Re: libqemuutil

[Thomas Huth]

On 16/03/2021 10.07, Markus Armbruster wrote:
> I rebased my "[PATCH v6 00/10] Configurable policy for handling
> deprecated interfaces" to master, and it surprisingly fails to link
> several utility programs.  Here's the first error:
>
>      gcc  -o tests/bench/benchmark-crypto-hmac 
> tests/bench/benchmark-crypto-hmac.p/benchmark-crypto-hmac.c.o -Wl,--as-needed 
> -Wl,--no-undefined -pie -Wl,--whole-archive libcrypto.fa libauthz.fa libqom.fa 
> -Wl,--no-whole-archive -Wl,--warn-common -Wl,-z,relro -Wl,-z,now -m64 
> -fstack-protector-strong -Wl,--start-group libqemuutil.a 
> subprojects/libvhost-user/libvhost-user-glib.a 
> subprojects/libvhost-user/libvhost-user.a libcrypto.fa libauthz.fa libqom.fa 
> -pthread -lgthread-2.0 -lglib-2.0 -lgio-2.0 -lgobject-2.0 -lglib-2.0 -lgio-2.0 
> -lgobject-2.0 -lglib-2.0 -lgnutls -lutil -lm -lgthread-2.0 -lglib-2.0 -lnettle 
> -lgnutls -lpam -Wl,--end-group
>      /usr/bin/ld: libqemuutil.a(util_main-loop.c.o): in function 
> `qemu_set_fd_handler':
>      /work/armbru/qemu/bld-x86/../util/main-loop.c:581: multiple definition of 
> `qemu_set_fd_handler'; 
> libqemuutil.a(stubs_set-fd-handler.c.o):/work/armbru/qemu/bld-x86/../stubs/set-fd-handler.c:8:
>  first defined here
>      collect2: error: ld returned 1 exit status
>
> Both master and PATCH 01 still link fine, PATCH 02 doesn't.  PATCH 02
> doesn't go anywhere near qemu_set_fd_handler().
>
> Turns out libqemuutil.a contains two definitions of
> qemu_set_fd_handler().  In master:
>
>      $ nm --defined-only libqemuutil.a | awk '/:$/ { f=$0 } / qemu_set_fd_handler/ { if 
> (f) { print f; f="" } print $0 }'
>      util_main-loop.c.o:
>      00000000000007fe T qemu_set_fd_handler
>      stubs_set-fd-handler.c.o:
>      0000000000000000 T qemu_set_fd_handler
>
> This is obviously unhealthy.
>
> I suspect the linker happens to pick the one that makes things work,
> until something in my patch makes it pick the other one.
>
> Is qemu_set_fd_handler() the only one?  Nope:
>
>      $ nm --defined-only bld-x86/libqemuutil.a | awk '/ T / { print $NF }' | 
> sort | uniq -c | grep -v '^ *1 '
>            2 qemu_set_fd_handler
>            2 yank_generic_iochannel
>            2 yank_register_function
>            2 yank_register_instance
>            2 yank_unregister_function
>            2 yank_unregister_instance
>
> I didn't run into this issue when I posted my series last Friday.  The
> issue now blocks its merge, and today is the soft freeze.  Help!

A very, very quick-n-dirty band-aid is likely to mark the function in stubs 
as weak:

diff --git a/stubs/set-fd-handler.c b/stubs/set-fd-handler.c
--- a/stubs/set-fd-handler.c
+++ b/stubs/set-fd-handler.c
@@ -1,6 +1,7 @@
 #include "qemu/osdep.h"
 #include "qemu/main-loop.h"

+__attribute__((weak))
 void qemu_set_fd_handler(int fd,
                          IOHandler *fd_read,
                          IOHandler *fd_write,

 ... should IMHO be good enough for the soft freeze. In the long run, you 
might want to analyze the problem more thoroughly, of course. I had similar 
problems in the past already, and solved them by moving the stubs around. See:

 b0476d6602adbf818132dc896b585e01f47eaf96
 stubs: Move qemu_timer_notify_cb() and remove
 qemu_notify_event() stub

 8c2787629eee73ca8ce4f100cff4f4946583b4e8
 stubs: Move qemu_fd_register stub to util/main-loop.c

HTH,
 Thomas