On Tue, Aug 14, 2012 at 01:38:15PM +0200, Pavel Březina wrote:
> On 08/13/2012 11:42 AM, Jakub Hrozek wrote:
>> On Mon, Aug 13, 2012 at 11:10:36AM +0200, Jakub Hrozek wrote:
>>> On Fri, Aug 10, 2012 at 09:26:49AM +0200, Pavel Březina wrote:
>>>> On 08/08/2012 04:24 PM, Jakub Hrozek wrote:
>>>>> On Wed, Aug 08, 2012 at 02:23:04PM +0200, Pavel Březina wrote:
>>>>>> This bug was probably introduced with the subdomain patches. The
>>>>>> problem was that sss_dp_get_domains_send() is called even for the
local
>>>>>> provider. There are certainly many possible solutions of this
issue. I
>>>>>> decided to modify sss_dp_issue_request() to call the callback
>>>>>> immediately if it is issued for local domain. This way, I
believe, we
>>>>>> can avoid further problems with issuing request for local
domains.
>>>>>
>>>>> I don't think this is the best solution.
>>>>>
>>>>> First, it leaves the check_provider of the per-domain structures we
use
>>>>> in responders around while duplicating the same functionality in the
DP.
>>>>> At the very least, the patch should remove the
"check_provider" tests from
>>>>> the existing loops.
>>>>>
>>>>> What I think would be even better solution is to
>>>>> 1) in the RC just fix the subdomains code so that it shortcuts
or
>>>>> doesn't run at all for LOCAL lookups.
>>>>
>>>> Patch is attached.
>>>
>>> Thank you, this works for me.
>>>
>>> I'm seeing this warning when LOCAL users are enumerated (as opposed to a
>>> direct lookup):
>>> (Mon Aug 13 11:06:58 2012) [sssd[nss]] [setent_notify] (0x0010): BUG: a
>>> callback did not free its request. May leak memory
>>>
>>> Given that the LOCAL domain is mostly useful for testing right now, I've
>>> just filed a ticket:
>>>
https://fedorahosted.org/sssd/ticket/1477
>>>
>>> Ack
>>
>> Nack, upon further testing. I'm seeing segfaults when looking up
>> non-LOCAL users:
>>
>> Core was generated by `/usr/libexec/sssd/sssd_nss -d 0x17f0'.
>> Program terminated with signal 11, Segmentation fault.
>> #0 0x000000000043acb1 in get_domains_next (req=0xf95ca0) at
src/responder/common/responder_get_domains.c:124
>> 124 tevent_req_post(req, info->rctx->ev);
>> Missing separate debuginfos, use: debuginfo-install
nss-softokn-freebl-3.13.5-1.fc17.x86_64 openldap-2.4.31-3.fc17.x86_64
>> (gdb) bt
>> #0 0x000000000043acb1 in get_domains_next (req=0xf95ca0) at
src/responder/common/responder_get_domains.c:124
>> #1 0x000000000043b461 in sss_dp_get_domains_callback (subreq=0xf968b0) at
src/responder/common/responder_get_domains.c:215
>> #2 0x0000000000439dd4 in sss_dp_internal_get_done (pending=0xf944e0,
ptr=0xf99220) at src/responder/common/responder_dp.c:731
>> #3 0x000000342fa0c42a in complete_pending_call_and_unlock
(connection=connection@entry=0xf95e40, pending=0xf944e0, message=message@entry=0xf98a80)
at dbus-connection.c:2308
>> #4 0x000000342fa0f5ea in dbus_connection_dispatch (connection=0xf95e40) at
dbus-connection.c:4593
>> #5 0x0000000000460432 in sbus_dispatch (ev=0xf89390, te=0xf98bc0, tv=...,
data=0xf979a0) at src/sbus/sssd_dbus_connection.c:104
>> #6 0x000000385b407600 in tevent_common_loop_timer_delay (ev=ev@entry=0xf89390)
at ../tevent_timed.c:254
>> #7 0x000000385b406cac in std_event_loop_once (ev=0xf89390,
location=<optimized out>) at ../tevent_standard.c:560
>> #8 0x000000385b403ed0 in _tevent_loop_once (ev=ev@entry=0xf89390,
location=location@entry=0x492f97 "src/util/server.c:554") at ../tevent.c:506
>> #9 0x000000385b40405b in tevent_common_loop_wait (ev=0xf89390, location=0x492f97
"src/util/server.c:554") at ../tevent.c:607
>> #10 0x0000000000467bf4 in server_loop (main_ctx=0xf8a4a0) at
src/util/server.c:554
>> #11 0x0000000000408d5c in main (argc=3, argv=0x7fff2dde9cc8) at
src/responder/nss/nsssrv.c:422
>> (gdb) p *info
>> $1 = {rctx = 0x3133623833347830, dom = 0x306d61736c3a313a, hint =
0x696e403334346336 <Address 0x696e403334346336 out of bounds>, force = 114, state =
0x0}
>>
>> It looks like "info" was not initialized.
>
> Thanks.
>
> The problem was that callback is invoked immediately when
> tevent_req_done() is called. And because req is freed in this
> callback, info was pointing to invalid data.
>
> New patch is attached.
I think it would be much cleaner to remove the tevent_req_post
completely from get_domains_next() and only call it in
sss_dp_get_domains_send() if get_domains_next() returned EOK.
I wouldn't expect tevent_req_post to be called in any other function
that *_send. Also checking the SSSD code shows that we only call
tevent_req_post from *_send elsewhere.
As you wish. Patch is attached.