DfsBlockCache: Consolidate where ReadableChannel is opened

Opening a readable channel can be expensive and the number of channels
can be limited in DFS. Ensure that caller of
BlockBasedFile.readOneBlock() is responsible for opening and closing
the file, and that the ReadableChannel is reused in the request. As a side
effect, this makes the code easier to read, with better use of
try-with-resources.

The downside is that this means a readable channel is always opened, even
when the entire pack is already available for copying from cache. This
should be an acceptable cost: it's rare enough not to overload the server
and from a client latency perspective, the latency cost is in the noise
relative to the data transfer cost involved in a clone. If this turns out
to be a problem in practice, we can reintroduce that optimization in a
followup change.

Change-Id: I340428ee4bacd2dce019d5616ef12339a0c85f0b
Signed-off-by: Minh Thai <mthai@google.com>
Signed-off-by: Jonathan Nieder <jrn@google.com>
This commit is contained in:
Minh Thai 2018-12-18 15:06:59 -08:00 committed by Jonathan Nieder
parent 8bc9acf264
commit 4207b8d6ae
4 changed files with 78 additions and 90 deletions

View File

@ -48,7 +48,6 @@
import java.nio.ByteBuffer;
import java.text.MessageFormat;
import org.eclipse.jgit.annotations.Nullable;
import org.eclipse.jgit.errors.PackInvalidException;
import org.eclipse.jgit.internal.storage.pack.PackExt;
@ -130,18 +129,18 @@ else if (size < cache.getBlockSize())
}
DfsBlock getOrLoadBlock(long pos, DfsReader ctx) throws IOException {
return cache.getOrLoad(this, pos, ctx, null);
try (ReadableChannel rc = ctx.db.openFile(desc, ext)) {
return cache.getOrLoad(this, pos, ctx, () -> rc);
}
}
DfsBlock readOneBlock(long pos, DfsReader ctx,
@Nullable ReadableChannel fileChannel) throws IOException {
DfsBlock readOneBlock(long pos, DfsReader ctx, ReadableChannel rc)
throws IOException {
if (invalid)
throw new PackInvalidException(getFileName());
ctx.stats.readBlock++;
long start = System.nanoTime();
ReadableChannel rc = fileChannel != null ? fileChannel
: ctx.db.openFile(desc, ext);
try {
int size = blockSize(rc);
pos = (pos / size) * size;
@ -189,9 +188,6 @@ DfsBlock readOneBlock(long pos, DfsReader ctx,
return new DfsBlock(key, pos, buf);
} finally {
if (rc != fileChannel) {
rc.close();
}
ctx.stats.readBlockMicros += elapsedMicros(start);
}
}

View File

@ -52,7 +52,6 @@
import java.util.function.Consumer;
import java.util.stream.LongStream;
import org.eclipse.jgit.annotations.Nullable;
import org.eclipse.jgit.internal.JGitText;
import org.eclipse.jgit.internal.storage.pack.PackExt;
@ -375,13 +374,13 @@ private static int tableSize(DfsBlockCacheConfig cfg) {
* @param ctx
* current thread's reader.
* @param fileChannel
* optional channel to read {@code pack}.
* supplier for channel to read {@code pack}.
* @return the object reference.
* @throws IOException
* the reference was not in the cache and could not be loaded.
*/
DfsBlock getOrLoad(BlockBasedFile file, long position, DfsReader ctx,
@Nullable ReadableChannel fileChannel) throws IOException {
ReadableChannelSupplier fileChannel) throws IOException {
final long requestedPosition = position;
position = file.alignToBlock(position);
@ -413,7 +412,8 @@ DfsBlock getOrLoad(BlockBasedFile file, long position, DfsReader ctx,
getStat(statMiss, key).incrementAndGet();
boolean credit = true;
try {
v = file.readOneBlock(requestedPosition, ctx, fileChannel);
v = file.readOneBlock(requestedPosition, ctx,
fileChannel.get());
credit = false;
} finally {
if (credit) {
@ -749,4 +749,16 @@ boolean has() {
interface RefLoader<T> {
Ref<T> load() throws IOException;
}
/**
* Supplier for readable channel
*/
@FunctionalInterface
public interface ReadableChannelSupplier {
/**
* @return ReadableChannel
* @throws IOException
*/
ReadableChannel get() throws IOException;
}
}

View File

@ -438,99 +438,77 @@ private byte[] decompress(long position, int sz, DfsReader ctx)
return dstbuf;
}
void copyPackAsIs(PackOutputStream out, DfsReader ctx)
throws IOException {
void copyPackAsIs(PackOutputStream out, DfsReader ctx) throws IOException {
// If the length hasn't been determined yet, pin to set it.
if (length == -1) {
ctx.pin(this, 0);
ctx.unpin();
}
if (cache.shouldCopyThroughCache(length)) {
copyPackThroughCache(out, ctx);
} else {
copyPackBypassCache(out, ctx);
try (ReadableChannel rc = ctx.db.openFile(desc, PACK)) {
int sz = ctx.getOptions().getStreamPackBufferSize();
if (sz > 0) {
rc.setReadAheadBytes(sz);
}
if (cache.shouldCopyThroughCache(length)) {
copyPackThroughCache(out, ctx, rc);
} else {
copyPackBypassCache(out, rc);
}
}
}
private void copyPackThroughCache(PackOutputStream out, DfsReader ctx)
throws IOException {
@SuppressWarnings("resource") // Explicitly closed in finally block
ReadableChannel rc = null;
try {
long position = 12;
long remaining = length - (12 + 20);
while (0 < remaining) {
DfsBlock b;
if (rc != null) {
b = cache.getOrLoad(this, position, ctx, rc);
} else {
b = cache.get(key, alignToBlock(position));
if (b == null) {
rc = ctx.db.openFile(desc, PACK);
int sz = ctx.getOptions().getStreamPackBufferSize();
if (sz > 0) {
rc.setReadAheadBytes(sz);
}
b = cache.getOrLoad(this, position, ctx, rc);
}
}
private void copyPackThroughCache(PackOutputStream out, DfsReader ctx,
ReadableChannel rc) throws IOException {
long position = 12;
long remaining = length - (12 + 20);
while (0 < remaining) {
DfsBlock b = cache.getOrLoad(this, position, ctx, () -> rc);
int ptr = (int) (position - b.start);
int n = (int) Math.min(b.size() - ptr, remaining);
b.write(out, position, n);
position += n;
remaining -= n;
}
}
private long copyPackBypassCache(PackOutputStream out, ReadableChannel rc)
throws IOException {
ByteBuffer buf = newCopyBuffer(out, rc);
long position = 12;
long remaining = length - (12 + 20);
boolean packHeadSkipped = false;
while (0 < remaining) {
DfsBlock b = cache.get(key, alignToBlock(position));
if (b != null) {
int ptr = (int) (position - b.start);
int n = (int) Math.min(b.size() - ptr, remaining);
b.write(out, position, n);
position += n;
remaining -= n;
rc.position(position);
packHeadSkipped = true;
continue;
}
} finally {
if (rc != null) {
rc.close();
buf.position(0);
int n = read(rc, buf);
if (n <= 0) {
throw packfileIsTruncated();
} else if (n > remaining) {
n = (int) remaining;
}
if (!packHeadSkipped) {
// Need skip the 'PACK' header for the first read
out.write(buf.array(), 12, n - 12);
packHeadSkipped = true;
} else {
out.write(buf.array(), 0, n);
}
position += n;
remaining -= n;
}
}
private long copyPackBypassCache(PackOutputStream out, DfsReader ctx)
throws IOException {
try (ReadableChannel rc = ctx.db.openFile(desc, PACK)) {
ByteBuffer buf = newCopyBuffer(out, rc);
if (ctx.getOptions().getStreamPackBufferSize() > 0) {
rc.setReadAheadBytes(ctx.getOptions().getStreamPackBufferSize());
}
long position = 12;
long remaining = length - (12 + 20);
boolean packHeadSkipped = false;
while (0 < remaining) {
DfsBlock b = cache.get(key, alignToBlock(position));
if (b != null) {
int ptr = (int) (position - b.start);
int n = (int) Math.min(b.size() - ptr, remaining);
b.write(out, position, n);
position += n;
remaining -= n;
rc.position(position);
packHeadSkipped = true;
continue;
}
buf.position(0);
int n = read(rc, buf);
if (n <= 0) {
throw packfileIsTruncated();
} else if (n > remaining) {
n = (int) remaining;
}
if (!packHeadSkipped) {
// Need skip the 'PACK' header for the first read
out.write(buf.array(), 12, n - 12);
packHeadSkipped = true;
} else {
out.write(buf.array(), 0, n);
}
position += n;
remaining -= n;
}
return position;
}
return position;
}
private ByteBuffer newCopyBuffer(PackOutputStream out, ReadableChannel rc) {

View File

@ -128,7 +128,9 @@ public ByteBuffer read(long pos, int cnt) throws IOException {
open().setReadAheadBytes(readAhead);
}
DfsBlock block = cache.getOrLoad(file, pos, ctx, ch);
DfsBlock block = cache.getOrLoad(file, pos, ctx, () -> {
return open();
});
if (block.start == pos && block.size() >= cnt) {
return block.zeroCopyByteBuffer(cnt);
}