Re: Re: xen: memory initialization/balloon fixes (#3)

[David Vrabel <david.vrabel@citrix.com>]

On 28/09/11 00:10, Konrad Rzeszutek Wilk wrote:
>>> (XEN) Xen-e820 RAM map:
>>> (XEN)  0000000000000000 - 000000000009d800 (usable)
>>
>> It's because it's not correctly handling the half-page of RAM at the end
>> of this region.
>>
>> I don't have access to any test boxes with a dodgy BIOS like this so can
>> you test this patch?  If it works I'll fold it in and post an updated
>> series.
> 
> It works. Albeit I think we are going to hit a problem with dmidecode
> if the DMI data is right in the reserved region
> 
> (http://lists.xensource.com/archives/html/xen-devel/2011-09/msg01299.html)
> 
> As in, if it starts in 9D800 - we consider 0->9d as RAM PFN, and 9e->100 as 1-1
> mapping.
> 
> I am thinking that perhaps the call to xen_set_phys_identity, where
> we call PFN_UP(x) should be replaced with PFN_DOWN(x). That way
> we would consider 0>9c as RAM PFN and 9D->100 as 1-1 mapping.

I almost did an equivalent change (see below) but discarded it as it
would have resulting in overlapping regions and attempting to
release/map some pages twice.

I think we will have to move the release/map until after the final e820
map has been sanitized so there are no overlapping regions.

I'll prepare another patch for this.

> That would imply a new patch to your series naturally.
>>
>> Can you remember why this page alignment was required?  I'd like to
> 
> The e820_* calls define how the memory subsystem will use it.
> It ended at some point assuming that the full page is RAM even thought
> it was only half-RAM and tried to use it and blew the machine up.
> 
> The fix was to make the calls to the e820_* with size and regions
> that were page-aligned.
> 
> Anyhow, here is what the bootup looks now:
> 
> [    0.000000] Freeing  9e-a0 pfn range: 2 pages freed
> [    0.000000] 1-1 mapping on 9e->a0
> [    0.000000] Freeing  a0-100 pfn range: 96 pages freed
> [    0.000000] 1-1 mapping on a0->100
> [    0.000000] Freeing  7fff0-80000 pfn range: 16 pages freed
> [    0.000000] 1-1 mapping on 7fff0->80000
> [    0.000000] Freeing  cfef0-cfef5 pfn range: 5 pages freed
> [    0.000000] 1-1 mapping on cfef0->cfef5
> [    0.000000] Freeing  cfef5-cff7f pfn range: 138 pages freed
> [    0.000000] 1-1 mapping on cfef5->cff7f
> [    0.000000] Freeing  cff7f-d0000 pfn range: 129 pages freed
> [    0.000000] 1-1 mapping on cff7f->d0000
> [    0.000000] Freeing  d0000-f0000 pfn range: 131072 pages freed
> [    0.000000] 1-1 mapping on d0000->f0000
> [    0.000000] Freeing  f0000-f4b58 pfn range: 19288 pages freed
> [    0.000000] 1-1 mapping on f0000->fec10
> [    0.000000] 1-1 mapping on fec10->fee01
> [    0.000000] 1-1 mapping on fee01->100000
> [    0.000000] Released 150746 pages of unused memory
> [    0.000000] Set 196994 page(s) to 1-1 mapping
> [    0.000000] BIOS-provided physical RAM map:
> [    0.000000]  Xen: 0000000000000000 - 000000000009d000 (usable)
> [    0.000000]  Xen: 000000000009d800 - 0000000000100000 (reserved)
> [    0.000000]  Xen: 0000000000100000 - 000000007fff0000 (usable)
> [    0.000000]  Xen: 000000007fff0000 - 0000000080000000 (reserved)
> 
> 
>> update the comment with the reason because the bare-metal x86 memory
>> init code doesn't appear to fixup the memory map in this way.
>>
>> diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
>> index 986661b..e473c4c 100644
>> --- a/arch/x86/xen/setup.c
>> +++ b/arch/x86/xen/setup.c
>> @@ -178,6 +178,19 @@ static unsigned long __init xen_get_max_pages(void)
>>  	return min(max_pages, MAX_DOMAIN_PAGES);
>>  }
>>
>> +static void xen_e820_add_region(u64 start, u64 size, int type)
>> +{
>> +	u64 end = start + size;
>> +
>> +	/* Align RAM regions to page boundaries. */
>> +	if (type == E820_RAM || type == E820_UNUSABLE) {
> 
> Hm, do we care about E820_UNUSABLE to be page aligned?
> If so, please comment why.

Er. We don't really but I think this if needs to be:

    /*
     * Page align regions.
     *
     * Reduce RAM regions and expand other (reserved) regions.
     */
    if (type == E820_RAM || type == E820_UNUSABLE) {
	start = PAGE_ALIGN(start);
        end  &= ~((u64)PAGE_SIZE - 1);
    } else {
        start &= ~((u64)PAGE_SIZE - 1);
        end = PAGE_ALIGN(start);
    }

So reserved regions also become page aligned (which is part of the fix
for the dmidecode bug).

>> +		start = PAGE_ALIGN(start);
> 
> Is that actually safe? Say it starts a 9ffff? We would
> end up using 9f000 which is not right.

PAGE_ALIGN() (and ALIGN()) round upwards.

David