep = grabLedgerEntryPage(ledger, pageEntry);
}
return lep.getOffset(offsetInPage * LedgerEntryPage.getIndexEntrySize());
} finally {
if (lep != null) {
lep.releasePage();
}
}
}
/**
* Get a clean page and provision it for the specified ledger and firstEntry within the ledger
*
* @param ledgerId Ledger id
* @param firstEntry Id of the first entry in the page
* @returns LedgerEntryPage if present
*/
LedgerEntryPage grabCleanPage(long ledgerId, long firstEntry) {
LedgerEntryPage lep = null;
while (lruCleanPageMap.size() > 0) {
lep = null;
synchronized (lruCleanPageMap) {
Iterator<Map.Entry<EntryKey, LedgerEntryPage>> iterator =
lruCleanPageMap.entrySet().iterator();
Map.Entry<EntryKey, LedgerEntryPage> entry = null;
while (iterator.hasNext()) {
entry = iterator.next();
iterator.remove();
if (entry.getValue().isClean() && !entry.getValue().inUse()) {
lep = entry.getValue();
break;
}
}
if (null == lep) {
LOG.debug("Did not find eligible page in the first pass");
return null;
}
}
// We found a candidate page, lets see if we can reclaim it before its re-used
ConcurrentMap<Long, LedgerEntryPage> pageMap = pages.get(lep.getLedger());
// Remove from map only if nothing has changed since we checked this lep.
// Its possible for the ledger to have been deleted or the page to have already
// been reclaimed. The page map is the definitive source of information, if anything
// has changed we should leave this page along and continue iterating to find
// another suitable page.
if ((null != pageMap) && (pageMap.remove(lep.getFirstEntry(), lep))) {
if (!lep.isClean()) {
// Someone wrote to this page while we were reclaiming it.
pageMap.put(lep.getFirstEntry(), lep);
lep = null;
} else {
// Do some bookkeeping on the page table
pages.remove(lep.getLedger(), EMPTY_PAGE_MAP);
// We can now safely reset this lep and return it.
lep.usePage();
lep.zeroPage();
lep.setLedgerAndFirstEntry(ledgerId, firstEntry);
return lep;
}
} else {
lep = null;
}
}
return lep;
}
/**
* Traverse the pages for a given ledger in memory and find the highest entry amongst these
* pages
*
* @param ledgerId Ledger id
* @returns last entry in the in memory pages
*/
private long getLastEntryInMem(long ledgerId) {
long lastEntry = 0;
// Find the last entry in the cache
ConcurrentMap<Long, LedgerEntryPage> map = pages.get(ledgerId);
if (map != null) {
for (LedgerEntryPage lep : map.values()) {
if (lep.getMaxPossibleEntry() < lastEntry) {
continue;
}
lep.usePage();
long highest = lep.getLastEntry();
if (highest > lastEntry) {
lastEntry = highest;
}
lep.releasePage();
}
}
return lastEntry;
}
void updatePage(LedgerEntryPage lep) throws IOException {
if (!lep.isClean()) {
throw new IOException("Trying to update a dirty page");
}
FileInfo fi = null;
try {
fi = getFileInfo(lep.getLedger(), null);
long pos = lep.getFirstEntryPosition();
if (pos >= fi.size()) {
lep.zeroPage();
} else {
lep.readPage(fi);
}
} finally {
if (fi != null) {
fi.release();
}
}
}
/**
* Add a LedgerEntryPage to the page map
*
* @param lep Ledger Entry Page object
*/
private LedgerEntryPage putPage(LedgerEntryPage lep) {
// Do a get here to avoid too many new ConcurrentHashMaps() as putIntoTable is called
// frequently.
ConcurrentMap<Long, LedgerEntryPage> map = pages.get(lep.getLedger());
if (null == map) {
ConcurrentMap<Long, LedgerEntryPage> mapToPut =
new ConcurrentHashMap<Long, LedgerEntryPage>();
map = pages.putIfAbsent(lep.getLedger(), mapToPut);
if (null == map) {
map = mapToPut;
}
}
LedgerEntryPage oldPage = map.putIfAbsent(lep.getFirstEntry(), lep);
if (null == oldPage) {
oldPage = lep;
// Also include this in the clean page map if it qualifies.
// Note: This is done for symmetry and correctness, however it should never
// get exercised since we shouldn't attempt a put without the page being in use
addToCleanPagesList(lep);
}
return oldPage;
}
/**
* Gets the list of pages in memory that have been changed and hence need to be written as a
* part of the flush operation that is being issued
*
* @param ledgerId Ledger id
* @returns last entry in the in memory pages.
*/
private LinkedList<Long> getFirstEntryListToBeFlushed(long ledgerId) {
ConcurrentMap<Long, LedgerEntryPage> pageMap = pages.get(ledgerId);
if (pageMap == null || pageMap.isEmpty()) {
return null;
}
LinkedList<Long> firstEntryList = new LinkedList<Long>();
for (ConcurrentMap.Entry<Long, LedgerEntryPage> entry : pageMap.entrySet()) {
LedgerEntryPage lep = entry.getValue();
if (lep.isClean()) {
if (!lep.inUse()) {
addToCleanPagesList(lep);
}
if (LOG.isTraceEnabled()) {
LOG.trace("Page is clean " + lep);
}
} else {
firstEntryList.add(lep.getFirstEntry());
}
}
return firstEntryList;
}
long getPersistEntryBeyondInMem(long ledgerId, long lastEntryInMem) throws IOException {
FileInfo fi = null;
long lastEntry = lastEntryInMem;
try {
fi = getFileInfo(ledgerId, null);
long size = fi.size();
// make sure the file size is aligned with index entry size
// otherwise we may read incorret data
if (0 != size % LedgerEntryPage.getIndexEntrySize()) {
LOG.warn("Index file of ledger {} is not aligned with index entry size.", ledgerId);
size = size - size % LedgerEntryPage.getIndexEntrySize();
}
// we may not have the last entry in the cache
if (size > lastEntry * LedgerEntryPage.getIndexEntrySize()) {
ByteBuffer bb = ByteBuffer.allocate(pageSize);
long position = size - pageSize;
if (position < 0) {
position = 0;
}
fi.read(bb, position);
bb.flip();
long startingEntryId = position / LedgerEntryPage.getIndexEntrySize();
for (int i = entriesPerPage - 1; i >= 0; i--) {
if (bb.getLong(i * LedgerEntryPage.getIndexEntrySize()) != 0) {
if (lastEntry < startingEntryId + i) {
lastEntry = startingEntryId + i;
}
break;
}
}
}
} finally {
if (fi != null) {
fi.release();
}
}
return lastEntry;
}