Hello,
Commit 04e35f4495dd560db30c25efca4eecae8ec8c375 "exec: avoid RLIMIT_STACK
races with prlimit()" that made it into 4.14.4 effectively changes the default
hard RLIMIT_STACK on machines with SELinux (seen on Fedora 27).
selinux_bprm_set_creds() sets bprm->secureexec for any SELinux domain
transition that does not have the "noatsecure" permission. The secureexec
logic thus kicks in for virtually every process launched by PID 1 systemd
(init_t), including gettys, display managers, etc.
I can see that 8 MiB "should be enough for everyone" using normal software,
but sadly the HPC stuff around here tends to need a little more (due to a
deficiency in gfortran).
Minimal example (the actual types are not too important):
# /bin/ulimit -Hs
unlimited
# runcon -r system_r -t sysadm_t runcon -t rpm_script_t /bin/ulimit -Hs
8192
Of course this can be somewhat worked around by adjusting the SELinux policy
(allowing blanket noatsecure permission for init_t and possibly others) or by
pam_limits (for components using PAM). Unfortunately, systemd's LimitSTACK= is
also broken (calls setrlimit before exec). Anyway, I wasn't expecting any of
that in connection with the 4.14.3->.4 upgrade.
--
Best regards,
Tomáš Trnka
Software for Chemistry & Materials
> Of course this can be somewhat worked around by adjusting the SELinux policy
> (allowing blanket noatsecure permission for init_t and possibly others) or
> by pam_limits (for components using PAM).
Correction: pam_limits also usually doesn't help here, as it's often followed
by another secureexec (for example when login (local_login_t) executes the
shell with transition to unconfined_t).
2T
On Tue, Dec 12, 2017 at 2:58 AM, Tomáš Trnka <[email protected]> wrote:
> Hello,
>
> Commit 04e35f4495dd560db30c25efca4eecae8ec8c375 "exec: avoid RLIMIT_STACK
> races with prlimit()" that made it into 4.14.4 effectively changes the default
> hard RLIMIT_STACK on machines with SELinux (seen on Fedora 27).
>
> selinux_bprm_set_creds() sets bprm->secureexec for any SELinux domain
> transition that does not have the "noatsecure" permission. The secureexec
> logic thus kicks in for virtually every process launched by PID 1 systemd
> (init_t), including gettys, display managers, etc.
Uuugh. Okay, we need to revert that commit. I'll send a patch for 4.15
(with a fix for -stable too).
I will design an alternative, which was considered much earlier:
keeping a copy of the rlimits in the bprm during exec so it can't
change out from under the execing process. This will avoid needing to
set the hard limit, avoid the locking race that commit was trying to
fix, etc.
This is an interesting state for the system to be in, though, it means
AT_SECURE is being set for virtually all processes too? I would expect
that might break a lot too (but clearly it hasn't).
>
> I can see that 8 MiB "should be enough for everyone" using normal software,
> but sadly the HPC stuff around here tends to need a little more (due to a
> deficiency in gfortran).
>
> Minimal example (the actual types are not too important):
>
> # /bin/ulimit -Hs
> unlimited
> # runcon -r system_r -t sysadm_t runcon -t rpm_script_t /bin/ulimit -Hs
> 8192
>
> Of course this can be somewhat worked around by adjusting the SELinux policy
> (allowing blanket noatsecure permission for init_t and possibly others) or by
> pam_limits (for components using PAM). Unfortunately, systemd's LimitSTACK= is
> also broken (calls setrlimit before exec). Anyway, I wasn't expecting any of
> that in connection with the 4.14.3->.4 upgrade.
>
> --
> Best regards,
>
> Tomáš Trnka
> Software for Chemistry & Materials
Thanks for the report and examples!
-Kees
--
Kees Cook
Pixel Security
On Tuesday, 12 December 2017 20:23:47 CET Kees Cook wrote:
> This is an interesting state for the system to be in, though, it means
> AT_SECURE is being set for virtually all processes too? I would expect
> that might break a lot too (but clearly it hasn't).
Not really. AT_SECURE is set only for the exec that triggers a domain
transition, but unlike the rlimits it's not inherited by descendants (as long
as they stay within the same SELinux domain).
2T
On 12/12/2017 11:23 AM, Kees Cook wrote:
> On Tue, Dec 12, 2017 at 2:58 AM, Tomáš Trnka <[email protected]> wrote:
>> Hello,
>>
>> Commit 04e35f4495dd560db30c25efca4eecae8ec8c375 "exec: avoid RLIMIT_STACK
>> races with prlimit()" that made it into 4.14.4 effectively changes the default
>> hard RLIMIT_STACK on machines with SELinux (seen on Fedora 27).
>>
>> selinux_bprm_set_creds() sets bprm->secureexec for any SELinux domain
>> transition that does not have the "noatsecure" permission. The secureexec
>> logic thus kicks in for virtually every process launched by PID 1 systemd
>> (init_t), including gettys, display managers, etc.
>
> Uuugh. Okay, we need to revert that commit. I'll send a patch for 4.15
> (with a fix for -stable too).
>
> I will design an alternative, which was considered much earlier:
> keeping a copy of the rlimits in the bprm during exec so it can't
> change out from under the execing process. This will avoid needing to
> set the hard limit, avoid the locking race that commit was trying to
> fix, etc.
>
> This is an interesting state for the system to be in, though, it means
> AT_SECURE is being set for virtually all processes too? I would expect
> that might break a lot too (but clearly it hasn't).
>
>>
>> I can see that 8 MiB "should be enough for everyone" using normal software,
>> but sadly the HPC stuff around here tends to need a little more (due to a
>> deficiency in gfortran).
>>
>> Minimal example (the actual types are not too important):
>>
>> # /bin/ulimit -Hs
>> unlimited
>> # runcon -r system_r -t sysadm_t runcon -t rpm_script_t /bin/ulimit -Hs
>> 8192
>>
>> Of course this can be somewhat worked around by adjusting the SELinux policy
>> (allowing blanket noatsecure permission for init_t and possibly others) or by
>> pam_limits (for components using PAM). Unfortunately, systemd's LimitSTACK= is
>> also broken (calls setrlimit before exec). Anyway, I wasn't expecting any of
>> that in connection with the 4.14.3->.4 upgrade.
>>
>> --
>> Best regards,
>>
>> Tomáš Trnka
>> Software for Chemistry & Materials
>
> Thanks for the report and examples!
>
> -Kees
>
FWIW, the issue I reported offline yesterday
https://bugzilla.redhat.com/show_bug.cgi?id=1524083 still happens with
selinux disabled. The conclusion there is still that trafficserver
needs to be fixed.
Thanks,
Laura
On Tue, Dec 12, 2017 at 11:52 AM, Laura Abbott <[email protected]> wrote:
> On 12/12/2017 11:23 AM, Kees Cook wrote:
>>
>> On Tue, Dec 12, 2017 at 2:58 AM, Tomáš Trnka <[email protected]> wrote:
>>>
>>> Hello,
>>>
>>> Commit 04e35f4495dd560db30c25efca4eecae8ec8c375 "exec: avoid RLIMIT_STACK
>>> races with prlimit()" that made it into 4.14.4 effectively changes the
>>> default
>>> hard RLIMIT_STACK on machines with SELinux (seen on Fedora 27).
>>>
>>> selinux_bprm_set_creds() sets bprm->secureexec for any SELinux domain
>>> transition that does not have the "noatsecure" permission. The secureexec
>>> logic thus kicks in for virtually every process launched by PID 1 systemd
>>> (init_t), including gettys, display managers, etc.
>>
>>
>> Uuugh. Okay, we need to revert that commit. I'll send a patch for 4.15
>> (with a fix for -stable too).
>>
>> I will design an alternative, which was considered much earlier:
>> keeping a copy of the rlimits in the bprm during exec so it can't
>> change out from under the execing process. This will avoid needing to
>> set the hard limit, avoid the locking race that commit was trying to
>> fix, etc.
>>
>> This is an interesting state for the system to be in, though, it means
>> AT_SECURE is being set for virtually all processes too? I would expect
>> that might break a lot too (but clearly it hasn't).
>>
>>>
>>> I can see that 8 MiB "should be enough for everyone" using normal
>>> software,
>>> but sadly the HPC stuff around here tends to need a little more (due to a
>>> deficiency in gfortran).
>>>
>>> Minimal example (the actual types are not too important):
>>>
>>> # /bin/ulimit -Hs
>>> unlimited
>>> # runcon -r system_r -t sysadm_t runcon -t rpm_script_t /bin/ulimit -Hs
>>> 8192
>>>
>>> Of course this can be somewhat worked around by adjusting the SELinux
>>> policy
>>> (allowing blanket noatsecure permission for init_t and possibly others)
>>> or by
>>> pam_limits (for components using PAM). Unfortunately, systemd's
>>> LimitSTACK= is
>>> also broken (calls setrlimit before exec). Anyway, I wasn't expecting any
>>> of
>>> that in connection with the 4.14.3->.4 upgrade.
>>>
>>> --
>>> Best regards,
>>>
>>> Tomáš Trnka
>>> Software for Chemistry & Materials
>>
>>
>> Thanks for the report and examples!
>>
>> -Kees
>>
>
> FWIW, the issue I reported offline yesterday
> https://bugzilla.redhat.com/show_bug.cgi?id=1524083 still happens with
> selinux disabled. The conclusion there is still that trafficserver
> needs to be fixed.
I've sent a revert regardless. I think the bprm needs to hold a copy
of the rlimits so they can't be manipulated during exec. This will
keep the hardlimit from being messed with, etc.
-Kees
--
Kees Cook
Pixel Security
On 12/12/2017 11:56 AM, Kees Cook wrote:
> On Tue, Dec 12, 2017 at 11:52 AM, Laura Abbott <[email protected]> wrote:
>> On 12/12/2017 11:23 AM, Kees Cook wrote:
>>>
>>> On Tue, Dec 12, 2017 at 2:58 AM, Tomáš Trnka <[email protected]> wrote:
>>>>
>>>> Hello,
>>>>
>>>> Commit 04e35f4495dd560db30c25efca4eecae8ec8c375 "exec: avoid RLIMIT_STACK
>>>> races with prlimit()" that made it into 4.14.4 effectively changes the
>>>> default
>>>> hard RLIMIT_STACK on machines with SELinux (seen on Fedora 27).
>>>>
>>>> selinux_bprm_set_creds() sets bprm->secureexec for any SELinux domain
>>>> transition that does not have the "noatsecure" permission. The secureexec
>>>> logic thus kicks in for virtually every process launched by PID 1 systemd
>>>> (init_t), including gettys, display managers, etc.
>>>
>>>
>>> Uuugh. Okay, we need to revert that commit. I'll send a patch for 4.15
>>> (with a fix for -stable too).
>>>
>>> I will design an alternative, which was considered much earlier:
>>> keeping a copy of the rlimits in the bprm during exec so it can't
>>> change out from under the execing process. This will avoid needing to
>>> set the hard limit, avoid the locking race that commit was trying to
>>> fix, etc.
>>>
>>> This is an interesting state for the system to be in, though, it means
>>> AT_SECURE is being set for virtually all processes too? I would expect
>>> that might break a lot too (but clearly it hasn't).
>>>
>>>>
>>>> I can see that 8 MiB "should be enough for everyone" using normal
>>>> software,
>>>> but sadly the HPC stuff around here tends to need a little more (due to a
>>>> deficiency in gfortran).
>>>>
>>>> Minimal example (the actual types are not too important):
>>>>
>>>> # /bin/ulimit -Hs
>>>> unlimited
>>>> # runcon -r system_r -t sysadm_t runcon -t rpm_script_t /bin/ulimit -Hs
>>>> 8192
>>>>
>>>> Of course this can be somewhat worked around by adjusting the SELinux
>>>> policy
>>>> (allowing blanket noatsecure permission for init_t and possibly others)
>>>> or by
>>>> pam_limits (for components using PAM). Unfortunately, systemd's
>>>> LimitSTACK= is
>>>> also broken (calls setrlimit before exec). Anyway, I wasn't expecting any
>>>> of
>>>> that in connection with the 4.14.3->.4 upgrade.
>>>>
>>>> --
>>>> Best regards,
>>>>
>>>> Tomáš Trnka
>>>> Software for Chemistry & Materials
>>>
>>>
>>> Thanks for the report and examples!
>>>
>>> -Kees
>>>
>>
>> FWIW, the issue I reported offline yesterday
>> https://bugzilla.redhat.com/show_bug.cgi?id=1524083 still happens with
>> selinux disabled. The conclusion there is still that trafficserver
>> needs to be fixed.
>
> I've sent a revert regardless. I think the bprm needs to hold a copy
> of the rlimits so they can't be manipulated during exec. This will
> keep the hardlimit from being messed with, etc.
>
> -Kees
>
Understood. I just wanted to clarify that the issues were separate
since there was still some question about where the secureexec
was coming from.
Thanks,
Laura