Message ID | 20210315170636.704201-1-zhlcindy@gmail.com |
---|---|
State | New |
Headers | show |
Series | [1/2] Fix the segment fault when calling yank_register_instance | expand |
Hi On Mon, Mar 15, 2021 at 9:22 PM Li Zhang <zhlcindy@gmail.com> wrote: > From: Li Zhang <li.zhang@cloud.ionos.com> > > When executing the QMP commands "chardev-change" to change the > backend device to socket, it will cause a segment fault because > it assumes chr->label as non-NULL in function yank_register_instance. > The function qmp_chardev_change calls chardev_new, which label > is NULL when creating a new chardev. The label will be passed to > yank_register_instance which causes a segment fault. The callchain > is as the following: > chardev_new -> > qemu_char_open -> > cc->open -> > qmp_chardev_open_socket -> > yank_register_instance > > Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com> > --- > chardev/char-socket.c | 8 +++++--- > 1 file changed, 5 insertions(+), 3 deletions(-) > > diff --git a/chardev/char-socket.c b/chardev/char-socket.c > index c8bced76b7..26d5172682 100644 > --- a/chardev/char-socket.c > +++ b/chardev/char-socket.c > @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev *chr, > qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); > } > > - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), errp)) > { > - return; > + if (chr->label) { > + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), > errp)) { > + return; > + } > + s->registered_yank = true; > } > - s->registered_yank = true; > > /* be isn't opened until we get a connection */ > *be_opened = false > Looks wrong to me, the new chardev will get the same label, and it should still be possible to call the yank functions then. The registration logic needs to be reworked during chardev-change.
Hi Marc-André, The new chardev can get the same label. It is assigned after the function ChardevReturn *qmp_chardev_change(const char *id, ChardevBackend *backend, Error **errp) { ..... chr_new = chardev_new(NULL, object_class_get_name(OBJECT_CLASS(cc)), backend, chr->gcontext, errp); if (!chr_new) { return NULL; } chr_new->label = g_strdup(id); if (chr->be_open && !chr_new->be_open) { qemu_chr_be_event(chr, CHR_EVENT_CLOSED); closed_sent = true; } chr->be = NULL; qemu_chr_fe_init(be, chr_new, &error_abort); ..... } It passes parameter NULL in chardev_new, I think it may be because the old chardev isn't released yet. It will cause duplicated problems. I need to consider this logic to see if it can be changed. Thanks Li On Mon, Mar 15, 2021 at 7:51 PM Marc-André Lureau < marcandre.lureau@gmail.com> wrote: > Hi > > On Mon, Mar 15, 2021 at 9:22 PM Li Zhang <zhlcindy@gmail.com> wrote: > >> From: Li Zhang <li.zhang@cloud.ionos.com> >> >> When executing the QMP commands "chardev-change" to change the >> backend device to socket, it will cause a segment fault because >> it assumes chr->label as non-NULL in function yank_register_instance. >> The function qmp_chardev_change calls chardev_new, which label >> is NULL when creating a new chardev. The label will be passed to >> yank_register_instance which causes a segment fault. The callchain >> is as the following: >> chardev_new -> >> qemu_char_open -> >> cc->open -> >> qmp_chardev_open_socket -> >> yank_register_instance >> >> Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com> >> --- >> chardev/char-socket.c | 8 +++++--- >> 1 file changed, 5 insertions(+), 3 deletions(-) >> >> diff --git a/chardev/char-socket.c b/chardev/char-socket.c >> index c8bced76b7..26d5172682 100644 >> --- a/chardev/char-socket.c >> +++ b/chardev/char-socket.c >> @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev *chr, >> qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); >> } >> >> - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >> errp)) { >> - return; >> + if (chr->label) { >> + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >> errp)) { >> + return; >> + } >> + s->registered_yank = true; >> } >> - s->registered_yank = true; >> >> /* be isn't opened until we get a connection */ >> *be_opened = false >> > > Looks wrong to me, the new chardev will get the same label, and it should > still be possible to call the yank functions then. The registration logic > needs to be reworked during chardev-change. > > -- > Marc-André Lureau >
Hi Marc-André, By looking into chardev and yank_register_function logic, this old chardev is registered according to the chardev label. So it's been in yank_instance_list. yank instance only has a chardev label, and the new chardev's label is the same as the old chardev. So it doesn't need to register it again when changing the chardev backend. Otherwise, it will report duplicated yank instances. I think the chardev logic has no problems. And it works with yank functions. Thanks Li On Mon, Mar 15, 2021 at 7:51 PM Marc-André Lureau < marcandre.lureau@gmail.com> wrote: > Hi > > On Mon, Mar 15, 2021 at 9:22 PM Li Zhang <zhlcindy@gmail.com> wrote: > >> From: Li Zhang <li.zhang@cloud.ionos.com> >> >> When executing the QMP commands "chardev-change" to change the >> backend device to socket, it will cause a segment fault because >> it assumes chr->label as non-NULL in function yank_register_instance. >> The function qmp_chardev_change calls chardev_new, which label >> is NULL when creating a new chardev. The label will be passed to >> yank_register_instance which causes a segment fault. The callchain >> is as the following: >> chardev_new -> >> qemu_char_open -> >> cc->open -> >> qmp_chardev_open_socket -> >> yank_register_instance >> >> Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com> >> --- >> chardev/char-socket.c | 8 +++++--- >> 1 file changed, 5 insertions(+), 3 deletions(-) >> >> diff --git a/chardev/char-socket.c b/chardev/char-socket.c >> index c8bced76b7..26d5172682 100644 >> --- a/chardev/char-socket.c >> +++ b/chardev/char-socket.c >> @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev *chr, >> qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); >> } >> >> - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >> errp)) { >> - return; >> + if (chr->label) { >> + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >> errp)) { >> + return; >> + } >> + s->registered_yank = true; >> } >> - s->registered_yank = true; >> >> /* be isn't opened until we get a connection */ >> *be_opened = false >> > > Looks wrong to me, the new chardev will get the same label, and it should > still be possible to call the yank functions then. The registration logic > needs to be reworked during chardev-change. > > -- > Marc-André Lureau >
Hi On Tue, Mar 16, 2021 at 6:46 PM Li Zhang <li.zhang@cloud.ionos.com> wrote: > Hi Marc-André, > > By looking into chardev and yank_register_function logic, this old > chardev is registered according to the chardev label. > So it's been in yank_instance_list. yank instance only has a chardev > label, and the new chardev's label is the same as the old chardev. > So it doesn't need to register it again when changing the chardev backend. > Otherwise, it will report duplicated yank instances. > I think the chardev logic has no problems. And it works with yank > functions. > > The previous instance is being removed with yank_unregister_instance() during object_unparent(). The new instance is not registered. That scenario deserves tests. (it's a shame there are no tests for yank ...) > Thanks > Li > > On Mon, Mar 15, 2021 at 7:51 PM Marc-André Lureau < > marcandre.lureau@gmail.com> wrote: > >> Hi >> >> On Mon, Mar 15, 2021 at 9:22 PM Li Zhang <zhlcindy@gmail.com> wrote: >> >>> From: Li Zhang <li.zhang@cloud.ionos.com> >>> >>> When executing the QMP commands "chardev-change" to change the >>> backend device to socket, it will cause a segment fault because >>> it assumes chr->label as non-NULL in function yank_register_instance. >>> The function qmp_chardev_change calls chardev_new, which label >>> is NULL when creating a new chardev. The label will be passed to >>> yank_register_instance which causes a segment fault. The callchain >>> is as the following: >>> chardev_new -> >>> qemu_char_open -> >>> cc->open -> >>> qmp_chardev_open_socket -> >>> yank_register_instance >>> >>> Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com> >>> --- >>> chardev/char-socket.c | 8 +++++--- >>> 1 file changed, 5 insertions(+), 3 deletions(-) >>> >>> diff --git a/chardev/char-socket.c b/chardev/char-socket.c >>> index c8bced76b7..26d5172682 100644 >>> --- a/chardev/char-socket.c >>> +++ b/chardev/char-socket.c >>> @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev *chr, >>> qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); >>> } >>> >>> - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >>> errp)) { >>> - return; >>> + if (chr->label) { >>> + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >>> errp)) { >>> + return; >>> + } >>> + s->registered_yank = true; >>> } >>> - s->registered_yank = true; >>> >>> /* be isn't opened until we get a connection */ >>> *be_opened = false >>> >> >> Looks wrong to me, the new chardev will get the same label, and it should >> still be possible to call the yank functions then. The registration logic >> needs to be reworked during chardev-change. >> >> -- >> Marc-André Lureau >> >
Hi Marc-André, Hi Marc-André, Ah, you are right. For some scenarios, it is not registered and registered for some scenarios. If the previous chardev is not socket, it won't be registered either. There are still problems. On Tue, Mar 16, 2021 at 4:25 PM Marc-André Lureau < marcandre.lureau@gmail.com> wrote: > Hi > > On Tue, Mar 16, 2021 at 6:46 PM Li Zhang <li.zhang@cloud.ionos.com> wrote: > >> Hi Marc-André, >> >> By looking into chardev and yank_register_function logic, this old >> chardev is registered according to the chardev label. >> So it's been in yank_instance_list. yank instance only has a chardev >> label, and the new chardev's label is the same as the old chardev. >> So it doesn't need to register it again when changing the chardev >> backend. Otherwise, it will report duplicated yank instances. >> I think the chardev logic has no problems. And it works with yank >> functions. >> >> > The previous instance is being removed with yank_unregister_instance() > during object_unparent(). The new instance is not registered. > > That scenario deserves tests. (it's a shame there are no tests for yank > ...) > > > >> Thanks >> Li >> >> On Mon, Mar 15, 2021 at 7:51 PM Marc-André Lureau < >> marcandre.lureau@gmail.com> wrote: >> >>> Hi >>> >>> On Mon, Mar 15, 2021 at 9:22 PM Li Zhang <zhlcindy@gmail.com> wrote: >>> >>>> From: Li Zhang <li.zhang@cloud.ionos.com> >>>> >>>> When executing the QMP commands "chardev-change" to change the >>>> backend device to socket, it will cause a segment fault because >>>> it assumes chr->label as non-NULL in function yank_register_instance. >>>> The function qmp_chardev_change calls chardev_new, which label >>>> is NULL when creating a new chardev. The label will be passed to >>>> yank_register_instance which causes a segment fault. The callchain >>>> is as the following: >>>> chardev_new -> >>>> qemu_char_open -> >>>> cc->open -> >>>> qmp_chardev_open_socket -> >>>> yank_register_instance >>>> >>>> Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com> >>>> --- >>>> chardev/char-socket.c | 8 +++++--- >>>> 1 file changed, 5 insertions(+), 3 deletions(-) >>>> >>>> diff --git a/chardev/char-socket.c b/chardev/char-socket.c >>>> index c8bced76b7..26d5172682 100644 >>>> --- a/chardev/char-socket.c >>>> +++ b/chardev/char-socket.c >>>> @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev >>>> *chr, >>>> qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); >>>> } >>>> >>>> - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >>>> errp)) { >>>> - return; >>>> + if (chr->label) { >>>> + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), >>>> errp)) { >>>> + return; >>>> + } >>>> + s->registered_yank = true; >>>> } >>>> - s->registered_yank = true; >>>> >>>> /* be isn't opened until we get a connection */ >>>> *be_opened = false >>>> >>> >>> Looks wrong to me, the new chardev will get the same label, and it >>> should still be possible to call the yank functions then. The registration >>> logic needs to be reworked during chardev-change. >>> >>> -- >>> Marc-André Lureau >>> >> > > -- > Marc-André Lureau >
On Mon, 15 Mar 2021 18:06:35 +0100 Li Zhang <zhlcindy@gmail.com> wrote: > From: Li Zhang <li.zhang@cloud.ionos.com> > > When executing the QMP commands "chardev-change" to change the > backend device to socket, it will cause a segment fault because > it assumes chr->label as non-NULL in function yank_register_instance. > The function qmp_chardev_change calls chardev_new, which label > is NULL when creating a new chardev. The label will be passed to > yank_register_instance which causes a segment fault. The callchain > is as the following: > chardev_new -> > qemu_char_open -> > cc->open -> > qmp_chardev_open_socket -> > yank_register_instance Oh, I didn't consider the chardev-change case. I'll look into it. Regards, Lukas Straub --
Looks like a bug fix. Lukas, can you take care of it in time for 6.0? Li Zhang <zhlcindy@gmail.com> writes: > From: Li Zhang <li.zhang@cloud.ionos.com> > > When executing the QMP commands "chardev-change" to change the > backend device to socket, it will cause a segment fault because > it assumes chr->label as non-NULL in function yank_register_instance. > The function qmp_chardev_change calls chardev_new, which label > is NULL when creating a new chardev. The label will be passed to > yank_register_instance which causes a segment fault. The callchain > is as the following: > chardev_new -> > qemu_char_open -> > cc->open -> > qmp_chardev_open_socket -> > yank_register_instance > > Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com> > --- > chardev/char-socket.c | 8 +++++--- > 1 file changed, 5 insertions(+), 3 deletions(-) > > diff --git a/chardev/char-socket.c b/chardev/char-socket.c > index c8bced76b7..26d5172682 100644 > --- a/chardev/char-socket.c > +++ b/chardev/char-socket.c > @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev *chr, > qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); > } > > - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), errp)) { > - return; > + if (chr->label) { > + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), errp)) { > + return; > + } > + s->registered_yank = true; > } > - s->registered_yank = true; > > /* be isn't opened until we get a connection */ > *be_opened = false;
On Fri, 26 Mar 2021 15:41:11 +0100 Markus Armbruster <armbru@redhat.com> wrote: > Looks like a bug fix. Lukas, can you take care of it in time for 6.0? > Yeah, this patch only fixes a symptom, but not the core cause of the bug. I have already written patches that fix the bugs inclusive test-cases: https://lore.kernel.org/qemu-devel/cover.1616744509.git.lukasstraub2@web.de/ Regards, Lukas Straub --
Lukas Straub <lukasstraub2@web.de> writes: > On Fri, 26 Mar 2021 15:41:11 +0100 > Markus Armbruster <armbru@redhat.com> wrote: > >> Looks like a bug fix. Lukas, can you take care of it in time for 6.0? >> > > Yeah, this patch only fixes a symptom, but not the core cause of the bug. > I have already written patches that fix the bugs inclusive test-cases: > https://lore.kernel.org/qemu-devel/cover.1616744509.git.lukasstraub2@web.de/ Awesome! I didn't make the connection. I trust you'll take care of getting the fixes merged :)
diff --git a/chardev/char-socket.c b/chardev/char-socket.c index c8bced76b7..26d5172682 100644 --- a/chardev/char-socket.c +++ b/chardev/char-socket.c @@ -1421,10 +1421,12 @@ static void qmp_chardev_open_socket(Chardev *chr, qemu_chr_set_feature(chr, QEMU_CHAR_FEATURE_FD_PASS); } - if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), errp)) { - return; + if (chr->label) { + if (!yank_register_instance(CHARDEV_YANK_INSTANCE(chr->label), errp)) { + return; + } + s->registered_yank = true; } - s->registered_yank = true; /* be isn't opened until we get a connection */ *be_opened = false;