On 07/22/2018 11:57 AM, M. Vefa Bicakci wrote: > On 07/21/2018 07:17 PM, M. Vefa Bicakci wrote: >> On 07/21/2018 05:25 PM, Boris Ostrovsky wrote: >>> On 07/21/2018 03:49 PM, M. Vefa Bicakci wrote: >>>> diff --git a/arch/x86/xen/enlighten_pv.c b/arch/x86/xen/enlighten_pv.c >>>> index 439a94bf89ad..87afb000142a 100644 >>>> --- a/arch/x86/xen/enlighten_pv.c >>>> +++ b/arch/x86/xen/enlighten_pv.c >>>> @@ -1257,6 +1257,7 @@ asmlinkage __visible void __init >>>> xen_start_kernel(void) >>>> /* Work out if we support NX */ >>>> get_cpu_cap(&boot_cpu_data); >>>> + get_cpu_address_sizes(&boot_cpu_data); >>>> x86_configure_nx(); >>> >>> >>> Have you observed any problems without this call? get_cpu_cap() is only >>> called here to set X86_FEATURE_NX, and is then called again, together >>> with get_cpu_address_sizes(), from early_identify_cpu(). >> >> Thank you for the reviews! Without the call to get_cpu_address_sizes, >> paravirtualized virtual machines do not boot up kernels with versions >> 4.17 and up at all; this includes dom0 and domU. No domU logs are >> generated in dom0's /var/log/xen/console/ directory either, despite >> having earlyprintk=xen on the kernel command line for my test domU. > > Hello Boris, > > I debugged this further with a debugging version of Xen (so that I can > get early kernel print-outs via the "xen_raw_console_write" function), > and I found the root cause of the boot up failure. > > In summary, the issue is due to the following call path in version > 4.17 (and higher, I assume), which the kernel goes through /only/ when > CONFIG_DEBUG_VIRTUAL is enabled: > > enlighten_pv.c::xen_start_kernel > mmu_pv.c::xen_reserve_special_pages > page.h::__pa > physaddr.c::__phys_addr > physaddr.h::phys_addr_valid // uses boot_cpu_data.x86_phys_bits > > The return value of phys_addr_valid is used with the VIRTUAL_BUG_ON > macro, > which evaluates to BUG_ON in case CONFIG_DEBUG_VIRTUAL is enabled. Ah, that's why it hasn't been detected. > > It looks like the call to get_cpu_address_size is required in the > xen_start_kernel function. Perhaps there is a more elegant way to > resolve this issue as well. > > Another approach could be to check in the phys_addr_valid function > whether > boot_cpu_data.x86_phys_bits has been initialized or not, I think, but > I am > not sure about the correctness of this approach. No, I think your patch is good. The only thing I'd suggest is to move the call a few lines down. The way it is placed now may create impression that we are calling get_cpu_address_sizes() to figure out NX support. -boris
