[pfx] Postfix is incompatible with LMDB 1.0

Christian Pfeiffer via Postfix-users <[email protected]> Thu, 02 Jul 2026 17:37:57 +0200
Newsgroups	gmane.mail.postfix.user
Message-ID	<[email protected]>
--===============8941839574997696508==
Content-Type: multipart/alternative;
 boundary="----=_=-_OpenGroupware_org_NGMime-4283-1783006677.054223-3------"

------=_=-_OpenGroupware_org_NGMime-4283-1783006677.054223-3------
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Content-Length: 2488


Hi,
I'm writing this in the context of ArchLinux pushing LMDB 1.0 into test=
ing.
It turns out that aside from a need to convert the LMDB databases by us=
ing mdb=5Fdump before the upgrade and mdb=5Fload to recreate them after=
wards (or running newaliases and postmap -F again for Postfix specifica=
lly), there are some compatibility problems.
The first issue is that with LMDB 1.0, running either postmap or newali=
ases now causes the following message
postmap: fatal: error updating lmdb:./aliases: MDB=5FBAD=5FTXN: Transac=
tion must abort, has a child, or is invalid
I looked at it with GDB, the cause for this is a change in LMDB 1.0, na=
mely this commit:
ITS#8174 LMDB: fixes for mdb=5Fdrop(MAIN=5FDBI) =C2=B7 LMDB/lmdb@d0fcfd=
3
Dropping the MAIN=5FDBI now causes the transaction to be marked MDB=5FT=
XN=5FDIRTY and MDB=5FTXN=5FDROPPED, which means any further usage of th=
e transaction that was used to drop MAIN=5FDBI will cause MDB=5FBAD=5FT=
XN.
For Postfix, all databases are dropped by slmdb=5Fprepare with this sta=
tement: postfix/postfix/src/util/slmdb.c at v3.11.4 =C2=B7 vdukhovni/po=
stfix
This causes lmdb/libraries/liblmdb/mdb.c at mdb.master3 =C2=B7 LMDB/lmd=
b to mark the transaction as dirty & dropped, and then when slmdb=5Fput=
 is called, it will call mdb=5Fput with the same transaction handle her=
e postfix/postfix/src/util/slmdb.c at master =C2=B7 vdukhovni/postfix c=
ausing MDB=5FBAD=5FTXN to be returned from here lmdb/libraries/liblmdb/=
mdb.c at mdb.master3 =C2=B7 LMDB/lmdb .
Now, I've made a temporary fix for Arch: Postfix LMDB 1.0 fix .=C2=A0
I'm not sure this is the right approach, and it doesn't seem to fix all=
 problems yet, as calling newaliases again on an already existing LMDB =
now causes duplicate key warnings, so I don't think this is right or co=
mplete just yet. My approach was to make slmdb=5Fprepare to drop the da=
tabase, then abort the transaction as is required, being a new one, and=
 then reopen the database with mdb=5Fdbi=5Fopen.
For reference, this is the downstream discussion at Arch: version 1.0.0=
 in testing breaks postfix (#5) =C2=B7 Issue =C2=B7 archlinux/packaging=
/packages/lmdb It covers LMDB 1.0 in general (other apps like bogofilte=
r also have issues not yet identified), but the discussion regarding Po=
stfix is in there, too.
But obviously, I would want to get this fixed upstream rather than doin=
g any downstream patching experiments.
Regards,
Christian Pfeiffer

------=_=-_OpenGroupware_org_NGMime-4283-1783006677.054223-3------
Content-Type: text/html; charset=utf-8
Content-Transfer-Encoding: quoted-printable
Content-Length: 3567

<html><p>Hi,</p><p>I'm writing this in the context of ArchLinux pushing=
 LMDB 1.0 into testing.</p><p>It turns out that aside from a need to co=
nvert the LMDB databases by using <code>mdb=5Fdump</code> before the up=
grade and <code>mdb=5Fload</code> to recreate them afterwards (or runni=
ng <code>newaliases</code> and <code>postmap -F</code> again for Postfi=
x specifically), there are some compatibility problems.</p><p>The first=
 issue is that with LMDB 1.0, running either <code>postmap</code> or <c=
ode>newaliases</code> now causes the following message</p><blockquote><=
p>postmap: fatal: error updating lmdb:./aliases: MDB=5FBAD=5FTXN: Trans=
action must abort, has a child, or is invalid</p></blockquote><p>I look=
ed at it with GDB, the cause for this is a change in LMDB 1.0, namely t=
his commit:</p><p><a href=3D"https://github.com/LMDB/lmdb/commit/d0fcfd=
3d02321d04b56a23e6c86c80309f4ecef5">ITS#8174 LMDB: fixes for mdb=5Fdrop=
(MAIN=5FDBI) =C2=B7 LMDB/lmdb@d0fcfd3</a></p><p>Dropping the <code>MAIN=
=5FDBI</code> now causes the transaction to be marked <code>MDB=5FTXN=5F=
DIRTY</code> and <code>MDB=5FTXN=5FDROPPED</code>, which means any furt=
her usage of the transaction that was used to drop <code>MAIN=5FDBI</co=
de> will cause <code>MDB=5FBAD=5FTXN</code>.</p><p>For Postfix, all dat=
abases are dropped by <code>slmdb=5Fprepare</code> with this statement:=
 <a href=3D"https://github.com/vdukhovni/postfix/blob/v3.11.4/postfix/s=
rc/util/slmdb.c#L399">postfix/postfix/src/util/slmdb.c at v3.11.4 =C2=B7=
 vdukhovni/postfix</a></p><p>This causes <a href=3D"https://github.com/=
LMDB/lmdb/blob/mdb.master3/libraries/liblmdb/mdb.c#L12544">lmdb/librari=
es/liblmdb/mdb.c at mdb.master3 =C2=B7 LMDB/lmdb</a> to mark the transa=
ction as dirty &amp; dropped, and then when <code>slmdb=5Fput</code> is=
 called, it will call <code>mdb=5Fput</code> with the same transaction =
handle here <a href=3D"https://github.com/vdukhovni/postfix/blob/master=
/postfix/src/util/slmdb.c#L613">postfix/postfix/src/util/slmdb.c at mas=
ter =C2=B7 vdukhovni/postfix</a> causing <code>MDB=5FBAD=5FTXN</code> t=
o be returned from here <a href=3D"https://github.com/LMDB/lmdb/blob/md=
b.master3/libraries/liblmdb/mdb.c#L11099">lmdb/libraries/liblmdb/mdb.c =
at mdb.master3 =C2=B7 LMDB/lmdb</a> .</p><p>Now, I've made a temporary =
fix for Arch: <a href=3D"https://gist.github.com/ChrisTX/731122de698a13=
497d808f8e0072310b">Postfix LMDB 1.0 fix</a> .&nbsp;</p><p>I'm not sure=
 this is the right approach, and it doesn't seem to fix all problems ye=
t, as calling <code>newaliases</code> again on an already existing LMDB=
 now causes duplicate key warnings, so I don't think this is right or c=
omplete just yet. My approach was to make <code>slmdb=5Fprepare</code> =
to drop the database, then abort the transaction as is required, being =
a new one, and then reopen the database with <code>mdb=5Fdbi=5Fopen</co=
de>.</p><p>For reference, this is the downstream discussion at Arch: <a=
 href=3D"https://gitlab.archlinux.org/archlinux/packaging/packages/lmdb=
/-/work=5Fitems/5">version 1.0.0 in testing breaks postfix (#5) =C2=B7 =
Issue =C2=B7 archlinux/packaging/packages/lmdb</a> It covers LMDB 1.0 i=
n general (other apps like bogofilter also have issues not yet identifi=
ed), but the discussion regarding Postfix is in there, too.</p><p>But o=
bviously, I would want to get this fixed upstream rather than doing any=
 downstream patching experiments.</p><p>Regards,</p><p>Christian Pfeiff=
er</p></html>

------=_=-_OpenGroupware_org_NGMime-4283-1783006677.054223-3--------


--===============8941839574997696508==
Content-Type: text/plain; charset="us-ascii"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Content-Disposition: inline

_______________________________________________
Postfix-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]

--===============8941839574997696508==--