Since commit
c5cb8f3b taught stat() to follow symlinks, and since initdb
uses pg_mkdir_p(), and that examines parent directories, our humble
readlink() implementation can now be exposed to junction points not of
PostgreSQL origin. Those might be corrupted by our naive path mangling,
which doesn't really understand NT paths in general.
Simply decline to transform paths that don't look like a drive absolute
path. That means that readlink() returns the NT path directly when
checking a parent directory of PGDATA that happen to point to a drive
using "rooted" format. That works for the purposes of our stat()
emulation.
Reported-by: Roman Zharkov <[email protected]>
Reviewed-by: Roman Zharkov <[email protected]>
Discussion: https://postgr.es/m/
4590c37927d7b8ee84f9855d83229018%40postgrespro.ru
Discussion: https://postgr.es/m/CA%2BhUKG%2BajSQ_8eu2AogTncOnZ5me2D-Cn66iN_-wZnRjLN%2Bicg%40mail.gmail.com
Backpatched commit
f71007fb as above by Thomas Munro into releases 13 thru 15
Discussion: https://postgr.es/m/CA+hUKGLbnv+pe3q1fYOVkLD3pMra7GuihfMxUN-1831YH9RYQg@mail.gmail.com
}
/*
- * If the path starts with "\??\", which it will do in most (all?) cases,
- * strip those out.
+ * If the path starts with "\??\" followed by a "drive absolute" path
+ * (known to Windows APIs as RtlPathTypeDriveAbsolute), then strip that
+ * prefix. This undoes some of the transformation performed by
+ * pqsymlink(), to get back to a format that users are used to seeing. We
+ * don't know how to transform other path types that might be encountered
+ * outside PGDATA, so we just return them directly.
*/
- if (r > 4 && strncmp(buf, "\\??\\", 4) == 0)
+ if (r >= 7 &&
+ buf[0] == '\\' &&
+ buf[1] == '?' &&
+ buf[2] == '?' &&
+ buf[3] == '\\' &&
+ isalpha(buf[4]) &&
+ buf[5] == ':' &&
+ buf[6] == '\\')
{
memmove(buf, buf + 4, strlen(buf + 4) + 1);
r -= 4;