[libvirt] [PATCH v2 4/4] security: don't relabel chardev source if virtlogd is used as stdio handler

Fri Jun 16 02:12:51 UTC 2017

On 06/15/2017 10:40 AM, Pavel Hrdina wrote:
> On Thu, Jun 15, 2017 at 07:57:18AM -0400, John Ferlan wrote:
>>
>>
>> On 06/15/2017 03:11 AM, Pavel Hrdina wrote:
>>> On Tue, Jun 13, 2017 at 08:00:41PM -0400, John Ferlan wrote:
>>>>
>>>>
>>>> On 05/29/2017 10:31 AM, Pavel Hrdina wrote:
>>>>> In the case that virtlogd is used as stdio handler we pass to QEMU
>>>>> only FD to a PIPE connected to virtlogd instead of the file itself.
>>>>>
>>>>> Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1430988
>>>>>
>>>>> Signed-off-by: Pavel Hrdina <phrdina at redhat.com>
>>>>> ---
>>>>>
>>>>> Notes:
>>>>>     new in v2
>>>>>
>>>>>  src/lxc/lxc_process.c            |  6 ++---
>>>>>  src/qemu/qemu_security.c         |  9 +++++--
>>>>>  src/security/security_apparmor.c |  7 ++++--
>>>>>  src/security/security_dac.c      | 54 +++++++++++++++++++++++++++++++---------
>>>>>  src/security/security_driver.h   |  6 +++--
>>>>>  src/security/security_manager.c  | 12 ++++++---
>>>>>  src/security/security_manager.h  |  6 +++--
>>>>>  src/security/security_nop.c      |  6 +++--
>>>>>  src/security/security_selinux.c  | 53 ++++++++++++++++++++++++++++++---------
>>>>>  src/security/security_stack.c    | 12 ++++++---
>>>>>  tests/securityselinuxlabeltest.c |  2 +-
>>>>>  11 files changed, 127 insertions(+), 46 deletions(-)
>>>>>
>>>>
>>>> Why is it a (!chr_seclabel && chardevStdioLogd)?  More to the point why
>>>> is (!chr_seclabel) even matter?
>>>
>>> If you configure the label we shouldn't ignore it in some cases, that's
>>> just wrong.  If the label is set for the char device we will configure
>>> it every time even if it will fail to start the guest, it's a
>>> responsibility of the user to configure proper label if it is provided.
>>>
>>
>> When email doesn't convey the question... Ugh...  I'm also trying to
>> speed learn an area of the code and review at the same time.
>>
>> If I go back to commit id 'f8b08d0e' where the original implementation
>> to add labels for chardevs was done (but has been modified for patch 1
>> to change where the label is stored), I get the impression that a label
>> should be added either from something specifically supplied for the
>> <chardev> or to use the per domain one:
>>
>> "The source element may contain an optional seclabel to override the way
>> that labelling is done on the socket path. If this element is not
>> present, the security label is inherited from the per-domain setting."
>>
>> So if I look at the condition "(!chr_seclabel && chardevStdioLogd)"
>> added by this patch to decide whether or not to supply the label, I'm
>> left with the impression that if for this particular chardev a label
>> doesn't exist *and* the configuration option chardevStdioLogd is true,
>> then we're going to return happy status *and* not inherit the per-domain
>> setting.
>>
>> So the bug is then that applying the default domain label for a chardev
>> configured to use a stdio handle is incorrect?  Perhaps I didn't get
>> that from reading bz or the patch.
> 
> Yes, that's the bug.  If virtlogd is used to handle stdio for char
> devices we shouldn't relabel the @path to the default labels, the
> @path doesn't have to be accessible by the qemu process, it has to be
> accessible by the virtlogd process.
> 
>>>> IIUC, whether or not someone set the label for the chardev, for this
>>>> particular issue/config where the chardev has a <log file=$path>, we
>>>> don't want to set (or restore) the label. I feel like I'm missing
>>>> something subtle. Maybe a bit more explanation of the adjustment would
>>>> help me...
>>>
>>> This is not for the <log file=$path/> but for the <source path=$path/>.
>>> We don't relabel $path for <log file=$path/> at all.
>>>
>>
>> hmm.. ah, right... I kept scrolling back and forth in the bz and the
>> docs, but missed this in the bz:
>>
>> 3) Check the virtlogd.log:
>> error : virRotatingFileWriterEntryNew:113 : Unable to open file:
>> /var/log/libvirt/qemu/log: Permission denied
>>
>> I guess I got lost in the power of suggestion of reading the docs
>> regarding the "optional log file" that can be associated paragraph and
>> trying to learn on the fly so that you at least get a review in a
>> somewhat timely manner ;-)
>>
>>>> Wouldn't these changes end up selecting "any" chardev if
>>>> chardevStdioLogd ended up being set regardless of whether they were
>>>> actually using the log file?
>>>
>>> I don't know what you mean by this sentence?
>>>
>>
>> Well let's see, chardevStdioLogd is set to true when meeting the two
>> conditions a qemu.conf global "cfg->stdioLogd" && a per domain or
>> emulator image capability "QEMU_CAPS_CHARDEV_FILE_APPEND".
>>
>> So, conceivably chardevStdioLogd could be true for *any* domain as long
>> as those conditions are met, right?
> 
> Yes, the two conditions are checked while starting a new domain in
> qemuProcessPrepareDomain() and stored in the private date of that
> domain.
> 
>> If you have a domain that has chardev's which are not configured to use
>> the stdio handler, then the chardevStdioLogd could still be true, right?
> 
> No, if the @chardevStdioLogd is true all char devices for that domain
> will use virtlogd.
> 

This is the part I'm stuck on as to why - based on the previous
assertion. I'm just not visualizing all those various chardev configs.
Just so you know - I have to be out Friday, but will be back Monday. If
you get someone else to ACK this one in the mean time - that's fine.

John

> The only issue I've just found out is that the code path for chardev
> hot-plug isn't updated to use virtlogd when it should be so for
> hot-plugged char devices we pass the path directly to QEMU.
> 
> With this patch applied the hot-plug fails if virtlogd is used because
> we don't relabel the path but we pass it directly to QEMU, this needs to
> be fixed to not introduce a regression, sigh.
> 
> Pavel
> 
>> If that's the case and the chardev doesn't have a label associated, then
>> we just return happy status and we do not inherit the per domain
>> setting. Wouldn't that be incorrect?
>>
>> My concern is more we're making a change in a (mostly) common set of
>> functions for a (very) specific problem.
>>
>>
>>>> As an aside, I think there's an "oddity" when it comes to the Restore,
>>>> but I'm not sure how to put it into words exactly. If a guest is running
>>>> code prior to this set of changes, would it have successfully run a Set?
>>>> If so, then after applying this change and restarting, the label
>>>> wouldn't be reset, right?  What happens at guest shutdown - does the
>>>> label not get unset now?  Of course this is all "interaction" with
>>>> virtlogd restart that really throws a monkey wrench into things.
>>>
>>> No, that's not correct.  The @chardevStdioLogd is stored in the status
>>> XML (the one stored in /var/run/libvirt/qemu/$domain_name.xml).  So when
>>> the libvirtd is stopped and started with this patch applied the status
>>> XML doesn't have the @chardevStdioLogd stored in it so it will be false
>>> and we will reset the label.  The @chardevStdioLogd is updated only when
>>> the domain is started and we will store the value in the status XML
>>> only with new libvirtd and only in that case we will not set/restore
>>> the label.
>>>
>>
>> hmmm.. Reading the bz indicates the 'virtlogd' daemon restarting... This
>> is where all this gets a bit "odd" for me. Like I said it was a weird
>> thing to even try and explain, but I think you talked me off the ledge
>> of concern.
>>
>> John
>>
>>>> Also, why is the Smartcard chardev handling not using this
>>>
>>> The smartcard doesn't ever use virtlogd as stdio handler.
>>>
>>> Pavel
>>>
>>
>> --
>> libvir-list mailing list
>> libvir-list at redhat.com
>> https://www.redhat.com/mailman/listinfo/libvir-list