On 12/03/2025 9:57 am, Roger Pau Monné wrote:
> On Wed, Mar 12, 2025 at 09:31:37AM +0100, Jan Beulich wrote:
>> On 11.03.2025 21:47, Andrew Cooper wrote:
>>> On 06/01/2025 11:54 am, Jan Beulich wrote:
>>>> On 06.01.2025 12:26, Andrew Cooper wrote:
>>>>> Regular data access into the trampoline is via the directmap.
>>>>>
>>>>> As now discussed quite extensively in asm/trampoline.h, the trampoline is
>>>>> arranged so that only the AP and S3 paths need an identity mapping, and
>>>>> that
>>>>> they fit within a single page.
>>>>>
>>>>> Right now, PFN_UP(trampoline_end - trampoline_start) is 2, causing more
>>>>> than
>>>>> expected of the trampoline to be mapped. Cut it down just the single
>>>>> page it
>>>>> ought to be.
>>>>>
>>>>> Signed-off-by: Andrew Cooper <andrew.coop...@citrix.com>
>>>> Acked-by: Jan Beulich <jbeul...@suse.com>
>>> Thanks. However,
>>>
>>>> on the basis that this improves things. However, ...
>>>>
>>>>> --- a/xen/arch/x86/x86_64/mm.c
>>>>> +++ b/xen/arch/x86/x86_64/mm.c
>>>>> @@ -718,14 +718,16 @@ void __init zap_low_mappings(void)
>>>>> {
>>>>> BUG_ON(num_online_cpus() != 1);
>>>>>
>>>>> - /* Remove aliased mapping of first 1:1 PML4 entry. */
>>>>> + /* Stop using l?_bootmap[] mappings. */
>>>>> l4e_write(&idle_pg_table[0], l4e_empty());
>>>>> flush_local(FLUSH_TLB_GLOBAL);
>>>>>
>>>>> - /* Replace with mapping of the boot trampoline only. */
>>>>> + /*
>>>>> + * Insert an identity mapping of the AP/S3 part of the trampoline,
>>>>> which
>>>>> + * is arranged to fit in a single page.
>>>>> + */
>>>>> map_pages_to_xen(trampoline_phys, maddr_to_mfn(trampoline_phys),
>>>>> - PFN_UP(trampoline_end - trampoline_start),
>>>>> - __PAGE_HYPERVISOR_RX);
>>>>> + 1, __PAGE_HYPERVISOR_RX);
>>>> ... literal numbers like this - however well they are commented - are
>>>> potentially problematic to locate in case something changes significantly.
>>>> The 1 here really would want connecting with the .equ establishing
>>>> wakeup_stack.
>>> how do you propose doing this?
>>>
>>> PFN_UP(wakeup_stack - trampoline_start) doesn't have the same obvious
>>> connection, and it would involve partly undoing 7d73c6f196a5 which hid
>>> the symbol recently.
>>>
>>> While 1 isn't ideal, it is next to a comment explaining what's going on,
>>> and it's not going to go stale in a silent way... (It's also not liable
>>> to go stale either.)
>> If in
>>
>> .equ wakeup_stack, trampoline_start + PAGE_SIZE
>>
>> PAGE_SIZE was replaced by a new (in asm/trampoline.h) TRAMPOLINE_PERM_SIZE,
>> you could use PFN_UP(TRAMPOLINE_PERM_SIZE) here to establish a connection.
>>
>> I have to admit I also don't really see why things going stale here would
>> (a) be unlikely and (b) be guaranteed to not go silently.
The size can't go to 0 or everything will break, and if it goes larger
than 1 (which it almost certainly never will), then APs and/or S3 will
break, and we've got both of these in CI.
Furthermore, the actual thing which matters is:
> /* Map the permanent trampoline page into l1_bootmap[]. */
> mov sym_esi(trampoline_phys), %ecx
> lea __PAGE_HYPERVISOR_RX(%ecx), %edx /* %edx = PTE to write */
> shr $PAGE_SHIFT, %ecx /* %ecx = Slot to write */
> mov %edx, sym_offs(l1_bootmap)(%esi, %ecx, 8)
which hardcodes 1 page, because there's almost certainly no chance this
will ever change.
>> We just don't
>> know what we may need to add to the trampoline, sooner or later.
> Maybe we could introduce trampoline_{ap?,runtime?}_{start,end} that
> covers this more narrow section of the trampoline code?
We already have one of those, and a linker assertion that it stays below
1k, so wakeup_stack is at least 3k.
The complexity is that the wakeup_stack overlays some init-only logic in
the placed trampoline. It's not something that exists concretely in the
Xen image.
~Andrew
