Re: Linux 2.6.27-rc5: System boot regression caused by commit a2bd7274b47124d2fc4dfdb8c0591f545ba749dd

["Yinghai Lu" <yhlu.kernel@gmail.com>]

On Fri, Aug 29, 2008 at 7:33 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
>
> On Fri, 29 Aug 2008, Yinghai Lu wrote:
>>
>> the root cause is:
>> before 2.6.26, call init_apic_mapping and will insert_resource for
>> lapic address.
>> and then call e820_resource_resouce (with request_resource) to
>> register e820 entries.
>
> So the problem there was that traditionally, e820_reserve_resource()
> expected to be the first one to populate any resources. That's changed,
> and that's why it now needs to use "insert_resource()" rather than
> "request_resource()".
>
>> so the lapic entry in the resource tree will prevent some entry in
>> e820 to be registered.
>> later request_resource for BAR res (==hpet) will succeed.
>>
>> from 2.6.26. we move lapic address registering to late_initcall, so
>> the entry is reserved in e820 getting into resource tree at first.
>> and later pci_resource_survey::request_resource for BAR res (==hpet,
>> 0xfed00000) will fail. so pci_assign_unsigned... will get new
>> res for the BAR, so it messed up hpet setting.
>
> So this didn't work _originally_ either?

orginally it works, because lapic address entry open the big hole for
near address.

>
>> solutions will be
>> 1. use quirk to protect hpet in BAR, Ingo said it is not generic.
>
> Yeah, I don't like it. The quirk I was talking about was the one about
> apparetly bogus MMIO address:
>
>>> pci 0000:00:00.0: BAR has MMCONFIG at e0000000-ffffffff
>>        PCI: Using MMCONFIG at e0000000 - efffffff
>>
>> Where did it get that bogus "ffffffff" end address?
>
> which you said came from another BAR. That isn't the HPET.

that is from Rafael's system mmconfig BAR in 00:00.0. that chipset is broken ...

>
>
>> 2. or the one you are reverted... check_bar_with_valid. (hpet, ioapic,
>> mmconfig) --> happenly reveal another problem with Rafael's
>> system/chipset.
>
> Yeah, no, that's horrid. I'm happy it's reverted.

if update res->end according mmconfig end, before insert it forcibly,
then could fix the chipset BAR problem too.

>
>> 3. or sticky resource... , but could have particallly overlapping
>
> And no, this doesn't work.
>
>> 4. or don't register reserved entries in e820.. Eric, Nacked.
>
> Yeah, no, we do want reserved entries from e820 to show up to at least
> stop late _dynamic_ allocations from taking over.
>
>> 5. or you sugges, regiser some reserved entries later...., and have
>> insert_resource_expand_to_fit...
>
> Yes. And I do think this is a workable model.

BTW, insert_resource_expand_to_fit need to be replaced with
insert_resource_split_to_fit....
test stub reveal expand will make __request_region not working for
some devices...because reserved_entries from e820 take
IORESOUCE_BUSY...

YH