Read files for a backup in pieces of 4 MiB instead of 8 KiB

PHP reads a file 8 KiB at a time by default. On a disk where every
request waits, such as a network or cloud volume, that made reading five
times slower: on the test server 5 MiB/s instead of 27 MiB/s. Without
PHP's read buffer every read of a blob is one request to the disk. The
same for packs on a local location.

CHANGELOG.md

@@ -40,6 +40,9 @@ shown under 1.0.0. The git tag of a release candidate goes after the date.
 - The key page warns against creating a new key when you are restoring a server.
 
 ### Changed
+- Reading files for a backup is much faster on disks where every request waits, such as network and
+  cloud volumes: files are read in pieces of 4 MiB instead of 8 KiB. On a test server the disk
+  delivered 27 MiB/s instead of 5 MiB/s.
 - Locations are backed up one after the other: a backup that has to wait shows as queued. A restore
   point has the time its backup really began, not the time it was queued.
 - Choosing a time no longer starts a backup right away when that time has already passed today.

lib/Backend/LocalBackend.php

@@ -67,6 +67,9 @@ final class LocalBackend implements IBackend {
 		if ($fh === false) {
 			throw new BackendException('Not found: ' . $path);
 		}
+		// Packs are read in large pieces: one request each, not 8 KiB at a time (see Repository::openForReading()).
+		stream_set_read_buffer($fh, 0);
+		stream_set_chunk_size($fh, 4 * 1048576);
 		return $fh;
 	}
 

lib/Job/BackupRun.php

@@ -120,7 +120,7 @@ final class BackupRun {
 					if ($cur === null || $cur['path'] !== $logical) {
 						$cur = ['path' => $logical, 'offset' => 0, 'blobs' => [], 'version' => $version];
 					}
-					$fh = @fopen($path, 'rb');
+					$fh = Repository::openForReading($path);
 					if ($fh === false) {
 						self::recordUnreadable($state, $stats, $logical);
 						$state['cur'] = null;

lib/Repository/Repository.php

@@ -193,6 +193,23 @@ final class Repository {
 		return $config;
 	}
 
+	/**
+	 * Open a local file to read it in whole blobs. PHP reads a file in pieces of 8 KiB by default;
+	 * on a disk where each request waits (a network or cloud volume) that made reading five times
+	 * slower: 5 MiB/s instead of 27 MiB/s on nc01. Without PHP's read buffer every fread() of a
+	 * blob is one request to the disk.
+	 *
+	 * @return resource|false
+	 */
+	public static function openForReading(string $path) {
+		$fh = @fopen($path, 'rb');
+		if ($fh !== false) {
+			stream_set_read_buffer($fh, 0);
+			stream_set_chunk_size($fh, self::BLOB_SIZE);
+		}
+		return $fh;
+	}
+
 	private static function configMatches(array $config, KeyRing $keys): bool {
 		try {
 			return isset($config['check'], $config['id'])
@@ -1148,7 +1165,10 @@ final class Repository {
 		for ($attempt = 1; $attempt <= 3; $attempt++) {
 			$before = self::fileVersion($path);
 			$blobs = [];
-			$fh = fopen($path, 'rb');
+			$fh = self::openForReading($path);
+			if ($fh === false) {
+				throw new RepositoryException('Cannot read ' . $path);
+			}
 			while (($data = self::readFull($fh, self::BLOB_SIZE)) !== '') {
 				$blobs[] = $this->storeBlob($data, $packs, $stats);
 			}

psalm-baseline.xml

@@ -367,7 +367,6 @@
       <code><![CDATA[$encrypted]]></code>
       <code><![CDATA[$encrypted]]></code>
       <code><![CDATA[$encrypted]]></code>
-      <code><![CDATA[$fh]]></code>
       <code><![CDATA[$plain]]></code>
       <code><![CDATA[$plain]]></code>
       <code><![CDATA[$plain]]></code>