Recent MFC to 7 causes crash on VMware ESXi

[Kostik Belousov]

On Mon, Feb 08, 2010 at 09:49:00AM -0500, John Baldwin wrote:
> On Saturday 06 February 2010 4:47:16 pm Tom McLaughlin wrote:
> > John Baldwin wrote, On 02/05/2010 08:27 AM:
> > > On Thursday 04 February 2010 10:00:55 pm Tom McLaughlin wrote:
> > >> Hi all, a recent MFC to 7-STABLE has started to cause issues for my VMs
> > >> on VMware ESXi 3.5u4.  After loading the mpt driver for the LSI disk
> > >> controller the VM just shuts off.  The workaround is to change the disk
> > >> controller to the BusLogic type.  Still, it used to work up until last
> > >> week.  The change was made around January 26th and based on the commits
> > >> that day I'm guessing it's either r203047 or r203073
> > >>
> > >> I have the same issue with both amd64 and i386 VMs.  This affects HEAD
> > >> and 8-STABLE as well and first affected HEAD over the summer.  (I just
> > >> worked around it and went about my business at the time. :-/)  I've
> > >> attached a dmesg from a kernel before the problem and one from after it
> > >> started.
> > > 
> > > What if you set 'hw.clfush_disable=1' from the loader?
> > > 
> > 
> > Yes, that corrected it on all my VMs.  I've talked to people on ESXi 4
> > and they do not see the problem.  I have yet to try 3.5u5 to see if this
> > is a non-issue.  3.5 will be supported for awhile longer from VMware.
> > I'm going to try upgrading the box during the week.
> 
> I believe folks had to do this on HEAD/8.x as well.  Perhaps we can 
> automatically disable clflush if we are executing under VMware or Xen:
> 
> Index: amd64/amd64/initcpu.c
> ===================================================================
> --- amd64/amd64/initcpu.c	(revision 203430)
> +++ amd64/amd64/initcpu.c	(working copy)
> @@ -177,17 +177,16 @@
>  	if ((cpu_feature & CPUID_CLFSH) != 0)
>  		cpu_clflush_line_size = ((cpu_procinfo >> 8) & 0xff) * 8;
>  	/*
> -	 * XXXKIB: (temporary) hack to work around traps generated when
> -	 * CLFLUSHing APIC registers window.
> +	 * XXXKIB: (temporary) hack to work around traps generated
> +	 * when CLFLUSHing APIC registers window under virtualization
> +	 * environments.
>  	 */
>  	TUNABLE_INT_FETCH("hw.clflush_disable", &hw_clflush_disable);
> -	if (cpu_vendor_id == CPU_VENDOR_INTEL && !(cpu_feature & CPUID_SS) &&
> -	    hw_clflush_disable == -1)
> +	if (vm_guest != 0 /* VM_GUEST_NO */ && hw_clflush_disable == -1)
>  		cpu_feature &= ~CPUID_CLFSH;
>  	/*
>  	 * Allow to disable CLFLUSH feature manually by
> -	 * hw.clflush_disable tunable.  This may help Xen guest on some AMD
> -	 * CPUs.
> +	 * hw.clflush_disable tunable.
>  	 */
>  	if (hw_clflush_disable == 1)
>  		cpu_feature &= ~CPUID_CLFSH;
> Index: i386/i386/initcpu.c
> ===================================================================
> --- i386/i386/initcpu.c	(revision 203430)
> +++ i386/i386/initcpu.c	(working copy)
> @@ -724,17 +724,16 @@
>  	if ((cpu_feature & CPUID_CLFSH) != 0)
>  		cpu_clflush_line_size = ((cpu_procinfo >> 8) & 0xff) * 8;
>  	/*
> -	 * XXXKIB: (temporary) hack to work around traps generated when
> -	 * CLFLUSHing APIC registers window.
> +	 * XXXKIB: (temporary) hack to work around traps generated
> +	 * when CLFLUSHing APIC registers window under virtualization
> +	 * environments.
>  	 */
>  	TUNABLE_INT_FETCH("hw.clflush_disable", &hw_clflush_disable);
> -	if (cpu_vendor_id == CPU_VENDOR_INTEL && !(cpu_feature & CPUID_SS) &&
> -	    hw_clflush_disable == -1)
> +	if (vm_guest != 0 /* VM_GUEST_NO */ && hw_clflush_disable == -1)
>  		cpu_feature &= ~CPUID_CLFSH;
>  	/*
>  	 * Allow to disable CLFLUSH feature manually by
> -	 * hw.clflush_disable tunable.  This may help Xen guest on some AMD
> -	 * CPUs.
> +	 * hw.clflush_disable tunable.
>  	 */
>  	if (hw_clflush_disable == 1)
>  		cpu_feature &= ~CPUID_CLFSH;

It might be better to "or" old condition, i.e. Intel without SS, and
new one, vm_guest != 0, instead of replacing the old ?
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-stable/attachments/20100208/488f321d/attachment.pgp