On Thu, Dec 05, 2013 at 04:38:33PM +0100, Stefan Hajnoczi wrote: > rbd callbacks are called from non-QEMU threads. Up until now a pipe was > used to signal completion back to the QEMU iothread. > > The pipe writer code handles EAGAIN using select(2). The select(2) API > is not scalable since fd_set size is static. FD_SET() can write beyond > the end of fd_set if the file descriptor number is too high. (QEMU's > main loop uses poll(2) to avoid this issue with select(2).) > > Since the pipe itself is quite clumsy to use and QEMUBH is now > thread-safe, just schedule a BH from the rbd callback function. This > way we can simplify I/O completion in addition to eliminating the > potential FD_SET() crash when file descriptor numbers become too high. > > Crash scenario: QEMU already has 1024 file descriptors open. Hotplug an > rbd drive and get the pipe writer to take the select(2) code path. > > Signed-off-by: Stefan Hajnoczi <stefa...@redhat.com> > --- > Josh: This patch has not been tested. I have just compiled it. > > block/rbd.c | 130 > ++++++++++-------------------------------------------------- > 1 file changed, 22 insertions(+), 108 deletions(-)
Applied to my block-next tree: https://github.com/stefanha/qemu/commits/block-next Stefan
