Re: [PATCH] rumpdisk: Defer hardware initialization to avoid deadlock on stat.
Applied, thanks!
Milos Nikic, le mar. 27 janv. 2026 12:01:00 -0800, a ecrit:
> Currently, rumpdisk initializes the Rump kernel and probes the PCI bus
> immediately during main(). When rumpdisk is configured as a passive
> translator (e.g., for /dev/rumpusbdisk), a simple stat or ls on the node
> triggers translator startup. This causes a second Rump instance to attempt to
> probe hardware already owned by the primary disk driver, leading to a
> system-wide deadlock.
>
> To reproduce:
> stat /dev/rumpdiskusb
> (rump starts again ... deadlock)
>
> With this patch:
> stat /dev/rumpdiskusb
> (normal printout of stats, no deadlock)
> ---
> rumpdisk/main.c | 14 +-
> 1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/rumpdisk/main.c b/rumpdisk/main.c
> index 3676b472..0bd753e2 100644
> --- a/rumpdisk/main.c
> +++ b/rumpdisk/main.c
> @@ -40,6 +40,15 @@
>
> mach_port_t bootstrap_resume_task = MACH_PORT_NULL;
>
> +static pthread_once_t rump_hw_initialized = PTHREAD_ONCE_INIT;
> +
> +static void
> +do_rump_hw_init (void)
> +{
> + mach_print("rumpdisk: Initializing Rump block device layer...\n");
> + rump_register_block ();
> +}
> +
> static const struct argp_option options[] = {
>{"host-priv-port", 'h', "PORT", 0, "Host private port PORT"},
>{"device-master-port",'d', "PORT", 0, "Device master port PORT"},
> @@ -102,6 +111,7 @@ static const struct argp *rumpdisk_argp_bootup =
> &rumpdisk_argp;
> static int __thread wired = 0;
> static int rumpdisk_demuxer (mach_msg_header_t *inp, mach_msg_header_t *outp)
> {
> + pthread_once (&rump_hw_initialized, do_rump_hw_init);
>/* FIXME: we are not wired while receiving our first message. */
>if (!wired)
> {
> @@ -150,7 +160,6 @@ main (int argc, char **argv)
>error(1, err, "Missing parameters for bootstrap");
> }
>
> - rump_register_block ();
>machdev_trivfs_init (argc, argv, bootstrap_resume_task, RUMPNAME, "/dev/"
> RUMPNAME, &bootstrap);
>
>/* Make sure we will not swap out, in case we drive the disk used for
> @@ -162,6 +171,9 @@ main (int argc, char **argv)
>if (err != KERN_SUCCESS)
> error (1, err, "cannot get vm_privilege");
>
> + if (bootstrap_resume_task != MACH_PORT_NULL)
> +pthread_once (&rump_hw_initialized, do_rump_hw_init);
> +
>machdev_device_init ();
>err = pthread_create (&t, NULL, rumpdisk_multithread_server, NULL);
>if (err)
> --
> 2.52.0
>
>
--
Samuel
quit When the quit statement is read, the bc processor
is terminated, regardless of where the quit state-
ment is found. For example, "if (0 == 1) quit"
will cause bc to terminate.
(Seen in the manpage for "bc". Note the "if" statement's logic)
Re: [PATCH] rumpdisk: Defer hardware initialization to avoid deadlock on stat.
It seems like i mistyped in reproduction steps...
The easiest way to deadlock is just to do this inside the live Hurd
stat /dev/rumpusbdisk
With this patch, that won't be happening any more
Thanks,
Milos
On Tue, Jan 27, 2026 at 12:01 PM Milos Nikic wrote:
> Currently, rumpdisk initializes the Rump kernel and probes the PCI bus
> immediately during main(). When rumpdisk is configured as a passive
> translator (e.g., for /dev/rumpusbdisk), a simple stat or ls on the node
> triggers translator startup. This causes a second Rump instance to attempt
> to probe hardware already owned by the primary disk driver, leading to a
> system-wide deadlock.
>
> To reproduce:
> stat /dev/rumpdiskusb
> (rump starts again ... deadlock)
>
> With this patch:
> stat /dev/rumpdiskusb
> (normal printout of stats, no deadlock)
> ---
> rumpdisk/main.c | 14 +-
> 1 file changed, 13 insertions(+), 1 deletion(-)
>
> diff --git a/rumpdisk/main.c b/rumpdisk/main.c
> index 3676b472..0bd753e2 100644
> --- a/rumpdisk/main.c
> +++ b/rumpdisk/main.c
> @@ -40,6 +40,15 @@
>
> mach_port_t bootstrap_resume_task = MACH_PORT_NULL;
>
> +static pthread_once_t rump_hw_initialized = PTHREAD_ONCE_INIT;
> +
> +static void
> +do_rump_hw_init (void)
> +{
> + mach_print("rumpdisk: Initializing Rump block device layer...\n");
> + rump_register_block ();
> +}
> +
> static const struct argp_option options[] = {
>{"host-priv-port", 'h', "PORT", 0, "Host private port PORT"},
>{"device-master-port",'d', "PORT", 0, "Device master port PORT"},
> @@ -102,6 +111,7 @@ static const struct argp *rumpdisk_argp_bootup =
> &rumpdisk_argp;
> static int __thread wired = 0;
> static int rumpdisk_demuxer (mach_msg_header_t *inp, mach_msg_header_t
> *outp)
> {
> + pthread_once (&rump_hw_initialized, do_rump_hw_init);
>/* FIXME: we are not wired while receiving our first message. */
>if (!wired)
> {
> @@ -150,7 +160,6 @@ main (int argc, char **argv)
>error(1, err, "Missing parameters for bootstrap");
> }
>
> - rump_register_block ();
>machdev_trivfs_init (argc, argv, bootstrap_resume_task, RUMPNAME,
> "/dev/" RUMPNAME, &bootstrap);
>
>/* Make sure we will not swap out, in case we drive the disk used for
> @@ -162,6 +171,9 @@ main (int argc, char **argv)
>if (err != KERN_SUCCESS)
> error (1, err, "cannot get vm_privilege");
>
> + if (bootstrap_resume_task != MACH_PORT_NULL)
> +pthread_once (&rump_hw_initialized, do_rump_hw_init);
> +
>machdev_device_init ();
>err = pthread_create (&t, NULL, rumpdisk_multithread_server, NULL);
>if (err)
> --
> 2.52.0
>
>
