Niklāvs reported a boot regression on an Alderlake machine and bisected it
to commit 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier").
By moving the invocation of arch_cpu_finalize_init() further down he
identified that efi_enter_virtual_mode() is the function which causes
the boot hang.
The main difference of the earlier invocation is that the boot CPU is
already fully initialized and mitigations and alternatives are applied.
But the only really interesting change turned out to be IBT, which is
now enabled before efi_enter_virtual_mode(). "ibt=off" on the kernel
command line cured the problem.
Inspection of the involved calls in efi_enter_virtual_mode() unearthed that
efi_set_virtual_address_map() is the only place in the kernel which invokes
an EFI call without the IBT safe wrapper. This went obviously unnoticed so
far as IBT was enabled later.
Use arch_efi_call_virt() instead of efi_call() to cure that.
Fixes: fe379fa4d199 ("x86/ibt: Disable IBT around firmware")
Fixes: 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier")
Reported-by: Niklāvs Koļesņikovs <[email protected]>
Signed-off-by: Thomas Gleixner <[email protected]>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=217602
---
I put two fixes tags in as the IBT one missed that and the earlier
invocation unearthed it and made it observable.
---
arch/x86/platform/efi/efi_64.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
--- a/arch/x86/platform/efi/efi_64.c
+++ b/arch/x86/platform/efi/efi_64.c
@@ -853,9 +853,9 @@ efi_set_virtual_address_map(unsigned lon
/* Disable interrupts around EFI calls: */
local_irq_save(flags);
- status = efi_call(efi.runtime->set_virtual_address_map,
- memory_map_size, descriptor_size,
- descriptor_version, virtual_map);
+ status = arch_efi_call_virt(efi.runtime, set_virtual_address_map,
+ memory_map_size, descriptor_size,
+ descriptor_version, virtual_map);
local_irq_restore(flags);
efi_fpu_end();
On Thu, 29 Jun 2023 at 21:35, Thomas Gleixner <[email protected]> wrote:
>
> Niklāvs reported a boot regression on an Alderlake machine and bisected it
> to commit 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier").
>
> By moving the invocation of arch_cpu_finalize_init() further down he
> identified that efi_enter_virtual_mode() is the function which causes
> the boot hang.
>
> The main difference of the earlier invocation is that the boot CPU is
> already fully initialized and mitigations and alternatives are applied.
>
> But the only really interesting change turned out to be IBT, which is
> now enabled before efi_enter_virtual_mode(). "ibt=off" on the kernel
> command line cured the problem.
>
> Inspection of the involved calls in efi_enter_virtual_mode() unearthed that
> efi_set_virtual_address_map() is the only place in the kernel which invokes
> an EFI call without the IBT safe wrapper. This went obviously unnoticed so
> far as IBT was enabled later.
>
> Use arch_efi_call_virt() instead of efi_call() to cure that.
>
> Fixes: fe379fa4d199 ("x86/ibt: Disable IBT around firmware")
> Fixes: 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier")
> Reported-by: Niklāvs Koļesņikovs <[email protected]>
> Signed-off-by: Thomas Gleixner <[email protected]>
> Link: https://bugzilla.kernel.org/show_bug.cgi?id=217602
Reviewed-by: Ard Biesheuvel <[email protected]>
I take it you'll send this straight to Linus?
> ---
> I put two fixes tags in as the IBT one missed that and the earlier
> invocation unearthed it and made it observable.
> ---
> arch/x86/platform/efi/efi_64.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> --- a/arch/x86/platform/efi/efi_64.c
> +++ b/arch/x86/platform/efi/efi_64.c
> @@ -853,9 +853,9 @@ efi_set_virtual_address_map(unsigned lon
>
> /* Disable interrupts around EFI calls: */
> local_irq_save(flags);
> - status = efi_call(efi.runtime->set_virtual_address_map,
> - memory_map_size, descriptor_size,
> - descriptor_version, virtual_map);
> + status = arch_efi_call_virt(efi.runtime, set_virtual_address_map,
> + memory_map_size, descriptor_size,
> + descriptor_version, virtual_map);
> local_irq_restore(flags);
>
> efi_fpu_end();
On Thu, Jun 29 2023 at 23:42, Ard Biesheuvel wrote:
> On Thu, 29 Jun 2023 at 21:35, Thomas Gleixner <[email protected]> wrote:
>> Niklāvs reported a boot regression on an Alderlake machine and bisected it
>> to commit 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier").
>>
>> By moving the invocation of arch_cpu_finalize_init() further down he
>> identified that efi_enter_virtual_mode() is the function which causes
>> the boot hang.
>>
>> The main difference of the earlier invocation is that the boot CPU is
>> already fully initialized and mitigations and alternatives are applied.
>>
>> But the only really interesting change turned out to be IBT, which is
>> now enabled before efi_enter_virtual_mode(). "ibt=off" on the kernel
>> command line cured the problem.
>>
>> Inspection of the involved calls in efi_enter_virtual_mode() unearthed that
>> efi_set_virtual_address_map() is the only place in the kernel which invokes
>> an EFI call without the IBT safe wrapper. This went obviously unnoticed so
>> far as IBT was enabled later.
>>
>> Use arch_efi_call_virt() instead of efi_call() to cure that.
>>
>> Fixes: fe379fa4d199 ("x86/ibt: Disable IBT around firmware")
>> Fixes: 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier")
>> Reported-by: Niklāvs Koļesņikovs <[email protected]>
>> Signed-off-by: Thomas Gleixner <[email protected]>
>> Link: https://bugzilla.kernel.org/show_bug.cgi?id=217602
>
> Reviewed-by: Ard Biesheuvel <[email protected]>
>
> I take it you'll send this straight to Linus?
I can do that.
Thanks,
tglx
The following commit has been merged into the x86/urgent branch of tip:
Commit-ID: 0303c9729afc4094ef53e552b7b8cff7436028d6
Gitweb: https://git.kernel.org/tip/0303c9729afc4094ef53e552b7b8cff7436028d6
Author: Thomas Gleixner <[email protected]>
AuthorDate: Thu, 29 Jun 2023 21:35:19 +02:00
Committer: Thomas Gleixner <[email protected]>
CommitterDate: Fri, 30 Jun 2023 15:26:24 +02:00
x86/efi: Make efi_set_virtual_address_map IBT safe
Niklāvs reported a boot regression on an Alderlake machine and bisected it
to commit 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier").
By moving the invocation of arch_cpu_finalize_init() further down he
identified that efi_enter_virtual_mode() is the function which causes the
boot hang.
The main difference of the earlier invocation is that the boot CPU is
already fully initialized and mitigations and alternatives are applied.
But the only really interesting change turned out to be IBT, which is now
enabled before efi_enter_virtual_mode(). "ibt=off" on the kernel command
line cured the problem.
Inspection of the involved calls in efi_enter_virtual_mode() unearthed that
efi_set_virtual_address_map() is the only place in the kernel which invokes
an EFI call without the IBT safe wrapper. This went obviously unnoticed so
far as IBT was enabled later.
Use arch_efi_call_virt() instead of efi_call() to cure that.
Fixes: fe379fa4d199 ("x86/ibt: Disable IBT around firmware")
Fixes: 9df9d2f0471b ("init: Invoke arch_cpu_finalize_init() earlier")
Reported-by: Niklāvs Koļesņikovs <[email protected]>
Signed-off-by: Thomas Gleixner <[email protected]>
Reviewed-by: Ard Biesheuvel <[email protected]>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=217602
Link: https://lore.kernel.org/r/87jzvm12q0.ffs@tglx
---
arch/x86/platform/efi/efi_64.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/arch/x86/platform/efi/efi_64.c b/arch/x86/platform/efi/efi_64.c
index 232acf4..77f7ac3 100644
--- a/arch/x86/platform/efi/efi_64.c
+++ b/arch/x86/platform/efi/efi_64.c
@@ -853,9 +853,9 @@ efi_set_virtual_address_map(unsigned long memory_map_size,
/* Disable interrupts around EFI calls: */
local_irq_save(flags);
- status = efi_call(efi.runtime->set_virtual_address_map,
- memory_map_size, descriptor_size,
- descriptor_version, virtual_map);
+ status = arch_efi_call_virt(efi.runtime, set_virtual_address_map,
+ memory_map_size, descriptor_size,
+ descriptor_version, virtual_map);
local_irq_restore(flags);
efi_fpu_end();