On Fri, Mar 08, 2013 at 09:10:24PM +0100, Michal Židek wrote:
>> On Mon, Mar 04, 2013 at 04:13:31PM +0100, Michal Židek wrote:
>>> On 03/04/2013 02:53 PM, Pavel Březina wrote:
>>>> On 03/01/2013 04:13 PM, Michal Židek wrote:
>>>>> Patch 1: Right now, we have SELinux procesing split between the
provider
>>>>> and responder. This patch moves it to provider only. So no partial
>>>>> results are sent to responder for further processing, which makes
the
>>>>> processing steps easier to follow. (Originally the idea was to move
the
>>>>> processing to the responder, but completely removing the procesing
from
>>>>> provider and let it just update db turned to be more invasive than
>>>>> completely removing the procesing from responder.) If this patch is
>>>>> accepted, SELinux policy needs to be modified to allow sssd_be to
write
>>>>> the selinux login files.
>>>>>
https://fedorahosted.org/sssd/ticket/1743
>>>>>
>>
>> Even though we would have to tailor the SELinux policy, I'm not opposed
>> to moving the logic to provider. Originally I proposed responder because
>> I simply wanted all the processing to happen on one place.
>>
>> The whole intent of the refactor is to get rid of passing some
>> information from provider by writing data to sysdb in provider and
>> reading them in responder. If we get rid of that architectural bug, I'm
>> fine with any solution.
>>
>> FWIW, the other alternative that we were discussing with Michal was to
>> done the processing in provider and then send the resulting string to
>> responder via DBus. But it seemed kind of pointless to send the result
>> with DBus just to avoid moving the write of the file.
>>
>>>>>
>>>>> Patch 2, 3: Reuse cached selinux maps. Patch 2 shows how it is done
with
>>>>> refresh interval fixed to 120 seconds. Patch 3 adds ipa option to
>>>>> configure this interval (I didn't want to merge these patches for
better
>>>>> readability). Default value is 30 seconds.
>>>>>
https://fedorahosted.org/sssd/ticket/1744
>>>>>
>>>>>
>>>>> Patches are attached.
>>>>>
>>>>> Thanks
>>>>> Michal
>>>>
>>>> Hi,
>>>> switch the order of the second and third patch. So the flow is: add new
>>>> option, then use it.
>>>
>>> I don't think this is good idea. If the option is present, the logic
>>> to implement it's effect should be present too. It is also easier to
>>> test the patches one by one this way (first test the mechanism with
>>> fixed time interval, then make it configurable and test if interval
>>> is changing). For now I only changed the 120 seconds in the 2nd
>>> patch to 5 seconds (so it is not that aggressive).
>>>>
>>>> Do we want to use it anyway? Shouldn't be SELinux maps always
updated?
>>>>
>>>
>>> Well, we can still make 0 seconds the default value (or 5 seconds,
>>> which is used by IPA_HBAC_REFRESH). But selinux maps are not changed
>>> frequently on ipa side, so having this option is IMO good thing. In
>>> the new attached patches the default value is changed to 5.
>>>
>>> Michal
>>
>> Good start, but a nack, sorry.
>>
>> I'm seeing a segfaut and valgrind says it's becuase of a user-after-free
>> bug:
>> ==28566== Invalid read of size 4
>> ==28566== at 0x5264CFE: sysdb_attrs_get_el_ext (sysdb.c:313)
>> ==28566== by 0x5264EC6: sysdb_attrs_get_string (sysdb.c:356)
>> ==28566== by 0x5280A75: sysdb_store_selinux_entity (sysdb_selinux.c:102)
>> ==28566== by 0x528153D: sysdb_store_selinux_usermap (sysdb_selinux.c:181)
>> ==28566== by 0x122AABE0: ipa_save_user_maps (ipa_selinux_common.c:48)
>> ==28566== by 0x122A55E1: ipa_selinux_handler_done (ipa_selinux.c:260)
>> ==28566== by 0x122A9943: ipa_get_selinux_maps_done (ipa_selinux.c:1149)
>> ==28566== by 0x122AA8DB: ipa_selinux_get_maps_done (ipa_selinux_maps.c:203)
>> ==28566== by 0x12E2995C: sdap_get_generic_done (sdap_async.c:1558)
>> ==28566== by 0x12E29132: sdap_get_generic_ext_done (sdap_async.c:1449)
>> ==28566== by 0x12E2204B: sdap_process_message (sdap_async.c:366)
>> ==28566== by 0x12E21593: sdap_process_result (sdap_async.c:209)
>> ==28566== Address 0x17e416a0 is 80 bytes inside a block of size 96 free'd
>> ==28566== at 0x4C297A6: free (vg_replace_malloc.c:446)
>> ==28566== by 0x56E2880: _talloc_free_internal (talloc.c:876)
>> ==28566== by 0x122A5BB6: ipa_selinux_process_maps (ipa_selinux.c:355)
>> ==28566== by 0x122A4FFF: ipa_selinux_handler_done (ipa_selinux.c:224)
>> ==28566== by 0x122A9943: ipa_get_selinux_maps_done (ipa_selinux.c:1149)
>> ==28566== by 0x122AA8DB: ipa_selinux_get_maps_done (ipa_selinux_maps.c:203)
>> ==28566== by 0x12E2995C: sdap_get_generic_done (sdap_async.c:1558)
>> ==28566== by 0x12E29132: sdap_get_generic_ext_done (sdap_async.c:1449)
>> ==28566== by 0x12E2204B: sdap_process_message (sdap_async.c:366)
>> ==28566== by 0x12E21593: sdap_process_result (sdap_async.c:209)
>> ==28566== by 0x12E20BE5: sdap_ldap_next_result (sdap_async.c:159)
>> ==28566== by 0x54DAD3F: tevent_common_loop_timer_delay (in
/usr/lib64/libtevent.so.0.9.17)
>>
>> Other tests passed, the use-after-free happens when there is a specific
>> mapping rule only:
>>
>> Rule name: test_user1_specific_host
>> SELinux User: staff_u:s0-s0:c0.c1023
>> Enabled: TRUE
>> Users: tuser3
>> Hosts:
client.example.com
>>
>
> Fixed.
>
>> The code looks OK to me, I would just change two things:
>> 1) Split the function prepare_selinux_file_content() into one that
>> creates the order array and one that evaluates the matched rules based
>> on the order. Currently the function does two different tasks and is too
>> long.
>>
>> 2) Rename the prepare_selinux_file_content() function so that it is
>> clear it picks a single rule from a set of rules with same priority
>> based on the configured order.
>>
>
> I split the function into two new functions (create_order_array,
> choose_best_seuser) and added some comments.
>
> New patches attached.
>
> Thanks
> Michal
Nack,
> +/* Choose best selinux user based on given order and write
> + * the user to selinux login file. */
> +static errno_t choose_best_seuser(struct sysdb_attrs **usermaps,
> + struct pam_data *pd,
> + char **order_array, int order_count,
> + const char *default_user)
> +{
> + TALLOC_CTX *tmp_ctx;
> + char *file_content = NULL;
> + const char *tmp_str;
> + errno_t ret, err;
> + int i, j;
> +
> + /* If no maps match, we'll use the default SELinux user from the
> + * config */
> + file_content = talloc_strdup(tmp_ctx, default_user);
This is clearly wrong, tmp_ctx is never assigned to. The code may work
only by chance without optimizations becaues tmp_ctx would have been NULL..
Good catch. Sorry for that.
> + if (file_content == NULL) {
> + ret = ENOMEM;
> + goto done;
> + }
I also changed the behaviour if HAVE_SELINUX_LOGIN_DIR is not defined.
In that case, the processing is skipped and the ipa_selinux_handler just
returns success.
New patches attached.
Thanks
Michal