[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH RFC 0/N] xen: arm: rework early bring up



On Wed, 2013-09-18 at 21:00 +0100, Julien Grall wrote:
> On 09/18/2013 08:24 PM, Ian Campbell wrote:
> > On Wed, 2013-09-18 at 20:22 +0100, Ian Campbell wrote:
> >> On Wed, 2013-09-18 at 18:33 +0100, Julien Grall wrote:
> >>> On 09/17/2013 02:37 AM, Ian Campbell wrote:
> >>>> Hi,
> >>>>
> >>>> The following reworks early bring up on ARM to use a separate set of
> >>>> boot pagetables. This simplifies things by avoiding the need to bring up
> >>>> all CPUs in lock step, which in turn allows us to do secondary CPU
> >>>> bringup in C code.
> >>>>
> >>>> Unfortunately the main bulk of this change is a single large patch which
> >>>> is hard to decompose any further since it is basically pulling on the
> >>>> thread and then knitting a new jumper from it.
> >>>>
> >>>> With these changes Xen now absolutely requires that the bootloader calls
> >>>> the hypervisor in HYP mode, the previous workarounds have been removed.
> >>>> For use on models a bootwrapper is now required. See
> >>>>          git://xenbits.xen.org/people/ianc/boot-wrapper.git xen-arm32
> >>>>          git://xenbits.xen.org/people/ianc/boot-wrapper-aarch64.git 
> >>>> xen-arm64
> >>>>
> >>>> I have implemented support for CPU bringup on the fastmodel vexpress
> >>>> platforms (v7and v8) here, I suppose it should work OK on a real
> >>>> vexpress too but I've not tried it.
> >>>
> >>> I have just tried the patch series on the versatile express and it
> >>> doesn't work. I'm using the u-boot from Andre and all the cpus already
> >>> boot in hyp mode (checked without this patch series).
> >>
> >> :-(
> >>
> >> The code should essentially the same as the old kick cpus, but
> >> reimplemented in C. The only difference I can spot is that the C version
> >> is kicking a single CPU at a time while the ASM version is kicking
> >> everyone at once. Does using send_SGI_allbutself instead work as a hack?
> >>
> >> Perhaps our idea of the CPUID is wrong or something? Your patches to
> >> fixup the logical CPU id stuff might help?
> 
> I have found the issue, in send_SGI_mask we remove the offline CPUs with 
> the following line:
>    mask &= cpumask_bits(&cpu_online_map);
> 
> As the secondary cpu (assuming we have a board with only 2 cores) is not 
> yet online, we can't send SGI with our code.
> I'm very surprised that the code works on the fast models.

So am I, now.

OK. So online mask == 0x1 and we are trying to wake mask 0x2, so the
actual target list we end up using is 0... 

The gic manual says "When TargetListFilter is 0b00, if the CPUTargetList
field is 0x00 the Distributor does not forward the interrupt to any CPU
interface." (footnote to table 4-21), so it shouldn't be waking
anything... It may be a model bug

> To solve the issue, I think we need to create another helper which will 
> send an SGI wihout removing offline CPUs.

Looking briefly at the callchains leading to send_SGI_mask I don't see
any which deliberately pass an offline CPU (which would be pretty
non-sensical apart from this one case). So we could perhaps just remove
this check and make it the caller's problem.

Or maybe the right check is for cpu_possible mask or something similar
instead of online?

Ian.


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


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.