Re: Hurd and UFS

[Michael Teichgraeber]

Roland McGrath <roland@gnu.org> writes:

> [...] As to the SIGLOST crash, that is kind of strange.  That error
> is supposed to indicate that the RPC server appeared to die, in this
> case the storeio or filesystem. 

Storeio keeps running, but ...


> First you can try "trap '' SIGLOST" to ignore the signal [...]

... `mkfs.ufs' tells me about a `(ipc/send) invalid destination port',
when SIGLOST is trapped.


I have looked into __libc_lseek(), which is called from rdfs() in
mkfs.c. The interesting part of this quite short function is the macro
HURD_DPORT_USE that expands to about 40 lines. In one of the first
lines a hurd_critical_section is being entered, then the program steps
into an `else' path with the following commands:

,-------------------------------------------------------------------
|  ctty = _hurd_port_get (&__d->ctty, &__ctty_ulink );
|  port = _hurd_port_locked_get (&__d->port, &__ulink); /* port.h */
|
|  _hurd_critical_section_unlock (crit);
|  __result = ((__io_seek (port, offset, whence, &offset)));
|  _hurd_port_free (&__d-> port, &__ulink, port);
|
|  if (ctty != MACH_PORT_NULL)
|    _hurd_port_free (&__d-> ctty, &__ctty_ulink, ctty);
`-------------------------------------------------------------------

In __io_seek(), which is implemented in glibc/hurd/RPC_io_seek.c,
there's a __mach_msg() call, that returns EMACH_SEND_INVALID_DEST, if
mkfs.ufs is run as root, otherwise `success'; __io_seek() then returns
with this value, so that __result (above, in lseek.c) contains
EMACH_...

Finally hurd_dfail() will be called and the SIGLOST signal be raised.

The input of __mach_msg(), i.e. Mess.In, looks ok in both (root and
non-root) cases (local and remote port numbers differ slightly).
Maybe in _hurd_port_locked_get() something happens, that makes the
port invalid.

I don't know wheather this strange behaviour applies to other Hurd
systems too. My system contains Debian packages of: Hurd (Nov 5),
Gnumach (Oct 13) and glibc[-dbg]_2.2.4-3 (Oct 11).

Perhaps you could test whether `mkfs.ufs' fails for `root' on your
system too, with running

   dd if=/dev/zero of=./testfs bs=1024 count=1024
   mkfs.ufs --track=1 --sectors=63 ./testfs


Michael