Re: [Qemu-devel] [PATCH 1/6] qemu-bridge-helper: Fix misuse of isspace()

[Markus Armbruster]

Daniel P. Berrangé <address@hidden> writes:

> On Wed, May 15, 2019 at 08:34:17AM +0200, Markus Armbruster wrote:
>> Jason Wang <address@hidden> writes:
>> 
>> > On 2019/5/14 下午8:18, Markus Armbruster wrote:
>> >> Peter Maydell <address@hidden> writes:
>> >>
>> >>> On Mon, 13 May 2019 at 14:21, Markus Armbruster <address@hidden> wrote:
>> >>>> Perhaps I should do it just for this file while I touch it anyway.  The
>> >>>> question to ask: should parse_acl_file() obey the locale for whitespace
>> >>>> recognition?
>> >>> I vote for "no".
>> >>>
>> >>> Q: do we document the format of the ACL file anywhere ?
>> >> Support for it was added in commit bdef79a2994, v1.1.  Just code, no
>> >> documentation.
>> >>
>> >> Grepping for qemu-bridge-helper finds just qemu-options.hx.  Contains
>> >> -help output and some .texi that goes into qemu-doc and the manual page.
>> >> None of it mentions how qemu-bridge-helper is run, or the ACL file
>> >> feature, let alone what its format might be.
>> >>
>> >> I'm afraid all we have is the commit message.  Which doesn't really
>> >> define the file format, it merely gives a bunch of examples.
>> >>
>> >> As far as I can tell, qemu-bridge-helper is for use with -netdev tap and
>> >> -netdev bridge.
>> >>
>> >> Both variations of -netdev call net_bridge_run_helper() to run the
>> >> helper.  First argument is -netdev parameter "helper", default usually
>> >> "$prefix/libexec/qemu-bridge-helper".  Second argument is parameter
>> >> "br", default "br0".
>> >>
>> >> If @helper contains space or tab, net_bridge_run_helper() guesses its a
>> >> full command, else it guesses its the name of the executable.  Bad
>> >> magic.
>> >>
>> >> If it guesses name of executable, it execv()s this executable with
>> >> arguments "--use-vnet", "--fd=FD", "address@hidden".
>> >>
>> >> If it guesses full command, it appends "--use-vnet --fd=FD", where FD is
>> >> the helper's half of the socketpair used to connect QEMU and the helper.
>> >> It further appends "address@hidden", unless @helper contains "--br=".
>> >> More bad magic.
>> >>
>> >> It executes the resulting string with sh -c.  Magic cherry on top.
>> >>
>> >> When the helper fails, netdev creation fails.
>> >>
>> >> The helper we ship with QEMU unconditionally tries to read
>> >> "$prefix/etc/bridge.conf".  Fatal error if this file doesn't exist.
>> >> Errors in this file are fatal.  Errors in files it includes are not
>> >> fatal; instead, the remainder of the erroneous file is ignored.
>> >> *Boggle*
>> >>
>> >> As far as I can tell, libvirt runs qemu-bridge-helper itself (Paolo's
>> >> commit 2d80fbb14df).  Makes sense, because running QEMU with the
>> >> necessary privileges would be unwise, and so would be letting it execute
>> >> setuid helpers.  Also bypasses the bad magic in QEMU's
>> >> net_bridge_run_helper().
>> >
>> >
>> > I don't notice this before. Is this only for the convenience of
>> > development? I guess libvirt should have native support like adding
>> > port to bridge/OVS without the help any external command or script.
>> 
>> Commit 2d80fbb14df hints at the reason:
>> 
>>     <source type='bridge'> uses a helper application to do the necessary
>>     TUN/TAP setup to use an existing network bridge, thus letting
>>     unprivileged users use TUN/TAP interfaces.
>>     ~~~~~~~~~~~~~~~~~~
>> 
>> The code confirms:
>> 
>>     /* qemuInterfaceBridgeConnect:
>>      * @def: the definition of the VM
>>      * @driver: qemu driver data
>>      * @net: pointer to the VM's interface description
>>      * @tapfd: array of file descriptor return value for the new device
>>      * @tapfdsize: number of file descriptors in @tapfd
>>      *
>> ---> * Called *only* called if actualType is VIR_DOMAIN_NET_TYPE_NETWORK or
>> ---> * VIR_DOMAIN_NET_TYPE_BRIDGE (i.e. if the connection is made with a tap
>>      * device connecting to a bridge device)
>>      */
>>     int
>>     qemuInterfaceBridgeConnect(virDomainDefPtr def,
>>                                virQEMUDriverPtr driver,
>>                                virDomainNetDefPtr net,
>>                                int *tapfd,
>>                                size_t *tapfdSize)
>>     {
>>         [...]
>> --->    if (virQEMUDriverIsPrivileged(driver)) {
>>             [...]
>>         } else {
>>             if (qemuCreateInBridgePortWithHelper(cfg, brname,
>>                                                  &net->ifname,
>>                                                  tapfd, tap_create_flags) < 
>> 0) {
>>                 virDomainAuditNetDevice(def, net, tunpath, false);
>>                 goto cleanup;
>>             }
>>             [...]
>>         }
>>         [...]
>>     }
>> 
>> >> qemu-bridge-helper should have a manual page, and its handling of errors
>> >> in ACL include files needs work.  There's probably more; I just glanced
>> >> at it.  I'm not volunteering, though.  It lacks a maintainer.  Should we
>> >> add it to Jason's "Network device backends"?
>> >
>> >
>> > Yes.
>> >
>> >> -netdev's helper parameter is seriously underdocumented.  Document or
>> >> deprecate?
>> >
>> >
>> > I believe management should only use fd parameter of TAP. If we have
>> > other, it should be a duplication. So I suggest to deprecate the
>> > bridge helper and -netdev bridge.
>> 
>> Objections, anyone?
>
> Libvirt runs the qemu bridge helper command directly, and we have
> applications using this functionality.

Specifically, when libvirt lacks the privileges to set up a TAP fd, it
farms out the job to setuid qemu-bridge-helper.  Correct?

> I'd like libvirt to be able to avoid use of the QEMU bridge helper and
> instead have unprivileged libvirt talk to privileged libvirtd to open a
> TAP fd on its behalf. If we ever get that done, then libvirt would not
> need the qemu bridge helper command anymore.

We don't want to deprecate qemu-bridge-helper while libvirt has a
sensible use for it.

We can still deprecate -netdev tap parameter "helper" and -netdev bridge
entirely.

Once they're gone, qemu-bridge-helper wull have no user within QEMU.  We
could discuss moving it to libvirt then, but I doubt it'll be worth the
trouble.