published May 1, 2025
In the Linux kernel, the following vulnerability has been resolved:
9p/trans_fd: always use O_NONBLOCK read/write
syzbot is reporting hung task at p9_fd_close() [1], for p9_mux_poll_stop()
from p9_conn_destroy() from p9_fd_close() is failing to interrupt already
started kernel_read() from p9_fd_read() from p9_read_work() and/or
kernel_write() from p9_fd_write() from p9_write_work() requests.
Since p9_socket_open() sets O_NONBLOCK flag, p9_mux_poll_stop() does not
need to interrupt kernel_read()/kernel_write(). However, since p9_fd_open()
does not set O_NONBLOCK flag, but pipe blocks unless signal is pending,
p9_mux_poll_stop() needs to interrupt kernel_read()/kernel_write() when
the file descriptor refers to a pipe. In other words, pipe file descriptor
needs to be handled as if socket file descriptor.
We somehow need to interrupt kernel_read()/kernel_write() on pipes.
A minimal change, which this patch is doing, is to set O_NONBLOCK flag
from p9_fd_open(), for O_NONBLOCK flag does not affect reading/writing
of regular files. But this approach changes O_NONBLOCK flag on userspace-
supplied file descriptors (which might break userspace programs), and
O_NONBLOCK flag could be changed by userspace. It would be possible to set
O_NONBLOCK flag every time p9_fd_read()/p9_fd_write() is invoked, but still
remains small race window for clearing O_NONBLOCK flag.
If we don't want to manipulate O_NONBLOCK flag, we might be able to
surround kernel_read()/kernel_write() with set_thread_flag(TIF_SIGPENDING)
and recalc_sigpending(). Since p9_read_work()/p9_write_work() works are
processed by kernel threads which process global system_wq workqueue,
signals could not be delivered from remote threads when p9_mux_poll_stop()
from p9_conn_destroy() from p9_fd_close() is called. Therefore, calling
set_thread_flag(TIF_SIGPENDING)/recalc_sigpending() every time would be
needed if we count on signals for making kernel_read()/kernel_write()
non-blocking.
[Dominique: add comment at Christian's suggestion]