Hi,
I've been chasing a bug when using virtio console to exchange data
between guest and host. It manifests in a process inside the guest
getting stuck while writing to the serial port:
[<0>] wait_port_writable+0x139/0x2d0
[<0>] port_fops_write+0x88/0x130
[<0>] vfs_write+0xf3/0x450
[<0>] ksys_write+0x6d/0xe0
[<0>] do_syscall_64+0x9e/0x1a0
[<0>] entry_SYSCALL_64_after_hwframe+0x77/0x7
I managed to track this down to a race in virtio_console.c. The driver
doesn't properly account for a port that receives blocking writes and
is polled at the same time. I suspect that a similar problem exists in
port_fops_read().
Here is how it goes. We need two threads inside the guest, A and B.
- Thread A writes to the serial port until the virtqueue fills up.
- Thread A invokes port_fops_write() which calls wait_port_writable().
There are no used buffers to reclaim and the thread is suspended,
waiting on port->waitqueue.
- The host side of the device makes some progress, marks some buffers
as consumed / used and sends an interrupt to the guest.
- Before the guest can service the interrupt, thread B executes
port_fops_poll(). This calls into
reclaim_consumed_buffers() via will_write_block(), which removes all
used buffers.
- The interrupt gets serviced, calling into vring_interrupt(). Here
the check for more_used() returns false because the poll just went
through all the work. The handler returns without waking
port->waitqueue.
- port_fops_write() never returns.
I'm not exactly sure how to best fix this. Maybe it's enough to only
check outvq_full in poll? It'd also be nice to fix this on the read
side as well.
I only have a reproducer which requires a Go toolchain unfortunately:
https://github.com/lmb/vimto/issues/29
Best
Lorenz
