[Xen-devel] Deadlocks by p2m_lock and event_lock

["Hao, Xudong" <xudong.hao@xxxxxxxxx>]

Hi, Jan and Andres

Here we met a deadlocks issue by p2m_lock and event_lock on Xen. The issue 
appears from the series of patches "Synchronized p2m lookups, Xen 24770~", 
deadlocks may happen on a specifical case: Assign a PCIe device with MSI-x 
capability to HVM guest(in case of EPT).

I've dump the all processor registers when dom0 hang( attach the log). The 
deadlocks happen as follows:
====CPU0===
map_domain_pirq()    Grab event_lock
  /
Pci_enable_msi()
  /
msix_capability_init()
  /
p2m_change_entry_type_global()   Trying to acquire p2m_lock

====CPU9===
hvm_hap_nested_page_fault() -> get_gfn_type_access()   Grab p2m_lock
  /
handle_mmio()
  /
...
  /
notify_via_xen_event_channel()    Trying to acquire event_lock


The event_lock is used anywhere in Xen, I only have a patch of workaround this 
issue for proposal, but not for the final fix. Any good suggestion?

diff -r f61120046915 xen/arch/x86/irq.c
--- a/xen/arch/x86/irq.c        Wed Mar 07 11:50:31 2012 +0100
+++ b/xen/arch/x86/irq.c        Sat Mar 10 02:06:18 2012 +0800
@@ -1875,10 +1875,12 @@ int map_domain_pirq(
         if ( !cpu_has_apic )
             goto done;
 
+        spin_unlock(&d->event_lock);
         pdev = pci_get_pdev(msi->seg, msi->bus, msi->devfn);
         ret = pci_enable_msi(msi, &msi_desc);
         if ( ret )
             goto done;
+        spin_lock(&d->event_lock);
 
         spin_lock_irqsave(&desc->lock, flags);

Best Regards,
Xudong Hao

Attachment: registers_dump.log
Description: registers_dump.log

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel