Hello,
I've just committed a whole bunch of bugfixes to CVS. Everyone who was
experiencing crashes recently, could you please try again with this
version?
The bugs in detail:
- The trick of defining special symbols via a linker script to retrieve
the module virtual address range broke badly under certain combinations
of kernel / insmod versions, causing all kinds of strange effects.
I've now completely removed this hack, and instead get the module
address range by querying the corresponding kernel structure (the
__this_module structure) in kernels >= 2.1.18, and using a hack in
earlier kernels (the symbol mod_use_count_ is guaranteed to point
to the start of the module, and the end of the module is guaranteed
to be the first virtual address greater than the start where there
is no physical page mapped ...).
Thanks to Oleg Drokin for helping to debug this problem.
- Returning to process pending signals didn't always work correctly;
on a 2.2.5 kernel, our pending_signals() routine caused strange
effects, and the return value of -ERESTARTSYS appears not to be
understood by some libc versions. Thus, I've switched to using
the Linux standard current_got_fatal_signal() routine for the test
(this is available on all interesting kernel versions), and return
directly -EINTR, which appears to always work ...
If anyone knows what's up with the ERESTARTSYS/EINTR mixup,
please let me know.
- Setting tss->io to 0 does *not* disallow I/O port access! Instead,
the tss structure is itself used as I/O permission bitmap, and
depending on the values of some of the fields, critical system ports
are accessible by the guest. This caused the timer hardware to
get into strange states after running the preemptive kernel, which
I only managed to fix by a reboot :-/
The fix is to set tss->io to some value greater than the TSS limit.
- The NT flag hack (necessary for proper IRET simulation) was missing
from the preemptive kernel.
- Some symbols in nexus.S were not defined as .globl, although they were
accessed from within monitor.c (apparently, some broken insmod versions
resolve those symbols anyway!).
Thanks to Julian Hall for pointing this out.
With those fixes in place, I'm able to run all three guests on both
of my systems (Linux 2.0.36 and Linux 2.2.5); the preemptive kernel
aborts on a not yet implemented 'out' instruction. Please let me know
whether it works for you!
Bye,
Ulrich
--
Ulrich Weigand,
IMMD 1, Universitaet Erlangen-Nuernberg,
Martensstr. 3, D-91058 Erlangen, Phone: +49 9131 85-7688
