On Mon, May 26, 2014 at 05:59:03PM +0200, Markus Armbruster wrote: > Kevin Wolf <kw...@redhat.com> writes: > > > Am 26.05.2014 um 17:02 hat Markus Armbruster geschrieben: > >> Stefan Hajnoczi <stefa...@gmail.com> writes: > >> > >> > On Mon, May 26, 2014 at 09:44:03AM +0800, Le Tan wrote: > >> >> Replace fprintf(stderr,...) with error_report() in files block/*, > >> >> block.c, > >> >> block-migration.c and blockdev.c. The trailing "\n"s of the @fmt > >> >> argument > >> >> have been removed because @fmt of error_report() should not > >> >> contain newline. > >> >> > >> >> Signed-off-by: Le Tan <tamlokv...@gmail.com> > >> >> --- > >> >> block-migration.c | 6 +-- > >> >> block.c | 4 +- > >> >> block/qcow2-refcount.c | 115 > >> >> +++++++++++++++++++++--------------------- > >> >> block/qcow2.c | 16 +++--- > >> >> block/raw-posix.c | 8 ++- > >> >> block/raw-win32.c | 6 +-- > >> >> block/ssh.c | 2 +- > >> >> block/vdi.c | 14 +++--- > >> >> block/vmdk.c | 15 +++--- > >> >> block/vpc.c | 4 +- > >> >> block/vvfat.c | 129 > >> >> ++++++++++++++++++++++++------------------------ > >> >> blockdev.c | 6 +-- > >> >> 12 files changed, 159 insertions(+), 166 deletions(-) > >> > > >> > There is one thing that worries me about this: > >> > > >> > error_report() assumes that the QEMU global mutex is held in order to > >> > access the "current monitor". > >> > >> Global variable cur_mon, non-null while we're executing a monitor > >> command. > > > > The important part here is that it's indeed global and not thread-local. > > > >> > This is problematic for code in the read/write/flush path (like qcow2 > >> > refcount allocation) since it can be invoked from a virtio-blk > >> > data-plane thread (where the QEMU global mutex is not held). > >> > >> error_report() reports to the current monitor when "in monitor context", > >> i.e. while executing a monitor command, i.e. while cur_mon isn't null. > >> > >> Can we ever be in monitor context (cur_mon not null) without holding the > >> global mutex? > > > > The right question is: Can a thread (= the main loop thread) ever be in > > monitor context while another thread (= dataplane thread) is executing > > block driver code and doesn't hold the global mutex? > > > > If I understand dataplane correctly, the whole point of it is that the > > answer to this is yes. > > Well, the answer used to be "no". Once upon a time because there was > just a single thread, later on because only one thread ever executed > "interesting" code. > > I'm *not* saying this should remain the case. I'm trying to find out > what needs to be done around cur_mon when we break the assumption behind > it. > > Would making cur_mon thread-local suffice?
It would suffice. I think there is no code that invokes error_report() from another thread on purpose (i.e. it knows the main thread is waiting), so making it thread-local should not introduce a regression. Stefan
