Messages in this thread Patch in this message | ![/](/images/icornerl.gif) | | Subject | Re: [GIT PULL] Small Xen bugfixes | From | Ian Campbell <> | Date | Mon, 1 Nov 2010 16:36:28 +0000 |
| |
On Sun, 2010-10-31 at 16:28 +0000, Jeremy Fitzhardinge wrote: > On 10/31/2010 02:13 AM, Ian Campbell wrote: > >> The 3rd is certainly simplest, at the cost of wasting a trivial amount > >> of memory. > > Doesn't Linux avoid using the lowest 1M anyway? (obviously apart from > > the start of day probing for firmware tables etc). > > No, it tries to use most of it I think. It will tend to avoid the low > 64k (maybe more) to avoid BIOS bugs.
It'll be interesting to see what effect Vista's avoidance of the whole region (so I hear) has on BIOS vendors... (I think we can all guess)
> >> Unfortunately it crashes early. Sigh, will try and sort it > >> out this afternoon. > > Strange! > > I didn't get a chance to poke at it again, but in retrospect, I think > there are various "must succeed" allocations in low memory. We don't > need those allocations (things like AP boot trampoline, etc), but we > don't bother to stub them out or prevent them from happening. Reducing > the system to one with *no* allocatable memory below 1M is just too > strange, and would be a continuous source of problems in the future.
Agreed, we should try and mimic native as far as possible in this regard or I fear we will see a never ending stream of little quirks and oddities related to this sort of thing.
> Of the other two options, I think your original approach is going to be > simplest. E820 gap filling wouldn't be too bad, but we'd end up having > to add a bit of gap-tracking logic to the E820 loop which isn't > currently there.
It would also make us susceptible to perhaps being a bit fragile in the face of unexpectedly insane e820s coming from the BIOS.
> Ignoring sub-1M gaps is simpler (and it needn't be > conditional on xen_initial_domain(), because we would never expect to > see anything strange sub-1M in a domU, and if there is, we should still > be careful of it in case something odd is going on).
Absolutely.
I wonder if we shouldn't also do the following (note: untested). Since Xen avoids using the sub-1M region for anything I think it is reasonable to give the whole lot over to domain 0 for the purposes of finding firmware table stashed in odd locations etc.
Ian.
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c index ebb74ec..ab086e5 100644 --- a/arch/x86/xen/mmu.c +++ b/arch/x86/xen/mmu.c @@ -2321,7 +2321,7 @@ __init void xen_ident_map_ISA(void) xen_raw_printk("Xen: setup ISA identity maps\n"); - for (pa = ISA_START_ADDRESS; pa < ISA_END_ADDRESS; pa += PAGE_SIZE) { + for (pa = 0; pa < ISA_END_ADDRESS; pa += PAGE_SIZE) { pte_t pte = mfn_pte(PFN_DOWN(pa), PAGE_KERNEL_IO); if (HYPERVISOR_update_va_mapping(PAGE_OFFSET + pa, pte, 0))
| ![\](/images/icornerr.gif) |