bogus assert in logicalmsg_desc

Started by Tomas Vondraover 3 years ago3 messages
#1Tomas Vondra
tomas.vondra@enterprisedb.com
1 attachment(s)

Hi,

while experimenting with logical messages, I ran into this assert in
logicalmsg_desc:

Assert(prefix[xlrec->prefix_size] != '\0');

This seems to be incorrect, because LogLogicalMessage does this:

xlrec.prefix_size = strlen(prefix) + 1;

So prefix_size includes the null byte, so the assert points out at the
first payload byte. And of course, the check should be "==" because we
expect the byte to be \0, not the other way around.

It's pretty simple to make this crash by writing a logical message where
the first payload byte is \0, e.g. like this:

select pg_logical_emit_message(true, 'a'::text, '\x00'::bytea);

and then running pg_waldump on the WAL segment.

Attached is a patch addressing this. This was added in 14, so we should
backpatch to that version.

regards

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

Attachments:

logicalmsgdesc-fix.patchtext/x-patch; charset=UTF-8; name=logicalmsgdesc-fix.patchDownload
diff --git a/src/backend/access/rmgrdesc/logicalmsgdesc.c b/src/backend/access/rmgrdesc/logicalmsgdesc.c
index 099e11a84e7..08e03aa30d1 100644
--- a/src/backend/access/rmgrdesc/logicalmsgdesc.c
+++ b/src/backend/access/rmgrdesc/logicalmsgdesc.c
@@ -28,7 +28,7 @@ logicalmsg_desc(StringInfo buf, XLogReaderState *record)
 		char	   *message = xlrec->message + xlrec->prefix_size;
 		char	   *sep = "";
 
-		Assert(prefix[xlrec->prefix_size] != '\0');
+		Assert(prefix[xlrec->prefix_size - 1] == '\0');
 
 		appendStringInfo(buf, "%s, prefix \"%s\"; payload (%zu bytes): ",
 						 xlrec->transactional ? "transactional" : "non-transactional",
#2Masahiko Sawada
sawada.mshk@gmail.com
In reply to: Tomas Vondra (#1)
Re: bogus assert in logicalmsg_desc

On Mon, Aug 15, 2022 at 1:17 AM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:

Hi,

while experimenting with logical messages, I ran into this assert in
logicalmsg_desc:

Assert(prefix[xlrec->prefix_size] != '\0');

This seems to be incorrect, because LogLogicalMessage does this:

xlrec.prefix_size = strlen(prefix) + 1;

So prefix_size includes the null byte, so the assert points out at the
first payload byte. And of course, the check should be "==" because we
expect the byte to be \0, not the other way around.

It's pretty simple to make this crash by writing a logical message where
the first payload byte is \0, e.g. like this:

select pg_logical_emit_message(true, 'a'::text, '\x00'::bytea);

and then running pg_waldump on the WAL segment.

Attached is a patch addressing this. This was added in 14, so we should
backpatch to that version.

+1

The patch looks good to me.

Regards,

--
Masahiko Sawada
EDB: https://www.enterprisedb.com/

#3Richard Guo
guofenglinux@gmail.com
In reply to: Tomas Vondra (#1)
Re: bogus assert in logicalmsg_desc

On Mon, Aug 15, 2022 at 12:17 AM Tomas Vondra <tomas.vondra@enterprisedb.com>
wrote:

So prefix_size includes the null byte, so the assert points out at the
first payload byte. And of course, the check should be "==" because we
expect the byte to be \0, not the other way around.

Yes, indeed. There is even a comment emphasizing the trailing null byte
in LogLogicalMessage.

/* trailing zero is critical; see logicalmsg_desc */

Attached is a patch addressing this. This was added in 14, so we should
backpatch to that version.

+1 for the patch.

Thanks
Richard