[PATCH 1/1] pg_rewind: Fetch small files according to new size.

agora inbox for [email protected]  
help / color / mirror / Atom feed

[PATCH 1/1] pg_rewind: Fetch small files according to new size.
23+ messages / 2 participants
[nested] [flat]

* [PATCH 1/1] pg_rewind: Fetch small files according to new size.
@ 2020-11-13 09:04 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2020-11-13 09:04 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 47beba277a4..f544d9828a2 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index fa1b6e80ec3..c1ad2c49c5f 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that we size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 2bbf8e74385..fe828c9e840 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -525,10 +525,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index e87f239a47a..a8dc9fda0da 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.20.1


--------------70A2CAC3C3DFA320063B7D3B--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* [PATCH v2 1/1] pg_rewind: Fetch small files according to new size.
@ 2021-01-22 13:16 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Heikki Linnakangas @ 2021-01-22 13:16 UTC (permalink / raw)

There's a race condition if a file changes in the source system after we
have collected the file list. If the file becomes larger, we only fetched
up to its original size. That can easily result in a truncated file.
That's not a problem for relation files, files in pg_xact, etc. because
any actions on them will be replayed from the WAL. However, configuration
files are affected.

This commit mitigates the race condition by fetching small files in
whole, even if they have grown. This is not a full fix: we still believe
the original file size for files larger than 1 MB. That should be enough
for configuration files, and doing more than that would require bigger
changes to the chunking logic in in libpq_source.c.

That mitigates the race condition if the file is modified between the
original scan of files and copying the file, but there's still a race
condition if a file is changed while it's being copied. That's a much
smaller window, though, and pg_basebackup has the same issue.

I ran into this while playing with pg_auto_failover, which frequently
uses ALTER SYSTEM, which update postgresql.auto.conf. Often, pg_rewind
would fail, because the postgresql.auto.conf file changed concurrently
and a partial version of it was copied to the target. The partial
file would fail to parse, preventing the server from starting up.

Reviewed-by: Cary Huang
Discussion: https://www.postgresql.org/message-id/f67feb24-5833-88cb-1020-19a4a2b83ac7%40iki.fi
---
 src/bin/pg_rewind/libpq_source.c  | 32 +++++++++++++
 src/bin/pg_rewind/local_source.c  | 76 +++++++++++++++++++++++++++----
 src/bin/pg_rewind/pg_rewind.c     |  5 +-
 src/bin/pg_rewind/rewind_source.h | 13 ++++++
 4 files changed, 112 insertions(+), 14 deletions(-)

diff --git a/src/bin/pg_rewind/libpq_source.c b/src/bin/pg_rewind/libpq_source.c
index 86d2adcaee9..ff16add16f5 100644
--- a/src/bin/pg_rewind/libpq_source.c
+++ b/src/bin/pg_rewind/libpq_source.c
@@ -63,6 +63,7 @@ static void process_queued_fetch_requests(libpq_source *src);
 /* public interface functions */
 static void libpq_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
+static void libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len);
 static void libpq_queue_fetch_range(rewind_source *source, const char *path,
 									off_t off, size_t len);
 static void libpq_finish_fetch(rewind_source *source);
@@ -88,6 +89,7 @@ init_libpq_source(PGconn *conn)
 
 	src->common.traverse_files = libpq_traverse_files;
 	src->common.fetch_file = libpq_fetch_file;
+	src->common.queue_fetch_file = libpq_queue_fetch_file;
 	src->common.queue_fetch_range = libpq_queue_fetch_range;
 	src->common.finish_fetch = libpq_finish_fetch;
 	src->common.get_current_wal_insert_lsn = libpq_get_current_wal_insert_lsn;
@@ -307,6 +309,36 @@ libpq_traverse_files(rewind_source *source, process_file_callback_t callback)
 	PQclear(res);
 }
 
+/*
+ * Queue up a request to fetch a file from remote system.
+ */
+static void
+libpq_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	/*
+	 * Truncate the target file immediately, and queue a request to fetch it
+	 * from the source. If the file is small, smaller than MAX_CHUNK_SIZE,
+	 * request fetching a full-sized chunk anyway, so that if the file has
+	 * become larger in the source system, after we scanned the source
+	 * directory, we still fetch the whole file. This only works for files up
+	 * to MAX_CHUNK_SIZE, but that's good enough for small configuration files
+	 * and such that are changed every now and then, but not WAL-logged.
+	 * For larger files, we fetch up to the original size.
+	 *
+	 * Even with that mechanism, there is an inherent race condition if the
+	 * file is modified at the same instant that we're copying it, so that we
+	 * might copy a torn version of the file with one half from the old
+	 * version and another half from the new. But pg_basebackup has the same
+	 * problem, and it hasn't been problem in practice.
+	 *
+	 * It might seem more natural to truncate the file later, when we receive
+	 * it from the source server, but then we'd need to track which
+	 * fetch-requests are for a whole file.
+	 */
+	open_target_file(path, true);
+	libpq_queue_fetch_range(source, path, 0, Max(len, MAX_CHUNK_SIZE));
+}
+
 /*
  * Queue up a request to fetch a piece of a file from remote system.
  */
diff --git a/src/bin/pg_rewind/local_source.c b/src/bin/pg_rewind/local_source.c
index 9c3491c3fba..1899d1cc4ae 100644
--- a/src/bin/pg_rewind/local_source.c
+++ b/src/bin/pg_rewind/local_source.c
@@ -29,8 +29,10 @@ static void local_traverse_files(rewind_source *source,
 								 process_file_callback_t callback);
 static char *local_fetch_file(rewind_source *source, const char *path,
 							  size_t *filesize);
-static void local_fetch_file_range(rewind_source *source, const char *path,
-								   off_t off, size_t len);
+static void local_queue_fetch_file(rewind_source *source, const char *path,
+								   size_t len);
+static void local_queue_fetch_range(rewind_source *source, const char *path,
+									off_t off, size_t len);
 static void local_finish_fetch(rewind_source *source);
 static void local_destroy(rewind_source *source);
 
@@ -43,7 +45,8 @@ init_local_source(const char *datadir)
 
 	src->common.traverse_files = local_traverse_files;
 	src->common.fetch_file = local_fetch_file;
-	src->common.queue_fetch_range = local_fetch_file_range;
+	src->common.queue_fetch_file = local_queue_fetch_file;
+	src->common.queue_fetch_range = local_queue_fetch_range;
 	src->common.finish_fetch = local_finish_fetch;
 	src->common.get_current_wal_insert_lsn = NULL;
 	src->common.destroy = local_destroy;
@@ -65,12 +68,65 @@ local_fetch_file(rewind_source *source, const char *path, size_t *filesize)
 	return slurpFile(((local_source *) source)->datadir, path, filesize);
 }
 
+/*
+ * Copy a file from source to target.
+ *
+ * 'len' is the expected length of the file.
+ */
+static void
+local_queue_fetch_file(rewind_source *source, const char *path, size_t len)
+{
+	const char *datadir = ((local_source *) source)->datadir;
+	PGAlignedBlock buf;
+	char		srcpath[MAXPGPATH];
+	int			srcfd;
+	size_t		written_len;
+
+	snprintf(srcpath, sizeof(srcpath), "%s/%s", datadir, path);
+
+	/* Open source file for reading. */
+	srcfd = open(srcpath, O_RDONLY | PG_BINARY, 0);
+	if (srcfd < 0)
+		pg_fatal("could not open source file \"%s\": %m",
+				 srcpath);
+
+	/* Truncate and open the target file for writing. */
+	open_target_file(path, true);
+
+	written_len = 0;
+	for (;;)
+	{
+		ssize_t		read_len;
+
+		read_len = read(srcfd, buf.data, sizeof(buf));
+
+		if (read_len < 0)
+			pg_fatal("could not read file \"%s\": %m", srcpath);
+		else if (read_len == 0)
+			break;	/* EOF reached */
+
+		write_target_range(buf.data, written_len, read_len);
+		written_len += read_len;
+	}
+
+	/*
+	 * A local source is not expected to change while we're rewinding, so check
+	 * that the size of the file matches our earlier expectation.
+	 */
+	if (written_len != len)
+		pg_fatal("size of source file \"%s\" changed concurrently: " UINT64_FORMAT " bytes expected, " UINT64_FORMAT " copied",
+				 srcpath, len, written_len);
+
+	if (close(srcfd) != 0)
+		pg_fatal("could not close file \"%s\": %m", srcpath);
+}
+
 /*
  * Copy a file from source to target, starting at 'off', for 'len' bytes.
  */
 static void
-local_fetch_file_range(rewind_source *source, const char *path, off_t off,
-					   size_t len)
+local_queue_fetch_range(rewind_source *source, const char *path, off_t off,
+						size_t len)
 {
 	const char *datadir = ((local_source *) source)->datadir;
 	PGAlignedBlock buf;
@@ -94,14 +150,14 @@ local_fetch_file_range(rewind_source *source, const char *path, off_t off,
 	while (end - begin > 0)
 	{
 		ssize_t		readlen;
-		size_t		len;
+		size_t		thislen;
 
 		if (end - begin > sizeof(buf))
-			len = sizeof(buf);
+			thislen = sizeof(buf);
 		else
-			len = end - begin;
+			thislen = end - begin;
 
-		readlen = read(srcfd, buf.data, len);
+		readlen = read(srcfd, buf.data, thislen);
 
 		if (readlen < 0)
 			pg_fatal("could not read file \"%s\": %m", srcpath);
@@ -120,7 +176,7 @@ static void
 local_finish_fetch(rewind_source *source)
 {
 	/*
-	 * Nothing to do, local_fetch_file_range() copies the ranges immediately.
+	 * Nothing to do, local_queue_fetch_range() copies the ranges immediately.
 	 */
 }
 
diff --git a/src/bin/pg_rewind/pg_rewind.c b/src/bin/pg_rewind/pg_rewind.c
index 359a6a587cb..9030a1505e3 100644
--- a/src/bin/pg_rewind/pg_rewind.c
+++ b/src/bin/pg_rewind/pg_rewind.c
@@ -537,10 +537,7 @@ perform_rewind(filemap_t *filemap, rewind_source *source,
 				break;
 
 			case FILE_ACTION_COPY:
-				/* Truncate the old file out of the way, if any */
-				open_target_file(entry->path, true);
-				source->queue_fetch_range(source, entry->path,
-										  0, entry->source_size);
+				source->queue_fetch_file(source, entry->path, entry->source_size);
 				break;
 
 			case FILE_ACTION_TRUNCATE:
diff --git a/src/bin/pg_rewind/rewind_source.h b/src/bin/pg_rewind/rewind_source.h
index 2da92dbff94..799b7c120ea 100644
--- a/src/bin/pg_rewind/rewind_source.h
+++ b/src/bin/pg_rewind/rewind_source.h
@@ -47,6 +47,19 @@ typedef struct rewind_source
 	void		(*queue_fetch_range) (struct rewind_source *, const char *path,
 									  off_t offset, size_t len);
 
+	/*
+	 * Like queue_fetch_range(), but requests replacing the whole local file
+	 * from the source system. 'len' is the expected length of the file,
+	 * although when the source is a live server, the file may change
+	 * concurrently. The implementation is not obliged to copy more than 'len'
+	 * bytes, even if the file is larger. However, to avoid copying a
+	 * truncated version of the file, which can cause trouble if e.g. a
+	 * configuration file is modified concurrently, the implementation should
+	 * try to copy the whole file, even if it's larger than expected.
+	 */
+	void		(*queue_fetch_file) (struct rewind_source *, const char *path,
+									 size_t len);
+
 	/*
 	 * Execute all requests queued up with queue_fetch_range().
 	 */
-- 
2.29.2


--------------79EE74DA33DF3081397629A9--





^ permalink  raw  reply  [nested|flat] 23+ messages in thread

* basebackups seem to have serious issues with FILE_COPY in CREATE DATABASE
@ 2024-06-24 14:12 Tomas Vondra <[email protected]>
  0 siblings, 0 replies; 23+ messages in thread

From: Tomas Vondra @ 2024-06-24 14:12 UTC (permalink / raw)
  To: PostgreSQL Hackers <[email protected]>

Hi,

While doing some additional testing of (incremental) backups, I ran into
a couple regular failures. After pulling my hair for a couple days, I
realized the issue seems to affect regular backups, and incremental
backups (which I've been trying to test) are likely innocent.

I'm using a simple (and admittedly not very pretty) bash scripts that
takes and verified backups, concurrently with this workload:

1) initialize a cluster

2) initialize pgbench in database 'db'

3) run short pgbench on 'db'

4) maybe do vacuum [full] on 'db'

5) drop a database 'db_copy' if it exists

6) create a database 'db_copy' by copying 'db' using one of the
   available strategies (file_copy, wal_log)

7) run short pgbench on 'db_copy'

8) maybe do vacuum [full] on 'db_copy'

And concurrently with this, it takes a basebackup, starts a cluster on
it (on a different port, ofc), and does various checks on that:

a) verify checksums using pg_checksums (cluster has them enabled)

b) run amcheck on tables/indexes on both databases

c) SQL check (we expect all tables to be 'consistent' as if we did a
PITR - in particular sum(balance) is expected to be the same value on
all pgbench tables) on both databases

I believe those are reasonable expectations - that we get a database
with valid checksums, with non-broken tables/indexes, and that the
database looks as a snapshot taken at a single instant.

Unfortunately it doesn't take long for the tests to start failing with
various strange symptoms on the db_copy database (I'm yet to see an
issue on the 'db' database):

i) amcheck fails with 'heap tuple lacks matching index tuple'

    ERROR:  heap tuple (116195,22) from table "pgbench_accounts" lacks
    matching index tuple within index "pgbench_accounts_pkey"
    HINT:  Retrying verification using the function
    bt_index_parent_check() might provide a more specific error.

    I've seen this with other tables/indexes too, e.g. system catalogs
    pg_statitics or toast tables, but 'accounts' is most common.

ii) amcheck fails with 'could not open file'

    ERROR:  could not open file "base/18121/18137": No such file or
    directory
    LINE 9:  lateral verify_heapam(relation => c.oid, on_error_stop =>
    f...
                     ^
    ERROR:  could not open file "base/18121/18137": No such file or
    directory

iii) failures in the SQL check, with different tables have different
balance sums

    SQL check fails (db_copy) (account 156142 branches 136132 tellers
                               136132 history -42826)

    Sometimes this is preceded by amcheck issue, but not always.

I guess this is not the behavior we expect :-(

I've reproduced all of this on PG16 - I haven't tried with older
releases, but I have no reason to assume pre-16 releases are not affected.

With incremental backups I've observed a couple more symptoms, but those
are most likely just fallout of this - not realizing the initial state
is a bit wrong, and making it worse by applying the increments.

The important observation is that this only happens if a database is
created while the backup is running, and that it only happens with the
FILE_COPY strategy - I've never seen this with WAL_LOG (which is the
default since PG15).

I don't recall any reports of similar issues from pre-15 releases, where
FILE_COPY was the only available option - I'm not sure why is that.
Either it didn't have this issue back then, or maybe people happen to
not create databases concurrently with a backup very often. It's a race
condition / timing issue, essentially.

I have no ambition to investigate this part of the code much deeper, or
invent a fix myself, at least not in foreseeable future. But it seems
like something we probably should fix - subtly broken backups are not a
great thing.

I see there have been a couple threads proposing various improvements to
FILE_COPY, that might make it more efficient/faster, namely using the
filesystem cloning [1] or switching pg_upgrade to use it [2]. But having
something that's (maybe) faster but not quite correct does not seem like
a winning strategy to me ...

Alternatively, if we don't have clear desire to fix it, maybe the right
solution would be get rid of it?

regards

[1]
https://www.postgresql.org/message-id/[email protected]...

[2] https://www.postgresql.org/message-id/Zl9ta3FtgdjizkJ5%40nathan

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Attachments:

  [application/x-compressed-tar] basebackup-test-scripts.tgz (13.9K, ../../[email protected]/2-basebackup-test-scripts.tgz)
  download

^ permalink  raw  reply  [nested|flat] 23+ messages in thread

end of thread, other threads:[~2024-06-24 14:12 UTC | newest]

Thread overview: 23+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2020-11-13 09:04 [PATCH 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2021-01-22 13:16 [PATCH v2 1/1] pg_rewind: Fetch small files according to new size. Heikki Linnakangas <[email protected]>
2024-06-24 14:12 basebackups seem to have serious issues with FILE_COPY in CREATE DATABASE Tomas Vondra <[email protected]>

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox