Re: [PATCH] ACPI: x86: Call acpi_boot_table_init() after acpi_table_upgrade()

[Rafael J. Wysocki]

On Wed, Apr 14, 2021 at 10:13 AM Mike Rapoport <rppt@xxxxxxxxxxxxx> wrote:
>
> On Wed, Apr 14, 2021 at 09:42:01AM +0200, David Hildenbrand wrote:
> > On 13.04.21 19:53, Rafael J. Wysocki wrote:
> > > On Tue, Apr 13, 2021 at 7:43 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > > >
> > > > On 13.04.21 16:01, Rafael J. Wysocki wrote:
> > > > > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > > > >
> > > > > Commit 1a1c130ab757 ("ACPI: tables: x86: Reserve memory occupied by
> > > > > ACPI tables") attempted to address an issue with reserving the memory
> > > > > occupied by ACPI tables, but it broke the initrd-based table override
> > > > > mechanism relied on by multiple users.
> > > > >
> > > > > To restore the initrd-based ACPI table override functionality, move
> > > > > the acpi_boot_table_init() invocation in setup_arch() on x86 after
> > > > > the acpi_table_upgrade() one.
> > > > >
> > > > > Fixes: 1a1c130ab757 ("ACPI: tables: x86: Reserve memory occupied by ACPI tables")
> > > > > Reported-by: Hans de Goede <hdegoede@xxxxxxxxxx>
> > > > > Tested-by: Hans de Goede <hdegoede@xxxxxxxxxx>
> > > > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > > > > ---
> > > > >
> > > > > George, can you please check if this reintroduces the issue addressed by
> > > > > the above commit for you?
> > > > >
> > > > > ---
> > > > >    arch/x86/kernel/setup.c |    5 ++---
> > > > >    1 file changed, 2 insertions(+), 3 deletions(-)
> > > > >
> > > > > Index: linux-pm/arch/x86/kernel/setup.c
> > > > > ===================================================================
> > > > > --- linux-pm.orig/arch/x86/kernel/setup.c
> > > > > +++ linux-pm/arch/x86/kernel/setup.c
> > > > > @@ -1045,9 +1045,6 @@ void __init setup_arch(char **cmdline_p)
> > > > >
> > > > >        cleanup_highmap();
> > > > >
> > > > > -     /* Look for ACPI tables and reserve memory occupied by them. */
> > > > > -     acpi_boot_table_init();
> > > > > -
> > > > >        memblock_set_current_limit(ISA_END_ADDRESS);
> > > > >        e820__memblock_setup();
> > > > >
> > > > > @@ -1132,6 +1129,8 @@ void __init setup_arch(char **cmdline_p)
> > > > >        reserve_initrd();
> > > > >
> > > > >        acpi_table_upgrade();
> > > > > +     /* Look for ACPI tables and reserve memory occupied by them. */
> > > > > +     acpi_boot_table_init();
> > > > >
> > > > >        vsmp_init();
> > > >
> > > > This is fairly late; especially, it's after actual allocations -- see
> > > > e820__memblock_alloc_reserved_mpc_new().
> > > >
> > > > Can't the table upgrade mechanism fix up when adjusting something?
> > >
> > > Not at this point of the cycle I'm afraid.
> > >
> > > > Some details on what actually breaks would be helpful.
> > >
> > > Generally speaking, the table overrides that come from the initrd are
> > > not taken into account if acpi_boot_table_init() runs before
> > > acpi_table_upgrade() and the latter cannot run before
> > > reserve_initrd().
> >
> > I see. (looking at Documentation/acpi/initrd_table_override.txt I understand
> > what acpi table overrides are for :) )
> >
> > >
> > > Honestly, I'm not sure how much effort it would take to untangle this ATM.
> >
> > Also true; ideally, we wouldn't have any allocations (find+reserve) before
> > ordinary reservations are done.
> >
> > However, I have no idea if we can move
> > e820__memblock_alloc_reserved_mpc_new() and reserve_real_mode() around
> > easily. Also, reserve_initrd()->relocate_initrd() does allocations.
>
> Even if we can move e820__memblock_alloc_reserved_mpc_new() and
> reserve_real_mode(), the allocation in reserve_initrd() has to be before
> the tables override, we would only reduce the probability of allocating an
> ACPI page.
>
> I think what we can do is to override the ACPI tables separately from their
> initial parsing. Rafael, what do you say?

Well, possibly.  With one caveat that parsing a table that's going to
be overridden subsequently may not be a good idea.

Anyway, things like that can only be done in the next cycle or later.