On 2012-05-09 16:27, Michael Tokarev wrote:
On 09.05.2012 23:21, Jan Kiszka wrote:
On i386, glibc only saves/restores the signal mask via sigprocmask,
excluding RT signal. A Linux bug in the compat version of this syscall
corrupts the RT signal state, which will cause lockups of QEMU's VCPU
threads.
This should obviously be fixed in kernel, for benefit of all (not only
qemu), do you have any details here?
compat_sys_sigprocmask reads 32-bit sigmask from user space, i.e.
excluding RT signal, but calls sys_sigprocmask that takes a 64-bit
sigset. So the RT signals are unblocked. I'm testing a simple patch ATM,
will post it to LKML once this works.
Signed-off-by: Jan Kiszka<address@hidden>
---
I'm not sure where to fall back to. The existing code uses gthread,
likely because it is the safer harbor. So I picked it as well.
Can't we resort to the SIGUSR1 workaround for the time being, while
no RT signals are in actual use, and just have the time to let the
kernel side to fix the things up before some actual RTsig user will
emerge in qemu? I think it is a bit more conservative approach,
especially having in mind the minority of users this issue affects
(only 32/64 mixed environment). I'd favor for this variant, and
it looks like I'm the "main" 32/64bit user of qemu in this world :)
Most conservative is definitely this patch, not switching to SIGUSR1,
hoping that no other RT signal user shows up until current kernel are no
longer in use.