On Thu, 1 Sep 2011, Stefan Bader wrote:
> After much joy with this, I thought I post this to a bigger audience. After
> having migrated to Xen 4.1.1, booting HVM guests had several issues. Some
> related to interrupts not being set up correctly (which Stefano has posted
> patches) and even with those 3.0 guests seem to hang for me while 2.6.38 or
> older kernels were ok.
>
> After digging deeply into this, I think I found the issue. However, if that is
> true, it seems rather lucky if pv spinlocks in HVM worked for anybody.
>
> The xen_hvm_smp_init() call will change the smp_ops hooks. One of which is
> smp_prepare_cpus. This is done in start_kernel and at this point in time,
> there
> is no change to the pv_lock_ops and they point to the ticket versions.
> Later in start_kernel, check_bugs is called and part of that takes the
> pv_lock_ops and patches the kernel with the correct jumps.
> _After_ that, the kernel_init is called and that in turn does the
> smp_prepare_cpus which now changes the pv_lock_ops again, *but not* run any
> patching again. So the _raw_spin_*lock calls still use the ticket calls.
>
> start_kernel
> setup_arch -> xen_hvm_smp_init (set smp_ops)
> ...
> check_bugs -> alternative_instructions (patches pv_locks sites)
> ...
> rest_init (triggers kernel_init as a thread)
> kernel_init
> ...
> smp_prepare_cpus (calls xen_init_spinlocks through smp_ops hook)
>
> To make things even more fun, there is the possibility that some spinlock
> functions are forced to be inlined and others are not (CONFIG_INLINE_SPIN_*).
> In
> our special case only two versions of spin_unlock were inlined. Put that
> together into a pot with modules, shake well, and there is the fun. Basically
> on
> load time, the non-inline calls remain pointing to the unmodified ticket
> implementation (main kernel). But the unlock calls (which are inlined) get
> modified because the loaded module gets patched up. One can imagine that this
> does not work too well.
>
> Anyway, even without the secondary issue, I am sure that just replacing the
> functions in pv_lock_ops without the spinlock calls getting actually modified
> is
> not the intended behaviour.
>
> Unfortunately I have not yet been able to make it work. Any attempt to move
> xen_init_spinlocks to be called before check_bugs or adding a call to
> alternative_instructions results in another hang on boot. At least the latter
> method results in a more usable dump for crash, which shows that on spinlock
> was
> taken (slow) and two spurious taken ones (this is more to play for me).
>
> But then, maybe someone has some ideas there...
we have two possible ways to solve this problem:
1) we wait for Jeremy's pv ticketlock series to be applied, that should
make the datastructure internally used by native ticketlock and pv
tickerlock identical, so a lock taken by the native ticket spin lock
function can be freed by the xen pv ticket spin unlock function.
2) we add a very early hook to call xen_init_spinlocks() before any
spinlocks have been taken, that means earlier than
init_hypervisor_platform()
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home