XDR related breakage in libvirt v6.6.0 when using libvirt-lxc
Hi, I expect that this falls under the "with meson now everything is different anyway" umbrella but wanted to let you know about this as it affects v6.6 in at least Ubuntu/Debian. The following recent patch has broken libvirt-lxc for us: commit d7147b3797380de2d159ce6324536f3e1f2d97e3 Author: Pavel Hrdina <phrdina@redhat.com> Date: Fri Jun 19 00:44:07 2020 +0200 m4: virt-xdr: rewrite XDR check I was tracking that down for [1] since the tests [4] failed on me. [2] holds the backtrace. In Debian the tests are skipped which explains why they were not seen there: smoke-lxc SKIP Test requires machine-level isolation but testbed does not provide that What happens is that the libvirt_lxc segfaults when using XDR functions. dmesg shows: [582093.524644] libvirt_lxc[261446]: segfault at 0 ip 0000000000000000 sp 00007ffdd2345598 error 14 in libvirt_lxc[5587e42aa000+8000] [582093.524650] Code: Bad RIP value. There are quite some uncertainties left, but on the surface it seems that it links with libtirpc but then instead of calling libtirpc: src/xdr.c:929:xdr_uint64_t(xdrs, ullp) it ends (gdb tells us in [2]) in glibc glibc: sunrpc/xdr_intXX_t.c:62:xdr_uint64_t (XDR *xdrs, uint64_t *uip) And the return from that function breaks it badly (instruction pointer at 0x0 -> segfault) Bisecting pointed to the referred commit which brings libtirpc into the mix. The former builds had xdr detected, but not with libtirpc. configure: xdr: yes (CFLAGS='' LIBS='') The new config now does configure: xdr: yes (CFLAGS='-I/usr/include/tirpc' LIBS='-ltirpc') And the resulting libvirt_lxc reflects that v6.0.0 $ lddtree /usr/lib/libvirt/libvirt_lxc | grep tirpc v6.6.0 $ lddtree /usr/lib/libvirt/libvirt_lxc | grep tirpc libtirpc.so.3 => /lib/x86_64-linux-gnu/libtirpc.so.3 This seems to lead to the bad jump and the crash eventually. Meanwhile reverting d7147b37 "m4: virt-xdr: rewrite XDR check" on top of v6.6.0 resolves the issue back to the former state. For anyone that wants to recreate this, I also attached a bisect script [3] which includes the test case you'd need. [1]: https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/1892826 [2]: https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/1892826/comments/4 [3]: https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/1892826/+attachment/5... [4]: https://objectstorage.prodstack4-5.canonical.com/v1/AUTH_77e2ada1e7a84929a74... -- Christian Ehrhardt Staff Engineer, Ubuntu Server Canonical Ltd
On Tue, Aug 25, 2020 at 03:16:50PM +0200, Christian Ehrhardt wrote:
Hi, I expect that this falls under the "with meson now everything is different anyway" umbrella but wanted to let you know about this as it affects v6.6 in at least Ubuntu/Debian.
The following recent patch has broken libvirt-lxc for us: commit d7147b3797380de2d159ce6324536f3e1f2d97e3 Author: Pavel Hrdina <phrdina@redhat.com> Date: Fri Jun 19 00:44:07 2020 +0200 m4: virt-xdr: rewrite XDR check
I was tracking that down for [1] since the tests [4] failed on me. [2] holds the backtrace. In Debian the tests are skipped which explains why they were not seen there: smoke-lxc SKIP Test requires machine-level isolation but testbed does not provide that
What happens is that the libvirt_lxc segfaults when using XDR functions.
dmesg shows: [582093.524644] libvirt_lxc[261446]: segfault at 0 ip 0000000000000000 sp 00007ffdd2345598 error 14 in libvirt_lxc[5587e42aa000+8000] [582093.524650] Code: Bad RIP value.
There are quite some uncertainties left, but on the surface it seems that it links with libtirpc but then instead of calling libtirpc: src/xdr.c:929:xdr_uint64_t(xdrs, ullp) it ends (gdb tells us in [2]) in glibc glibc: sunrpc/xdr_intXX_t.c:62:xdr_uint64_t (XDR *xdrs, uint64_t *uip)
And the return from that function breaks it badly (instruction pointer at 0x0 -> segfault)
Right so that's a serious problem with clashing symbols between tirpc and glibc. In Fedora/RHEL it is impossible to build against glibc for the XDR symbols for a long time now. Glibc maintainers want everyone to be using tirpc. The symbols are still exported from glibc, but they should only be used by legacy apps built against older glibc. Symbol versioning should ensure libvirt_lxc always resolves to the libtirpc library $ eu-readelf -a /usr/lib64/libc.so.6 | grep xdr_uint64 | grep GLOBAL 2017: 00000000001349c0 226 FUNC GLOBAL DEFAULT 15 xdr_uint64_t@GLIBC_2.2.5 $ eu-readelf -a /usr/lib64/libtirpc.so | grep xdr_uint64 | grep GLOBAL 344: 000000000001ce20 9 FUNC GLOBAL DEFAULT 14 xdr_uint64_t@@TIRPC_0.3.0 $ eu-readelf -a /usr/libexec/libvirt_lxc | grep xdr_uint64 0x0000000000024a30 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 149: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@TIRPC_0.3.0 (13) This shows libvirt_lxc will only resolve to libtirpc. I see the Ubuntu package for glibc is passing --enable-obsolete-rpc which allows apps to continue to build against glibc for RPC :-( So I suspect somehow libvirt has ended up using tirpc headers, but the linker probably resolved symbols to glibc. I don't know how the linker decides which library to resolve symbols to when multiple provided the same symbol with different versions. Possibly tries in order ? I do recall that there were lots of problems with having both glibc and libtirpc used in Fedora before glibc introduced the abilty to disable RPC via --disable-obsolete-rpc to Did I mention that --enable-obsolete-rpc is a bad idea yet :-P FWIW, you're going to forced to stop using this arg because it has been deleted entirely in glibc 2.32, so there's no way to compile against glibc for XDR. Only existing built binaries will work. Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
On Tue, Aug 25, 2020 at 4:07 PM Daniel P. Berrangé <berrange@redhat.com> wrote:
On Tue, Aug 25, 2020 at 03:16:50PM +0200, Christian Ehrhardt wrote:
Hi, I expect that this falls under the "with meson now everything is different anyway" umbrella but wanted to let you know about this as it affects v6.6 in at least Ubuntu/Debian.
The following recent patch has broken libvirt-lxc for us: commit d7147b3797380de2d159ce6324536f3e1f2d97e3 Author: Pavel Hrdina <phrdina@redhat.com> Date: Fri Jun 19 00:44:07 2020 +0200 m4: virt-xdr: rewrite XDR check
I was tracking that down for [1] since the tests [4] failed on me. [2] holds the backtrace. In Debian the tests are skipped which explains why they were not seen there: smoke-lxc SKIP Test requires machine-level isolation but testbed does not provide that
What happens is that the libvirt_lxc segfaults when using XDR functions.
dmesg shows: [582093.524644] libvirt_lxc[261446]: segfault at 0 ip 0000000000000000 sp 00007ffdd2345598 error 14 in libvirt_lxc[5587e42aa000+8000] [582093.524650] Code: Bad RIP value.
There are quite some uncertainties left, but on the surface it seems that it links with libtirpc but then instead of calling libtirpc: src/xdr.c:929:xdr_uint64_t(xdrs, ullp) it ends (gdb tells us in [2]) in glibc glibc: sunrpc/xdr_intXX_t.c:62:xdr_uint64_t (XDR *xdrs, uint64_t *uip)
And the return from that function breaks it badly (instruction pointer at 0x0 -> segfault)
Right so that's a serious problem with clashing symbols between tirpc and glibc.
In Fedora/RHEL it is impossible to build against glibc for the XDR symbols for a long time now. Glibc maintainers want everyone to be using tirpc. The symbols are still exported from glibc, but they should only be used by legacy apps built against older glibc.
Symbol versioning should ensure libvirt_lxc always resolves to the libtirpc library
$ eu-readelf -a /usr/lib64/libc.so.6 | grep xdr_uint64 | grep GLOBAL 2017: 00000000001349c0 226 FUNC GLOBAL DEFAULT 15 xdr_uint64_t@GLIBC_2.2.5
$ eu-readelf -a /usr/lib64/libtirpc.so | grep xdr_uint64 | grep GLOBAL 344: 000000000001ce20 9 FUNC GLOBAL DEFAULT 14 xdr_uint64_t@@TIRPC_0.3.0
$ eu-readelf -a /usr/libexec/libvirt_lxc | grep xdr_uint64 0x0000000000024a30 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 149: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@TIRPC_0.3.0 (13)
ubuntu@groovy:~$ eu-readelf -a /lib/x86_64-linux-gnu/libc.so.6 | grep xdr_uint64 | grep GLOBAL 2019: 0000000000159ed0 228 FUNC GLOBAL DEFAULT 16 xdr_uint64_t@@GLIBC_2.2.5 ubuntu@groovy:~$ eu-readelf -a /lib/x86_64-linux-gnu/libtirpc.so.3.0.0 | grep xdr_uint64 | grep GLOBAL 343: 000000000001ae20 9 FUNC GLOBAL DEFAULT 15 xdr_uint64_t@@TIRPC_0.3.0 Ubuntu v6.0 builds ubuntu@groovy:~$ eu-readelf -a /usr/lib/libvirt/libvirt_lxc | grep xdr_uint64 0x0000000000026820 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 99: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@GLIBC_2.2.5 (4) [ 1c02] xdr_uint64_t Ubuntu v6.6 builds ubuntu@groovy:~$ eu-readelf -a /usr/lib/libvirt/libvirt_lxc | grep xdr_uint64 0x00000000000268d0 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 104: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@GLIBC_2.2.5 (4) [ 1a81] xdr_uint64_t They miss the version 3.0 entry - interesting. libvirt 6.6 build from git on the same system: $ eu-readelf -a libvirt/build/src/.libs/libvirt_lxc | grep xdr_uint64 0x0000000000028968 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 99: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@GLIBC_2.2.5 (3) 598: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@@GLIBC_2.2.5 [ 31df] xdr_uint64_t@@GLIBC_2.2.5 [ 18f4] xdr_uint64_t That is with configure: xdr: yes (CFLAGS='-I/usr/include/tirpc' LIBS='-ltirpc') So something is wrong at build time when glibc AND tirpc provide that symbol.
This shows libvirt_lxc will only resolve to libtirpc.
I see the Ubuntu package for glibc is passing --enable-obsolete-rpc which allows apps to continue to build against glibc for RPC :-(
So I suspect somehow libvirt has ended up using tirpc headers, but the linker probably resolved symbols to glibc.
As I wrote above my builds don't get the 3.0 entry in libvirt_lxc which seems to be the reason to then jump to the wrong one.
I don't know how the linker decides which library to resolve symbols to when multiple provided the same symbol with different versions. Possibly tries in order ? I do recall that there were lots of problems with having both glibc and libtirpc used in Fedora before glibc introduced the abilty to disable RPC via --disable-obsolete-rpc to
Did I mention that --enable-obsolete-rpc is a bad idea yet :-P
You are probably right, but that will be a different bug for a different day.
FWIW, you're going to be forced to stop using this arg because it has been deleted entirely in glibc 2.32, so there's no way to compile against glibc for XDR. Only existing built binaries will work.
By then at least it won't be able to link in the wrong one anymore :-) And 2.32 is planned sometime soon for Ubuntu [1], so maybe I can do the revert for a week and then drop it on a rebuild. [1]: https://discourse.ubuntu.com/t/groovy-gorilla-release-schedule/15531
Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
-- Christian Ehrhardt Staff Engineer, Ubuntu Server Canonical Ltd
On Tue, Aug 25, 2020 at 04:26:01PM +0200, Christian Ehrhardt wrote:
On Tue, Aug 25, 2020 at 4:07 PM Daniel P. Berrangé <berrange@redhat.com> wrote:
On Tue, Aug 25, 2020 at 03:16:50PM +0200, Christian Ehrhardt wrote:
Hi, I expect that this falls under the "with meson now everything is different anyway" umbrella but wanted to let you know about this as it affects v6.6 in at least Ubuntu/Debian.
The following recent patch has broken libvirt-lxc for us: commit d7147b3797380de2d159ce6324536f3e1f2d97e3 Author: Pavel Hrdina <phrdina@redhat.com> Date: Fri Jun 19 00:44:07 2020 +0200 m4: virt-xdr: rewrite XDR check
I was tracking that down for [1] since the tests [4] failed on me. [2] holds the backtrace. In Debian the tests are skipped which explains why they were not seen there: smoke-lxc SKIP Test requires machine-level isolation but testbed does not provide that
What happens is that the libvirt_lxc segfaults when using XDR functions.
dmesg shows: [582093.524644] libvirt_lxc[261446]: segfault at 0 ip 0000000000000000 sp 00007ffdd2345598 error 14 in libvirt_lxc[5587e42aa000+8000] [582093.524650] Code: Bad RIP value.
There are quite some uncertainties left, but on the surface it seems that it links with libtirpc but then instead of calling libtirpc: src/xdr.c:929:xdr_uint64_t(xdrs, ullp) it ends (gdb tells us in [2]) in glibc glibc: sunrpc/xdr_intXX_t.c:62:xdr_uint64_t (XDR *xdrs, uint64_t *uip)
And the return from that function breaks it badly (instruction pointer at 0x0 -> segfault)
Right so that's a serious problem with clashing symbols between tirpc and glibc.
In Fedora/RHEL it is impossible to build against glibc for the XDR symbols for a long time now. Glibc maintainers want everyone to be using tirpc. The symbols are still exported from glibc, but they should only be used by legacy apps built against older glibc.
Symbol versioning should ensure libvirt_lxc always resolves to the libtirpc library
$ eu-readelf -a /usr/lib64/libc.so.6 | grep xdr_uint64 | grep GLOBAL 2017: 00000000001349c0 226 FUNC GLOBAL DEFAULT 15 xdr_uint64_t@GLIBC_2.2.5
$ eu-readelf -a /usr/lib64/libtirpc.so | grep xdr_uint64 | grep GLOBAL 344: 000000000001ce20 9 FUNC GLOBAL DEFAULT 14 xdr_uint64_t@@TIRPC_0.3.0
$ eu-readelf -a /usr/libexec/libvirt_lxc | grep xdr_uint64 0x0000000000024a30 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 149: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@TIRPC_0.3.0 (13)
ubuntu@groovy:~$ eu-readelf -a /lib/x86_64-linux-gnu/libc.so.6 | grep xdr_uint64 | grep GLOBAL 2019: 0000000000159ed0 228 FUNC GLOBAL DEFAULT 16 xdr_uint64_t@@GLIBC_2.2.5 ubuntu@groovy:~$ eu-readelf -a /lib/x86_64-linux-gnu/libtirpc.so.3.0.0 | grep xdr_uint64 | grep GLOBAL 343: 000000000001ae20 9 FUNC GLOBAL DEFAULT 15 xdr_uint64_t@@TIRPC_0.3.0
Ubuntu v6.0 builds ubuntu@groovy:~$ eu-readelf -a /usr/lib/libvirt/libvirt_lxc | grep xdr_uint64 0x0000000000026820 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 99: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@GLIBC_2.2.5 (4) [ 1c02] xdr_uint64_t
Ubuntu v6.6 builds ubuntu@groovy:~$ eu-readelf -a /usr/lib/libvirt/libvirt_lxc | grep xdr_uint64 0x00000000000268d0 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 104: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@GLIBC_2.2.5 (4) [ 1a81] xdr_uint64_t
They miss the version 3.0 entry - interesting.
libvirt 6.6 build from git on the same system: $ eu-readelf -a libvirt/build/src/.libs/libvirt_lxc | grep xdr_uint64 0x0000000000028968 X86_64_JUMP_SLOT 000000000000000000 +0 xdr_uint64_t 99: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@GLIBC_2.2.5 (3) 598: 0000000000000000 0 FUNC GLOBAL DEFAULT UNDEF xdr_uint64_t@@GLIBC_2.2.5 [ 31df] xdr_uint64_t@@GLIBC_2.2.5 [ 18f4] xdr_uint64_t
That is with configure: xdr: yes (CFLAGS='-I/usr/include/tirpc' LIBS='-ltirpc')
So something is wrong at build time when glibc AND tirpc provide that symbol.
The 'struct XDR' contents is public and defined differently in glibc vs tirpc. So when libvirt has been built using libtirpc include files but linked to glibc, it could end up accessing bad fields in the struct if we use any of the inline'd accessor functions. Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
participants (2)
-
Christian Ehrhardt -
Daniel P. Berrangé