Обсуждение: Ignoring symlinks when recovering time zone identifiers in initdb

There's a story making the rounds that the removal of the
/usr/share/zoneinfo/US/Pacific etc. in Debian broke installations.
These identifiers should not be used, but they may get picked by
initdb if they are shorter than the alternatives.  The length-based
tie breaking was introduced in commit e3846a00c2f ("Prefer timezone
name "UTC" over alternative spellings.").

Would it make sense to ignore symbolic links when searching for the
time zone identifier under /usr/share/zoneinfo in scan_available_timezones?

Something like this (untested):

diff --git a/src/bin/initdb/findtimezone.c b/src/bin/initdb/findtimezone.c
index 2b2ae39adf3..e492e904993 100644
--- a/src/bin/initdb/findtimezone.c
+++ b/src/bin/initdb/findtimezone.c
@@ -677,7 +677,11 @@ scan_available_timezones(char *tzdir, char *tzdirsub, struct tztry *tt,
         snprintf(tzdir + tzdir_orig_len, MAXPGPATH - tzdir_orig_len,
                  "/%s", name);
 
-        if (stat(tzdir, &statbuf) != 0)
+        /*
+         * Ignore symbolic links, so that shorter aliases in directories such
+         * as "US" are not preferred over the "America" directory.
+         */
+        if (lstat(tzdir, &statbuf) != 0)
         {
 #ifdef DEBUG_IDENTIFY_TIMEZONE
             fprintf(stderr, "could not stat \"%s\": %m\n",
@@ -693,7 +697,7 @@ scan_available_timezones(char *tzdir, char *tzdirsub, struct tztry *tt,
             scan_available_timezones(tzdir, tzdirsub, tt,
                                      bestscore, bestzonename);
         }
-        else
+        else if (S_ISREG(statbuf.st_mode))
         {
             /* Load and test this file */
             int            score = score_timezone(tzdirsub, tt);

Florian Weimer <fw@deneb.enyo.de> writes:
> There's a story making the rounds that the removal of the
> /usr/share/zoneinfo/US/Pacific etc. in Debian broke installations.
> These identifiers should not be used, but they may get picked by
> initdb if they are shorter than the alternatives.

Ugh.  (Fortunately, I don't think fixing this is any harder than
editing the timezone entries in postgresql.conf.)

> The length-based
> tie breaking was introduced in commit e3846a00c2f ("Prefer timezone
> name "UTC" over alternative spellings.").

Just for the record, it's far older than that.  Looks like it
dates to 2def4552e from 2004.

> Would it make sense to ignore symbolic links when searching for the
> time zone identifier under /usr/share/zoneinfo in scan_available_timezones?

That seems only vaguely related to the problem.  If a distro
arbitrarily decides they can remove some zoneinfo entries,
that's going to break someone.  Also, it's common to implement
zone links as hardlinks not symlinks (Red Hat, for one, does
it that way), in which case this heuristic won't help at all.
I'd just as soon not introduce a behavior that will result in
choosing different zone names on different platforms even
when they have logically-equivalent zoneinfo contents.

            regards, tom lane

I wrote:
> Florian Weimer <fw@deneb.enyo.de> writes:
>> There's a story making the rounds that the removal of the
>> /usr/share/zoneinfo/US/Pacific etc. in Debian broke installations.
>> These identifiers should not be used, but they may get picked by
>> initdb if they are shorter than the alternatives.

> Ugh.  (Fortunately, I don't think fixing this is any harder than
> editing the timezone entries in postgresql.conf.)

BTW, I wonder why 23bd3cec6 ("Attempt to identify system timezone by
reading /etc/localtime symlink") didn't help here.  It looks to me
like Debian follows the convention of /etc/localtime being a symlink,
so we should have chosen whichever zone is the system's setting
and never fallen back to our zone-name heuristics at all.

            regards, tom lane