arm: EFI: Disallow EFI bootmgr when providing PSCI
RFC PATCH 1/1 arm: EFI: Disallow EFI bootmgr when providing PSCI
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Commit f3866909e350 ("distro_bootcmd: call EFI bootmgr even without having /EFI/boot") changed the order of EFI probing, so the EFI bootmgr is now *always* run, resulting in the default distro boot commands now *always* failing, even in the total absence of any UEFI directories or boot files.
So use the newly added build option to disable the EFI bootmgr, which makes those boards boot again using the distro boot commands. Explicitly calling "bootefi bootmgr" still breaks the boot, though.
---
Hi,
the above is the result of my analysis, happy to stand corrected in case I missed something. I know that this is not a proper solution, but it's an effective stop-gap measure to fix all those boards. It looks like a proper solution would either be:
- Let the EFI bootmgr run in the current security state.
- Install the PSCI handlers early in U-Boot.
Both solutions sound rather involved, so probably require more time. But we need to fix this breakage now.
Cheers,
Andre
lib/efi_loader/Kconfig | 1 + 1 file changed, 1 insertion(+)
lib/efi_loader/Kconfig +1 -0
@@ -30,6 +30,7 @@ if EFI_LOADER config CMD_BOOTEFI_BOOTMGR bool "UEFI Boot Manager" default y + depends on !ARMV7_PSCI help Select this option if you want to select the UEFI binary to be booted via UEFI variables Boot####, BootOrder, and BootNext. This enables the
Signature · 1 line
2.17.5
RFC PATCH 1/1 arm: EFI: Disallow EFI bootmgr when providing PSCI
Quoted 30 lines · Andre Przywara
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Commit
f3866909e350("distro_bootcmd: call EFI bootmgr even without having /EFI/boot") changed the order of EFI probing, so the EFI bootmgr is now *always* run, resulting in the default distro boot commands now *always* failing, even in the total absence of any UEFI directories or boot files.So use the newly added build option to disable the EFI bootmgr, which makes those boards boot again using the distro boot commands. Explicitly calling "bootefi bootmgr" still breaks the boot, though.
Signed-off-by Andre Przywara <[email protected]>Reported-by Jernej Skrabec <[email protected]>---
Hi,the above is the result of my analysis, happy to stand corrected in case I missed something. I know that this is not a proper solution, but it's an effective stop-gap measure to fix all those boards. It looks like a proper solution would either be:
- Let the EFI bootmgr run in the current security state.
- Install the PSCI handlers early in U-Boot.
Quoted 8 lines
Both solutions sound rather involved, so probably require more time. But we need to fix this breakage now.
Cheers,
Andrelib/efi_loader/Kconfig | 1 + 1 file changed, 1 insertion(+)
RFC PATCH 1/1 arm: EFI: Disallow EFI bootmgr when providing PSCI
Quoted 8 lines · Simon Glass
Quoted 7 lines · Andre Przywara
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
thank you for reporting the issue.
I have an Orange Pi PC with a 32 bit Allwinner CPU. orangepi_pc_defconfig has CONFIG_ARMV7_PSCI=y.
I use origin/master (e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.
But I observe that it takes around 60 seconds between SetVirtualAddressMap() and the first kernel log output.
EFI stub: Exiting boot services and installing virtual address map...
EHCI failed to shut down host controller.
<<< 60 seconds waiting without output >>>>
[ 0.000000] Booting Linux on physical CPU 0x0
I have seen this regression since some time last year.
Reverting patch f3866909e350 does not solve the problem. Reverting to U-Boot v2020.01 does not solve the problem.
Reverting the kernel from v5.9 to 5.4 solves the problem both for U-Boot v2020.01 as well as for U-Boot v2021.01.
I have poked around with some pre-built kernels from
http://snapshot.debian.org/package/linux:
Linux 5.9.11 - 1 minute delay
Linux 5.8.14 - 1 minute delay
Linux 5.7.17 - no delay
Linux 5.6.14 - no delay
Linux 5.5.17 - no delay
Linux 5.4.19 - no delay
It seems that some change in Linux is causing the regression. Could you, please, try to analyze it in more depth.
Best regards
Heinrich
Quoted 34 lines
Quoted 23 lines
Commit
f3866909e350("distro_bootcmd: call EFI bootmgr even without having /EFI/boot") changed the order of EFI probing, so the EFI bootmgr is now *always* run, resulting in the default distro boot commands now *always* failing, even in the total absence of any UEFI directories or boot files.So use the newly added build option to disable the EFI bootmgr, which makes those boards boot again using the distro boot commands. Explicitly calling "bootefi bootmgr" still breaks the boot, though.
Signed-off-by Andre Przywara <[email protected]>Reported-by Jernej Skrabec <[email protected]>---
Hi,the above is the result of my analysis, happy to stand corrected in case I missed something. I know that this is not a proper solution, but it's an effective stop-gap measure to fix all those boards. It looks like a proper solution would either be:
- Let the EFI bootmgr run in the current security state.
- Install the PSCI handlers early in U-Boot.Quoted 8 lines
Both solutions sound rather involved, so probably require more time. But we need to fix this breakage now.
Cheers,
Andrelib/efi_loader/Kconfig | 1 + 1 file changed, 1 insertion(+)Reviewed-by Simon Glass <[email protected]>
RFC PATCH 1/1 arm: EFI: Disallow EFI bootmgr when providing PSCI
Hi!
Dne nedelja, 24. januar 2021 ob 09:27:02 CET je Heinrich Schuchardt napisal(a):
Quoted 2 lines · Simon Glass
Quoted 1 line
On Fri, 22 Jan 2021 at 05:05, Andre Przywara <[email protected]>
Quoted 9 lines · wrote:
Quoted 6 lines
Quoted 6 lines
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
Quoted 7 lines
thank you for reporting the issue.
I have an Orange Pi PC with a 32 bit Allwinner CPU.
orangepi_pc_defconfighasCONFIG_ARMV7_PSCI=y.I use origin/master (
e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.
I observed issue on OrangePi Plus2E which has eMMC. At that time, there was Android on it. If I disabled eMMC support in U-Boot, boot went just fine.
Quoted 3 lines
But I observe that it takes around 60 seconds between
SetVirtualAddressMap()and the first kernel log output.
I guess this is another issue. It never booted in my case. It stopped right after "Starting kernel...".
Best regards,
Jernej
Quoted 65 lines
EFI stub: Exiting boot services and installing virtual address map...
EHCI failed to shut down host controller.
<<< 60 seconds waiting without output >>>>[ 0.000000] Booting Linux on physical CPU 0x0I have seen this regression since some time last year.
Reverting patch
f3866909e350does not solve the problem. Reverting to U-Boot v2020.01 does not solve the problem.Reverting the kernel from v5.9 to 5.4 solves the problem both for U-Boot v2020.01 as well as for U-Boot v2021.01.
I have poked around with some pre-built kernels from
http://snapshot.debian.org/package/linux:Linux 5.9.11 - 1 minute delay
Linux 5.8.14 - 1 minute delay
Linux 5.7.17 - no delay
Linux 5.6.14 - no delay
Linux 5.5.17 - no delay
Linux 5.4.19 - no delayIt seems that some change in Linux is causing the regression. Could you, please, try to analyze it in more depth.
Best regards
Heinrich
Quoted 32 lines
Quoted 22 lines
Commit
f3866909e350("distro_bootcmd: call EFI bootmgr even without having /EFI/boot") changed the order of EFI probing, so the EFI bootmgr is now *always* run, resulting in the default distro boot commands now *always* failing, even in the total absence of any UEFI directories or boot files.So use the newly added build option to disable the EFI bootmgr, which makes those boards boot again using the distro boot commands. Explicitly calling "bootefi bootmgr" still breaks the boot, though.
Signed-off-by Andre Przywara <[email protected]>Reported-by Jernej Skrabec <[email protected]>---
Hi,the above is the result of my analysis, happy to stand corrected in case I missed something. I know that this is not a proper solution, but it's an effective stop-gap measure to fix all those boards. It looks like a proper solution would either be:
- Let the EFI bootmgr run in the current security state.
- Install the PSCI handlers early in U-Boot.Quoted 8 lines
Both solutions sound rather involved, so probably require more time. But we need to fix this breakage now.
Cheers,
Andrelib/efi_loader/Kconfig | 1 + 1 file changed, 1 insertion(+)Reviewed-by Simon Glass <[email protected]>
Dne nedelja, 24. januar 2021 ob 09:33:15 CET je Jernej ?krabec napisal(a):
Quoted 25 lines
Hi!
Dne nedelja, 24. januar 2021 ob 09:27:02 CET je Heinrich Schuchardt
napisal(a):
Quoted 2 lines · Simon Glass
Quoted 1 line
On Fri, 22 Jan 2021 at 05:05, Andre Przywara <[email protected]>
Quoted 16 lines · wrote:
Quoted 6 lines
Quoted 6 lines
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
thank you for reporting the issue.
I have an Orange Pi PC with a 32 bit Allwinner CPU.
orangepi_pc_defconfighasCONFIG_ARMV7_PSCI=y.I use origin/master (
e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.
Just one clarification - issue here is that "bootefi bootmgr" command when unsuccessful breaks booting with bootm command.
Quoted 78 lines
I observed issue on OrangePi Plus2E which has eMMC. At that time, there was Android on it. If I disabled eMMC support in U-Boot, boot went just fine.
Quoted 2 lines
But I observe that it takes around 60 seconds between
SetVirtualAddressMap()and the first kernel log output.I guess this is another issue. It never booted in my case. It stopped right after "Starting kernel...".
Best regards,
JernejQuoted 65 lines
EFI stub: Exiting boot services and installing virtual address map...
EHCI failed to shut down host controller.
<<< 60 seconds waiting without output >>>>[ 0.000000] Booting Linux on physical CPU 0x0I have seen this regression since some time last year.
Reverting patch
f3866909e350does not solve the problem. Reverting to U-Boot v2020.01 does not solve the problem.Reverting the kernel from v5.9 to 5.4 solves the problem both for U-Boot v2020.01 as well as for U-Boot v2021.01.
I have poked around with some pre-built kernels from
http://snapshot.debian.org/package/linux:Linux 5.9.11 - 1 minute delay
Linux 5.8.14 - 1 minute delay
Linux 5.7.17 - no delay
Linux 5.6.14 - no delay
Linux 5.5.17 - no delay
Linux 5.4.19 - no delayIt seems that some change in Linux is causing the regression. Could you, please, try to analyze it in more depth.
Best regards
Heinrich
Quoted 33 lines
Quoted 23 lines
Commit
f3866909e350("distro_bootcmd: call EFI bootmgr even without having /EFI/boot") changed the order of EFI probing, so the EFI bootmgr is now *always* run, resulting in the default distro boot commands now *always* failing, even in the total absence of any UEFI directories or boot files.So use the newly added build option to disable the EFI bootmgr, which makes those boards boot again using the distro boot commands. Explicitly calling "bootefi bootmgr" still breaks the boot, though.
Signed-off-by Andre Przywara <[email protected]>Reported-by Jernej Skrabec <[email protected]>---
Hi,the above is the result of my analysis, happy to stand corrected in case I missed something. I know that this is not a proper solution, but it's an effective stop-gap measure to fix all those boards. It looks
like a proper solution would either be:
- Let the EFI bootmgr run in the current security state.
- Install the PSCI handlers early in U-Boot.Quoted 8 lines
Both solutions sound rather involved, so probably require more time. But we need to fix this breakage now.
Cheers,
Andrelib/efi_loader/Kconfig | 1 + 1 file changed, 1 insertion(+)Reviewed-by Simon Glass <[email protected]>
Quoted 21 lines · Jernej ?krabec
Dne nedelja, 24. januar 2021 ob 09:33:15 CET je Jernej ?krabec napisal(a):
Quoted 20 lines
Hi!
Dne nedelja, 24. januar 2021 ob 09:27:02 CET je Heinrich Schuchardt
napisal(a):
Quoted 2 lines · Simon Glass
Quoted 1 line
On Fri, 22 Jan 2021 at 05:05, Andre Przywara <[email protected]>
Quoted 11 lines · wrote:
Quoted 6 lines
Quoted 6 lines
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
thank you for reporting the issue.
Quoted 8 lines
Quoted 5 lines
Quoted 5 lines
I have an Orange Pi PC with a 32 bit Allwinner CPU.
orangepi_pc_defconfighasCONFIG_ARMV7_PSCI=y.I use origin/master (
e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.Just one clarification - issue here is that "bootefi bootmgr" command when unsuccessful breaks booting with bootm command.
If I press the enter key to get into the console circumventing distro-boot, booting via bootz works. If I wait until distro_boot is finished, booting via bootz fails. Both with Linux 5.7.17.
This confirms your finding that there is a problem with the initialization of the UEFI sub-system.
lib/efi_loader/efi_setup.c:192 is the only place where we call switch_to_non_secure_mode().
With the line removed:
* Booting via bootz is successful.
* The EFI stub shows: Entering in SVC mode with MMU enabled
* Booting via bootefi fails
switch_to_non_secure_mode() is safe to be called repeatedly. So we could move the switch_to_non_secure_mode() call to do_bootefi_exec(). This is after the boot manager has searched for a bootable image.
With the change (see diff below):
* Booting via bootz is successful.
* EFI stub shows: Entering in HYP mode with MMU enabled
* Booting via bootefi is successful.
The downside of the change is that executing bootz will still fail after a UEFI binary returns to U-Boot.
Running a shell in secure mode seems unwise. So we should dig a bit deeper:
Where in the code is the PSCI handler installed and where occurs the switch to non-secure state when booting via bootz/bootm? Can we move this to before distro-boot?
Best regards
Heinrich
cmd/bootefi.c +1 -0
@@ -8,6 +8,7 @@#define LOG_CATEGORY LOGC_EFI#include <common.h> +#include <bootm.h> #include <charset.h> #include <command.h> #include <dm.h>
@@ -338,6 +339,9 @@ static efi_status_t do_bootefi_exec(efi_handle_t
handle, void *load_options)
efi_uintn_t exit_data_size = 0;
u16 *exit_data = NULL;+ /* On ARM switch from EL3 or secure mode to EL2 or non-secure
mode */
+ switch_to_non_secure_mode();
+
/* Call our payload! */
ret = EFI_CALL(efi_start_image(handle, &exit_data_size,
&exit_data));
if (ret != EFI_SUCCESS) {lib/efi_loader/efi_setup.c +0 -1
@@ -188,9 +188,6 @@ efi_status_t efi_init_obj_list(void)/* Allow unaligned memory access */ allow_unaligned (); -/* On ARM switch from EL3 or secure mode to EL2 or non-secure
mode */
- switch_to_non_secure_mode();
-
/* Initialize root node */
ret = efi_root_node_register();
if (ret != EFI_SUCCESS)On Sun, 24 Jan 2021 11:44:35 +0100
Heinrich Schuchardt <[email protected]> wrote:
Hi Heinrich,
many thanks for digging into this!
Quoted 28 lines · Jernej ?krabec
Quoted 27 lines
Dne nedelja, 24. januar 2021 ob 09:33:15 CET je Jernej ?krabec napisal(a):
Quoted 25 lines
Hi!
Dne nedelja, 24. januar 2021 ob 09:27:02 CET je Heinrich Schuchardt
napisal(a):
Quoted 2 lines · Simon Glass
Quoted 1 line
On Fri, 22 Jan 2021 at 05:05, Andre Przywara <[email protected]>
Quoted 16 lines · wrote:
Quoted 6 lines
Quoted 6 lines
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
thank you for reporting the issue.
I have an Orange Pi PC with a 32 bit Allwinner CPU.
orangepi_pc_defconfighasCONFIG_ARMV7_PSCI=y.I use origin/master (
e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.
Quoted 6 lines
Quoted 2 lines
Just one clarification - issue here is that "bootefi bootmgr" command when unsuccessful breaks booting with bootm command.
If I press the enter key to get into the console circumventing distro-boot, booting via bootz works. If I wait until
distro_bootis finished, booting via bootz fails. Both with Linux 5.7.17.
Yes, that was exactly Jernej's and my observation.
Quoted 11 lines
This confirms your finding that there is a problem with the initialization of the UEFI sub-system.
lib/efi_loader/efi_setup.c:192is the only place where we callswitch_to_non_secure_mode().With the line removed:
* Booting via bootz is successful.
* The EFI stub shows: Entering in SVC mode with MMU enabled
* Booting via bootefi fails
Ah, thanks for giving this a try. I know next to nothing about U-Boot's UEFI internals, so didn't dare to touch this code.
Quoted 9 lines
switch_to_non_secure_mode()is safe to be called repeatedly. So we could move theswitch_to_non_secure_mode()call todo_bootefi_exec(). This is after the boot manager has searched for a bootable image.With the change (see diff below):
* Booting via bootz is successful.
* EFI stub shows: Entering in HYP mode with MMU enabled
* Booting via bootefi is successful.
Oh nice! I wasn't sure how much the UEFI initialisation actually relies on non-secure mode, and if switching to non-secure *after* the initialisation would change things.
Quoted 2 lines
The downside of the change is that executing bootz will still fail after a UEFI binary returns to U-Boot.
OK, that is not nice, but not really a big issue. At least it's a lot better than disabling bootmgr altogether or no bootz after a bootmgr call.
If you deem this change being not very intrusive, I would very much prefer this over my patch here.
Quoted 5 lines
Running a shell in secure mode seems unwise. So we should dig a bit deeper:
Where in the code is the PSCI handler installed and where occurs the switch to non-secure state when booting via bootz/bootm? Can we move this to before distro-boot?
Well, yes, I think installing the PSCI handlers (and doing everything that requires secure state) early is the best way forward, but this is quite some change, and I would rather plug this problem now.
At the moment all of the v7 PSCI code is run as late as possible, so U-Boot can run in secure state. I agree this *sounds* scary, but running in secure is actually quite common for many ARM32 machines (even Linux sometimes runs with the NS bit cleared).
For v7 Allwinner SoCs specifically we need access to the secure-only SID registers for the MAC address generation, also secure SRAM becomes inaccessible in non-secure world (as expected, but in contrast to the ARMv8 chips).
So I think eventually we will need to bite the bullet and teach U-Boot to cope with non-secure in sunxi-v7, but this needs some time and requires possibly intrusive changes.
Cheers,
Andre
Quoted 41 lines
cmd/bootefi.c +2
-0@@ -8,6 +8,7 @@#define LOG_CATEGORY LOGC_EFI#include <common.h> +#include <bootm.h> #include <charset.h> #include <command.h> #include <dm.h> @@ -338,6 +339,9 @@ static efi_status_t do_bootefi_exec(efi_handle_t handle, void *load_options)efi_uintn_t exit_data_size =0 ;u16 *exit_data =NULL ; +/* On ARM switch from EL3 or secure mode to EL2 or non-secure mode */
cmd/bootefi.c +2
-0+switch_to_non_secure_mode (); +/* Call our payload! */ ret = EFI_CALL(efi_start_image(handle, &exit_data_size,&exit_data)); if (ret != EFI_SUCCESS) {lib/efi_loader/efi_setup.c +0
-1@@ -188,9 +188,6 @@ efi_status_t efi_init_obj_list(void)/* Allow unaligned memory access */ allow_unaligned (); -/* On ARM switch from EL3 or secure mode to EL2 or non-secure mode */
lib/efi_loader/efi_setup.c +0
-2-switch_to_non_secure_mode (); -/* Initialize root node */ ret =efi_root_node_register ();if (ret !=EFI_SUCCESS )
Quoted 73 lines · Andre Przywara
On Sun, 24 Jan 2021 11:44:35 +0100
Heinrich Schuchardt <[email protected]> wrote:Hi Heinrich,
many thanks for digging into this!
Quoted 28 lines · Jernej ?krabec
Quoted 27 lines
Dne nedelja, 24. januar 2021 ob 09:33:15 CET je Jernej ?krabec napisal(a):
Quoted 25 lines
Hi!
Dne nedelja, 24. januar 2021 ob 09:27:02 CET je Heinrich Schuchardt
napisal(a):
Quoted 2 lines · Simon Glass
Quoted 1 line
On Fri, 22 Jan 2021 at 05:05, Andre Przywara <[email protected]>
Quoted 16 lines · wrote:
Quoted 6 lines
Quoted 6 lines
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
thank you for reporting the issue.
I have an Orange Pi PC with a 32 bit Allwinner CPU.
orangepi_pc_defconfighasCONFIG_ARMV7_PSCI=y.I use origin/master (
e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.Quoted 6 lines
Quoted 2 lines
Just one clarification - issue here is that "bootefi bootmgr" command when unsuccessful breaks booting with bootm command.
If I press the enter key to get into the console circumventing distro-boot, booting via bootz works. If I wait until
distro_bootis finished, booting via bootz fails. Both with Linux 5.7.17.Yes, that was exactly Jernej's and my observation.
Quoted 11 lines
This confirms your finding that there is a problem with the initialization of the UEFI sub-system.
lib/efi_loader/efi_setup.c:192is the only place where we callswitch_to_non_secure_mode().With the line removed:
* Booting via bootz is successful.
* The EFI stub shows: Entering in SVC mode with MMU enabled
* Booting via bootefi failsAh, thanks for giving this a try. I know next to nothing about U-Boot's UEFI internals, so didn't dare to touch this code.
Quoted 9 lines
switch_to_non_secure_mode()is safe to be called repeatedly. So we could move theswitch_to_non_secure_mode()call todo_bootefi_exec(). This is after the boot manager has searched for a bootable image.With the change (see diff below):
* Booting via bootz is successful.
* EFI stub shows: Entering in HYP mode with MMU enabled
* Booting via bootefi is successful.Oh nice! I wasn't sure how much the UEFI initialisation actually relies on non-secure mode, and if switching to non-secure *after* the initialisation would change things.
Quoted 8 lines
Quoted 2 lines
The downside of the change is that executing bootz will still fail after a UEFI binary returns to U-Boot.
OK, that is not nice, but not really a big issue. At least it's a lot better than disabling bootmgr altogether or no bootz after a bootmgr call.
If you deem this change being not very intrusive, I would very much prefer this over my patch here.
Here is the patch:
[PATCH 1/1] efi_loader: switch to non-secure mode later
https://lists.denx.de/pipermail/u-boot/2021-January/438533.html
Quoted 7 lines
Quoted 6 lines
Running a shell in secure mode seems unwise. So we should dig a bit deeper:
Where in the code is the PSCI handler installed and where occurs the switch to non-secure state when booting via bootz/bootm? Can we move this to before distro-boot?
Quoted 8 lines
Well, yes, I think installing the PSCI handlers (and doing everything that requires secure state) early is the best way forward, but this is quite some change, and I would rather plug this problem now.
At the moment all of the v7 PSCI code is run as late as possible, so U-Boot can run in secure state. I agree this *sounds* scary, but running in secure is actually quite common for many ARM32 machines (even Linux sometimes runs with the NS bit cleared).
If we wanted security, I guess, we would have to move this stuff to TF-A.
Quoted 5 lines
For v7 Allwinner SoCs specifically we need access to the secure-only SID registers for the MAC address generation, also secure SRAM becomes inaccessible in non-secure world (as expected, but in contrast to the ARMv8 chips).
Where would I find that code where we generate the MAC address?
Is this the only thing requiring running in secure mode?
Or do you need secure-mode when setting the MAC address in sun8i_eth_write_hwaddr(), _sunxi_write_hwaddr()?
Best regards
Heinrich
Quoted 50 lines
So I think eventually we will need to bite the bullet and teach U-Boot to cope with non-secure in sunxi-v7, but this needs some time and requires possibly intrusive changes.
Cheers,
AndreQuoted 41 lines
cmd/bootefi.c +2
-0@@ -8,6 +8,7 @@#define LOG_CATEGORY LOGC_EFI#include <common.h> +#include <bootm.h> #include <charset.h> #include <command.h> #include <dm.h> @@ -338,6 +339,9 @@ static efi_status_t do_bootefi_exec(efi_handle_t handle, void *load_options)efi_uintn_t exit_data_size =0 ;u16 *exit_data =NULL ; +/* On ARM switch from EL3 or secure mode to EL2 or non-secure mode */
cmd/bootefi.c +2
-0+switch_to_non_secure_mode (); +/* Call our payload! */ ret = EFI_CALL(efi_start_image(handle, &exit_data_size,&exit_data)); if (ret != EFI_SUCCESS) {lib/efi_loader/efi_setup.c +0
-1@@ -188,9 +188,6 @@ efi_status_t efi_init_obj_list(void)/* Allow unaligned memory access */ allow_unaligned (); -/* On ARM switch from EL3 or secure mode to EL2 or non-secure mode */
lib/efi_loader/efi_setup.c +0
-2-switch_to_non_secure_mode (); -/* Initialize root node */ ret =efi_root_node_register ();if (ret !=EFI_SUCCESS )
Quoted 83 lines · On Sun, 24 Jan 2021 15:45:23 +0100 Heinrich Schuchardt <[email protected]>…
Quoted 73 lines · Andre Przywara
On Sun, 24 Jan 2021 11:44:35 +0100
Heinrich Schuchardt <[email protected]> wrote:Hi Heinrich,
many thanks for digging into this!
Quoted 28 lines · Jernej ?krabec
Quoted 27 lines
Dne nedelja, 24. januar 2021 ob 09:33:15 CET je Jernej ?krabec napisal(a):
Quoted 25 lines
Hi!
Dne nedelja, 24. januar 2021 ob 09:27:02 CET je Heinrich Schuchardt
napisal(a):
Quoted 2 lines · Simon Glass
Quoted 1 line
On Fri, 22 Jan 2021 at 05:05, Andre Przywara <[email protected]>
Quoted 16 lines · wrote:
Quoted 6 lines
Quoted 6 lines
When "bootefi bootmgr" is run, it switches the CPU into non-secure state. This breaks platforms like 32-bit Allwinner boards that rely on running in secure state until late in the process, when they install the PSCI handler in secure memory and drop into non-secure state. They hang just before entering the kernel, after the "Starting the kernel" message.
Dear Andre,
thank you for reporting the issue.
I have an Orange Pi PC with a 32 bit Allwinner CPU.
orangepi_pc_defconfighasCONFIG_ARMV7_PSCI=y.I use origin/master (
e716c9022970dac9b) and the Orange PI boots successfully using GRUB EFI into Linux 5.9.Quoted 6 lines
Quoted 2 lines
Just one clarification - issue here is that "bootefi bootmgr" command when unsuccessful breaks booting with bootm command.
If I press the enter key to get into the console circumventing distro-boot, booting via bootz works. If I wait until
distro_bootis finished, booting via bootz fails. Both with Linux 5.7.17.Yes, that was exactly Jernej's and my observation.
Quoted 11 lines
This confirms your finding that there is a problem with the initialization of the UEFI sub-system.
lib/efi_loader/efi_setup.c:192is the only place where we callswitch_to_non_secure_mode().With the line removed:
* Booting via bootz is successful.
* The EFI stub shows: Entering in SVC mode with MMU enabled
* Booting via bootefi failsAh, thanks for giving this a try. I know next to nothing about U-Boot's UEFI internals, so didn't dare to touch this code.
Quoted 9 lines
switch_to_non_secure_mode()is safe to be called repeatedly. So we could move theswitch_to_non_secure_mode()call todo_bootefi_exec(). This is after the boot manager has searched for a bootable image.With the change (see diff below):
* Booting via bootz is successful.
* EFI stub shows: Entering in HYP mode with MMU enabled
* Booting via bootefi is successful.Oh nice! I wasn't sure how much the UEFI initialisation actually relies on non-secure mode, and if switching to non-secure *after* the initialisation would change things.
Quoted 8 lines
Quoted 2 lines
The downside of the change is that executing bootz will still fail after a UEFI binary returns to U-Boot.
OK, that is not nice, but not really a big issue. At least it's a lot better than disabling bootmgr altogether or no bootz after a bootmgr call.
If you deem this change being not very intrusive, I would very much prefer this over my patch here.
Quoted 4 lines
Here is the patch:
[PATCH 1/1]
efi_loader: switch to non-secure mode later
https://lists.denx.de/pipermail/u-boot/2021-January/438533.html
Thanks, I will give this a spin ASAP.
Quoted 10 lines
Quoted 10 lines
Quoted 5 lines
Running a shell in secure mode seems unwise. So we should dig a bit deeper:
Where in the code is the PSCI handler installed and where occurs the switch to non-secure state when booting via bootz/bootm? Can we move this to before distro-boot?
Well, yes, I think installing the PSCI handlers (and doing everything that requires secure state) early is the best way forward, but this is quite some change, and I would rather plug this problem now.
Quoted 6 lines
Quoted 4 lines
At the moment all of the v7 PSCI code is run as late as possible, so U-Boot can run in secure state. I agree this *sounds* scary, but running in secure is actually quite common for many ARM32 machines (even Linux sometimes runs with the NS bit cleared).
If we wanted security, I guess, we would have to move this stuff to TF-A.
So yeah, Samuel is investigating TF-A for 32-bit Allwinner chip, AFAIK. But this will take some time, I guess, and won't probably never replace the existing PSCI implementation in U-Boot completely.
And actually the term "security" is somewhat misleading here, not sure who we protect against. On 32-bit Allwinner U-Boot is the first and only "firmware" code that is running, so it has all the rights to do that in secure state.
Quoted 6 lines
Quoted 4 lines
For v7 Allwinner SoCs specifically we need access to the secure-only SID registers for the MAC address generation, also secure SRAM becomes inaccessible in non-secure world (as expected, but in contrast to the ARMv8 chips).
Where would I find that code where we generate the MAC address?
The unique SoC serial number is read from the SID "device", the interesting part for us are some r/o ID registers. The code is in arch/arm/mach-sunxi/cpu_info.c:sunxi_get_sid(), the MAC address is generated from that in board/sunxi/board.c:setup_environment(). (Use latest master, as this code recently changed).
This method already bites us when 64-bit SoCs run with the secure boot fuse burnt. I have some code to read the MAC address from the DT instead (or rather to make that actually work), and TF-A can use the same algorithm to generate the MAC address (just didn't manage to use the ARMv8 CRC instructions for that yet).
When starting in secure and dropping early, we could just read the SID register very early and store it somewhere, so not a real problem.
Quoted 1 line
Is this the only thing requiring running in secure mode?
There might be other places where secure state is required, CNTFRQ writes require the highest EL, maybe some core clock setup as well?
So it's surely possible, I am just a bit wary of touching 32-bit code too much, as I have only an A20 and H3 to test on (plus an A10 I need to get running first).
Quoted 2 lines
Or do you need secure-mode when setting the MAC address in
sun8i_eth_write_hwaddr(),_sunxi_write_hwaddr()?
No, this is a pure EMAC operation, totally ignorant of the NS bit.
Thanks!
Andre
Quoted 52 lines
Quoted 51 lines
So I think eventually we will need to bite the bullet and teach U-Boot to cope with non-secure in sunxi-v7, but this needs some time and requires possibly intrusive changes.
Cheers,
AndreQuoted 41 lines
cmd/bootefi.c +2
-0@@ -8,6 +8,7 @@#define LOG_CATEGORY LOGC_EFI#include <common.h> +#include <bootm.h> #include <charset.h> #include <command.h> #include <dm.h> @@ -338,6 +339,9 @@ static efi_status_t do_bootefi_exec(efi_handle_t handle, void *load_options) efi_uintn_t exit_data_size = 0;u16 *exit_data =NULL ; +/* On ARM switch from EL3 or secure mode to EL2 or non-secure mode */
cmd/bootefi.c +2
-0+switch_to_non_secure_mode (); +/* Call our payload! */ ret = EFI_CALL(efi_start_image(handle, &exit_data_size,&exit_data)); if (ret != EFI_SUCCESS) {(quoted diff) +0
-0
b/lib/efi_loader/efi_setup.cindex5800cbf6d4..588fbda736100644lib/efi_loader/efi_setup.c +0
-1@@ -188,9 +188,6 @@ efi_status_t efi_init_obj_list(void)/* Allow unaligned memory access */ allow_unaligned (); -/* On ARM switch from EL3 or secure mode to EL2 or non-secure mode */
lib/efi_loader/efi_setup.c +0
-2-switch_to_non_secure_mode (); -/* Initialize root node */ ret =efi_root_node_register ();if (ret !=EFI_SUCCESS )