Re: [Samba] winbind causing huge timeouts/delays since 4.8
- Date: Tue, 26 Feb 2019 11:19:45 +0100
- From: Alexander Spannagel via samba <samba@xxxxxxxxxxxxxxx>
- Subject: Re: [Samba] winbind causing huge timeouts/delays since 4.8
Am 23.02.19 um 22:23 schrieb Rowland Penny via samba:
Agreed, but not feasible in our company as there are - for good reasons
- two big server farms spreading across multiple countries. One is based
on Linux and the other on Windows. The glue between those two farms is
build upon samba due to it's flexibility, strength and reliability.
On Sat, 23 Feb 2019 21:54:31 +0100
Alexander Spannagel via samba <samba@xxxxxxxxxxxxxxx> wrote:
Am 23.02.19 um 15:48 schrieb Rowland Penny via samba:
If you have, as you have, 'files sss winbind' in the the
passwd & group line in nsswitch.conf, means this:
First /etc/passwd or /etc/group is searched and if the user or
group is found, this info is returned.
Next sssd will be asked, 'do you know this user or group ?' if
found, the info is returned.
Finally winbind will be asked, 'do you know this user or
group ?' if found, the info is returned.
Lets take a user called 'fred', this user is in AD. The first
search will return nothing, so sssd is asked, this 'asks' AD
and returns the users info. Finally, wait that's it, we have
the info, there is no need to ask winbind for anything.
who does what.
That is incorrect. Alexander stated:
No. we use max. 3 auth providers: (1. and 2. on all unix
servers) 1. unix (local passwd)
for static OS/service accounts across all our env
2. sssd (with unix ldap servers as provider)
unix experienced user and application related service accounts
for windows users/services needing access to a group of unix
They don't - as stated above we use sssd for query/caching
entries from our ldap directory server and not Windows
- also this is possible, but makes more trouble and don't
provide what samba's smb/windbind does.
He clearly writes (in multiple emails) that sssd is configured
to use his unix ldap servers and not AD.
Maybe three sources of user databases is not regular, but I
fail to see why this should be a problem (provided that
usernames, uidNumbers and such are unique across the
And there is the problem, if 'fred' is in /etc/passwd, that user
will be used, but what if you meant fred in ldap or AD ?
We are aware of this possible clash and it's handled during users
If you only used one database, you would be 100% sure you wouldn't get
Beside what i mentioned earlier you are rigth that the RFC2307
attributes are still around, but reading a blog entry from 2016 in
technet it's cleraly stated that one should look on other alternatives
as it may go in future realease. Here the comment:
There is absolutely no point in having 4 databases (yes there
are 4, Unix, sssd, winbind and the ldap lines in smb.conf), they
could all be combined in AD.
No it won't work as our windows team doesn't accept schema changes
for unix in AD.
How shall I put this, your windows team is either not telling you the
entire truth or is stupid. The entire RFC2307 attributes are part of
the standard Windows schema. what isn't added is IDMU and this just
makes the Unix attributes tab work in ADUC. You can however use other
tools to create Unix users etc.
"I am using Windows Server IDMU/NIS Server role today, what should I do?
We recommend to start planning for alternatives, for example: native
LDAP, Samba Client, Kerberos or other non-Microsoft options. Existing
Windows Server 2012 R2 or earlier deployments will continue to be
supported in accordance with the Microsoft Support lifecycle."
Taken out of this blog entry:
introduced patch for Bug 13503 "getpwnam resolves local system accounts
to AD accounts" with 4.8.4 it may be the same problem. As mentioned
earlier removing patch for Bug 13503 also resolves seen problems.
The main problem is that the OP wants Samba changing to cope
with his mess, it might be a valid change, but the reason for
the change is invalid.
The intial reason we hitted this after upgrade from samba-4.7 to 4.8
is a script that frequentyl checks the system for changes and a final
"chown root.wheel FILE" freezes the system for approx a minute
(simliar to "wbinfo -i foo"). The winbind and also sssd log showd
that both were asked about a user "root.wheel" which is another
question, why the notation which usually (under linux) indicates
user.group and not an account with a dot in it's name - but more to
glibc related. Removing winbind from nss fixed the freeze, but isn't
an option. It leads to the point that asking winbind for an uknown
user without domain took a long time before it returns
You see, there you have 2 methods checking for the same thing, also the
problem you are referring to has been fixed in the latest Samba (I
think it has also been backported).
Well, I think the problem is you _assume_ users are in multiple
databases and we just don't know that. I think there is a good
change Alexander perfectly knows what he is doing and users are
unique across databases.
Nevertheless, at some point nss is clearly querying winbind,
which means nss did not find the user in either /etc/passwd nor
via sssd. In the case that winbind _is_ queried, Alexander is
experiencing, like he wrote, 'frequently system hangs/slowness
for a couple of seconds' and he observed that winbind is causing
So maybe we should set our focus on winbind instead of the
multiple database stuff and figure out why it behaves like this
since the upgrade from 4.7 to 4.8. I would say we should start
with fixing the winbind stuff in smb.conf. Right?
P.S. I am following this thread since I also noticed occasional
'hangs' when the system is querying winbind. This is Samba 4.8.7
on FreeBSD 11.2.As the function - that i suggested patches for - hasn't changed since
They don't and shouldn't provide same thing in our setup. The main jobs
in the unix farm can and should be processed while the windows farm may
in trouble and vice versa.
The huge delays are seen, when user isn't known to sssd and winbind
tries to look that user without explicitly a domain given and the option
"winbind use default domain" is on it's default of "No" in smb.conf.
I am quite prepared to help in getting winbind working correctly,
but this will require the OP changing their smb.conf considerably
and removing sssd. We do not support sssd, it is not a Samba
product (for want of a better name). Samba on a Unix domain member
is designed around 3 binaries, smbd, nmbd and winbind, the latter
can do just about anything sssd can do, so why use sssd ? Now you
say that I am making assumptions, well about this one, probably
somewhere in the mix there will be Windows domain members and the
users in ldap are unlikely to be known to them.
I consider sssd as 'just another' user database, like /etc/passwd
(which Samba apparently does support) and I personally cannot see
any difference there, but I respect your opinion.
Where is it documented winbind should be the only service which
should be used with nss? If it is not documented, maybe it should.
I am not saying that sssd shouldn't be used, just Samba does not
support it. If you want to use sssd, then do so, just don't expect
to get help with using it, we don't produce it, so don't know it.
What I will say is this, there is no need to use both on a Unix
domain member, they both do the same thing.
I'm with you and don't expect that you support sssd.
On the otherside windind shouln't require to be the only one in
nss-setup as i didn't ever heard, that only a certain amount can be
taken into the stack. Before nss_sss we used nss_ldap alongside with
nss_winbind without issues.
There is no real problem with using multiple methods in nsswitch.conf,
there is just no real point, the first one to produce a result wins
and as sssd & winbind do the same thing, you don't need both.
He also has these:
idmap config * : rangesize = 1000000
idmap config * : range = 1000000-19999999
idmap config * : backend = autorid
The '*' domain is meant for the Well Known SIDs and anything
outside the Samba domain. I would have expected something like
idmap config * : backend = tdb
idmap config * : range = 3000-7999
idmap config OPS : backend = rid
idmap config OPS : range = 10000-999999
That should also be fixed.
We use this as we have a multi-domain setup on windows side and this
is a suggested setup from wiki.samba.org:
Cannot argue with that fact, it is there, but it also says it is meant
to be used with the 'DOMAIN' domain not the '*' domain, looks like I
will have to make that more prominent.
I'll try to somehow reconfig idmap as you suggested taking care of
all the trees in our forest and will report back if that changes the
I reconfigured idmap config auccesfuly like this:
idmap config * : range = 3000-7999
idmap config * : backend = tdb
idmap config ops : range = 1000000-1999999
idmap config ops : backend = rid
idmap config ops2 : range = 2000000-2999999
idmap config ops2 : backend = rid
But it didn't change the problem with the delayed response or timeout of
winbind calling "wbinfo -i foo".
One good point about the suggestion - It point to a fix of a "problem"
seen with autorid:
The used ID ranges from the pool used for different domains seem to vary
for servers being member of one of the other domains. Changing from
autorid to rid looks like the way to fix that as we can configure
dedicated ranges from the pool per domain and so uid/gid resloution will
become consitant across all of our domains. :)
I did some more tested about nsswitch.conf:
1. removing sss -> no more hangs (independand of shadow configured with
winbind or not)
passwd: files winbind
shadow: files # winbind
group: files winbind
2. removing winbind -> no more hangs
passwd: files sss
shadow: files sss
group: files sss
3. exchange winbind/sss - no change, still delays
passwd: files winbind sss
shadow: files sss
group: files winbind sss
It still looks that either sssd or winbindd (or both) relays/calls each
other, when no domain is given. I will try to get debug logging
correlated to see if and if yes which service calls the other.
To unsubscribe from this list go to the following URL and read the