On Sat, Feb 26, 2011 at 01:29:01PM +0100, Jan Kiszka wrote: > > at /var/tmp/portage/app-emulation/qemu-kvm-0.14.0/work/qemu-kvm-0.14.0/qemu-kvm.c:1466 > > #12 0x00007ffff77bb944 in start_thread () from /lib/libpthread.so.0 > > #13 0x00007ffff5e491dd in clone () from /lib/libc.so.6 > > (gdb) > > That's a spice bug. In fact, there are a lot of > qemu_mutex_lock/unlock_iothread in that subsystem. I bet at least a few > of them can cause even more subtle problems. > > Two general issues with dropping the global mutex like this: > - The caller of mutex_unlock is responsible for maintaining > cpu_single_env across the unlocked phase (that's related to the > abort above). > - Dropping the lock in the middle of a callback is risky. That may > enable re-entrances of code sections that weren't designed for this > (I'm skeptic about the side effects of > qemu_spice_vm_change_state_handler - why dropping the lock here?). > > Spice requires a careful review regarding such issues. Or it should > pioneer with introducing its own lock so that we can handle at least > related I/O activities over the VCPUs without holding the global mutex > (but I bet it's not the simplest candidate for such a new scheme). > > Jan > Agree with the concern regarding spice. Regarding global mutex, TCG and KVM execution behaviour can become more similar wrt locking by dropping qemu_global_mutex during generation and execution of TBs. Of course for memory or PIO accesses from vcpu context qemu_global_mutex must be acquired. With that in place, it becomes easier to justify further improvements regarding parallelization, such as using a read-write lock for l1_phys_map / phys_page_find_alloc. 21.62% sh 3d38920b3f [.] 0x00003d38920b3f 6.38% sh qemu-system-x86_64 [.] phys_page_find_alloc 4.90% sh qemu-system-x86_64 [.] tb_find_fast 4.34% sh qemu-system-x86_64 [.] tlb_flush -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
