[U-Boot] [RFC 0/6] efi_loader: support runtime variable access via cache

EBBR v1.0 section 2.5.3 says Even when SetVariable() is not supported during runtime services, firmware should cache variable names and values in EfiRuntimeServicesData memory so that GetVariable() and GetNextVeriableName() can behave as specified.
This is an experimental patch set and the aim is to enable this feature.
Cache buffer is in the same format as U-Boot environment hash table, but we cannot use functions in hashtable.c partly because most of U-Boot code are not available at UEFI runtime and partly because entries in a table are allocated by malloc(), which are again not available at UEFI runtime. It is quite painful to modify exiting U-Boot code so as to make it executable at UEFI runtime.
So I implemented a limited version of hsearch_r(). This may be a discussion. Given that there are not so many UEFI variables, we may want to use a simpler table format for caching.
Known issues: * Currently I test this feature with a test which is temporarily embedded in ExitBootServices. * After SetVirtualAddressMap, it won't work (TODO). ConvertPointer was implemented here just for this future work. * So how can we test the feature? We can't use "printenv" command as its relocation won't take place. * I see some Travis CI errors. This is probably due to no storage configured for UEFI variables.
Patch#1 to #4 are preparatory patches. Patch#5 is mainly for testing. Patch#6 is core part of this patch set.
AKASHI Takahiro (5): efi_loader: runtime: make SetVirtualAddressMap configurable efi: add RuntimeServicesSupported variable efi_loader: support convert_pointer at runtime cmd: efidebug: add "boot exit" sub-command efi_loader: variable: support runtime variable access via cache
Alexander Graf (1): efi_loader: Patch non-runtime code out at ExitBootServices already
cmd/efidebug.c | 63 ++++ include/efi_api.h | 15 + include/efi_loader.h | 22 ++ lib/efi_loader/Kconfig | 24 ++ lib/efi_loader/efi_boottime.c | 11 +- lib/efi_loader/efi_runtime.c | 155 ++++++++-- lib/efi_loader/efi_setup.c | 5 + lib/efi_loader/efi_variable.c | 467 +++++++++++++++++++++++++++++ test/py/tests/test_efi_selftest.py | 4 +- 9 files changed, 735 insertions(+), 31 deletions(-)

OS does not always need to call SetVirtualAddressMap. (Ard confirmed this on arm64 linux.) So let this API configurable. If disabled, it will return EFI_UNSUPPORTED as UEFI specification requires.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org --- lib/efi_loader/Kconfig | 7 +++++++ lib/efi_loader/efi_runtime.c | 8 ++++++++ 2 files changed, 15 insertions(+)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index 8bf4b1754d06..bb9c7582b14d 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -44,6 +44,13 @@ config EFI_SET_TIME Provide the SetTime() runtime service at boottime. This service can be used by an EFI application to adjust the real time clock.
+config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP + bool "runtime service: SetVirtualAddressMap" + default n + help + Enable SetVirtualAddressMap runtime service. This API will be + called by OS just before it enters into virtual address mode. + config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 9c50955c9bd0..60442cb21d37 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -374,10 +374,12 @@ static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, .patchto = efi_reset_system, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP }, { /* invalidate_*cache_all are gone */ .ptr = &efi_runtime_services.set_virtual_address_map, .patchto = &efi_unimplemented, +#endif }, { /* RTC accessors are gone */ .ptr = &efi_runtime_services.get_time, @@ -512,6 +514,7 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) invalidate_icache_all(); }
+#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP /** * efi_set_virtual_address_map() - change from physical to virtual mapping * @@ -619,6 +622,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(
return EFI_EXIT(EFI_INVALID_PARAMETER); } +#endif /* CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP */
/** * efi_add_runtime_mmio() - add memory-mapped IO region @@ -796,7 +800,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { .set_time = &efi_set_time_boottime, .get_wakeup_time = (void *)&efi_unimplemented, .set_wakeup_time = (void *)&efi_unimplemented, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP .set_virtual_address_map = &efi_set_virtual_address_map, +#else + .set_virtual_address_map = (void *)&efi_unimplemented, +#endif .convert_pointer = (void *)&efi_invalid_parameter, .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name,

On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
OS does not always need to call SetVirtualAddressMap. (Ard confirmed this on arm64 linux.) So let this API configurable. If disabled, it will return EFI_UNSUPPORTED as UEFI specification requires.
Currently we do not support this scenario. Alex's patch should go in first.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 7 +++++++ lib/efi_loader/efi_runtime.c | 8 ++++++++ 2 files changed, 15 insertions(+)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index 8bf4b1754d06..bb9c7582b14d 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -44,6 +44,13 @@ config EFI_SET_TIME Provide the SetTime() runtime service at boottime. This service can be used by an EFI application to adjust the real time clock.
+config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- bool "runtime service: SetVirtualAddressMap"
- default n
- help
Enable SetVirtualAddressMap runtime service. This API will be
called by OS just before it enters into virtual address mode.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 9c50955c9bd0..60442cb21d37 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -374,10 +374,12 @@ static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, .patchto = efi_reset_system, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP }, { /* invalidate_*cache_all are gone */ .ptr = &efi_runtime_services.set_virtual_address_map, .patchto = &efi_unimplemented, +#endif }, { /* RTC accessors are gone */ .ptr = &efi_runtime_services.get_time, @@ -512,6 +514,7 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) invalidate_icache_all(); }
+#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP /**
- efi_set_virtual_address_map() - change from physical to virtual mapping
@@ -619,6 +622,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(
return EFI_EXIT(EFI_INVALID_PARAMETER); } +#endif /* CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP */
/**
- efi_add_runtime_mmio() - add memory-mapped IO region
@@ -796,7 +800,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { .set_time = &efi_set_time_boottime, .get_wakeup_time = (void *)&efi_unimplemented, .set_wakeup_time = (void *)&efi_unimplemented, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP .set_virtual_address_map = &efi_set_virtual_address_map, +#else
- .set_virtual_address_map = (void *)&efi_unimplemented,
Depending on the ABI it is not save to use a function with another set of parameters.
Best regards
Heinrich
+#endif .convert_pointer = (void *)&efi_invalid_parameter, .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name,

From: Heinrich Schuchardt xypron.glpk@gmx.de Date: Sat, 15 Jun 2019 21:46:02 +0200
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
OS does not always need to call SetVirtualAddressMap. (Ard confirmed this on arm64 linux.) So let this API configurable. If disabled, it will return EFI_UNSUPPORTED as UEFI specification requires.
Currently we do not support this scenario. Alex's patch should go in first.
OpenBSD/arm64 will always call this function. It does this in order to randomize the virtual addresses used by runtime services to make it harder for an attacker to call into UEFI runtime services. Note that the UEFI 2.7 standard provides no indication that this interface might be optional, so I don't think OpenBSD is doing anything wrong here.
I think it is unwise to make this API configurable. But disabling it by default like this diff does would be a seriously bad idea. It means U-Boot would no longer be backwards compatible with UEFI 2.7.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 7 +++++++ lib/efi_loader/efi_runtime.c | 8 ++++++++ 2 files changed, 15 insertions(+)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index 8bf4b1754d06..bb9c7582b14d 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -44,6 +44,13 @@ config EFI_SET_TIME Provide the SetTime() runtime service at boottime. This service can be used by an EFI application to adjust the real time clock.
+config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- bool "runtime service: SetVirtualAddressMap"
- default n
- help
Enable SetVirtualAddressMap runtime service. This API will be
called by OS just before it enters into virtual address mode.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 9c50955c9bd0..60442cb21d37 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -374,10 +374,12 @@ static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, .patchto = efi_reset_system, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP }, { /* invalidate_*cache_all are gone */ .ptr = &efi_runtime_services.set_virtual_address_map, .patchto = &efi_unimplemented, +#endif }, { /* RTC accessors are gone */ .ptr = &efi_runtime_services.get_time, @@ -512,6 +514,7 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) invalidate_icache_all(); }
+#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP /**
- efi_set_virtual_address_map() - change from physical to virtual mapping
@@ -619,6 +622,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(
return EFI_EXIT(EFI_INVALID_PARAMETER); } +#endif /* CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP */
/**
- efi_add_runtime_mmio() - add memory-mapped IO region
@@ -796,7 +800,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { .set_time = &efi_set_time_boottime, .get_wakeup_time = (void *)&efi_unimplemented, .set_wakeup_time = (void *)&efi_unimplemented, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP .set_virtual_address_map = &efi_set_virtual_address_map, +#else
- .set_virtual_address_map = (void *)&efi_unimplemented,
Depending on the ABI it is not save to use a function with another set of parameters.
Best regards
Heinrich
+#endif .convert_pointer = (void *)&efi_invalid_parameter, .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name,
U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot

On 6/15/19 11:14 PM, Mark Kettenis wrote:
From: Heinrich Schuchardt xypron.glpk@gmx.de Date: Sat, 15 Jun 2019 21:46:02 +0200
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
OS does not always need to call SetVirtualAddressMap. (Ard confirmed this on arm64 linux.) So let this API configurable. If disabled, it will return EFI_UNSUPPORTED as UEFI specification requires.
Currently we do not support this scenario. Alex's patch should go in first.
OpenBSD/arm64 will always call this function. It does this in order
OpenBSD/arm32 does not call it. I have no clue why the 32bit and 64bit code have diverged.
to randomize the virtual addresses used by runtime services to make it harder for an attacker to call into UEFI runtime services. Note that the UEFI 2.7 standard provides no indication that this interface might be optional, so I don't think OpenBSD is doing anything wrong here.
I think it is unwise to make this API configurable. But disabling it by default like this diff does would be a seriously bad idea. It means U-Boot would no longer be backwards compatible with UEFI 2.7.
The current UEFI specification is 2.8. It foresees that SetVirtualAddressMap() can be marked as unsupported but that does not imply that we have to make it customizable. And disabling it by default does not make much sense while most operating systems use it.
Best regards
Heinrich
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 7 +++++++ lib/efi_loader/efi_runtime.c | 8 ++++++++ 2 files changed, 15 insertions(+)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index 8bf4b1754d06..bb9c7582b14d 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -44,6 +44,13 @@ config EFI_SET_TIME Provide the SetTime() runtime service at boottime. This service can be used by an EFI application to adjust the real time clock.
+config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- bool "runtime service: SetVirtualAddressMap"
- default n
- help
Enable SetVirtualAddressMap runtime service. This API will be
called by OS just before it enters into virtual address mode.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 9c50955c9bd0..60442cb21d37 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -374,10 +374,12 @@ static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, .patchto = efi_reset_system, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP }, { /* invalidate_*cache_all are gone */ .ptr = &efi_runtime_services.set_virtual_address_map, .patchto = &efi_unimplemented, +#endif }, { /* RTC accessors are gone */ .ptr = &efi_runtime_services.get_time, @@ -512,6 +514,7 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) invalidate_icache_all(); }
+#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP /** * efi_set_virtual_address_map() - change from physical to virtual mapping * @@ -619,6 +622,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(
return EFI_EXIT(EFI_INVALID_PARAMETER);
} +#endif /* CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP */
/** * efi_add_runtime_mmio() - add memory-mapped IO region @@ -796,7 +800,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { .set_time = &efi_set_time_boottime, .get_wakeup_time = (void *)&efi_unimplemented, .set_wakeup_time = (void *)&efi_unimplemented, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP .set_virtual_address_map = &efi_set_virtual_address_map, +#else
- .set_virtual_address_map = (void *)&efi_unimplemented,
Depending on the ABI it is not save to use a function with another set of parameters.
Best regards
Heinrich
+#endif .convert_pointer = (void *)&efi_invalid_parameter, .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name,
U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot

From: Heinrich Schuchardt xypron.glpk@gmx.de Date: Sun, 16 Jun 2019 23:52:49 +0200
On 6/15/19 11:14 PM, Mark Kettenis wrote:
From: Heinrich Schuchardt xypron.glpk@gmx.de Date: Sat, 15 Jun 2019 21:46:02 +0200
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
OS does not always need to call SetVirtualAddressMap. (Ard confirmed this on arm64 linux.) So let this API configurable. If disabled, it will return EFI_UNSUPPORTED as UEFI specification requires.
Currently we do not support this scenario. Alex's patch should go in first.
OpenBSD/arm64 will always call this function. It does this in order
OpenBSD/arm32 does not call it. I have no clue why the 32bit and 64bit code have diverged.
Beyond the device drivers, there isn't much shared code between OpenBSD/armv7 and OpenBSD/arm64. And the code that handles EFI runtime services is tied quite closely to the memory management code which is quite different between AArch32 and AArch64. The primary reason for adding support for EFI runtime services to OpenBSD/amd64 was to support getting and setting the time on machines with "fullblown" UEFI support. Machines that have armv7 CPUs *and* fullblown UEFI support are pretty much non-existent so I didn't implement support for armv7 yet.
to randomize the virtual addresses used by runtime services to make it harder for an attacker to call into UEFI runtime services. Note that the UEFI 2.7 standard provides no indication that this interface might be optional, so I don't think OpenBSD is doing anything wrong here.
I think it is unwise to make this API configurable. But disabling it by default like this diff does would be a seriously bad idea. It means U-Boot would no longer be backwards compatible with UEFI 2.7.
The current UEFI specification is 2.8. It foresees that SetVirtualAddressMap() can be marked as unsupported but that does not imply that we have to make it customizable. And disabling it by default does not make much sense while most operating systems use it.
Right! Thanks,
Mark
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 7 +++++++ lib/efi_loader/efi_runtime.c | 8 ++++++++ 2 files changed, 15 insertions(+)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index 8bf4b1754d06..bb9c7582b14d 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -44,6 +44,13 @@ config EFI_SET_TIME Provide the SetTime() runtime service at boottime. This service can be used by an EFI application to adjust the real time clock.
+config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- bool "runtime service: SetVirtualAddressMap"
- default n
- help
Enable SetVirtualAddressMap runtime service. This API will be
called by OS just before it enters into virtual address mode.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 9c50955c9bd0..60442cb21d37 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -374,10 +374,12 @@ static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, .patchto = efi_reset_system, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP }, { /* invalidate_*cache_all are gone */ .ptr = &efi_runtime_services.set_virtual_address_map, .patchto = &efi_unimplemented, +#endif }, { /* RTC accessors are gone */ .ptr = &efi_runtime_services.get_time, @@ -512,6 +514,7 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) invalidate_icache_all(); }
+#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP /** * efi_set_virtual_address_map() - change from physical to virtual mapping * @@ -619,6 +622,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(
return EFI_EXIT(EFI_INVALID_PARAMETER);
} +#endif /* CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP */
/** * efi_add_runtime_mmio() - add memory-mapped IO region @@ -796,7 +800,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { .set_time = &efi_set_time_boottime, .get_wakeup_time = (void *)&efi_unimplemented, .set_wakeup_time = (void *)&efi_unimplemented, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP .set_virtual_address_map = &efi_set_virtual_address_map, +#else
- .set_virtual_address_map = (void *)&efi_unimplemented,
Depending on the ABI it is not save to use a function with another set of parameters.
Best regards
Heinrich
+#endif .convert_pointer = (void *)&efi_invalid_parameter, .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name,
U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot

On Sat, Jun 15, 2019 at 09:46:02PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
OS does not always need to call SetVirtualAddressMap. (Ard confirmed this on arm64 linux.) So let this API configurable. If disabled, it will return EFI_UNSUPPORTED as UEFI specification requires.
Currently we do not support this scenario. Alex's patch should go in first.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 7 +++++++ lib/efi_loader/efi_runtime.c | 8 ++++++++ 2 files changed, 15 insertions(+)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index 8bf4b1754d06..bb9c7582b14d 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -44,6 +44,13 @@ config EFI_SET_TIME Provide the SetTime() runtime service at boottime. This service can be used by an EFI application to adjust the real time clock.
+config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- bool "runtime service: SetVirtualAddressMap"
- default n
- help
Enable SetVirtualAddressMap runtime service. This API will be
called by OS just before it enters into virtual address mode.
config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 9c50955c9bd0..60442cb21d37 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -374,10 +374,12 @@ static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, .patchto = efi_reset_system, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP }, { /* invalidate_*cache_all are gone */ .ptr = &efi_runtime_services.set_virtual_address_map, .patchto = &efi_unimplemented, +#endif }, { /* RTC accessors are gone */ .ptr = &efi_runtime_services.get_time, @@ -512,6 +514,7 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) invalidate_icache_all(); }
+#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP /**
- efi_set_virtual_address_map() - change from physical to virtual mapping
@@ -619,6 +622,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(
return EFI_EXIT(EFI_INVALID_PARAMETER); } +#endif /* CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP */
/**
- efi_add_runtime_mmio() - add memory-mapped IO region
@@ -796,7 +800,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { .set_time = &efi_set_time_boottime, .get_wakeup_time = (void *)&efi_unimplemented, .set_wakeup_time = (void *)&efi_unimplemented, +#ifdef CONFIG_RUNTIME_SET_VIRTUAL_ADDRESS_MAP .set_virtual_address_map = &efi_set_virtual_address_map, +#else
- .set_virtual_address_map = (void *)&efi_unimplemented,
Depending on the ABI it is not save to use a function with another set of parameters.
I don't think it's a good practice, but there already are a couple of precedents: get_wakeup_time set_wakeup_time get_next_high_mono_count, and convert_pointer
-Takahiro Akashi
Best regards
Heinrich
+#endif .convert_pointer = (void *)&efi_invalid_parameter, .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name,

This variable is defined in UEFI specification 2.8, section 8.1. Its value should be updated whenever we add any usable runtime services function.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org --- include/efi_api.h | 15 +++++++++++++++ include/efi_loader.h | 3 +++ lib/efi_loader/efi_runtime.c | 28 ++++++++++++++++++++++++++++ lib/efi_loader/efi_setup.c | 5 +++++ 4 files changed, 51 insertions(+)
diff --git a/include/efi_api.h b/include/efi_api.h index 65584dd2d82a..d7d95edd4dfc 100644 --- a/include/efi_api.h +++ b/include/efi_api.h @@ -213,6 +213,21 @@ struct efi_capsule_header { u32 capsule_image_size; };
+#define EFI_RT_SUPPORTED_GET_TIME 0x0001 +#define EFI_RT_SUPPORTED_SET_TIME 0x0002 +#define EFI_RT_SUPPORTED_GET_WAKEUP_TIME 0x0004 +#define EFI_RT_SUPPORTED_SET_WAKEUP_TIME 0x0008 +#define EFI_RT_SUPPORTED_GET_VARIABLE 0x0010 +#define EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME 0x0020 +#define EFI_RT_SUPPORTED_SET_VARIABLE 0x0040 +#define EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP 0x0080 +#define EFI_RT_SUPPORTED_CONVERT_POINTER 0x0100 +#define EFI_RT_SUPPORTED_GET_NEXT_HIGH_MONOTONIC_COUNT 0x0200 +#define EFI_RT_SUPPORTED_RESET_SYSTEM 0x0400 +#define EFI_RT_SUPPORTED_UPDATE_CAPSULE 0x0800 +#define EFI_RT_SUPPORTED_QUERY_CAPSULE_CAPABILITIES 0x1000 +#define EFI_RT_SUPPORTED_QUERY_VARIABLE_INFO 0x2000 + struct efi_runtime_services { struct efi_table_hdr hdr; efi_status_t (EFIAPI *get_time)(struct efi_time *time, diff --git a/include/efi_loader.h b/include/efi_loader.h index 23ce73226762..7bd8002e303e 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -573,6 +573,9 @@ static inline int guidcmp(const efi_guid_t *g1, const efi_guid_t *g2) #define __efi_runtime_data __attribute__ ((section (".data.efi_runtime"))) #define __efi_runtime __attribute__ ((section (".text.efi_runtime")))
+/* Indicate supported runtime services */ +efi_status_t efi_init_runtime_supported(void); + /* Update CRC32 in table header */ void __efi_runtime efi_update_table_header_crc32(struct efi_table_hdr *table);
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 60442cb21d37..cf202bb9ec07 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -89,6 +89,34 @@ struct elf_rela { * handle a good number of runtime callbacks */
+efi_status_t efi_init_runtime_supported(void) +{ + u16 efi_runtime_services_supported; + + /* + * This value must be synced with efi_runtime_detach_list + * as well as efi_runtime_services. + */ + efi_runtime_services_supported = EFI_RT_SUPPORTED_RESET_SYSTEM; +#ifdef CONFIG_EFI_GET_TIME + efi_runtime_services_supported |= EFI_RT_SUPPORTED_GET_TIME; +#endif +#ifdef CONFIG_EFI_SET_TIME + efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_TIME; +#endif +#ifdef CONFIG_EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP + efi_runtime_services_supported |= + EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP; +#endif + + return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", + &efi_global_variable_guid, + EFI_VARIABLE_BOOTSERVICE_ACCESS | + EFI_VARIABLE_RUNTIME_ACCESS, + sizeof(efi_runtime_services_supported), + &efi_runtime_services_supported)); +} + /** * efi_update_table_header_crc32() - Update crc32 in table header * diff --git a/lib/efi_loader/efi_setup.c b/lib/efi_loader/efi_setup.c index 45d6aca051f3..75fa344060d5 100644 --- a/lib/efi_loader/efi_setup.c +++ b/lib/efi_loader/efi_setup.c @@ -123,6 +123,11 @@ efi_status_t efi_init_obj_list(void) if (ret != EFI_SUCCESS) goto out;
+ /* Indicate supported runtime services */ + ret = efi_init_runtime_supported(); + if (ret != EFI_SUCCESS) + goto out; + /* Initialize system table */ ret = efi_initialize_system_table(); if (ret != EFI_SUCCESS)

On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
This variable is defined in UEFI specification 2.8, section 8.1. Its value should be updated whenever we add any usable runtime services function.
It is also required by the EBBR specification.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_api.h | 15 +++++++++++++++ include/efi_loader.h | 3 +++ lib/efi_loader/efi_runtime.c | 28 ++++++++++++++++++++++++++++ lib/efi_loader/efi_setup.c | 5 +++++ 4 files changed, 51 insertions(+)
diff --git a/include/efi_api.h b/include/efi_api.h index 65584dd2d82a..d7d95edd4dfc 100644 --- a/include/efi_api.h +++ b/include/efi_api.h @@ -213,6 +213,21 @@ struct efi_capsule_header { u32 capsule_image_size; };
+#define EFI_RT_SUPPORTED_GET_TIME 0x0001 +#define EFI_RT_SUPPORTED_SET_TIME 0x0002 +#define EFI_RT_SUPPORTED_GET_WAKEUP_TIME 0x0004 +#define EFI_RT_SUPPORTED_SET_WAKEUP_TIME 0x0008 +#define EFI_RT_SUPPORTED_GET_VARIABLE 0x0010 +#define EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME 0x0020 +#define EFI_RT_SUPPORTED_SET_VARIABLE 0x0040 +#define EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP 0x0080 +#define EFI_RT_SUPPORTED_CONVERT_POINTER 0x0100 +#define EFI_RT_SUPPORTED_GET_NEXT_HIGH_MONOTONIC_COUNT 0x0200 +#define EFI_RT_SUPPORTED_RESET_SYSTEM 0x0400 +#define EFI_RT_SUPPORTED_UPDATE_CAPSULE 0x0800 +#define EFI_RT_SUPPORTED_QUERY_CAPSULE_CAPABILITIES 0x1000 +#define EFI_RT_SUPPORTED_QUERY_VARIABLE_INFO 0x2000
struct efi_runtime_services { struct efi_table_hdr hdr; efi_status_t (EFIAPI *get_time)(struct efi_time *time, diff --git a/include/efi_loader.h b/include/efi_loader.h index 23ce73226762..7bd8002e303e 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -573,6 +573,9 @@ static inline int guidcmp(const efi_guid_t *g1, const efi_guid_t *g2) #define __efi_runtime_data __attribute__ ((section (".data.efi_runtime"))) #define __efi_runtime __attribute__ ((section (".text.efi_runtime")))
+/* Indicate supported runtime services */ +efi_status_t efi_init_runtime_supported(void);
/* Update CRC32 in table header */ void __efi_runtime efi_update_table_header_crc32(struct efi_table_hdr *table);
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 60442cb21d37..cf202bb9ec07 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -89,6 +89,34 @@ struct elf_rela {
- handle a good number of runtime callbacks
*/
+efi_status_t efi_init_runtime_supported(void) +{
- u16 efi_runtime_services_supported;
- /*
* This value must be synced with efi_runtime_detach_list
* as well as efi_runtime_services.
*/
- efi_runtime_services_supported = EFI_RT_SUPPORTED_RESET_SYSTEM;
This support is system dependent, e.g. on RK3288 systems it does not exist.
+#ifdef CONFIG_EFI_GET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_GET_TIME;
We do not support this at runtime.
+#endif +#ifdef CONFIG_EFI_SET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_TIME;
We do not support this at runtime.
+#endif +#ifdef CONFIG_EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP;
Our support is incomplete as we have not implemented ConvertPointer().
For unsupported services we will have to change the return value to EFI_UNSUPPORTED. But that will be a separate patch.
Best regards
Heinrich
+#endif
- return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported",
&efi_global_variable_guid,
EFI_VARIABLE_BOOTSERVICE_ACCESS |
EFI_VARIABLE_RUNTIME_ACCESS,
sizeof(efi_runtime_services_supported),
&efi_runtime_services_supported));
+}
/**
- efi_update_table_header_crc32() - Update crc32 in table header
diff --git a/lib/efi_loader/efi_setup.c b/lib/efi_loader/efi_setup.c index 45d6aca051f3..75fa344060d5 100644 --- a/lib/efi_loader/efi_setup.c +++ b/lib/efi_loader/efi_setup.c @@ -123,6 +123,11 @@ efi_status_t efi_init_obj_list(void) if (ret != EFI_SUCCESS) goto out;
- /* Indicate supported runtime services */
- ret = efi_init_runtime_supported();
- if (ret != EFI_SUCCESS)
goto out;
- /* Initialize system table */ ret = efi_initialize_system_table(); if (ret != EFI_SUCCESS)

On Thu, Jun 13, 2019 at 07:56:19AM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
This variable is defined in UEFI specification 2.8, section 8.1. Its value should be updated whenever we add any usable runtime services function.
It is also required by the EBBR specification.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_api.h | 15 +++++++++++++++ include/efi_loader.h | 3 +++ lib/efi_loader/efi_runtime.c | 28 ++++++++++++++++++++++++++++ lib/efi_loader/efi_setup.c | 5 +++++ 4 files changed, 51 insertions(+)
diff --git a/include/efi_api.h b/include/efi_api.h index 65584dd2d82a..d7d95edd4dfc 100644 --- a/include/efi_api.h +++ b/include/efi_api.h @@ -213,6 +213,21 @@ struct efi_capsule_header { u32 capsule_image_size; };
+#define EFI_RT_SUPPORTED_GET_TIME 0x0001 +#define EFI_RT_SUPPORTED_SET_TIME 0x0002 +#define EFI_RT_SUPPORTED_GET_WAKEUP_TIME 0x0004 +#define EFI_RT_SUPPORTED_SET_WAKEUP_TIME 0x0008 +#define EFI_RT_SUPPORTED_GET_VARIABLE 0x0010 +#define EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME 0x0020 +#define EFI_RT_SUPPORTED_SET_VARIABLE 0x0040 +#define EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP 0x0080 +#define EFI_RT_SUPPORTED_CONVERT_POINTER 0x0100 +#define EFI_RT_SUPPORTED_GET_NEXT_HIGH_MONOTONIC_COUNT 0x0200 +#define EFI_RT_SUPPORTED_RESET_SYSTEM 0x0400 +#define EFI_RT_SUPPORTED_UPDATE_CAPSULE 0x0800 +#define EFI_RT_SUPPORTED_QUERY_CAPSULE_CAPABILITIES 0x1000 +#define EFI_RT_SUPPORTED_QUERY_VARIABLE_INFO 0x2000
struct efi_runtime_services { struct efi_table_hdr hdr; efi_status_t (EFIAPI *get_time)(struct efi_time *time, diff --git a/include/efi_loader.h b/include/efi_loader.h index 23ce73226762..7bd8002e303e 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -573,6 +573,9 @@ static inline int guidcmp(const efi_guid_t *g1, const efi_guid_t *g2) #define __efi_runtime_data __attribute__ ((section (".data.efi_runtime"))) #define __efi_runtime __attribute__ ((section (".text.efi_runtime")))
+/* Indicate supported runtime services */ +efi_status_t efi_init_runtime_supported(void);
/* Update CRC32 in table header */ void __efi_runtime efi_update_table_header_crc32(struct efi_table_hdr *table);
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 60442cb21d37..cf202bb9ec07 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -89,6 +89,34 @@ struct elf_rela {
- handle a good number of runtime callbacks
*/
+efi_status_t efi_init_runtime_supported(void) +{
- u16 efi_runtime_services_supported;
- /*
* This value must be synced with efi_runtime_detach_list
* as well as efi_runtime_services.
*/
- efi_runtime_services_supported = EFI_RT_SUPPORTED_RESET_SYSTEM;
This support is system dependent, e.g. on RK3288 systems it does not exist.
efi_reset_system() is defined as a weak function and so there is no easy way to determine whether this API is supported or not.
+#ifdef CONFIG_EFI_GET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_GET_TIME;
We do not support this at runtime.
Okay, I will drop it.
+#endif +#ifdef CONFIG_EFI_SET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_TIME;
We do not support this at runtime.
ditto
+#endif +#ifdef CONFIG_EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP;
Our support is incomplete as we have not implemented ConvertPointer().
Incomplete? My patch#3 adds ConvertPointer().
-Takahiro Akashi
For unsupported services we will have to change the return value to EFI_UNSUPPORTED. But that will be a separate patch.
Best regards
Heinrich
+#endif
- return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported",
&efi_global_variable_guid,
EFI_VARIABLE_BOOTSERVICE_ACCESS |
EFI_VARIABLE_RUNTIME_ACCESS,
sizeof(efi_runtime_services_supported),
&efi_runtime_services_supported));
+}
/**
- efi_update_table_header_crc32() - Update crc32 in table header
diff --git a/lib/efi_loader/efi_setup.c b/lib/efi_loader/efi_setup.c index 45d6aca051f3..75fa344060d5 100644 --- a/lib/efi_loader/efi_setup.c +++ b/lib/efi_loader/efi_setup.c @@ -123,6 +123,11 @@ efi_status_t efi_init_obj_list(void) if (ret != EFI_SUCCESS) goto out;
- /* Indicate supported runtime services */
- ret = efi_init_runtime_supported();
- if (ret != EFI_SUCCESS)
goto out;
- /* Initialize system table */ ret = efi_initialize_system_table(); if (ret != EFI_SUCCESS)

On 6/13/19 9:06 AM, AKASHI Takahiro wrote:
On Thu, Jun 13, 2019 at 07:56:19AM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
This variable is defined in UEFI specification 2.8, section 8.1. Its value should be updated whenever we add any usable runtime services function.
It is also required by the EBBR specification.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_api.h | 15 +++++++++++++++ include/efi_loader.h | 3 +++ lib/efi_loader/efi_runtime.c | 28 ++++++++++++++++++++++++++++ lib/efi_loader/efi_setup.c | 5 +++++ 4 files changed, 51 insertions(+)
diff --git a/include/efi_api.h b/include/efi_api.h index 65584dd2d82a..d7d95edd4dfc 100644 --- a/include/efi_api.h +++ b/include/efi_api.h @@ -213,6 +213,21 @@ struct efi_capsule_header { u32 capsule_image_size; };
+#define EFI_RT_SUPPORTED_GET_TIME 0x0001 +#define EFI_RT_SUPPORTED_SET_TIME 0x0002 +#define EFI_RT_SUPPORTED_GET_WAKEUP_TIME 0x0004 +#define EFI_RT_SUPPORTED_SET_WAKEUP_TIME 0x0008 +#define EFI_RT_SUPPORTED_GET_VARIABLE 0x0010 +#define EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME 0x0020 +#define EFI_RT_SUPPORTED_SET_VARIABLE 0x0040 +#define EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP 0x0080 +#define EFI_RT_SUPPORTED_CONVERT_POINTER 0x0100 +#define EFI_RT_SUPPORTED_GET_NEXT_HIGH_MONOTONIC_COUNT 0x0200 +#define EFI_RT_SUPPORTED_RESET_SYSTEM 0x0400 +#define EFI_RT_SUPPORTED_UPDATE_CAPSULE 0x0800 +#define EFI_RT_SUPPORTED_QUERY_CAPSULE_CAPABILITIES 0x1000 +#define EFI_RT_SUPPORTED_QUERY_VARIABLE_INFO 0x2000
struct efi_runtime_services { struct efi_table_hdr hdr; efi_status_t (EFIAPI *get_time)(struct efi_time *time, diff --git a/include/efi_loader.h b/include/efi_loader.h index 23ce73226762..7bd8002e303e 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -573,6 +573,9 @@ static inline int guidcmp(const efi_guid_t *g1, const efi_guid_t *g2) #define __efi_runtime_data __attribute__ ((section (".data.efi_runtime"))) #define __efi_runtime __attribute__ ((section (".text.efi_runtime")))
+/* Indicate supported runtime services */ +efi_status_t efi_init_runtime_supported(void);
/* Update CRC32 in table header */ void __efi_runtime efi_update_table_header_crc32(struct efi_table_hdr *table);
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 60442cb21d37..cf202bb9ec07 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -89,6 +89,34 @@ struct elf_rela {
- handle a good number of runtime callbacks
*/
+efi_status_t efi_init_runtime_supported(void) +{
- u16 efi_runtime_services_supported;
- /*
* This value must be synced with efi_runtime_detach_list
* as well as efi_runtime_services.
*/
- efi_runtime_services_supported = EFI_RT_SUPPORTED_RESET_SYSTEM;
This support is system dependent, e.g. on RK3288 systems it does not exist.
efi_reset_system() is defined as a weak function and so there is no easy way to determine whether this API is supported or not.
+#ifdef CONFIG_EFI_GET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_GET_TIME;
We do not support this at runtime.
Okay, I will drop it.
+#endif +#ifdef CONFIG_EFI_SET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_TIME;
We do not support this at runtime.
ditto
+#endif +#ifdef CONFIG_EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP;
Our support is incomplete as we have not implemented ConvertPointer().
Incomplete? My patch#3 adds ConvertPointer().
That is a later patch. So I did not consider it for this patch. Furthermore the implementation of ConvertPointer depends on CONFIG_EFI_RUNTIME_CONVERT_POINTER.
Best regards
Heinrich
-Takahiro Akashi
For unsupported services we will have to change the return value to EFI_UNSUPPORTED. But that will be a separate patch.
Best regards
Heinrich
+#endif
- return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported",
&efi_global_variable_guid,
EFI_VARIABLE_BOOTSERVICE_ACCESS |
EFI_VARIABLE_RUNTIME_ACCESS,
sizeof(efi_runtime_services_supported),
&efi_runtime_services_supported));
+}
/**
- efi_update_table_header_crc32() - Update crc32 in table header
diff --git a/lib/efi_loader/efi_setup.c b/lib/efi_loader/efi_setup.c index 45d6aca051f3..75fa344060d5 100644 --- a/lib/efi_loader/efi_setup.c +++ b/lib/efi_loader/efi_setup.c @@ -123,6 +123,11 @@ efi_status_t efi_init_obj_list(void) if (ret != EFI_SUCCESS) goto out;
- /* Indicate supported runtime services */
- ret = efi_init_runtime_supported();
- if (ret != EFI_SUCCESS)
goto out;
- /* Initialize system table */ ret = efi_initialize_system_table(); if (ret != EFI_SUCCESS)

On 6/13/19 11:17 AM, Heinrich Schuchardt wrote:
On 6/13/19 9:06 AM, AKASHI Takahiro wrote:
On Thu, Jun 13, 2019 at 07:56:19AM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
This variable is defined in UEFI specification 2.8, section 8.1. Its value should be updated whenever we add any usable runtime services function.
It is also required by the EBBR specification.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_api.h | 15 +++++++++++++++ include/efi_loader.h | 3 +++ lib/efi_loader/efi_runtime.c | 28 ++++++++++++++++++++++++++++ lib/efi_loader/efi_setup.c | 5 +++++ 4 files changed, 51 insertions(+)
diff --git a/include/efi_api.h b/include/efi_api.h index 65584dd2d82a..d7d95edd4dfc 100644 --- a/include/efi_api.h +++ b/include/efi_api.h @@ -213,6 +213,21 @@ struct efi_capsule_header { u32 capsule_image_size; };
+#define EFI_RT_SUPPORTED_GET_TIME 0x0001 +#define EFI_RT_SUPPORTED_SET_TIME 0x0002 +#define EFI_RT_SUPPORTED_GET_WAKEUP_TIME 0x0004 +#define EFI_RT_SUPPORTED_SET_WAKEUP_TIME 0x0008 +#define EFI_RT_SUPPORTED_GET_VARIABLE 0x0010 +#define EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME 0x0020 +#define EFI_RT_SUPPORTED_SET_VARIABLE 0x0040 +#define EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP 0x0080 +#define EFI_RT_SUPPORTED_CONVERT_POINTER 0x0100 +#define EFI_RT_SUPPORTED_GET_NEXT_HIGH_MONOTONIC_COUNT 0x0200 +#define EFI_RT_SUPPORTED_RESET_SYSTEM 0x0400 +#define EFI_RT_SUPPORTED_UPDATE_CAPSULE 0x0800 +#define EFI_RT_SUPPORTED_QUERY_CAPSULE_CAPABILITIES 0x1000 +#define EFI_RT_SUPPORTED_QUERY_VARIABLE_INFO 0x2000
struct efi_runtime_services { struct efi_table_hdr hdr; efi_status_t (EFIAPI *get_time)(struct efi_time *time, diff --git a/include/efi_loader.h b/include/efi_loader.h index 23ce73226762..7bd8002e303e 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -573,6 +573,9 @@ static inline int guidcmp(const efi_guid_t *g1, const efi_guid_t *g2) #define __efi_runtime_data __attribute__ ((section (".data.efi_runtime"))) #define __efi_runtime __attribute__ ((section (".text.efi_runtime")))
+/* Indicate supported runtime services */ +efi_status_t efi_init_runtime_supported(void);
/* Update CRC32 in table header */ void __efi_runtime efi_update_table_header_crc32(struct efi_table_hdr *table);
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index 60442cb21d37..cf202bb9ec07 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -89,6 +89,34 @@ struct elf_rela {
- handle a good number of runtime callbacks
*/
+efi_status_t efi_init_runtime_supported(void) +{
- u16 efi_runtime_services_supported;
- /*
* This value must be synced with efi_runtime_detach_list
* as well as efi_runtime_services.
*/
- efi_runtime_services_supported = EFI_RT_SUPPORTED_RESET_SYSTEM;
This support is system dependent, e.g. on RK3288 systems it does not exist.
efi_reset_system() is defined as a weak function and so there is no easy way to determine whether this API is supported or not.
+#ifdef CONFIG_EFI_GET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_GET_TIME;
We do not support this at runtime.
Okay, I will drop it.
+#endif +#ifdef CONFIG_EFI_SET_TIME
- efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_TIME;
We do not support this at runtime.
ditto
+#endif +#ifdef CONFIG_EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP;
Our support is incomplete as we have not implemented ConvertPointer().
Incomplete? My patch#3 adds ConvertPointer().
That is a later patch. So I did not consider it for this patch. Furthermore the implementation of ConvertPointer depends on CONFIG_EFI_RUNTIME_CONVERT_POINTER.
EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP and EFI_RT_SUPPORTED_CONVERT_POINTER are separate values.
You are right in switching EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP on here.
Best regards
Heinrich
-Takahiro Akashi
For unsupported services we will have to change the return value to EFI_UNSUPPORTED. But that will be a separate patch.
Best regards
Heinrich
+#endif
- return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported",
&efi_global_variable_guid,
EFI_VARIABLE_BOOTSERVICE_ACCESS |
EFI_VARIABLE_RUNTIME_ACCESS,
sizeof(efi_runtime_services_supported),
&efi_runtime_services_supported));
+}
/**
- efi_update_table_header_crc32() - Update crc32 in table header
diff --git a/lib/efi_loader/efi_setup.c b/lib/efi_loader/efi_setup.c index 45d6aca051f3..75fa344060d5 100644 --- a/lib/efi_loader/efi_setup.c +++ b/lib/efi_loader/efi_setup.c @@ -123,6 +123,11 @@ efi_status_t efi_init_obj_list(void) if (ret != EFI_SUCCESS) goto out;
- /* Indicate supported runtime services */
- ret = efi_init_runtime_supported();
- if (ret != EFI_SUCCESS)
goto out;
- /* Initialize system table */ ret = efi_initialize_system_table(); if (ret != EFI_SUCCESS)

With this patch, ConvertPointer runtime service is enabled. This function will be useful only after SetVirtualAddressMap is called and before it exits according to UEFI specification.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org --- lib/efi_loader/Kconfig | 8 ++++ lib/efi_loader/efi_runtime.c | 81 ++++++++++++++++++++++++++---------- 2 files changed, 66 insertions(+), 23 deletions(-)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index bb9c7582b14d..e2ef43157568 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -51,6 +51,14 @@ config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP Enable SetVirtualAddressMap runtime service. This API will be called by OS just before it enters into virtual address mode.
+config EFI_RUNTIME_CONVERT_POINTER + bool "runtime service: ConvertPointer" + default n + help + Enable ConvertPointer runtime service. This API will be expected + to be called by UEFI drivers in relocating themselves to virtual + address space. + config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index cf202bb9ec07..ff3684a4b692 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -27,7 +27,6 @@ LIST_HEAD(efi_runtime_mmio);
static efi_status_t __efi_runtime EFIAPI efi_unimplemented(void); static efi_status_t __efi_runtime EFIAPI efi_device_error(void); -static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void);
/* * TODO(sjg@chromium.org): These defines and structures should come from the ELF @@ -108,6 +107,10 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP; #endif +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER + efi_runtime_services_supported |= + EFI_RT_SUPPORTED_CONVERT_POINTER; +#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -392,6 +395,39 @@ efi_status_t __weak __efi_runtime EFIAPI efi_set_time(struct efi_time *time) return EFI_UNSUPPORTED; }
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +static struct efi_mem_desc *efi_virtmap __efi_runtime_data; +static int efi_virtmap_num __efi_runtime_data; + +static efi_status_t __efi_runtime EFIAPI efi_convert_pointer(unsigned long dbg, + void **address) +{ + struct efi_mem_desc *map; + efi_physical_addr_t addr; + int i; + + if (!efi_virtmap) + return EFI_UNSUPPORTED; + + if (!address) + return EFI_INVALID_PARAMETER; + + for (i = 0, map = efi_virtmap; i < efi_virtmap_num; i++, map++) { + addr = (efi_physical_addr_t)*address; + if (addr >= map->physical_start && + (addr < (map->physical_start + + (map->num_pages << EFI_PAGE_SHIFT)))) { + *address = (void *)map->virtual_start; + *address += addr - map->physical_start; + + return EFI_SUCCESS; + } + } + + return EFI_NOT_FOUND; +} +#endif + struct efi_runtime_detach_list_struct { void *ptr; void *patchto; @@ -599,6 +635,10 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( return EFI_EXIT(EFI_INVALID_PARAMETER); }
+ efi_virtmap = virtmap; + efi_virtmap_num = n; + +#if 0 /* FIXME: This code is fragile as calloc is used in add_runtime_mmio */ /* Rebind mmio pointers */ for (i = 0; i < n; i++) { struct efi_mem_desc *map = (void*)virtmap + @@ -622,14 +662,14 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( *lmmio->ptr = (void *)new_addr; } } - if ((map_start <= (uintptr_t)systab.tables) && - (map_end >= (uintptr_t)systab.tables)) { - char *ptr = (char *)systab.tables; - - ptr += off; - systab.tables = (struct efi_configuration_table *)ptr; - } } +#endif + + /* FIXME */ + efi_convert_pointer(0, (void **)&systab.tables); + + /* All fixes must be done before this line */ + efi_virtmap = NULL;
/* Move the actual runtime code over */ for (i = 0; i < n; i++) { @@ -644,6 +684,11 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( /* Once we're virtual, we can no longer handle complex callbacks */ efi_runtime_detach(new_offset); + + /* + * FIXME: + * We can no longer update RuntimeServicesSupported. + */ return EFI_EXIT(EFI_SUCCESS); } } @@ -733,20 +778,6 @@ static efi_status_t __efi_runtime EFIAPI efi_device_error(void) return EFI_DEVICE_ERROR; }
-/** - * efi_invalid_parameter() - replacement function, returns EFI_INVALID_PARAMETER - * - * This function is used after SetVirtualAddressMap() is called as replacement - * for services that are not available anymore due to constraints of the U-Boot - * implementation. - * - * Return: EFI_INVALID_PARAMETER - */ -static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void) -{ - return EFI_INVALID_PARAMETER; -} - /** * efi_update_capsule() - process information from operating system * @@ -833,7 +864,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { #else .set_virtual_address_map = (void *)&efi_unimplemented, #endif - .convert_pointer = (void *)&efi_invalid_parameter, +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER + .convert_pointer = &efi_convert_pointer, +#else + .convert_pointer = (void *)&efi_unimplemented, +#endif .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name, .set_variable = efi_set_variable,

On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, ConvertPointer runtime service is enabled. This function will be useful only after SetVirtualAddressMap is called and before it exits according to UEFI specification.
ConvertPointer() is called by drivers upon calling the notification functions of events registered as EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE.
We still lack support for these events.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 8 ++++ lib/efi_loader/efi_runtime.c | 81 ++++++++++++++++++++++++++---------- 2 files changed, 66 insertions(+), 23 deletions(-)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index bb9c7582b14d..e2ef43157568 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -51,6 +51,14 @@ config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP Enable SetVirtualAddressMap runtime service. This API will be called by OS just before it enters into virtual address mode.
+config EFI_RUNTIME_CONVERT_POINTER
- bool "runtime service: ConvertPointer"
- default n
- help
Enable ConvertPointer runtime service. This API will be expected
to be called by UEFI drivers in relocating themselves to virtual
address space.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index cf202bb9ec07..ff3684a4b692 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -27,7 +27,6 @@ LIST_HEAD(efi_runtime_mmio);
static efi_status_t __efi_runtime EFIAPI efi_unimplemented(void); static efi_status_t __efi_runtime EFIAPI efi_device_error(void); -static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void);
/*
- TODO(sjg@chromium.org): These defines and structures should come from the ELF
@@ -108,6 +107,10 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP; #endif +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_CONVERT_POINTER;
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -392,6 +395,39 @@ efi_status_t __weak __efi_runtime EFIAPI efi_set_time(struct efi_time *time) return EFI_UNSUPPORTED; }
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +static struct efi_mem_desc *efi_virtmap __efi_runtime_data; +static int efi_virtmap_num __efi_runtime_data;
Please, put a description of the function and its parameters here.
+static efi_status_t __efi_runtime EFIAPI efi_convert_pointer(unsigned long dbg,
void **address)
+{
- struct efi_mem_desc *map;
- efi_physical_addr_t addr;
- int i;
- if (!efi_virtmap)
return EFI_UNSUPPORTED;
- if (!address)
return EFI_INVALID_PARAMETER;
- for (i = 0, map = efi_virtmap; i < efi_virtmap_num; i++, map++) {
addr = (efi_physical_addr_t)*address;
This line should be before the loop.
if (addr >= map->physical_start &&
(addr < (map->physical_start
%s/(addr/addr/ No need for that extra parenthesis.
+ (map->num_pages << EFI_PAGE_SHIFT)))) {
*address = (void *)map->virtual_start;
I guess on 32bit this will end in a warning? Wouldn't you need to convert to uintptr_t first?
*address += addr - map->physical_start;
I think a single assignment is enough. Here you are updating a 32bit pointer with the difference of two u64. I would expect a warning.
*address = (void *)(uintptr_t) (addr - map->physical_start + map->virtual_start);
Or do all calculation with addr before copying to *address.
return EFI_SUCCESS;
}
- }
- return EFI_NOT_FOUND;
+} +#endif
- struct efi_runtime_detach_list_struct { void *ptr; void *patchto;
@@ -599,6 +635,10 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( return EFI_EXIT(EFI_INVALID_PARAMETER); }
- efi_virtmap = virtmap;
- efi_virtmap_num = n;
+#if 0 /* FIXME: This code is fragile as calloc is used in add_runtime_mmio */ /* Rebind mmio pointers */ for (i = 0; i < n; i++) { struct efi_mem_desc *map = (void*)virtmap + @@ -622,14 +662,14 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( *lmmio->ptr = (void *)new_addr; } }
if ((map_start <= (uintptr_t)systab.tables) &&
(map_end >= (uintptr_t)systab.tables)) {
char *ptr = (char *)systab.tables;
ptr += off;
systab.tables = (struct efi_configuration_table *)ptr;
}
This looks like an unrelated change.
Put it into a separate patch, please.
Best regards
Heinrich
} +#endif
/* FIXME */
efi_convert_pointer(0, (void **)&systab.tables);
/* All fixes must be done before this line */
efi_virtmap = NULL;
/* Move the actual runtime code over */ for (i = 0; i < n; i++) {
@@ -644,6 +684,11 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( /* Once we're virtual, we can no longer handle complex callbacks */ efi_runtime_detach(new_offset);
/*
* FIXME:
* We can no longer update RuntimeServicesSupported.
} }*/ return EFI_EXIT(EFI_SUCCESS);
@@ -733,20 +778,6 @@ static efi_status_t __efi_runtime EFIAPI efi_device_error(void) return EFI_DEVICE_ERROR; }
-/**
- efi_invalid_parameter() - replacement function, returns EFI_INVALID_PARAMETER
- This function is used after SetVirtualAddressMap() is called as replacement
- for services that are not available anymore due to constraints of the U-Boot
- implementation.
- Return: EFI_INVALID_PARAMETER
- */
-static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void) -{
- return EFI_INVALID_PARAMETER;
-}
- /**
- efi_update_capsule() - process information from operating system
@@ -833,7 +864,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { #else .set_virtual_address_map = (void *)&efi_unimplemented, #endif
- .convert_pointer = (void *)&efi_invalid_parameter,
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER
- .convert_pointer = &efi_convert_pointer,
+#else
- .convert_pointer = (void *)&efi_unimplemented,
I feel uneasy using a function efi_unimplemented() with a different number of parameters here. Depending on the ABI this may lead to a crash.
Best regards
Heinrich
+#endif .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name, .set_variable = efi_set_variable,

On Sat, Jun 15, 2019 at 09:41:31PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, ConvertPointer runtime service is enabled. This function will be useful only after SetVirtualAddressMap is called and before it exits according to UEFI specification.
ConvertPointer() is called by drivers upon calling the notification functions of events registered as EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE.
We still lack support for these events.
So? What do you want me to do here?
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 8 ++++ lib/efi_loader/efi_runtime.c | 81 ++++++++++++++++++++++++++---------- 2 files changed, 66 insertions(+), 23 deletions(-)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index bb9c7582b14d..e2ef43157568 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -51,6 +51,14 @@ config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP Enable SetVirtualAddressMap runtime service. This API will be called by OS just before it enters into virtual address mode.
+config EFI_RUNTIME_CONVERT_POINTER
- bool "runtime service: ConvertPointer"
- default n
- help
Enable ConvertPointer runtime service. This API will be expected
to be called by UEFI drivers in relocating themselves to virtual
address space.
config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index cf202bb9ec07..ff3684a4b692 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -27,7 +27,6 @@ LIST_HEAD(efi_runtime_mmio);
static efi_status_t __efi_runtime EFIAPI efi_unimplemented(void); static efi_status_t __efi_runtime EFIAPI efi_device_error(void); -static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void);
/*
- TODO(sjg@chromium.org): These defines and structures should come from the ELF
@@ -108,6 +107,10 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP; #endif +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_CONVERT_POINTER;
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -392,6 +395,39 @@ efi_status_t __weak __efi_runtime EFIAPI efi_set_time(struct efi_time *time) return EFI_UNSUPPORTED; }
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +static struct efi_mem_desc *efi_virtmap __efi_runtime_data; +static int efi_virtmap_num __efi_runtime_data;
Please, put a description of the function and its parameters here.
Okay.
+static efi_status_t __efi_runtime EFIAPI efi_convert_pointer(unsigned long dbg,
void **address)
+{
- struct efi_mem_desc *map;
- efi_physical_addr_t addr;
- int i;
- if (!efi_virtmap)
return EFI_UNSUPPORTED;
- if (!address)
return EFI_INVALID_PARAMETER;
- for (i = 0, map = efi_virtmap; i < efi_virtmap_num; i++, map++) {
addr = (efi_physical_addr_t)*address;
This line should be before the loop.
if (addr >= map->physical_start &&
(addr < (map->physical_start
%s/(addr/addr/ No need for that extra parenthesis.
+ (map->num_pages << EFI_PAGE_SHIFT)))) {
*address = (void *)map->virtual_start;
I guess on 32bit this will end in a warning? Wouldn't you need to convert to uintptr_t first?
*address += addr - map->physical_start;
I think a single assignment is enough. Here you are updating a 32bit pointer with the difference of two u64. I would expect a warning.
I will check.
*address = (void *)(uintptr_t) (addr - map->physical_start + map->virtual_start);
Or do all calculation with addr before copying to *address.
return EFI_SUCCESS;
}
- }
- return EFI_NOT_FOUND;
+} +#endif
struct efi_runtime_detach_list_struct { void *ptr; void *patchto; @@ -599,6 +635,10 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( return EFI_EXIT(EFI_INVALID_PARAMETER); }
- efi_virtmap = virtmap;
- efi_virtmap_num = n;
+#if 0 /* FIXME: This code is fragile as calloc is used in add_runtime_mmio */ /* Rebind mmio pointers */ for (i = 0; i < n; i++) { struct efi_mem_desc *map = (void*)virtmap + @@ -622,14 +662,14 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( *lmmio->ptr = (void *)new_addr; } }
if ((map_start <= (uintptr_t)systab.tables) &&
(map_end >= (uintptr_t)systab.tables)) {
char *ptr = (char *)systab.tables;
ptr += off;
systab.tables = (struct efi_configuration_table *)ptr;
}
This looks like an unrelated change.
It does. This code should be replaced by:
- /* FIXME */
- efi_convert_pointer(0, (void **)&systab.tables);
-Takahiro Akashi
Put it into a separate patch, please.
Best regards
Heinrich
} +#endif
/* FIXME */
efi_convert_pointer(0, (void **)&systab.tables);
/* All fixes must be done before this line */
efi_virtmap = NULL;
/* Move the actual runtime code over */ for (i = 0; i < n; i++) {
@@ -644,6 +684,11 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( /* Once we're virtual, we can no longer handle complex callbacks */ efi_runtime_detach(new_offset);
/*
* FIXME:
* We can no longer update RuntimeServicesSupported.
} }*/ return EFI_EXIT(EFI_SUCCESS);
@@ -733,20 +778,6 @@ static efi_status_t __efi_runtime EFIAPI efi_device_error(void) return EFI_DEVICE_ERROR; }
-/**
- efi_invalid_parameter() - replacement function, returns EFI_INVALID_PARAMETER
- This function is used after SetVirtualAddressMap() is called as replacement
- for services that are not available anymore due to constraints of the U-Boot
- implementation.
- Return: EFI_INVALID_PARAMETER
- */
-static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void) -{
- return EFI_INVALID_PARAMETER;
-}
/**
- efi_update_capsule() - process information from operating system
@@ -833,7 +864,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { #else .set_virtual_address_map = (void *)&efi_unimplemented, #endif
- .convert_pointer = (void *)&efi_invalid_parameter,
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER
- .convert_pointer = &efi_convert_pointer,
+#else
- .convert_pointer = (void *)&efi_unimplemented,
I feel uneasy using a function efi_unimplemented() with a different number of parameters here. Depending on the ABI this may lead to a crash.
Best regards
Heinrich
+#endif .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name, .set_variable = efi_set_variable,

On 6/17/19 3:15 AM, AKASHI Takahiro wrote:
On Sat, Jun 15, 2019 at 09:41:31PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, ConvertPointer runtime service is enabled. This function will be useful only after SetVirtualAddressMap is called and before it exits according to UEFI specification.
ConvertPointer() is called by drivers upon calling the notification functions of events registered as EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE.
We still lack support for these events.
So? What do you want me to do here?
We will have to address this in a future patch.
Regards Heinrich
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig | 8 ++++ lib/efi_loader/efi_runtime.c | 81 ++++++++++++++++++++++++++---------- 2 files changed, 66 insertions(+), 23 deletions(-)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index bb9c7582b14d..e2ef43157568 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -51,6 +51,14 @@ config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP Enable SetVirtualAddressMap runtime service. This API will be called by OS just before it enters into virtual address mode.
+config EFI_RUNTIME_CONVERT_POINTER
- bool "runtime service: ConvertPointer"
- default n
- help
Enable ConvertPointer runtime service. This API will be expected
to be called by UEFI drivers in relocating themselves to virtual
address space.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index cf202bb9ec07..ff3684a4b692 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -27,7 +27,6 @@ LIST_HEAD(efi_runtime_mmio);
static efi_status_t __efi_runtime EFIAPI efi_unimplemented(void); static efi_status_t __efi_runtime EFIAPI efi_device_error(void); -static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void);
/*
- TODO(sjg@chromium.org): These defines and structures should come from the ELF
@@ -108,6 +107,10 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP; #endif +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER
- efi_runtime_services_supported |=
EFI_RT_SUPPORTED_CONVERT_POINTER;
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -392,6 +395,39 @@ efi_status_t __weak __efi_runtime EFIAPI efi_set_time(struct efi_time *time) return EFI_UNSUPPORTED; }
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +static struct efi_mem_desc *efi_virtmap __efi_runtime_data; +static int efi_virtmap_num __efi_runtime_data;
Please, put a description of the function and its parameters here.
Okay.
+static efi_status_t __efi_runtime EFIAPI efi_convert_pointer(unsigned long dbg,
void **address)
+{
- struct efi_mem_desc *map;
- efi_physical_addr_t addr;
- int i;
- if (!efi_virtmap)
return EFI_UNSUPPORTED;
- if (!address)
return EFI_INVALID_PARAMETER;
- for (i = 0, map = efi_virtmap; i < efi_virtmap_num; i++, map++) {
addr = (efi_physical_addr_t)*address;
This line should be before the loop.
if (addr >= map->physical_start &&
(addr < (map->physical_start
%s/(addr/addr/ No need for that extra parenthesis.
+ (map->num_pages << EFI_PAGE_SHIFT)))) {
*address = (void *)map->virtual_start;
I guess on 32bit this will end in a warning? Wouldn't you need to convert to uintptr_t first?
*address += addr - map->physical_start;
I think a single assignment is enough. Here you are updating a 32bit pointer with the difference of two u64. I would expect a warning.
I will check.
*address = (void *)(uintptr_t) (addr - map->physical_start + map->virtual_start);
Or do all calculation with addr before copying to *address.
return EFI_SUCCESS;
}
- }
- return EFI_NOT_FOUND;
+} +#endif
- struct efi_runtime_detach_list_struct { void *ptr; void *patchto;
@@ -599,6 +635,10 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( return EFI_EXIT(EFI_INVALID_PARAMETER); }
- efi_virtmap = virtmap;
- efi_virtmap_num = n;
+#if 0 /* FIXME: This code is fragile as calloc is used in add_runtime_mmio */ /* Rebind mmio pointers */ for (i = 0; i < n; i++) { struct efi_mem_desc *map = (void*)virtmap + @@ -622,14 +662,14 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( *lmmio->ptr = (void *)new_addr; } }
if ((map_start <= (uintptr_t)systab.tables) &&
(map_end >= (uintptr_t)systab.tables)) {
char *ptr = (char *)systab.tables;
ptr += off;
systab.tables = (struct efi_configuration_table *)ptr;
}
This looks like an unrelated change.
It does. This code should be replaced by:
- /* FIXME */
- efi_convert_pointer(0, (void **)&systab.tables);
-Takahiro Akashi
Put it into a separate patch, please.
Best regards
Heinrich
} +#endif
/* FIXME */
efi_convert_pointer(0, (void **)&systab.tables);
/* All fixes must be done before this line */
efi_virtmap = NULL;
/* Move the actual runtime code over */ for (i = 0; i < n; i++) {
@@ -644,6 +684,11 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( /* Once we're virtual, we can no longer handle complex callbacks */ efi_runtime_detach(new_offset);
/*
* FIXME:
* We can no longer update RuntimeServicesSupported.
} }*/ return EFI_EXIT(EFI_SUCCESS);
@@ -733,20 +778,6 @@ static efi_status_t __efi_runtime EFIAPI efi_device_error(void) return EFI_DEVICE_ERROR; }
-/**
- efi_invalid_parameter() - replacement function, returns EFI_INVALID_PARAMETER
- This function is used after SetVirtualAddressMap() is called as replacement
- for services that are not available anymore due to constraints of the U-Boot
- implementation.
- Return: EFI_INVALID_PARAMETER
- */
-static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void) -{
- return EFI_INVALID_PARAMETER;
-}
- /**
- efi_update_capsule() - process information from operating system
@@ -833,7 +864,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = { #else .set_virtual_address_map = (void *)&efi_unimplemented, #endif
- .convert_pointer = (void *)&efi_invalid_parameter,
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER
- .convert_pointer = &efi_convert_pointer,
+#else
- .convert_pointer = (void *)&efi_unimplemented,
I feel uneasy using a function efi_unimplemented() with a different number of parameters here. Depending on the ABI this may lead to a crash.
Best regards
Heinrich
+#endif .get_variable = efi_get_variable, .get_next_variable_name = efi_get_next_variable_name, .set_variable = efi_set_variable,

On 6/17/19 7:41 AM, Heinrich Schuchardt wrote:
On 6/17/19 3:15 AM, AKASHI Takahiro wrote:
On Sat, Jun 15, 2019 at 09:41:31PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, ConvertPointer runtime service is enabled. This function will be useful only after SetVirtualAddressMap is called and before it exits according to UEFI specification.
ConvertPointer() is called by drivers upon calling the notification functions of events registered as EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE.
We still lack support for these events.
So? What do you want me to do here?
We will have to address this in a future patch.
Now that EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE is implemented and a unit test for ConvertPointer() provided we should proceed with implementing ConvertPointer().
Regards Heinrich
Regards Heinrich
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
lib/efi_loader/Kconfig      | 8 ++++  lib/efi_loader/efi_runtime.c | 81 ++++++++++++++++++++++++++----------  2 files changed, 66 insertions(+), 23 deletions(-)
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index bb9c7582b14d..e2ef43157568 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -51,6 +51,14 @@ config EFI_RUNTIME_SET_VIRTUAL_ADDRESS_MAP        Enable SetVirtualAddressMap runtime service. This API will be        called by OS just before it enters into virtual address mode.
+config EFI_RUNTIME_CONVERT_POINTER +Â Â Â bool "runtime service: ConvertPointer" +Â Â Â default n +Â Â Â help +Â Â Â Â Â Enable ConvertPointer runtime service. This API will be expected +Â Â Â Â Â to be called by UEFI drivers in relocating themselves to virtual +Â Â Â Â Â address space.
config EFI_DEVICE_PATH_TO_TEXT Â Â Â Â Â bool "Device path to text protocol" Â Â Â Â Â default y diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index cf202bb9ec07..ff3684a4b692 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -27,7 +27,6 @@ LIST_HEAD(efi_runtime_mmio);
static efi_status_t __efi_runtime EFIAPI efi_unimplemented(void); Â static efi_status_t __efi_runtime EFIAPI efi_device_error(void); -static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void);
/* Â Â * TODO(sjg@chromium.org): These defines and structures should come from the ELF @@ -108,6 +107,10 @@ efi_status_t efi_init_runtime_supported(void) Â Â Â Â Â efi_runtime_services_supported |= Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â EFI_RT_SUPPORTED_SET_VIRTUAL_ADDRESS_MAP; Â #endif +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +Â Â Â efi_runtime_services_supported |= +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â EFI_RT_SUPPORTED_CONVERT_POINTER; +#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â &efi_global_variable_guid, @@ -392,6 +395,39 @@ efi_status_t __weak __efi_runtime EFIAPI efi_set_time(struct efi_time *time) Â Â Â Â Â return EFI_UNSUPPORTED; Â }
+#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +static struct efi_mem_desc *efi_virtmap __efi_runtime_data; +static int efi_virtmap_num __efi_runtime_data;
Please, put a description of the function and its parameters here.
Okay.
+static efi_status_t __efi_runtime EFIAPI efi_convert_pointer(unsigned long dbg, +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â void **address) +{ +Â Â Â struct efi_mem_desc *map; +Â Â Â efi_physical_addr_t addr; +Â Â Â int i;
+Â Â Â if (!efi_virtmap) +Â Â Â Â Â Â Â return EFI_UNSUPPORTED;
+Â Â Â if (!address) +Â Â Â Â Â Â Â return EFI_INVALID_PARAMETER;
+Â Â Â for (i = 0, map = efi_virtmap; i < efi_virtmap_num; i++, map++) { +Â Â Â Â Â Â Â addr = (efi_physical_addr_t)*address;
This line should be before the loop.
+Â Â Â Â Â Â Â if (addr >= map->physical_start && +Â Â Â Â Â Â Â Â Â Â Â (addr < (map->physical_start
%s/(addr/addr/Â No need for that extra parenthesis.
+Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â + (map->num_pages << EFI_PAGE_SHIFT)))) { +Â Â Â Â Â Â Â Â Â Â Â *address = (void *)map->virtual_start;
I guess on 32bit this will end in a warning? Wouldn't you need to convert to uintptr_t first?
+Â Â Â Â Â Â Â Â Â Â Â *address += addr - map->physical_start;
I think a single assignment is enough. Here you are updating a 32bit pointer with the difference of two u64. I would expect a warning.
I will check.
*address = (void *)(uintptr_t) (addr - map->physical_start + map->virtual_start);
Or do all calculation with addr before copying to *address.
+Â Â Â Â Â Â Â Â Â Â Â return EFI_SUCCESS; +Â Â Â Â Â Â Â } +Â Â Â }
+Â Â Â return EFI_NOT_FOUND; +} +#endif
struct efi_runtime_detach_list_struct { Â Â Â Â Â void *ptr; Â Â Â Â Â void *patchto; @@ -599,6 +635,10 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( Â Â Â Â Â Â Â Â Â return EFI_EXIT(EFI_INVALID_PARAMETER); Â Â Â Â Â }
+Â Â Â efi_virtmap = virtmap; +Â Â Â efi_virtmap_num = n;
+#if 0 /* FIXME: This code is fragile as calloc is used in add_runtime_mmio */ Â Â Â Â Â /* Rebind mmio pointers */ Â Â Â Â Â for (i = 0; i < n; i++) { Â Â Â Â Â Â Â Â Â struct efi_mem_desc *map = (void*)virtmap + @@ -622,14 +662,14 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â *lmmio->ptr = (void *)new_addr; Â Â Â Â Â Â Â Â Â Â Â Â Â } Â Â Â Â Â Â Â Â Â } -Â Â Â Â Â Â Â if ((map_start <= (uintptr_t)systab.tables) && -Â Â Â Â Â Â Â Â Â Â Â (map_end >= (uintptr_t)systab.tables)) { -Â Â Â Â Â Â Â Â Â Â Â char *ptr = (char *)systab.tables;
-Â Â Â Â Â Â Â Â Â Â Â ptr += off; -Â Â Â Â Â Â Â Â Â Â Â systab.tables = (struct efi_configuration_table *)ptr; -Â Â Â Â Â Â Â }
This looks like an unrelated change.
It does. This code should be replaced by:
+Â Â Â /* FIXME */ +Â Â Â efi_convert_pointer(0, (void **)&systab.tables);
-Takahiro Akashi
Put it into a separate patch, please.
Best regards
Heinrich
} +#endif
+Â Â Â /* FIXME */ +Â Â Â efi_convert_pointer(0, (void **)&systab.tables);
+Â Â Â /* All fixes must be done before this line */ +Â Â Â efi_virtmap = NULL;
/* Move the actual runtime code over */      for (i = 0; i < n; i++) { @@ -644,6 +684,11 @@ static efi_status_t EFIAPI efi_set_virtual_address_map(              /* Once we're virtual, we can no longer handle                 complex callbacks */              efi_runtime_detach(new_offset);
+Â Â Â Â Â Â Â Â Â Â Â /* +Â Â Â Â Â Â Â Â Â Â Â Â * FIXME: +Â Â Â Â Â Â Â Â Â Â Â Â * We can no longer update RuntimeServicesSupported. +Â Â Â Â Â Â Â Â Â Â Â Â */ Â Â Â Â Â Â Â Â Â Â Â Â Â return EFI_EXIT(EFI_SUCCESS); Â Â Â Â Â Â Â Â Â } Â Â Â Â Â } @@ -733,20 +778,6 @@ static efi_status_t __efi_runtime EFIAPI efi_device_error(void) Â Â Â Â Â return EFI_DEVICE_ERROR; Â }
-/**
- efi_invalid_parameter() - replacement function, returns
EFI_INVALID_PARAMETER
- This function is used after SetVirtualAddressMap() is called as
replacement
- for services that are not available anymore due to constraints
of the U-Boot
- implementation.
- Return:Â Â Â EFI_INVALID_PARAMETER
- */
-static efi_status_t __efi_runtime EFIAPI efi_invalid_parameter(void) -{ -Â Â Â return EFI_INVALID_PARAMETER; -}
/**   * efi_update_capsule() - process information from operating system   * @@ -833,7 +864,11 @@ struct efi_runtime_services __efi_runtime_data efi_runtime_services = {  #else      .set_virtual_address_map = (void *)&efi_unimplemented,  #endif -   .convert_pointer = (void *)&efi_invalid_parameter, +#ifdef CONFIG_EFI_RUNTIME_CONVERT_POINTER +   .convert_pointer = &efi_convert_pointer, +#else +   .convert_pointer = (void *)&efi_unimplemented,
I feel uneasy using a function efi_unimplemented() with a different number of parameters here. Depending on the ABI this may lead to a crash.
Best regards
Heinrich
+#endif      .get_variable = efi_get_variable,      .get_next_variable_name = efi_get_next_variable_name,      .set_variable = efi_set_variable,

From: Alexander Graf agraf@suse.de
While discussing something completely different, Ard pointed out that it might be legal to omit calling SetVirtualAddressMap altogether.
There is even a patch on the Linux Kernel Mailing List that implements such behavior by now:
https://patchwork.kernel.org/patch/10782393/
While that sounds great, we currently rely on the SetVirtualAddressMap call to remove all references to code that would not work outside of boot services.
So let's patch out those bits already on the call to ExitBootServices, so that we can successfully run even when an OS chooses to omit any call to SetVirtualAddressMap.
Reported-by: Ard Biesheuvel ard.biesheuvel@linaro.org Signed-off-by: Alexander Graf agraf@suse.de
OpenBSD is not calling SetVirtualAddressMap on ARM 32 bit.
Adjust selftest: expect 'U-Boot' instead of 'resetting'.
Reviewed-by: Heinrich Schuchardt xypron.glpk@gmx.de --- include/efi_loader.h | 2 ++ lib/efi_loader/efi_boottime.c | 1 + lib/efi_loader/efi_runtime.c | 25 ++++++++++++++++++++----- test/py/tests/test_efi_selftest.py | 4 ++-- 4 files changed, 25 insertions(+), 7 deletions(-)
diff --git a/include/efi_loader.h b/include/efi_loader.h index 7bd8002e303e..93f7ece814a0 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -381,6 +381,8 @@ void efi_save_gd(void); void efi_restore_gd(void); /* Call this to relocate the runtime section to an address space */ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map); +/* Call this when we start to live in a runtime only world */ +void efi_runtime_detach(ulong offset); /* Call this to set the current device name */ void efi_set_bootdev(const char *dev, const char *devnr, const char *path); /* Add a new object to the object list. */ diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index 7d1d6e92138e..e4abaf3601d9 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1932,6 +1932,7 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, bootm_disable_interrupts();
/* Disable boot time services */ + efi_runtime_detach((ulong)gd->relocaddr); systab.con_in_handle = NULL; systab.con_in = NULL; systab.con_out_handle = NULL; diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index ff3684a4b692..fc5bdee80e00 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -433,7 +433,7 @@ struct efi_runtime_detach_list_struct { void *patchto; };
-static const struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { +static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { { /* do_reset is gone */ .ptr = &efi_runtime_services.reset_system, @@ -490,7 +490,15 @@ static bool efi_runtime_tobedetached(void *p) return false; }
-static void efi_runtime_detach(ulong offset) +/** + * efi_runtime_detach() - Remove any dependency on non-runtime sections + * + * This function patches all remaining code to be self-sufficient inside + * runtime sections. Any calls to non-runtime will be removed after this. + * + * @offset: relocaddr for pre-set_v_a_space, offset to VA after + */ +__efi_runtime void efi_runtime_detach(ulong offset) { int i; ulong patchoff = offset - (ulong)gd->relocaddr; @@ -506,6 +514,8 @@ static void efi_runtime_detach(ulong offset)
/* Update CRC32 */ efi_update_table_header_crc32(&efi_runtime_services.hdr); + + invalidate_icache_all(); }
/* Relocate EFI runtime to uboot_reloc_base = offset */ @@ -593,19 +603,25 @@ void efi_runtime_relocate(ulong offset, struct efi_mem_desc *map) * @virtmap: virtual address mapping information * Return: status code */ -static efi_status_t EFIAPI efi_set_virtual_address_map( +static __efi_runtime efi_status_t EFIAPI efi_set_virtual_address_map( unsigned long memory_map_size, unsigned long descriptor_size, uint32_t descriptor_version, struct efi_mem_desc *virtmap) { + static __efi_runtime_data bool is_patched; int n = memory_map_size / descriptor_size; int i; int rt_code_sections = 0;
+ if (is_patched) + return EFI_INVALID_PARAMETER; + EFI_ENTRY("%lx %lx %x %p", memory_map_size, descriptor_size, descriptor_version, virtmap);
+ is_patched = true; + /* * TODO: * Further down we are cheating. While really we should implement @@ -681,8 +697,7 @@ static efi_status_t EFIAPI efi_set_virtual_address_map( map->physical_start + gd->relocaddr;
efi_runtime_relocate(new_offset, map); - /* Once we're virtual, we can no longer handle - complex callbacks */ + /* We need to repatch callbacks for their new VA */ efi_runtime_detach(new_offset);
/* diff --git a/test/py/tests/test_efi_selftest.py b/test/py/tests/test_efi_selftest.py index 07e4db045284..f950911655a3 100644 --- a/test/py/tests/test_efi_selftest.py +++ b/test/py/tests/test_efi_selftest.py @@ -20,7 +20,7 @@ def test_efi_selftest(u_boot_console): if m != 0: raise Exception('Failures occurred during the EFI selftest') u_boot_console.run_command(cmd='', wait_for_echo=False, wait_for_prompt=False); - m = u_boot_console.p.expect(['resetting', 'U-Boot']) + m = u_boot_console.p.expect(['U-Boot']) if m != 0: raise Exception('Reset failed during the EFI selftest') u_boot_console.restart_uboot(); @@ -47,7 +47,7 @@ def test_efi_selftest_watchdog_reboot(u_boot_console): assert ''watchdog reboot'' in output u_boot_console.run_command(cmd='setenv efi_selftest watchdog reboot') u_boot_console.run_command(cmd='bootefi selftest', wait_for_prompt=False) - m = u_boot_console.p.expect(['resetting', 'U-Boot']) + m = u_boot_console.p.expect(['U-Boot']) if m != 0: raise Exception('Reset failed in 'watchdog reboot' test') u_boot_console.restart_uboot();

With "boot exit" sub-command, we will let the system to exit UEFI Boottime Services. This is mainly for testing.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org --- cmd/efidebug.c | 63 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 63 insertions(+)
diff --git a/cmd/efidebug.c b/cmd/efidebug.c index e65722625455..d6abf08c93cc 100644 --- a/cmd/efidebug.c +++ b/cmd/efidebug.c @@ -977,6 +977,66 @@ out: return ret; }
+/** + * do_efi_boot_exit() - exit boottime services + * + * @cmdtp: Command table + * @flag: Command flag + * @argc: Number of arguments + * @argv: Argument array + * Return: CMD_RET_SUCCESS on success, + * CMD_RET_USAGE or CMD_RET_RET_FAILURE on failure + * + * Implement efidebug "boot exit" sub-command. + * - boot exit + */ +static int do_efi_boot_exit(cmd_tbl_t *cmdtp, int flag, + int argc, char * const argv[]) +{ + struct efi_mem_desc *memmap; + efi_uintn_t map_size, map_key, desc_size; + u32 desc_version; + efi_status_t ret; + + if (argc != 1) + return CMD_RET_USAGE; + + memmap = NULL; + map_size = 0; + ret = EFI_CALL(BS->get_memory_map(&map_size, memmap, &map_key, + &desc_size, &desc_version)); + if (ret == EFI_BUFFER_TOO_SMALL) { + map_size += sizeof(struct efi_mem_desc); /* for my own */ + ret = EFI_CALL(BS->allocate_pool(EFI_LOADER_DATA, + map_size, (void *)&memmap)); + if (ret != EFI_SUCCESS) { + printf("Out of memory\n"); + return CMD_RET_FAILURE; + } + ret = EFI_CALL(BS->get_memory_map(&map_size, memmap, + &map_key, + &desc_size, &desc_version)); + } + if (ret != EFI_SUCCESS) { + printf("Getting UEFI memory map failed (%zu)\n", + ret & ~EFI_ERROR_MASK); + goto err; + } + + ret = EFI_CALL(BS->exit_boot_services(efi_root, map_key)); + if (ret == EFI_SUCCESS) + return CMD_RET_SUCCESS; + + printf("exiting Boottime Services failed (%zu)\n", + ret & ~EFI_ERROR_MASK); + +err: + if (memmap) + EFI_CALL(BS->free_pool(memmap)); + + return CMD_RET_FAILURE; +} + static cmd_tbl_t cmd_efidebug_boot_sub[] = { U_BOOT_CMD_MKENT(add, CONFIG_SYS_MAXARGS, 1, do_efi_boot_add, "", ""), U_BOOT_CMD_MKENT(rm, CONFIG_SYS_MAXARGS, 1, do_efi_boot_rm, "", ""), @@ -984,6 +1044,7 @@ static cmd_tbl_t cmd_efidebug_boot_sub[] = { U_BOOT_CMD_MKENT(next, CONFIG_SYS_MAXARGS, 1, do_efi_boot_next, "", ""), U_BOOT_CMD_MKENT(order, CONFIG_SYS_MAXARGS, 1, do_efi_boot_order, "", ""), + U_BOOT_CMD_MKENT(exit, CONFIG_SYS_MAXARGS, 1, do_efi_boot_exit, "", ""), };
/** @@ -1087,6 +1148,8 @@ static char efidebug_help_text[] = " - set UEFI BootNext variable\n" "efidebug boot order [<bootid#1> [<bootid#2> [<bootid#3> [...]]]]\n" " - set/show UEFI boot order\n" + "efidebug boot exit\n" + " - exit UEFI Boottime Services\n" "\n" "efidebug devices\n" " - show uefi devices\n"

With this patch, cache buffer for UEFI variables will be created so that we will still be able to access, at least retrieve, UEFI variables when we exit from boottime services,
This feature is a "should" behavior described in EBBR v1.0 section 2.5.3.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org --- include/efi_loader.h | 17 ++ lib/efi_loader/Kconfig | 9 + lib/efi_loader/efi_boottime.c | 10 +- lib/efi_loader/efi_runtime.c | 13 + lib/efi_loader/efi_variable.c | 467 ++++++++++++++++++++++++++++++++++ 5 files changed, 515 insertions(+), 1 deletion(-)
diff --git a/include/efi_loader.h b/include/efi_loader.h index 93f7ece814a0..acab657b9d70 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -620,6 +620,23 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name, const efi_guid_t *vendor, u32 attributes, efi_uintn_t data_size, const void *data);
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +efi_status_t efi_freeze_variable_table(void); + +/* runtime version of APIs */ +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name, + const efi_guid_t *vendor, + u32 *attributes, + efi_uintn_t *data_size, + void *data); +efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime( + efi_uintn_t *variable_name_size, + u16 *variable_name, + const efi_guid_t *vendor); +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */ + /* * See section 3.1.3 in the v2.7 UEFI spec for more details on * the layout of EFI_LOAD_OPTION. In short it is: diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index e2ef43157568..3f284795648f 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -59,6 +59,15 @@ config EFI_RUNTIME_CONVERT_POINTER to be called by UEFI drivers in relocating themselves to virtual address space.
+config EFI_RUNTIME_GET_VARIABLE_CACHING + bool "runtime_service: GetVariable: Enable runtime access via cache (read-only)" + default y + help + Select this option if you want to access UEFI variables at + runtime even though you cannot update values on the fly. + With or without this option, you can access UEFI variables + at boottime. + config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index e4abaf3601d9..14e343abbd43 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1892,6 +1892,9 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, efi_uintn_t map_key) { struct efi_event *evt; +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING + efi_status_t ret; +#endif
EFI_ENTRY("%p, %zx", image_handle, map_key);
@@ -1921,7 +1924,12 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, } }
- /* TODO: Should persist EFI variables here */ +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING + /* No more variable update */ + ret = efi_freeze_variable_table(); + if (ret != EFI_SUCCESS) + return EFI_EXIT(ret); +#endif
board_quiesce_devices();
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index fc5bdee80e00..b60f70f04613 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -111,6 +111,11 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_CONVERT_POINTER; #endif +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING + efi_runtime_services_supported |= + (EFI_RT_SUPPORTED_GET_VARIABLE | + EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME); +#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -469,10 +474,18 @@ static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { .patchto = NULL, }, { .ptr = &efi_runtime_services.get_variable, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING + .patchto = &efi_get_variable_runtime, +#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.get_next_variable_name, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING + .patchto = &efi_get_next_variable_name, +#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.set_variable, .patchto = &efi_device_error, diff --git a/lib/efi_loader/efi_variable.c b/lib/efi_loader/efi_variable.c index d9887be938c2..ee21892dd291 100644 --- a/lib/efi_loader/efi_variable.c +++ b/lib/efi_loader/efi_variable.c @@ -706,3 +706,470 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name,
return EFI_EXIT(ret); } + +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +/* + * runtime version of APIs + * We only support read-only variable access. + * The table is in U-Boot's hash table format, but has its own + * _ENTRY structure for specific use. + * + * Except for efi_freeze_variable_table(), which is to be called in + * exit_boot_services(), all the functions and data below must be + * placed in either RUNTIME_SERVICES_CODE or RUNTIME_SERVICES_DATA. + */ +typedef struct _ENTRY { + unsigned int used; /* hash value; 0 for not used */ + size_t name; /* name offset from itself */ + efi_guid_t vendor; + u32 attributes; + size_t data; /* data offset from itself */ + size_t data_size; +} _ENTRY; + +static inline u16 *entry_name(_ENTRY *e) { return (void *)e + e->name; } +static inline u16 *entry_data(_ENTRY *e) { return (void *)e + e->data; } + +static struct hsearch_data *efi_variable_table __efi_runtime_data; + +static size_t __efi_runtime u16_strlen_runtime(const u16 *s1) +{ + size_t n = 0; + + while (*s1) { + n++; + s1++; + } + + return n; +} + +static int __efi_runtime memcmp_runtime(const void *m1, const void *m2, + size_t n) +{ + while (n && *(u8 *)m1 == *(u8 *)m2) { + n--; + m1++; + m2++; + } + + if (n) + return *(u8 *)m1 - *(u8 *)m2; + + return 0; +} + +static void __efi_runtime memcpy_runtime(void *m1, const void *m2, size_t n) +{ + for (; n; n--, m1++, m2++) + *(u8 *)m1 = *(u8 *)m2; +} + +static int __efi_runtime efi_cmpkey(_ENTRY *e, const u16 *name, + const efi_guid_t *vendor) +{ + size_t name_len; + + name_len = u16_strlen_runtime(entry_name(e)); + + /* return zero if matched */ + return name_len != u16_strlen_runtime(name) || + memcmp_runtime(entry_name(e), name, name_len * 2) || + memcmp_runtime(e->vendor.b, vendor->b, sizeof(vendor)); +} + +/* simplified and slightly different version of hsearch_r() */ +static int __efi_runtime hsearch_runtime(const u16 *name, + const efi_guid_t *vendor, + ACTION action, + _ENTRY **retval, + struct hsearch_data *htab) +{ + unsigned int hval; + unsigned int count; + unsigned int len; + unsigned int idx, new; + + /* Compute an value for the given string. */ + len = u16_strlen_runtime(name); + hval = len; + count = len; + while (count-- > 0) { + hval <<= 4; + hval += name[count]; + } + + /* + * First hash function: + * simply take the modulo but prevent zero. + */ + hval %= htab->size; + if (hval == 0) + ++hval; + + /* The first index tried. */ + new = -1; /* not found */ + idx = hval; + + if (htab->table[idx].used) { + /* + * Further action might be required according to the + * action value. + */ + unsigned int hval2; + + if (htab->table[idx].used == hval && + !efi_cmpkey(&htab->table[idx], name, vendor)) { + if (action == FIND) { + *retval = &htab->table[idx]; + return idx; + } + /* we don't need to support overwrite */ + return -1; + } + + /* + * Second hash function: + * as suggested in [Knuth] + */ + hval2 = 1 + hval % (htab->size - 2); + + do { + /* + * Because SIZE is prime this guarantees to + * step through all available indices. + */ + if (idx <= hval2) + idx = htab->size + idx - hval2; + else + idx -= hval2; + + /* + * If we visited all entries leave the loop + * unsuccessfully. + */ + if (idx == hval) + break; + + /* If entry is found use it. */ + if (htab->table[idx].used == hval && + !efi_cmpkey(&htab->table[idx], name, vendor)) { + if (action == FIND) { + *retval = &htab->table[idx]; + return idx; + } + /* we don't need to support overwrite */ + return -1; + } + } while (htab->table[idx].used); + + if (!htab->table[idx].used) + new = idx; + } else { + new = idx; + } + + /* + * An empty bucket has been found. + * The following code should never be executed after + * exit_boot_services() + */ + if (action == ENTER) { + /* + * If table is full and another entry should be + * entered return with error. + */ + if (htab->filled == htab->size) { + *retval = NULL; + return 0; + } + + /* Create new entry */ + htab->table[new].used = hval; + ++htab->filled; + + /* return new entry */ + *retval = &htab->table[new]; + return 1; + } + + *retval = NULL; + return 0; +} + +/* from lib/hashtable.c */ +static inline int isprime(unsigned int number) +{ + /* no even number will be passed */ + unsigned int div = 3; + + while (div * div < number && number % div != 0) + div += 2; + + return number % div != 0; +} + +efi_status_t efi_freeze_variable_table(void) +{ + int var_num = 0; + size_t var_data_size = 0; + u16 *name; + efi_uintn_t name_buf_len, name_len; + efi_guid_t vendor; + u32 attributes; + u8 *mem_pool, *var_buf = NULL; + size_t table_size, var_size, var_buf_size; + _ENTRY *new = NULL; + efi_status_t ret; + + /* phase-1 loop */ + name_buf_len = 128; + name = malloc(name_buf_len); + if (!name) + return EFI_OUT_OF_RESOURCES; + name[0] = 0; + for (;;) { + name_len = name_buf_len; + ret = EFI_CALL(efi_get_next_variable_name(&name_len, name, + &vendor)); + if (ret == EFI_NOT_FOUND) { + break; + } else if (ret == EFI_BUFFER_TOO_SMALL) { + u16 *buf; + + name_buf_len = name_len; + buf = realloc(name, name_buf_len); + if (!buf) { + free(name); + return EFI_OUT_OF_RESOURCES; + } + name = buf; + name_len = name_buf_len; + ret = EFI_CALL(efi_get_next_variable_name(&name_len, + name, + &vendor)); + } + + if (ret != EFI_SUCCESS) + return ret; + + var_size = 0; + ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes, + &var_size, NULL)); + if (ret != EFI_BUFFER_TOO_SMALL) + return ret; + + if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS)) + continue; + + var_num++; + var_data_size += (u16_strlen_runtime(name) + 1) * sizeof(u16); + var_data_size += var_size; + /* mem_pool must 2-byte aligned for u16 variable name */ + if (var_data_size & 0x1) + var_data_size++; + } + + /* + * total of entries in hash table must be a prime number. + * The logic below comes from lib/hashtable.c + */ + var_num |= 1; /* make odd */ + while (!isprime(var_num)) + var_num += 2; + + /* We need table[var_num] for hsearch_runtime algo */ + table_size = sizeof(*efi_variable_table) + + sizeof(_ENTRY) * (var_num + 1) + var_data_size; + ret = efi_allocate_pool(EFI_RUNTIME_SERVICES_DATA, + table_size, (void **)&efi_variable_table); + if (ret != EFI_SUCCESS) + return ret; + + efi_variable_table->size = var_num; + efi_variable_table->table = (void *)efi_variable_table + + sizeof(*efi_variable_table); + mem_pool = (u8 *)efi_variable_table->table + + sizeof(_ENTRY) * (var_num + 1); + + var_buf_size = 128; + var_buf = malloc(var_buf_size); + if (!var_buf) { + ret = EFI_OUT_OF_RESOURCES; + goto err; + } + + /* phase-2 loop */ + name[0] = 0; + name_len = name_buf_len; + for (;;) { + name_len = name_buf_len; + ret = EFI_CALL(efi_get_next_variable_name(&name_len, name, + &vendor)); + if (ret == EFI_NOT_FOUND) + break; + else if (ret != EFI_SUCCESS) + goto err; + + var_size = var_buf_size; + ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes, + &var_size, var_buf)); + if (ret == EFI_BUFFER_TOO_SMALL) { + free(var_buf); + var_buf_size = var_size; + var_buf = malloc(var_buf_size); + if (!var_buf) { + ret = EFI_OUT_OF_RESOURCES; + goto err; + } + ret = EFI_CALL(efi_get_variable(name, &vendor, + &attributes, + &var_size, var_buf)); + } + if (ret != EFI_SUCCESS) + goto err; + + if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS)) + continue; + + if (hsearch_runtime(name, &vendor, ENTER, &new, + efi_variable_table) <= 0) { + /* This should not happen */ + ret = EFI_INVALID_PARAMETER; + goto err; + } + + /* allocate space from RUNTIME DATA */ + name_len = (u16_strlen_runtime(name) + 1) * sizeof(u16); + memcpy_runtime(mem_pool, name, name_len); + new->name = mem_pool - (u8 *)new; /* offset */ + mem_pool += name_len; + + memcpy_runtime(&new->vendor.b, &vendor.b, sizeof(vendor)); + + new->attributes = attributes; + + memcpy_runtime(mem_pool, var_buf, var_size); + new->data = mem_pool - (u8 *)new; /* offset */ + new->data_size = var_size; + mem_pool += var_size; + + /* mem_pool must 2-byte aligned for u16 variable name */ + if ((uintptr_t)mem_pool & 0x1) + mem_pool++; + } +#ifdef DEBUG + name[0] = 0; + name_len = name_buf_len; + for (;;) { + name_len = name_buf_len; + ret = efi_get_next_variable_name_runtime(&name_len, name, + &vendor); + if (ret == EFI_NOT_FOUND) + break; + else if (ret != EFI_SUCCESS) + goto err; + + var_size = var_buf_size; + ret = efi_get_variable_runtime(name, &vendor, &attributes, + &var_size, var_buf); + if (ret != EFI_SUCCESS) + goto err; + + printf("%ls_%pUl:\n", name, &vendor); + printf(" attributes: 0x%x\n", attributes); + printf(" value (size: 0x%lx)\n", var_size); + } +#endif + ret = EFI_SUCCESS; + +err: + free(name); + free(var_buf); + if (ret != EFI_SUCCESS && efi_variable_table) { + efi_free_pool(efi_variable_table); + efi_variable_table = NULL; + } + + return ret; +} + +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name, + const efi_guid_t *vendor, + u32 *attributes, + efi_uintn_t *data_size, + void *data) +{ + _ENTRY *new; + + if (!variable_name || !vendor || !data_size) + return EFI_EXIT(EFI_INVALID_PARAMETER); + + if (hsearch_runtime(variable_name, vendor, FIND, &new, + efi_variable_table) <= 0) + return EFI_NOT_FOUND; + + if (attributes) + *attributes = new->attributes; + if (*data_size < new->data_size) { + *data_size = new->data_size; + return EFI_BUFFER_TOO_SMALL; + } + + *data_size = new->data_size; + memcpy_runtime(data, entry_data(new), new->data_size); + + return EFI_SUCCESS; +} + +static int prev_idx __efi_runtime_data; + +efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime( + efi_uintn_t *variable_name_size, + u16 *variable_name, + const efi_guid_t *vendor) +{ + _ENTRY *e; + u16 *name; + efi_uintn_t name_size; + + if (!variable_name_size || !variable_name || !vendor) + return EFI_INVALID_PARAMETER; + + if (variable_name[0]) { + /* sanity check for previous variable */ + if (prev_idx < 0) + return EFI_INVALID_PARAMETER; + + e = &efi_variable_table->table[prev_idx]; + if (!e->used || efi_cmpkey(e, variable_name, vendor)) + return EFI_INVALID_PARAMETER; + } else { + prev_idx = -1; + } + + /* next variable */ + while (++prev_idx <= efi_variable_table->size) { + e = &efi_variable_table->table[prev_idx]; + if (e->used) + break; + } + if (prev_idx > efi_variable_table->size) + return EFI_NOT_FOUND; + + name = entry_name(e); + name_size = (u16_strlen_runtime(name) + 1) + * sizeof(u16); + if (*variable_name_size < name_size) { + *variable_name_size = name_size; + return EFI_BUFFER_TOO_SMALL; + } + + memcpy_runtime(variable_name, name, name_size); + memcpy_runtime((void *)&vendor->b, &e->vendor.b, sizeof(vendor)); + + return EFI_SUCCESS; +} +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */

On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, cache buffer for UEFI variables will be created so that we will still be able to access, at least retrieve, UEFI variables when we exit from boottime services,
This feature is a "should" behavior described in EBBR v1.0 section 2.5.3.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_loader.h | 17 ++ lib/efi_loader/Kconfig | 9 + lib/efi_loader/efi_boottime.c | 10 +- lib/efi_loader/efi_runtime.c | 13 + lib/efi_loader/efi_variable.c | 467 ++++++++++++++++++++++++++++++++++ 5 files changed, 515 insertions(+), 1 deletion(-)
Please, put the cache into a separate file.
diff --git a/include/efi_loader.h b/include/efi_loader.h index 93f7ece814a0..acab657b9d70 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -620,6 +620,23 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name, const efi_guid_t *vendor, u32 attributes, efi_uintn_t data_size, const void *data);
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +efi_status_t efi_freeze_variable_table(void);
+/* runtime version of APIs */ +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime. Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data);
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor);
+#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */
- /*
- See section 3.1.3 in the v2.7 UEFI spec for more details on
- the layout of EFI_LOAD_OPTION. In short it is:
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index e2ef43157568..3f284795648f 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -59,6 +59,15 @@ config EFI_RUNTIME_CONVERT_POINTER to be called by UEFI drivers in relocating themselves to virtual address space.
+config EFI_RUNTIME_GET_VARIABLE_CACHING
- bool "runtime_service: GetVariable: Enable runtime access via cache (read-only)"
- default y
- help
Select this option if you want to access UEFI variables at
runtime even though you cannot update values on the fly.
With or without this option, you can access UEFI variables
at boottime.
Updates of volatile variables should always be possible.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index e4abaf3601d9..14e343abbd43 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1892,6 +1892,9 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, efi_uintn_t map_key) { struct efi_event *evt; +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_status_t ret;
+#endif
EFI_ENTRY("%p, %zx", image_handle, map_key);
@@ -1921,7 +1924,12 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, } }
- /* TODO: Should persist EFI variables here */
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
Can we have weak functions for initializing and persisting the cache, e.g.
efi_status_t __weak efi_load_variable_cache(cache_entry *cache, size_t *size) { Â Â Â Â Â Â Â Â cache->len = 0; Â Â Â Â Â Â Â Â return EFI_SUCCESS; }
efi_status_t __runtime __weak efi_write_variable_cache(cache_entry *cache, size_t size) { Â Â Â Â Â Â Â Â return EFI_UNSUPPORTED; }
Then we can override these in whatever driver we implement.
- /* No more variable update */
- ret = efi_freeze_variable_table();
- if (ret != EFI_SUCCESS)
return EFI_EXIT(ret);
+#endif
board_quiesce_devices();
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index fc5bdee80e00..b60f70f04613 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -111,6 +111,11 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_CONVERT_POINTER; #endif +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_runtime_services_supported |=
(EFI_RT_SUPPORTED_GET_VARIABLE |
EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME);
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -469,10 +474,18 @@ static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { .patchto = NULL, }, { .ptr = &efi_runtime_services.get_variable, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_variable_runtime,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.get_next_variable_name, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_next_variable_name,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.set_variable, .patchto = &efi_device_error, diff --git a/lib/efi_loader/efi_variable.c b/lib/efi_loader/efi_variable.c index d9887be938c2..ee21892dd291 100644 --- a/lib/efi_loader/efi_variable.c +++ b/lib/efi_loader/efi_variable.c @@ -706,3 +706,470 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name,
return EFI_EXIT(ret); }
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +/*
- runtime version of APIs
- We only support read-only variable access.
- The table is in U-Boot's hash table format, but has its own
- _ENTRY structure for specific use.
- Except for efi_freeze_variable_table(), which is to be called in
- exit_boot_services(), all the functions and data below must be
- placed in either RUNTIME_SERVICES_CODE or RUNTIME_SERVICES_DATA.
- */
+typedef struct _ENTRY {
- unsigned int used; /* hash value; 0 for not used */
- size_t name; /* name offset from itself */
- efi_guid_t vendor;
- u32 attributes;
- size_t data; /* data offset from itself */
- size_t data_size;
+} _ENTRY;
+static inline u16 *entry_name(_ENTRY *e) { return (void *)e + e->name; } +static inline u16 *entry_data(_ENTRY *e) { return (void *)e + e->data; }
+static struct hsearch_data *efi_variable_table __efi_runtime_data;
+static size_t __efi_runtime u16_strlen_runtime(const u16 *s1)
Please, do not duplicate existing functions. If they have to be runtime simply change the existing function to __runtime.
+{
- size_t n = 0;
- while (*s1) {
n++;
s1++;
- }
- return n;
+}
+static int __efi_runtime memcmp_runtime(const void *m1, const void *m2,
size_t n)
I dislike duplicate code. Can't we simply define the existing memcmp function as __runtime?
+{
- while (n && *(u8 *)m1 == *(u8 *)m2) {
n--;
m1++;
m2++;
- }
- if (n)
return *(u8 *)m1 - *(u8 *)m2;
- return 0;
+}
+static void __efi_runtime memcpy_runtime(void *m1, const void *m2, size_t n) +{
Can't we simply define the existing memcpy function as __runtime?
- for (; n; n--, m1++, m2++)
*(u8 *)m1 = *(u8 *)m2;
+}
+static int __efi_runtime efi_cmpkey(_ENTRY *e, const u16 *name,
const efi_guid_t *vendor)
+{
- size_t name_len;
- name_len = u16_strlen_runtime(entry_name(e));
- /* return zero if matched */
- return name_len != u16_strlen_runtime(name) ||
memcmp_runtime(entry_name(e), name, name_len * 2) ||
memcmp_runtime(e->vendor.b, vendor->b, sizeof(vendor));
+}
+/* simplified and slightly different version of hsearch_r() */
These hash functions are so complicated that they really need a unit test testing them rigorously.
+static int __efi_runtime hsearch_runtime(const u16 *name,
const efi_guid_t *vendor,
ACTION action,
_ENTRY **retval,
struct hsearch_data *htab)
+{
- unsigned int hval;
- unsigned int count;
- unsigned int len;
- unsigned int idx, new;
- /* Compute an value for the given string. */
- len = u16_strlen_runtime(name);
Can't the same variable name exist for different GUIDs? Why is the GUID not considered in the hash?
- hval = len;
- count = len;
- while (count-- > 0) {
hval <<= 4;
hval += name[count];
- }
- /*
* First hash function:
* simply take the modulo but prevent zero.
*/
- hval %= htab->size;
- if (hval == 0)
++hval;
- /* The first index tried. */
- new = -1; /* not found */
- idx = hval;
- if (htab->table[idx].used) {
/*
* Further action might be required according to the
* action value.
*/
unsigned int hval2;
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
/*
* Second hash function:
* as suggested in [Knuth]
*/
hval2 = 1 + hval % (htab->size - 2);
do {
/*
* Because SIZE is prime this guarantees to
* step through all available indices.
*/
if (idx <= hval2)
idx = htab->size + idx - hval2;
else
idx -= hval2;
/*
* If we visited all entries leave the loop
* unsuccessfully.
*/
if (idx == hval)
break;
/* If entry is found use it. */
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
} while (htab->table[idx].used);
if (!htab->table[idx].used)
new = idx;
- } else {
new = idx;
- }
- /*
* An empty bucket has been found.
* The following code should never be executed after
* exit_boot_services()
*/
- if (action == ENTER) {
/*
* If table is full and another entry should be
* entered return with error.
*/
if (htab->filled == htab->size) {
*retval = NULL;
return 0;
}
/* Create new entry */
htab->table[new].used = hval;
++htab->filled;
/* return new entry */
*retval = &htab->table[new];
return 1;
- }
- *retval = NULL;
- return 0;
+}
+/* from lib/hashtable.c */ +static inline int isprime(unsigned int number) +{
- /* no even number will be passed */
- unsigned int div = 3;
- while (div * div < number && number % div != 0)
div += 2;
- return number % div != 0;
+}
+efi_status_t efi_freeze_variable_table(void)
Please, add comments to your functions. It is not self-evident what this function is meant to do.
I cannot imagine why a variable cache should be frozen. It is a living data structure until the system is switched off.
For a variable cache I expect that you allocate memory before handling the first variable and never again. At runtime you will not have chance to allocate memory anyway.
This function is way too long. Pleae, break it down.
Best regards
Heinrich
+{
- int var_num = 0;
- size_t var_data_size = 0;
- u16 *name;
- efi_uintn_t name_buf_len, name_len;
- efi_guid_t vendor;
- u32 attributes;
- u8 *mem_pool, *var_buf = NULL;
- size_t table_size, var_size, var_buf_size;
- _ENTRY *new = NULL;
- efi_status_t ret;
- /* phase-1 loop */
- name_buf_len = 128;
- name = malloc(name_buf_len);
- if (!name)
return EFI_OUT_OF_RESOURCES;
- name[0] = 0;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND) {
break;
} else if (ret == EFI_BUFFER_TOO_SMALL) {
u16 *buf;
name_buf_len = name_len;
buf = realloc(name, name_buf_len);
if (!buf) {
free(name);
return EFI_OUT_OF_RESOURCES;
}
name = buf;
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len,
name,
&vendor));
}
if (ret != EFI_SUCCESS)
return ret;
var_size = 0;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, NULL));
if (ret != EFI_BUFFER_TOO_SMALL)
return ret;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
var_num++;
var_data_size += (u16_strlen_runtime(name) + 1) * sizeof(u16);
var_data_size += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if (var_data_size & 0x1)
var_data_size++;
- }
- /*
* total of entries in hash table must be a prime number.
* The logic below comes from lib/hashtable.c
*/
- var_num |= 1; /* make odd */
- while (!isprime(var_num))
var_num += 2;
- /* We need table[var_num] for hsearch_runtime algo */
- table_size = sizeof(*efi_variable_table)
+ sizeof(_ENTRY) * (var_num + 1) + var_data_size;
- ret = efi_allocate_pool(EFI_RUNTIME_SERVICES_DATA,
table_size, (void **)&efi_variable_table);
- if (ret != EFI_SUCCESS)
return ret;
- efi_variable_table->size = var_num;
- efi_variable_table->table = (void *)efi_variable_table
+ sizeof(*efi_variable_table);
- mem_pool = (u8 *)efi_variable_table->table
+ sizeof(_ENTRY) * (var_num + 1);
- var_buf_size = 128;
- var_buf = malloc(var_buf_size);
- if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
- }
- /* phase-2 loop */
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, var_buf));
if (ret == EFI_BUFFER_TOO_SMALL) {
free(var_buf);
var_buf_size = var_size;
var_buf = malloc(var_buf_size);
if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
}
ret = EFI_CALL(efi_get_variable(name, &vendor,
&attributes,
&var_size, var_buf));
}
if (ret != EFI_SUCCESS)
goto err;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
if (hsearch_runtime(name, &vendor, ENTER, &new,
efi_variable_table) <= 0) {
/* This should not happen */
ret = EFI_INVALID_PARAMETER;
goto err;
}
/* allocate space from RUNTIME DATA */
name_len = (u16_strlen_runtime(name) + 1) * sizeof(u16);
memcpy_runtime(mem_pool, name, name_len);
new->name = mem_pool - (u8 *)new; /* offset */
mem_pool += name_len;
memcpy_runtime(&new->vendor.b, &vendor.b, sizeof(vendor));
new->attributes = attributes;
memcpy_runtime(mem_pool, var_buf, var_size);
new->data = mem_pool - (u8 *)new; /* offset */
new->data_size = var_size;
mem_pool += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if ((uintptr_t)mem_pool & 0x1)
mem_pool++;
- }
+#ifdef DEBUG
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = efi_get_next_variable_name_runtime(&name_len, name,
&vendor);
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = efi_get_variable_runtime(name, &vendor, &attributes,
&var_size, var_buf);
if (ret != EFI_SUCCESS)
goto err;
printf("%ls_%pUl:\n", name, &vendor);
printf(" attributes: 0x%x\n", attributes);
printf(" value (size: 0x%lx)\n", var_size);
- }
+#endif
- ret = EFI_SUCCESS;
+err:
- free(name);
- free(var_buf);
- if (ret != EFI_SUCCESS && efi_variable_table) {
efi_free_pool(efi_variable_table);
efi_variable_table = NULL;
- }
- return ret;
+}
+efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data)
+{
- _ENTRY *new;
- if (!variable_name || !vendor || !data_size)
return EFI_EXIT(EFI_INVALID_PARAMETER);
- if (hsearch_runtime(variable_name, vendor, FIND, &new,
efi_variable_table) <= 0)
return EFI_NOT_FOUND;
- if (attributes)
*attributes = new->attributes;
- if (*data_size < new->data_size) {
*data_size = new->data_size;
return EFI_BUFFER_TOO_SMALL;
- }
- *data_size = new->data_size;
- memcpy_runtime(data, entry_data(new), new->data_size);
- return EFI_SUCCESS;
+}
+static int prev_idx __efi_runtime_data;
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor)
+{
- _ENTRY *e;
- u16 *name;
- efi_uintn_t name_size;
- if (!variable_name_size || !variable_name || !vendor)
return EFI_INVALID_PARAMETER;
- if (variable_name[0]) {
/* sanity check for previous variable */
if (prev_idx < 0)
return EFI_INVALID_PARAMETER;
e = &efi_variable_table->table[prev_idx];
if (!e->used || efi_cmpkey(e, variable_name, vendor))
return EFI_INVALID_PARAMETER;
- } else {
prev_idx = -1;
- }
- /* next variable */
- while (++prev_idx <= efi_variable_table->size) {
e = &efi_variable_table->table[prev_idx];
if (e->used)
break;
- }
- if (prev_idx > efi_variable_table->size)
return EFI_NOT_FOUND;
- name = entry_name(e);
- name_size = (u16_strlen_runtime(name) + 1)
* sizeof(u16);
- if (*variable_name_size < name_size) {
*variable_name_size = name_size;
return EFI_BUFFER_TOO_SMALL;
- }
- memcpy_runtime(variable_name, name, name_size);
- memcpy_runtime((void *)&vendor->b, &e->vendor.b, sizeof(vendor));
- return EFI_SUCCESS;
+} +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */

On Sat, Jun 15, 2019 at 09:01:56PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, cache buffer for UEFI variables will be created so that we will still be able to access, at least retrieve, UEFI variables when we exit from boottime services,
This feature is a "should" behavior described in EBBR v1.0 section 2.5.3.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_loader.h | 17 ++ lib/efi_loader/Kconfig | 9 + lib/efi_loader/efi_boottime.c | 10 +- lib/efi_loader/efi_runtime.c | 13 + lib/efi_loader/efi_variable.c | 467 ++++++++++++++++++++++++++++++++++ 5 files changed, 515 insertions(+), 1 deletion(-)
Please, put the cache into a separate file.
Why?
diff --git a/include/efi_loader.h b/include/efi_loader.h index 93f7ece814a0..acab657b9d70 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -620,6 +620,23 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name, const efi_guid_t *vendor, u32 attributes, efi_uintn_t data_size, const void *data);
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +efi_status_t efi_freeze_variable_table(void);
+/* runtime version of APIs */ +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data);
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor);
+#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */
/*
- See section 3.1.3 in the v2.7 UEFI spec for more details on
- the layout of EFI_LOAD_OPTION. In short it is:
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index e2ef43157568..3f284795648f 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -59,6 +59,15 @@ config EFI_RUNTIME_CONVERT_POINTER to be called by UEFI drivers in relocating themselves to virtual address space.
+config EFI_RUNTIME_GET_VARIABLE_CACHING
- bool "runtime_service: GetVariable: Enable runtime access via cache (read-only)"
- default y
- help
Select this option if you want to access UEFI variables at
runtime even though you cannot update values on the fly.
With or without this option, you can access UEFI variables
at boottime.
Updates of volatile variables should always be possible.
Why "should"? Give me any use case. UEFI spec does not describe such a variant implementation at all.
config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index e4abaf3601d9..14e343abbd43 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1892,6 +1892,9 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, efi_uintn_t map_key) { struct efi_event *evt; +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_status_t ret;
+#endif
EFI_ENTRY("%p, %zx", image_handle, map_key);
@@ -1921,7 +1924,12 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, } }
- /* TODO: Should persist EFI variables here */
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
Can we have weak functions for initializing and persisting the cache, e.g.
efi_status_t __weak efi_load_variable_cache(cache_entry *cache, size_t *size) { Â Â Â Â Â Â Â Â cache->len = 0; Â Â Â Â Â Â Â Â return EFI_SUCCESS; }
efi_status_t __runtime __weak efi_write_variable_cache(cache_entry *cache, size_t size) { Â Â Â Â Â Â Â Â return EFI_UNSUPPORTED; }
Then we can override these in whatever driver we implement.
What is the difference between yours and my env_efi_load/save() with backing-storage driver?
- /* No more variable update */
- ret = efi_freeze_variable_table();
- if (ret != EFI_SUCCESS)
return EFI_EXIT(ret);
+#endif
board_quiesce_devices();
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index fc5bdee80e00..b60f70f04613 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -111,6 +111,11 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_CONVERT_POINTER; #endif +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_runtime_services_supported |=
(EFI_RT_SUPPORTED_GET_VARIABLE |
EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME);
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -469,10 +474,18 @@ static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { .patchto = NULL, }, { .ptr = &efi_runtime_services.get_variable, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_variable_runtime,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.get_next_variable_name, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_next_variable_name,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.set_variable, .patchto = &efi_device_error, diff --git a/lib/efi_loader/efi_variable.c b/lib/efi_loader/efi_variable.c index d9887be938c2..ee21892dd291 100644 --- a/lib/efi_loader/efi_variable.c +++ b/lib/efi_loader/efi_variable.c @@ -706,3 +706,470 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name,
return EFI_EXIT(ret); }
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +/*
- runtime version of APIs
- We only support read-only variable access.
- The table is in U-Boot's hash table format, but has its own
- _ENTRY structure for specific use.
- Except for efi_freeze_variable_table(), which is to be called in
- exit_boot_services(), all the functions and data below must be
- placed in either RUNTIME_SERVICES_CODE or RUNTIME_SERVICES_DATA.
- */
+typedef struct _ENTRY {
- unsigned int used; /* hash value; 0 for not used */
- size_t name; /* name offset from itself */
- efi_guid_t vendor;
- u32 attributes;
- size_t data; /* data offset from itself */
- size_t data_size;
+} _ENTRY;
+static inline u16 *entry_name(_ENTRY *e) { return (void *)e + e->name; } +static inline u16 *entry_data(_ENTRY *e) { return (void *)e + e->data; }
+static struct hsearch_data *efi_variable_table __efi_runtime_data;
+static size_t __efi_runtime u16_strlen_runtime(const u16 *s1)
Please, do not duplicate existing functions. If they have to be runtime simply change the existing function to __runtime.
We should do that if possible, but please note that [str|mem]xxx() functions are *architecture* dependent. Do you want to mark all the functions across all the architectures?
+{
- size_t n = 0;
- while (*s1) {
n++;
s1++;
- }
- return n;
+}
+static int __efi_runtime memcmp_runtime(const void *m1, const void *m2,
size_t n)
I dislike duplicate code. Can't we simply define the existing memcmp function as __runtime?
ditto
+{
- while (n && *(u8 *)m1 == *(u8 *)m2) {
n--;
m1++;
m2++;
- }
- if (n)
return *(u8 *)m1 - *(u8 *)m2;
- return 0;
+}
+static void __efi_runtime memcpy_runtime(void *m1, const void *m2, size_t n) +{
Can't we simply define the existing memcpy function as __runtime?
- for (; n; n--, m1++, m2++)
*(u8 *)m1 = *(u8 *)m2;
+}
+static int __efi_runtime efi_cmpkey(_ENTRY *e, const u16 *name,
const efi_guid_t *vendor)
+{
- size_t name_len;
- name_len = u16_strlen_runtime(entry_name(e));
- /* return zero if matched */
- return name_len != u16_strlen_runtime(name) ||
memcmp_runtime(entry_name(e), name, name_len * 2) ||
memcmp_runtime(e->vendor.b, vendor->b, sizeof(vendor));
+}
+/* simplified and slightly different version of hsearch_r() */
These hash functions are so complicated that they really need a unit test testing them rigorously.
Do you know that there are no any tests for hsearch_r()? Moreover, this function would better be exercised well if we could add *runtime* tests.
+static int __efi_runtime hsearch_runtime(const u16 *name,
const efi_guid_t *vendor,
ACTION action,
_ENTRY **retval,
struct hsearch_data *htab)
+{
- unsigned int hval;
- unsigned int count;
- unsigned int len;
- unsigned int idx, new;
- /* Compute an value for the given string. */
- len = u16_strlen_runtime(name);
Can't the same variable name exist for different GUIDs? Why is the GUID not considered in the hash?
efi_cmpkey() does take into consideration GUID as well as the name.
- hval = len;
- count = len;
- while (count-- > 0) {
hval <<= 4;
hval += name[count];
- }
- /*
* First hash function:
* simply take the modulo but prevent zero.
*/
- hval %= htab->size;
- if (hval == 0)
++hval;
- /* The first index tried. */
- new = -1; /* not found */
- idx = hval;
- if (htab->table[idx].used) {
/*
* Further action might be required according to the
* action value.
*/
unsigned int hval2;
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
/*
* Second hash function:
* as suggested in [Knuth]
*/
hval2 = 1 + hval % (htab->size - 2);
do {
/*
* Because SIZE is prime this guarantees to
* step through all available indices.
*/
if (idx <= hval2)
idx = htab->size + idx - hval2;
else
idx -= hval2;
/*
* If we visited all entries leave the loop
* unsuccessfully.
*/
if (idx == hval)
break;
/* If entry is found use it. */
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
} while (htab->table[idx].used);
if (!htab->table[idx].used)
new = idx;
- } else {
new = idx;
- }
- /*
* An empty bucket has been found.
* The following code should never be executed after
* exit_boot_services()
*/
- if (action == ENTER) {
/*
* If table is full and another entry should be
* entered return with error.
*/
if (htab->filled == htab->size) {
*retval = NULL;
return 0;
}
/* Create new entry */
htab->table[new].used = hval;
++htab->filled;
/* return new entry */
*retval = &htab->table[new];
return 1;
- }
- *retval = NULL;
- return 0;
+}
+/* from lib/hashtable.c */ +static inline int isprime(unsigned int number) +{
- /* no even number will be passed */
- unsigned int div = 3;
- while (div * div < number && number % div != 0)
div += 2;
- return number % div != 0;
+}
+efi_status_t efi_freeze_variable_table(void)
Please, add comments to your functions. It is not self-evident what this function is meant to do.
I cannot imagine why a variable cache should be frozen. It is a living data structure until the system is switched off.
I don't get your point.
For a variable cache I expect that you allocate memory before handling the first variable and never again. At runtime you will not have chance to allocate memory anyway.
I don't get your point.
This function is way too long. Pleae, break it down.
Freezing is implemented in 2-phase steps, and some complexity is inevitable. I suppose adding some comments would be enough.
-Takahiro Akashi
Best regards
Heinrich
+{
- int var_num = 0;
- size_t var_data_size = 0;
- u16 *name;
- efi_uintn_t name_buf_len, name_len;
- efi_guid_t vendor;
- u32 attributes;
- u8 *mem_pool, *var_buf = NULL;
- size_t table_size, var_size, var_buf_size;
- _ENTRY *new = NULL;
- efi_status_t ret;
- /* phase-1 loop */
- name_buf_len = 128;
- name = malloc(name_buf_len);
- if (!name)
return EFI_OUT_OF_RESOURCES;
- name[0] = 0;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND) {
break;
} else if (ret == EFI_BUFFER_TOO_SMALL) {
u16 *buf;
name_buf_len = name_len;
buf = realloc(name, name_buf_len);
if (!buf) {
free(name);
return EFI_OUT_OF_RESOURCES;
}
name = buf;
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len,
name,
&vendor));
}
if (ret != EFI_SUCCESS)
return ret;
var_size = 0;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, NULL));
if (ret != EFI_BUFFER_TOO_SMALL)
return ret;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
var_num++;
var_data_size += (u16_strlen_runtime(name) + 1) * sizeof(u16);
var_data_size += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if (var_data_size & 0x1)
var_data_size++;
- }
- /*
* total of entries in hash table must be a prime number.
* The logic below comes from lib/hashtable.c
*/
- var_num |= 1; /* make odd */
- while (!isprime(var_num))
var_num += 2;
- /* We need table[var_num] for hsearch_runtime algo */
- table_size = sizeof(*efi_variable_table)
+ sizeof(_ENTRY) * (var_num + 1) + var_data_size;
- ret = efi_allocate_pool(EFI_RUNTIME_SERVICES_DATA,
table_size, (void **)&efi_variable_table);
- if (ret != EFI_SUCCESS)
return ret;
- efi_variable_table->size = var_num;
- efi_variable_table->table = (void *)efi_variable_table
+ sizeof(*efi_variable_table);
- mem_pool = (u8 *)efi_variable_table->table
+ sizeof(_ENTRY) * (var_num + 1);
- var_buf_size = 128;
- var_buf = malloc(var_buf_size);
- if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
- }
- /* phase-2 loop */
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, var_buf));
if (ret == EFI_BUFFER_TOO_SMALL) {
free(var_buf);
var_buf_size = var_size;
var_buf = malloc(var_buf_size);
if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
}
ret = EFI_CALL(efi_get_variable(name, &vendor,
&attributes,
&var_size, var_buf));
}
if (ret != EFI_SUCCESS)
goto err;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
if (hsearch_runtime(name, &vendor, ENTER, &new,
efi_variable_table) <= 0) {
/* This should not happen */
ret = EFI_INVALID_PARAMETER;
goto err;
}
/* allocate space from RUNTIME DATA */
name_len = (u16_strlen_runtime(name) + 1) * sizeof(u16);
memcpy_runtime(mem_pool, name, name_len);
new->name = mem_pool - (u8 *)new; /* offset */
mem_pool += name_len;
memcpy_runtime(&new->vendor.b, &vendor.b, sizeof(vendor));
new->attributes = attributes;
memcpy_runtime(mem_pool, var_buf, var_size);
new->data = mem_pool - (u8 *)new; /* offset */
new->data_size = var_size;
mem_pool += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if ((uintptr_t)mem_pool & 0x1)
mem_pool++;
- }
+#ifdef DEBUG
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = efi_get_next_variable_name_runtime(&name_len, name,
&vendor);
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = efi_get_variable_runtime(name, &vendor, &attributes,
&var_size, var_buf);
if (ret != EFI_SUCCESS)
goto err;
printf("%ls_%pUl:\n", name, &vendor);
printf(" attributes: 0x%x\n", attributes);
printf(" value (size: 0x%lx)\n", var_size);
- }
+#endif
- ret = EFI_SUCCESS;
+err:
- free(name);
- free(var_buf);
- if (ret != EFI_SUCCESS && efi_variable_table) {
efi_free_pool(efi_variable_table);
efi_variable_table = NULL;
- }
- return ret;
+}
+efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data)
+{
- _ENTRY *new;
- if (!variable_name || !vendor || !data_size)
return EFI_EXIT(EFI_INVALID_PARAMETER);
- if (hsearch_runtime(variable_name, vendor, FIND, &new,
efi_variable_table) <= 0)
return EFI_NOT_FOUND;
- if (attributes)
*attributes = new->attributes;
- if (*data_size < new->data_size) {
*data_size = new->data_size;
return EFI_BUFFER_TOO_SMALL;
- }
- *data_size = new->data_size;
- memcpy_runtime(data, entry_data(new), new->data_size);
- return EFI_SUCCESS;
+}
+static int prev_idx __efi_runtime_data;
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor)
+{
- _ENTRY *e;
- u16 *name;
- efi_uintn_t name_size;
- if (!variable_name_size || !variable_name || !vendor)
return EFI_INVALID_PARAMETER;
- if (variable_name[0]) {
/* sanity check for previous variable */
if (prev_idx < 0)
return EFI_INVALID_PARAMETER;
e = &efi_variable_table->table[prev_idx];
if (!e->used || efi_cmpkey(e, variable_name, vendor))
return EFI_INVALID_PARAMETER;
- } else {
prev_idx = -1;
- }
- /* next variable */
- while (++prev_idx <= efi_variable_table->size) {
e = &efi_variable_table->table[prev_idx];
if (e->used)
break;
- }
- if (prev_idx > efi_variable_table->size)
return EFI_NOT_FOUND;
- name = entry_name(e);
- name_size = (u16_strlen_runtime(name) + 1)
* sizeof(u16);
- if (*variable_name_size < name_size) {
*variable_name_size = name_size;
return EFI_BUFFER_TOO_SMALL;
- }
- memcpy_runtime(variable_name, name, name_size);
- memcpy_runtime((void *)&vendor->b, &e->vendor.b, sizeof(vendor));
- return EFI_SUCCESS;
+} +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */

On 6/17/19 3:51 AM, AKASHI Takahiro wrote:
On Sat, Jun 15, 2019 at 09:01:56PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, cache buffer for UEFI variables will be created so that we will still be able to access, at least retrieve, UEFI variables when we exit from boottime services,
This feature is a "should" behavior described in EBBR v1.0 section 2.5.3.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_loader.h | 17 ++ lib/efi_loader/Kconfig | 9 + lib/efi_loader/efi_boottime.c | 10 +- lib/efi_loader/efi_runtime.c | 13 + lib/efi_loader/efi_variable.c | 467 ++++++++++++++++++++++++++++++++++ 5 files changed, 515 insertions(+), 1 deletion(-)
Please, put the cache into a separate file.
Why?
It is a separate set of functions. In C++ programming you wouldn't put two classes into the same file.
diff --git a/include/efi_loader.h b/include/efi_loader.h index 93f7ece814a0..acab657b9d70 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -620,6 +620,23 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name, const efi_guid_t *vendor, u32 attributes, efi_uintn_t data_size, const void *data);
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +efi_status_t efi_freeze_variable_table(void);
+/* runtime version of APIs */ +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Do we really need a multilevel hash table? I would not expect hundreds of variables.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
You could start the cache implementation with a less complicated data structure like a linked list.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
In case of write you could do a write-through in your cache if needed.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
The cache should support write and read. Only NV variables have to be written to a medium. If you do not support this currently just return some error code vor NV variables. But you could accept still accept changes to non-NV variables. This way we can test the code at runtime even before implementing runtime persistence.
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data);
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor);
+#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */
- /*
- See section 3.1.3 in the v2.7 UEFI spec for more details on
- the layout of EFI_LOAD_OPTION. In short it is:
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index e2ef43157568..3f284795648f 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -59,6 +59,15 @@ config EFI_RUNTIME_CONVERT_POINTER to be called by UEFI drivers in relocating themselves to virtual address space.
+config EFI_RUNTIME_GET_VARIABLE_CACHING
- bool "runtime_service: GetVariable: Enable runtime access via cache (read-only)"
- default y
- help
Select this option if you want to access UEFI variables at
runtime even though you cannot update values on the fly.
With or without this option, you can access UEFI variables
at boottime.
Updates of volatile variables should always be possible.
Why "should"? Give me any use case. UEFI spec does not describe such a variant implementation at all.
See above. I would like to be able to test setting variables at runtime even if persisting NV variables is not yet implemented.
- config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y
diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index e4abaf3601d9..14e343abbd43 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1892,6 +1892,9 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, efi_uintn_t map_key) { struct efi_event *evt; +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_status_t ret;
+#endif
EFI_ENTRY("%p, %zx", image_handle, map_key);
@@ -1921,7 +1924,12 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, } }
- /* TODO: Should persist EFI variables here */
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
Can we have weak functions for initializing and persisting the cache, e.g.
efi_status_t __weak efi_load_variable_cache(cache_entry *cache, size_t *size) { Â Â Â Â Â Â Â Â cache->len = 0; Â Â Â Â Â Â Â Â return EFI_SUCCESS; }
efi_status_t __runtime __weak efi_write_variable_cache(cache_entry *cache, size_t size) { Â Â Â Â Â Â Â Â return EFI_UNSUPPORTED; }
Then we can override these in whatever driver we implement.
What is the difference between yours and my env_efi_load/save() with backing-storage driver?
I cannot see that you clearly separate the functions API, cache, persistence.
I would like to see clearly define interfaces into which we can plug different persistence implementations.
- /* No more variable update */
- ret = efi_freeze_variable_table();
- if (ret != EFI_SUCCESS)
return EFI_EXIT(ret);
+#endif
board_quiesce_devices();
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index fc5bdee80e00..b60f70f04613 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -111,6 +111,11 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_CONVERT_POINTER; #endif +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_runtime_services_supported |=
(EFI_RT_SUPPORTED_GET_VARIABLE |
EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME);
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -469,10 +474,18 @@ static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { .patchto = NULL, }, { .ptr = &efi_runtime_services.get_variable, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_variable_runtime,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.get_next_variable_name, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_next_variable_name,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.set_variable, .patchto = &efi_device_error, diff --git a/lib/efi_loader/efi_variable.c b/lib/efi_loader/efi_variable.c index d9887be938c2..ee21892dd291 100644 --- a/lib/efi_loader/efi_variable.c +++ b/lib/efi_loader/efi_variable.c @@ -706,3 +706,470 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name,
return EFI_EXIT(ret); }
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +/*
- runtime version of APIs
- We only support read-only variable access.
- The table is in U-Boot's hash table format, but has its own
- _ENTRY structure for specific use.
- Except for efi_freeze_variable_table(), which is to be called in
- exit_boot_services(), all the functions and data below must be
- placed in either RUNTIME_SERVICES_CODE or RUNTIME_SERVICES_DATA.
- */
+typedef struct _ENTRY {
- unsigned int used; /* hash value; 0 for not used */
- size_t name; /* name offset from itself */
- efi_guid_t vendor;
- u32 attributes;
- size_t data; /* data offset from itself */
- size_t data_size;
+} _ENTRY;
+static inline u16 *entry_name(_ENTRY *e) { return (void *)e + e->name; } +static inline u16 *entry_data(_ENTRY *e) { return (void *)e + e->data; }
+static struct hsearch_data *efi_variable_table __efi_runtime_data;
+static size_t __efi_runtime u16_strlen_runtime(const u16 *s1)
Please, do not duplicate existing functions. If they have to be runtime simply change the existing function to __runtime.
We should do that if possible, but please note that [str|mem]xxx() functions are *architecture* dependent. Do you want to mark all the functions across all the architectures?
+{
- size_t n = 0;
- while (*s1) {
n++;
s1++;
- }
- return n;
+}
+static int __efi_runtime memcmp_runtime(const void *m1, const void *m2,
size_t n)
I dislike duplicate code. Can't we simply define the existing memcmp function as __runtime?
ditto
+{
- while (n && *(u8 *)m1 == *(u8 *)m2) {
n--;
m1++;
m2++;
- }
- if (n)
return *(u8 *)m1 - *(u8 *)m2;
- return 0;
+}
+static void __efi_runtime memcpy_runtime(void *m1, const void *m2, size_t n) +{
Can't we simply define the existing memcpy function as __runtime?
- for (; n; n--, m1++, m2++)
*(u8 *)m1 = *(u8 *)m2;
+}
+static int __efi_runtime efi_cmpkey(_ENTRY *e, const u16 *name,
const efi_guid_t *vendor)
+{
- size_t name_len;
- name_len = u16_strlen_runtime(entry_name(e));
- /* return zero if matched */
- return name_len != u16_strlen_runtime(name) ||
memcmp_runtime(entry_name(e), name, name_len * 2) ||
memcmp_runtime(e->vendor.b, vendor->b, sizeof(vendor));
+}
+/* simplified and slightly different version of hsearch_r() */
These hash functions are so complicated that they really need a unit test testing them rigorously.
Do you know that there are no any tests for hsearch_r()? Moreover, this function would better be exercised well if we could add *runtime* tests.
Have a look at test/env/hashtable.c
I think efi_selftest could work after SetVirtualMemoryMap if we do not change the map for memory actually used by U-Boot.
Or we use a minimum Linux kernel and put our tests into an init binary.
+static int __efi_runtime hsearch_runtime(const u16 *name,
const efi_guid_t *vendor,
ACTION action,
_ENTRY **retval,
struct hsearch_data *htab)
+{
- unsigned int hval;
- unsigned int count;
- unsigned int len;
- unsigned int idx, new;
- /* Compute an value for the given string. */
- len = u16_strlen_runtime(name);
Can't the same variable name exist for different GUIDs? Why is the GUID not considered in the hash?
efi_cmpkey() does take into consideration GUID as well as the name.
My question concerned the hash.
- hval = len;
- count = len;
- while (count-- > 0) {
hval <<= 4;
hval += name[count];
- }
- /*
* First hash function:
* simply take the modulo but prevent zero.
*/
- hval %= htab->size;
- if (hval == 0)
++hval;
- /* The first index tried. */
- new = -1; /* not found */
- idx = hval;
- if (htab->table[idx].used) {
/*
* Further action might be required according to the
* action value.
*/
unsigned int hval2;
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
/*
* Second hash function:
* as suggested in [Knuth]
*/
hval2 = 1 + hval % (htab->size - 2);
do {
/*
* Because SIZE is prime this guarantees to
* step through all available indices.
*/
if (idx <= hval2)
idx = htab->size + idx - hval2;
else
idx -= hval2;
/*
* If we visited all entries leave the loop
* unsuccessfully.
*/
if (idx == hval)
break;
/* If entry is found use it. */
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
} while (htab->table[idx].used);
if (!htab->table[idx].used)
new = idx;
- } else {
new = idx;
- }
- /*
* An empty bucket has been found.
* The following code should never be executed after
* exit_boot_services()
*/
- if (action == ENTER) {
/*
* If table is full and another entry should be
* entered return with error.
*/
if (htab->filled == htab->size) {
*retval = NULL;
return 0;
}
/* Create new entry */
htab->table[new].used = hval;
++htab->filled;
/* return new entry */
*retval = &htab->table[new];
return 1;
- }
- *retval = NULL;
- return 0;
+}
+/* from lib/hashtable.c */ +static inline int isprime(unsigned int number) +{
- /* no even number will be passed */
- unsigned int div = 3;
- while (div * div < number && number % div != 0)
div += 2;
- return number % div != 0;
+}
+efi_status_t efi_freeze_variable_table(void)
Please, add comments to your functions. It is not self-evident what this function is meant to do.
I cannot imagine why a variable cache should be frozen. It is a living data structure until the system is switched off.
I don't get your point.
Please, explain what you mean by freeze.
I suggest that the cache is read/write at all times.
Best regards
Heinrich
For a variable cache I expect that you allocate memory before handling the first variable and never again. At runtime you will not have chance to allocate memory anyway.
I don't get your point.
This function is way too long. Pleae, break it down.
Freezing is implemented in 2-phase steps, and some complexity is inevitable. I suppose adding some comments would be enough.
-Takahiro Akashi
Best regards
Heinrich
+{
- int var_num = 0;
- size_t var_data_size = 0;
- u16 *name;
- efi_uintn_t name_buf_len, name_len;
- efi_guid_t vendor;
- u32 attributes;
- u8 *mem_pool, *var_buf = NULL;
- size_t table_size, var_size, var_buf_size;
- _ENTRY *new = NULL;
- efi_status_t ret;
- /* phase-1 loop */
- name_buf_len = 128;
- name = malloc(name_buf_len);
- if (!name)
return EFI_OUT_OF_RESOURCES;
- name[0] = 0;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND) {
break;
} else if (ret == EFI_BUFFER_TOO_SMALL) {
u16 *buf;
name_buf_len = name_len;
buf = realloc(name, name_buf_len);
if (!buf) {
free(name);
return EFI_OUT_OF_RESOURCES;
}
name = buf;
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len,
name,
&vendor));
}
if (ret != EFI_SUCCESS)
return ret;
var_size = 0;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, NULL));
if (ret != EFI_BUFFER_TOO_SMALL)
return ret;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
var_num++;
var_data_size += (u16_strlen_runtime(name) + 1) * sizeof(u16);
var_data_size += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if (var_data_size & 0x1)
var_data_size++;
- }
- /*
* total of entries in hash table must be a prime number.
* The logic below comes from lib/hashtable.c
*/
- var_num |= 1; /* make odd */
- while (!isprime(var_num))
var_num += 2;
- /* We need table[var_num] for hsearch_runtime algo */
- table_size = sizeof(*efi_variable_table)
+ sizeof(_ENTRY) * (var_num + 1) + var_data_size;
- ret = efi_allocate_pool(EFI_RUNTIME_SERVICES_DATA,
table_size, (void **)&efi_variable_table);
- if (ret != EFI_SUCCESS)
return ret;
- efi_variable_table->size = var_num;
- efi_variable_table->table = (void *)efi_variable_table
+ sizeof(*efi_variable_table);
- mem_pool = (u8 *)efi_variable_table->table
+ sizeof(_ENTRY) * (var_num + 1);
- var_buf_size = 128;
- var_buf = malloc(var_buf_size);
- if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
- }
- /* phase-2 loop */
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, var_buf));
if (ret == EFI_BUFFER_TOO_SMALL) {
free(var_buf);
var_buf_size = var_size;
var_buf = malloc(var_buf_size);
if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
}
ret = EFI_CALL(efi_get_variable(name, &vendor,
&attributes,
&var_size, var_buf));
}
if (ret != EFI_SUCCESS)
goto err;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
if (hsearch_runtime(name, &vendor, ENTER, &new,
efi_variable_table) <= 0) {
/* This should not happen */
ret = EFI_INVALID_PARAMETER;
goto err;
}
/* allocate space from RUNTIME DATA */
name_len = (u16_strlen_runtime(name) + 1) * sizeof(u16);
memcpy_runtime(mem_pool, name, name_len);
new->name = mem_pool - (u8 *)new; /* offset */
mem_pool += name_len;
memcpy_runtime(&new->vendor.b, &vendor.b, sizeof(vendor));
new->attributes = attributes;
memcpy_runtime(mem_pool, var_buf, var_size);
new->data = mem_pool - (u8 *)new; /* offset */
new->data_size = var_size;
mem_pool += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if ((uintptr_t)mem_pool & 0x1)
mem_pool++;
- }
+#ifdef DEBUG
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = efi_get_next_variable_name_runtime(&name_len, name,
&vendor);
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = efi_get_variable_runtime(name, &vendor, &attributes,
&var_size, var_buf);
if (ret != EFI_SUCCESS)
goto err;
printf("%ls_%pUl:\n", name, &vendor);
printf(" attributes: 0x%x\n", attributes);
printf(" value (size: 0x%lx)\n", var_size);
- }
+#endif
- ret = EFI_SUCCESS;
+err:
- free(name);
- free(var_buf);
- if (ret != EFI_SUCCESS && efi_variable_table) {
efi_free_pool(efi_variable_table);
efi_variable_table = NULL;
- }
- return ret;
+}
+efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data)
+{
- _ENTRY *new;
- if (!variable_name || !vendor || !data_size)
return EFI_EXIT(EFI_INVALID_PARAMETER);
- if (hsearch_runtime(variable_name, vendor, FIND, &new,
efi_variable_table) <= 0)
return EFI_NOT_FOUND;
- if (attributes)
*attributes = new->attributes;
- if (*data_size < new->data_size) {
*data_size = new->data_size;
return EFI_BUFFER_TOO_SMALL;
- }
- *data_size = new->data_size;
- memcpy_runtime(data, entry_data(new), new->data_size);
- return EFI_SUCCESS;
+}
+static int prev_idx __efi_runtime_data;
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor)
+{
- _ENTRY *e;
- u16 *name;
- efi_uintn_t name_size;
- if (!variable_name_size || !variable_name || !vendor)
return EFI_INVALID_PARAMETER;
- if (variable_name[0]) {
/* sanity check for previous variable */
if (prev_idx < 0)
return EFI_INVALID_PARAMETER;
e = &efi_variable_table->table[prev_idx];
if (!e->used || efi_cmpkey(e, variable_name, vendor))
return EFI_INVALID_PARAMETER;
- } else {
prev_idx = -1;
- }
- /* next variable */
- while (++prev_idx <= efi_variable_table->size) {
e = &efi_variable_table->table[prev_idx];
if (e->used)
break;
- }
- if (prev_idx > efi_variable_table->size)
return EFI_NOT_FOUND;
- name = entry_name(e);
- name_size = (u16_strlen_runtime(name) + 1)
* sizeof(u16);
- if (*variable_name_size < name_size) {
*variable_name_size = name_size;
return EFI_BUFFER_TOO_SMALL;
- }
- memcpy_runtime(variable_name, name, name_size);
- memcpy_runtime((void *)&vendor->b, &e->vendor.b, sizeof(vendor));
- return EFI_SUCCESS;
+} +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */

On Mon, Jun 17, 2019 at 09:52:34PM +0200, Heinrich Schuchardt wrote:
On 6/17/19 3:51 AM, AKASHI Takahiro wrote:
On Sat, Jun 15, 2019 at 09:01:56PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, cache buffer for UEFI variables will be created so that we will still be able to access, at least retrieve, UEFI variables when we exit from boottime services,
This feature is a "should" behavior described in EBBR v1.0 section 2.5.3.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_loader.h | 17 ++ lib/efi_loader/Kconfig | 9 + lib/efi_loader/efi_boottime.c | 10 +- lib/efi_loader/efi_runtime.c | 13 + lib/efi_loader/efi_variable.c | 467 ++++++++++++++++++++++++++++++++++ 5 files changed, 515 insertions(+), 1 deletion(-)
Please, put the cache into a separate file.
Why?
It is a separate set of functions. In C++ programming you wouldn't put two classes into the same file.
? I don't get your point. This is not C++, but C.
diff --git a/include/efi_loader.h b/include/efi_loader.h index 93f7ece814a0..acab657b9d70 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -620,6 +620,23 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name, const efi_guid_t *vendor, u32 attributes, efi_uintn_t data_size, const void *data);
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +efi_status_t efi_freeze_variable_table(void);
+/* runtime version of APIs */ +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Do we really need a multilevel hash table? I would not expect hundreds of variables.
Please don't change your point suddenly. Here we are discussing whether "The cache should be used both at runtime and at boottime" or not.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
You could start the cache implementation with a less complicated data structure like a linked list.
This is totally a different issue. I listed this issue in my cover letter.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables.
In case of write you could do a write-through in your cache if needed.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
The cache should support write and read. Only NV variables have to be
Why? I don't think it make sense that we support volatile variable, but not non-volatile variable at runtime. Even UEFI specification doesn't describe such an irregular behavior. See EFI_RT_SUPPORTED_XXX definitions. Do you have a meaningful use case that you want to support?
written to a medium. If you do not support this currently just return some error code vor NV variables. But you could accept still accept changes to non-NV variables. This way we can test the code at runtime even before implementing runtime persistence.
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data);
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor);
+#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */
/*
- See section 3.1.3 in the v2.7 UEFI spec for more details on
- the layout of EFI_LOAD_OPTION. In short it is:
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index e2ef43157568..3f284795648f 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -59,6 +59,15 @@ config EFI_RUNTIME_CONVERT_POINTER to be called by UEFI drivers in relocating themselves to virtual address space.
+config EFI_RUNTIME_GET_VARIABLE_CACHING
- bool "runtime_service: GetVariable: Enable runtime access via cache (read-only)"
- default y
- help
Select this option if you want to access UEFI variables at
runtime even though you cannot update values on the fly.
With or without this option, you can access UEFI variables
at boottime.
Updates of volatile variables should always be possible.
Why "should"? Give me any use case. UEFI spec does not describe such a variant implementation at all.
See above. I would like to be able to test setting variables at runtime even if persisting NV variables is not yet implemented.
Just for testing? No real use case? Please note that if we *have* persistent storage at runtime, we don't need cache because we can access that storage.
config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index e4abaf3601d9..14e343abbd43 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1892,6 +1892,9 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, efi_uintn_t map_key) { struct efi_event *evt; +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_status_t ret;
+#endif
EFI_ENTRY("%p, %zx", image_handle, map_key);
@@ -1921,7 +1924,12 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, } }
- /* TODO: Should persist EFI variables here */
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
Can we have weak functions for initializing and persisting the cache, e.g.
efi_status_t __weak efi_load_variable_cache(cache_entry *cache, size_t *size) { Â Â Â Â Â Â Â Â cache->len = 0; Â Â Â Â Â Â Â Â return EFI_SUCCESS; }
efi_status_t __runtime __weak efi_write_variable_cache(cache_entry *cache, size_t size) { Â Â Â Â Â Â Â Â return EFI_UNSUPPORTED; }
Then we can override these in whatever driver we implement.
What is the difference between yours and my env_efi_load/save() with backing-storage driver?
I cannot see that you clearly separate the functions API, cache, persistence.
I don't get your point. For instance, efi_boottime.c has a lot of different type of functions. Do you want to separate them into different files?
I would like to see clearly define interfaces into which we can plug different persistence implementations.
Your assertion doesn't clarify why you want to separate API and cache.
Regarding persistent storage, as I said, efi_[get|set]_variable() will be completely replaced with the corresponding functions (at least, in case of Standalone MM services).
- /* No more variable update */
- ret = efi_freeze_variable_table();
- if (ret != EFI_SUCCESS)
return EFI_EXIT(ret);
+#endif
board_quiesce_devices();
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index fc5bdee80e00..b60f70f04613 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -111,6 +111,11 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_CONVERT_POINTER; #endif +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_runtime_services_supported |=
(EFI_RT_SUPPORTED_GET_VARIABLE |
EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME);
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -469,10 +474,18 @@ static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { .patchto = NULL, }, { .ptr = &efi_runtime_services.get_variable, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_variable_runtime,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.get_next_variable_name, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_next_variable_name,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.set_variable, .patchto = &efi_device_error, diff --git a/lib/efi_loader/efi_variable.c b/lib/efi_loader/efi_variable.c index d9887be938c2..ee21892dd291 100644 --- a/lib/efi_loader/efi_variable.c +++ b/lib/efi_loader/efi_variable.c @@ -706,3 +706,470 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name,
return EFI_EXIT(ret); }
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +/*
- runtime version of APIs
- We only support read-only variable access.
- The table is in U-Boot's hash table format, but has its own
- _ENTRY structure for specific use.
- Except for efi_freeze_variable_table(), which is to be called in
- exit_boot_services(), all the functions and data below must be
- placed in either RUNTIME_SERVICES_CODE or RUNTIME_SERVICES_DATA.
- */
+typedef struct _ENTRY {
- unsigned int used; /* hash value; 0 for not used */
- size_t name; /* name offset from itself */
- efi_guid_t vendor;
- u32 attributes;
- size_t data; /* data offset from itself */
- size_t data_size;
+} _ENTRY;
+static inline u16 *entry_name(_ENTRY *e) { return (void *)e + e->name; } +static inline u16 *entry_data(_ENTRY *e) { return (void *)e + e->data; }
+static struct hsearch_data *efi_variable_table __efi_runtime_data;
+static size_t __efi_runtime u16_strlen_runtime(const u16 *s1)
Please, do not duplicate existing functions. If they have to be runtime simply change the existing function to __runtime.
We should do that if possible, but please note that [str|mem]xxx() functions are *architecture* dependent. Do you want to mark all the functions across all the architectures?
Agree or not agree?
+{
- size_t n = 0;
- while (*s1) {
n++;
s1++;
- }
- return n;
+}
+static int __efi_runtime memcmp_runtime(const void *m1, const void *m2,
size_t n)
I dislike duplicate code. Can't we simply define the existing memcmp function as __runtime?
ditto
+{
- while (n && *(u8 *)m1 == *(u8 *)m2) {
n--;
m1++;
m2++;
- }
- if (n)
return *(u8 *)m1 - *(u8 *)m2;
- return 0;
+}
+static void __efi_runtime memcpy_runtime(void *m1, const void *m2, size_t n) +{
Can't we simply define the existing memcpy function as __runtime?
- for (; n; n--, m1++, m2++)
*(u8 *)m1 = *(u8 *)m2;
+}
+static int __efi_runtime efi_cmpkey(_ENTRY *e, const u16 *name,
const efi_guid_t *vendor)
+{
- size_t name_len;
- name_len = u16_strlen_runtime(entry_name(e));
- /* return zero if matched */
- return name_len != u16_strlen_runtime(name) ||
memcmp_runtime(entry_name(e), name, name_len * 2) ||
memcmp_runtime(e->vendor.b, vendor->b, sizeof(vendor));
+}
+/* simplified and slightly different version of hsearch_r() */
These hash functions are so complicated that they really need a unit test testing them rigorously.
Do you know that there are no any tests for hsearch_r()? Moreover, this function would better be exercised well if we could add *runtime* tests.
Have a look at test/env/hashtable.c
Okay, but why not re-use this test for my stripped version?
I think efi_selftest could work after SetVirtualMemoryMap if we do not change the map for memory actually used by U-Boot.
Maybe.
-Takahiro Akashi
Or we use a minimum Linux kernel and put our tests into an init binary.
+static int __efi_runtime hsearch_runtime(const u16 *name,
const efi_guid_t *vendor,
ACTION action,
_ENTRY **retval,
struct hsearch_data *htab)
+{
- unsigned int hval;
- unsigned int count;
- unsigned int len;
- unsigned int idx, new;
- /* Compute an value for the given string. */
- len = u16_strlen_runtime(name);
Can't the same variable name exist for different GUIDs? Why is the GUID not considered in the hash?
efi_cmpkey() does take into consideration GUID as well as the name.
My question concerned the hash.
- hval = len;
- count = len;
- while (count-- > 0) {
hval <<= 4;
hval += name[count];
- }
- /*
* First hash function:
* simply take the modulo but prevent zero.
*/
- hval %= htab->size;
- if (hval == 0)
++hval;
- /* The first index tried. */
- new = -1; /* not found */
- idx = hval;
- if (htab->table[idx].used) {
/*
* Further action might be required according to the
* action value.
*/
unsigned int hval2;
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
/*
* Second hash function:
* as suggested in [Knuth]
*/
hval2 = 1 + hval % (htab->size - 2);
do {
/*
* Because SIZE is prime this guarantees to
* step through all available indices.
*/
if (idx <= hval2)
idx = htab->size + idx - hval2;
else
idx -= hval2;
/*
* If we visited all entries leave the loop
* unsuccessfully.
*/
if (idx == hval)
break;
/* If entry is found use it. */
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
} while (htab->table[idx].used);
if (!htab->table[idx].used)
new = idx;
- } else {
new = idx;
- }
- /*
* An empty bucket has been found.
* The following code should never be executed after
* exit_boot_services()
*/
- if (action == ENTER) {
/*
* If table is full and another entry should be
* entered return with error.
*/
if (htab->filled == htab->size) {
*retval = NULL;
return 0;
}
/* Create new entry */
htab->table[new].used = hval;
++htab->filled;
/* return new entry */
*retval = &htab->table[new];
return 1;
- }
- *retval = NULL;
- return 0;
+}
+/* from lib/hashtable.c */ +static inline int isprime(unsigned int number) +{
- /* no even number will be passed */
- unsigned int div = 3;
- while (div * div < number && number % div != 0)
div += 2;
- return number % div != 0;
+}
+efi_status_t efi_freeze_variable_table(void)
Please, add comments to your functions. It is not self-evident what this function is meant to do.
I cannot imagine why a variable cache should be frozen. It is a living data structure until the system is switched off.
I don't get your point.
Please, explain what you mean by freeze.
I suggest that the cache is read/write at all times.
Best regards
Heinrich
For a variable cache I expect that you allocate memory before handling the first variable and never again. At runtime you will not have chance to allocate memory anyway.
I don't get your point.
This function is way too long. Pleae, break it down.
Freezing is implemented in 2-phase steps, and some complexity is inevitable. I suppose adding some comments would be enough.
-Takahiro Akashi
Best regards
Heinrich
+{
- int var_num = 0;
- size_t var_data_size = 0;
- u16 *name;
- efi_uintn_t name_buf_len, name_len;
- efi_guid_t vendor;
- u32 attributes;
- u8 *mem_pool, *var_buf = NULL;
- size_t table_size, var_size, var_buf_size;
- _ENTRY *new = NULL;
- efi_status_t ret;
- /* phase-1 loop */
- name_buf_len = 128;
- name = malloc(name_buf_len);
- if (!name)
return EFI_OUT_OF_RESOURCES;
- name[0] = 0;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND) {
break;
} else if (ret == EFI_BUFFER_TOO_SMALL) {
u16 *buf;
name_buf_len = name_len;
buf = realloc(name, name_buf_len);
if (!buf) {
free(name);
return EFI_OUT_OF_RESOURCES;
}
name = buf;
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len,
name,
&vendor));
}
if (ret != EFI_SUCCESS)
return ret;
var_size = 0;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, NULL));
if (ret != EFI_BUFFER_TOO_SMALL)
return ret;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
var_num++;
var_data_size += (u16_strlen_runtime(name) + 1) * sizeof(u16);
var_data_size += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if (var_data_size & 0x1)
var_data_size++;
- }
- /*
* total of entries in hash table must be a prime number.
* The logic below comes from lib/hashtable.c
*/
- var_num |= 1; /* make odd */
- while (!isprime(var_num))
var_num += 2;
- /* We need table[var_num] for hsearch_runtime algo */
- table_size = sizeof(*efi_variable_table)
+ sizeof(_ENTRY) * (var_num + 1) + var_data_size;
- ret = efi_allocate_pool(EFI_RUNTIME_SERVICES_DATA,
table_size, (void **)&efi_variable_table);
- if (ret != EFI_SUCCESS)
return ret;
- efi_variable_table->size = var_num;
- efi_variable_table->table = (void *)efi_variable_table
+ sizeof(*efi_variable_table);
- mem_pool = (u8 *)efi_variable_table->table
+ sizeof(_ENTRY) * (var_num + 1);
- var_buf_size = 128;
- var_buf = malloc(var_buf_size);
- if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
- }
- /* phase-2 loop */
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, var_buf));
if (ret == EFI_BUFFER_TOO_SMALL) {
free(var_buf);
var_buf_size = var_size;
var_buf = malloc(var_buf_size);
if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
}
ret = EFI_CALL(efi_get_variable(name, &vendor,
&attributes,
&var_size, var_buf));
}
if (ret != EFI_SUCCESS)
goto err;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
if (hsearch_runtime(name, &vendor, ENTER, &new,
efi_variable_table) <= 0) {
/* This should not happen */
ret = EFI_INVALID_PARAMETER;
goto err;
}
/* allocate space from RUNTIME DATA */
name_len = (u16_strlen_runtime(name) + 1) * sizeof(u16);
memcpy_runtime(mem_pool, name, name_len);
new->name = mem_pool - (u8 *)new; /* offset */
mem_pool += name_len;
memcpy_runtime(&new->vendor.b, &vendor.b, sizeof(vendor));
new->attributes = attributes;
memcpy_runtime(mem_pool, var_buf, var_size);
new->data = mem_pool - (u8 *)new; /* offset */
new->data_size = var_size;
mem_pool += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if ((uintptr_t)mem_pool & 0x1)
mem_pool++;
- }
+#ifdef DEBUG
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = efi_get_next_variable_name_runtime(&name_len, name,
&vendor);
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = efi_get_variable_runtime(name, &vendor, &attributes,
&var_size, var_buf);
if (ret != EFI_SUCCESS)
goto err;
printf("%ls_%pUl:\n", name, &vendor);
printf(" attributes: 0x%x\n", attributes);
printf(" value (size: 0x%lx)\n", var_size);
- }
+#endif
- ret = EFI_SUCCESS;
+err:
- free(name);
- free(var_buf);
- if (ret != EFI_SUCCESS && efi_variable_table) {
efi_free_pool(efi_variable_table);
efi_variable_table = NULL;
- }
- return ret;
+}
+efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data)
+{
- _ENTRY *new;
- if (!variable_name || !vendor || !data_size)
return EFI_EXIT(EFI_INVALID_PARAMETER);
- if (hsearch_runtime(variable_name, vendor, FIND, &new,
efi_variable_table) <= 0)
return EFI_NOT_FOUND;
- if (attributes)
*attributes = new->attributes;
- if (*data_size < new->data_size) {
*data_size = new->data_size;
return EFI_BUFFER_TOO_SMALL;
- }
- *data_size = new->data_size;
- memcpy_runtime(data, entry_data(new), new->data_size);
- return EFI_SUCCESS;
+}
+static int prev_idx __efi_runtime_data;
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor)
+{
- _ENTRY *e;
- u16 *name;
- efi_uintn_t name_size;
- if (!variable_name_size || !variable_name || !vendor)
return EFI_INVALID_PARAMETER;
- if (variable_name[0]) {
/* sanity check for previous variable */
if (prev_idx < 0)
return EFI_INVALID_PARAMETER;
e = &efi_variable_table->table[prev_idx];
if (!e->used || efi_cmpkey(e, variable_name, vendor))
return EFI_INVALID_PARAMETER;
- } else {
prev_idx = -1;
- }
- /* next variable */
- while (++prev_idx <= efi_variable_table->size) {
e = &efi_variable_table->table[prev_idx];
if (e->used)
break;
- }
- if (prev_idx > efi_variable_table->size)
return EFI_NOT_FOUND;
- name = entry_name(e);
- name_size = (u16_strlen_runtime(name) + 1)
* sizeof(u16);
- if (*variable_name_size < name_size) {
*variable_name_size = name_size;
return EFI_BUFFER_TOO_SMALL;
- }
- memcpy_runtime(variable_name, name, name_size);
- memcpy_runtime((void *)&vendor->b, &e->vendor.b, sizeof(vendor));
- return EFI_SUCCESS;
+} +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */

Heinrich,
On Tue, Jun 18, 2019 at 10:19:06AM +0900, AKASHI Takahiro wrote:
On Mon, Jun 17, 2019 at 09:52:34PM +0200, Heinrich Schuchardt wrote:
On 6/17/19 3:51 AM, AKASHI Takahiro wrote:
On Sat, Jun 15, 2019 at 09:01:56PM +0200, Heinrich Schuchardt wrote:
On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
With this patch, cache buffer for UEFI variables will be created so that we will still be able to access, at least retrieve, UEFI variables when we exit from boottime services,
This feature is a "should" behavior described in EBBR v1.0 section 2.5.3.
Signed-off-by: AKASHI Takahiro takahiro.akashi@linaro.org
include/efi_loader.h | 17 ++ lib/efi_loader/Kconfig | 9 + lib/efi_loader/efi_boottime.c | 10 +- lib/efi_loader/efi_runtime.c | 13 + lib/efi_loader/efi_variable.c | 467 ++++++++++++++++++++++++++++++++++ 5 files changed, 515 insertions(+), 1 deletion(-)
Please, put the cache into a separate file.
Why?
It is a separate set of functions. In C++ programming you wouldn't put two classes into the same file.
? I don't get your point. This is not C++, but C.
diff --git a/include/efi_loader.h b/include/efi_loader.h index 93f7ece814a0..acab657b9d70 100644 --- a/include/efi_loader.h +++ b/include/efi_loader.h @@ -620,6 +620,23 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name, const efi_guid_t *vendor, u32 attributes, efi_uintn_t data_size, const void *data);
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +efi_status_t efi_freeze_variable_table(void);
+/* runtime version of APIs */ +efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Do we really need a multilevel hash table? I would not expect hundreds of variables.
Please don't change your point suddenly. Here we are discussing whether "The cache should be used both at runtime and at boottime" or not.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
You could start the cache implementation with a less complicated data structure like a linked list.
This is totally a different issue. I listed this issue in my cover letter.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables.
Let me elaborate more: My tentative "secure boot" patch is self-contained on non-secure side and provides the functionality that UEFI specification describes, but there is one missing and essential aspect: Tamper-resistant. Standalone MM services is expected to provides this semantics with the same interfaces as [Get|Set]Variable for both volatile and non-volatile variables alike. There can be another implementation (with yet another interfaces to secure side), but it is currently the only implementation available as far as I know.
-Takahiro Akashi
In case of write you could do a write-through in your cache if needed.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
The cache should support write and read. Only NV variables have to be
Why? I don't think it make sense that we support volatile variable, but not non-volatile variable at runtime. Even UEFI specification doesn't describe such an irregular behavior. See EFI_RT_SUPPORTED_XXX definitions. Do you have a meaningful use case that you want to support?
written to a medium. If you do not support this currently just return some error code vor NV variables. But you could accept still accept changes to non-NV variables. This way we can test the code at runtime even before implementing runtime persistence.
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data);
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor);
+#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */
/*
- See section 3.1.3 in the v2.7 UEFI spec for more details on
- the layout of EFI_LOAD_OPTION. In short it is:
diff --git a/lib/efi_loader/Kconfig b/lib/efi_loader/Kconfig index e2ef43157568..3f284795648f 100644 --- a/lib/efi_loader/Kconfig +++ b/lib/efi_loader/Kconfig @@ -59,6 +59,15 @@ config EFI_RUNTIME_CONVERT_POINTER to be called by UEFI drivers in relocating themselves to virtual address space.
+config EFI_RUNTIME_GET_VARIABLE_CACHING
- bool "runtime_service: GetVariable: Enable runtime access via cache (read-only)"
- default y
- help
Select this option if you want to access UEFI variables at
runtime even though you cannot update values on the fly.
With or without this option, you can access UEFI variables
at boottime.
Updates of volatile variables should always be possible.
Why "should"? Give me any use case. UEFI spec does not describe such a variant implementation at all.
See above. I would like to be able to test setting variables at runtime even if persisting NV variables is not yet implemented.
Just for testing? No real use case? Please note that if we *have* persistent storage at runtime, we don't need cache because we can access that storage.
config EFI_DEVICE_PATH_TO_TEXT bool "Device path to text protocol" default y diff --git a/lib/efi_loader/efi_boottime.c b/lib/efi_loader/efi_boottime.c index e4abaf3601d9..14e343abbd43 100644 --- a/lib/efi_loader/efi_boottime.c +++ b/lib/efi_loader/efi_boottime.c @@ -1892,6 +1892,9 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, efi_uintn_t map_key) { struct efi_event *evt; +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_status_t ret;
+#endif
EFI_ENTRY("%p, %zx", image_handle, map_key);
@@ -1921,7 +1924,12 @@ static efi_status_t EFIAPI efi_exit_boot_services(efi_handle_t image_handle, } }
- /* TODO: Should persist EFI variables here */
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
Can we have weak functions for initializing and persisting the cache, e.g.
efi_status_t __weak efi_load_variable_cache(cache_entry *cache, size_t *size) { Â Â Â Â Â Â Â Â cache->len = 0; Â Â Â Â Â Â Â Â return EFI_SUCCESS; }
efi_status_t __runtime __weak efi_write_variable_cache(cache_entry *cache, size_t size) { Â Â Â Â Â Â Â Â return EFI_UNSUPPORTED; }
Then we can override these in whatever driver we implement.
What is the difference between yours and my env_efi_load/save() with backing-storage driver?
I cannot see that you clearly separate the functions API, cache, persistence.
I don't get your point. For instance, efi_boottime.c has a lot of different type of functions. Do you want to separate them into different files?
I would like to see clearly define interfaces into which we can plug different persistence implementations.
Your assertion doesn't clarify why you want to separate API and cache.
Regarding persistent storage, as I said, efi_[get|set]_variable() will be completely replaced with the corresponding functions (at least, in case of Standalone MM services).
- /* No more variable update */
- ret = efi_freeze_variable_table();
- if (ret != EFI_SUCCESS)
return EFI_EXIT(ret);
+#endif
board_quiesce_devices();
diff --git a/lib/efi_loader/efi_runtime.c b/lib/efi_loader/efi_runtime.c index fc5bdee80e00..b60f70f04613 100644 --- a/lib/efi_loader/efi_runtime.c +++ b/lib/efi_loader/efi_runtime.c @@ -111,6 +111,11 @@ efi_status_t efi_init_runtime_supported(void) efi_runtime_services_supported |= EFI_RT_SUPPORTED_CONVERT_POINTER; #endif +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
- efi_runtime_services_supported |=
(EFI_RT_SUPPORTED_GET_VARIABLE |
EFI_RT_SUPPORTED_GET_NEXT_VARIABLE_NAME);
+#endif
return EFI_CALL(efi_set_variable(L"RuntimeServicesSupported", &efi_global_variable_guid, @@ -469,10 +474,18 @@ static struct efi_runtime_detach_list_struct efi_runtime_detach_list[] = { .patchto = NULL, }, { .ptr = &efi_runtime_services.get_variable, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_variable_runtime,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.get_next_variable_name, +#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING
.patchto = &efi_get_next_variable_name,
+#else .patchto = &efi_device_error, +#endif }, { .ptr = &efi_runtime_services.set_variable, .patchto = &efi_device_error, diff --git a/lib/efi_loader/efi_variable.c b/lib/efi_loader/efi_variable.c index d9887be938c2..ee21892dd291 100644 --- a/lib/efi_loader/efi_variable.c +++ b/lib/efi_loader/efi_variable.c @@ -706,3 +706,470 @@ efi_status_t EFIAPI efi_set_variable(u16 *variable_name,
return EFI_EXIT(ret); }
+#ifdef CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING +/*
- runtime version of APIs
- We only support read-only variable access.
- The table is in U-Boot's hash table format, but has its own
- _ENTRY structure for specific use.
- Except for efi_freeze_variable_table(), which is to be called in
- exit_boot_services(), all the functions and data below must be
- placed in either RUNTIME_SERVICES_CODE or RUNTIME_SERVICES_DATA.
- */
+typedef struct _ENTRY {
- unsigned int used; /* hash value; 0 for not used */
- size_t name; /* name offset from itself */
- efi_guid_t vendor;
- u32 attributes;
- size_t data; /* data offset from itself */
- size_t data_size;
+} _ENTRY;
+static inline u16 *entry_name(_ENTRY *e) { return (void *)e + e->name; } +static inline u16 *entry_data(_ENTRY *e) { return (void *)e + e->data; }
+static struct hsearch_data *efi_variable_table __efi_runtime_data;
+static size_t __efi_runtime u16_strlen_runtime(const u16 *s1)
Please, do not duplicate existing functions. If they have to be runtime simply change the existing function to __runtime.
We should do that if possible, but please note that [str|mem]xxx() functions are *architecture* dependent. Do you want to mark all the functions across all the architectures?
Agree or not agree?
+{
- size_t n = 0;
- while (*s1) {
n++;
s1++;
- }
- return n;
+}
+static int __efi_runtime memcmp_runtime(const void *m1, const void *m2,
size_t n)
I dislike duplicate code. Can't we simply define the existing memcmp function as __runtime?
ditto
+{
- while (n && *(u8 *)m1 == *(u8 *)m2) {
n--;
m1++;
m2++;
- }
- if (n)
return *(u8 *)m1 - *(u8 *)m2;
- return 0;
+}
+static void __efi_runtime memcpy_runtime(void *m1, const void *m2, size_t n) +{
Can't we simply define the existing memcpy function as __runtime?
- for (; n; n--, m1++, m2++)
*(u8 *)m1 = *(u8 *)m2;
+}
+static int __efi_runtime efi_cmpkey(_ENTRY *e, const u16 *name,
const efi_guid_t *vendor)
+{
- size_t name_len;
- name_len = u16_strlen_runtime(entry_name(e));
- /* return zero if matched */
- return name_len != u16_strlen_runtime(name) ||
memcmp_runtime(entry_name(e), name, name_len * 2) ||
memcmp_runtime(e->vendor.b, vendor->b, sizeof(vendor));
+}
+/* simplified and slightly different version of hsearch_r() */
These hash functions are so complicated that they really need a unit test testing them rigorously.
Do you know that there are no any tests for hsearch_r()? Moreover, this function would better be exercised well if we could add *runtime* tests.
Have a look at test/env/hashtable.c
Okay, but why not re-use this test for my stripped version?
I think efi_selftest could work after SetVirtualMemoryMap if we do not change the map for memory actually used by U-Boot.
Maybe.
-Takahiro Akashi
Or we use a minimum Linux kernel and put our tests into an init binary.
+static int __efi_runtime hsearch_runtime(const u16 *name,
const efi_guid_t *vendor,
ACTION action,
_ENTRY **retval,
struct hsearch_data *htab)
+{
- unsigned int hval;
- unsigned int count;
- unsigned int len;
- unsigned int idx, new;
- /* Compute an value for the given string. */
- len = u16_strlen_runtime(name);
Can't the same variable name exist for different GUIDs? Why is the GUID not considered in the hash?
efi_cmpkey() does take into consideration GUID as well as the name.
My question concerned the hash.
- hval = len;
- count = len;
- while (count-- > 0) {
hval <<= 4;
hval += name[count];
- }
- /*
* First hash function:
* simply take the modulo but prevent zero.
*/
- hval %= htab->size;
- if (hval == 0)
++hval;
- /* The first index tried. */
- new = -1; /* not found */
- idx = hval;
- if (htab->table[idx].used) {
/*
* Further action might be required according to the
* action value.
*/
unsigned int hval2;
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
/*
* Second hash function:
* as suggested in [Knuth]
*/
hval2 = 1 + hval % (htab->size - 2);
do {
/*
* Because SIZE is prime this guarantees to
* step through all available indices.
*/
if (idx <= hval2)
idx = htab->size + idx - hval2;
else
idx -= hval2;
/*
* If we visited all entries leave the loop
* unsuccessfully.
*/
if (idx == hval)
break;
/* If entry is found use it. */
if (htab->table[idx].used == hval &&
!efi_cmpkey(&htab->table[idx], name, vendor)) {
if (action == FIND) {
*retval = &htab->table[idx];
return idx;
}
/* we don't need to support overwrite */
return -1;
}
} while (htab->table[idx].used);
if (!htab->table[idx].used)
new = idx;
- } else {
new = idx;
- }
- /*
* An empty bucket has been found.
* The following code should never be executed after
* exit_boot_services()
*/
- if (action == ENTER) {
/*
* If table is full and another entry should be
* entered return with error.
*/
if (htab->filled == htab->size) {
*retval = NULL;
return 0;
}
/* Create new entry */
htab->table[new].used = hval;
++htab->filled;
/* return new entry */
*retval = &htab->table[new];
return 1;
- }
- *retval = NULL;
- return 0;
+}
+/* from lib/hashtable.c */ +static inline int isprime(unsigned int number) +{
- /* no even number will be passed */
- unsigned int div = 3;
- while (div * div < number && number % div != 0)
div += 2;
- return number % div != 0;
+}
+efi_status_t efi_freeze_variable_table(void)
Please, add comments to your functions. It is not self-evident what this function is meant to do.
I cannot imagine why a variable cache should be frozen. It is a living data structure until the system is switched off.
I don't get your point.
Please, explain what you mean by freeze.
I suggest that the cache is read/write at all times.
Best regards
Heinrich
For a variable cache I expect that you allocate memory before handling the first variable and never again. At runtime you will not have chance to allocate memory anyway.
I don't get your point.
This function is way too long. Pleae, break it down.
Freezing is implemented in 2-phase steps, and some complexity is inevitable. I suppose adding some comments would be enough.
-Takahiro Akashi
Best regards
Heinrich
+{
- int var_num = 0;
- size_t var_data_size = 0;
- u16 *name;
- efi_uintn_t name_buf_len, name_len;
- efi_guid_t vendor;
- u32 attributes;
- u8 *mem_pool, *var_buf = NULL;
- size_t table_size, var_size, var_buf_size;
- _ENTRY *new = NULL;
- efi_status_t ret;
- /* phase-1 loop */
- name_buf_len = 128;
- name = malloc(name_buf_len);
- if (!name)
return EFI_OUT_OF_RESOURCES;
- name[0] = 0;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND) {
break;
} else if (ret == EFI_BUFFER_TOO_SMALL) {
u16 *buf;
name_buf_len = name_len;
buf = realloc(name, name_buf_len);
if (!buf) {
free(name);
return EFI_OUT_OF_RESOURCES;
}
name = buf;
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len,
name,
&vendor));
}
if (ret != EFI_SUCCESS)
return ret;
var_size = 0;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, NULL));
if (ret != EFI_BUFFER_TOO_SMALL)
return ret;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
var_num++;
var_data_size += (u16_strlen_runtime(name) + 1) * sizeof(u16);
var_data_size += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if (var_data_size & 0x1)
var_data_size++;
- }
- /*
* total of entries in hash table must be a prime number.
* The logic below comes from lib/hashtable.c
*/
- var_num |= 1; /* make odd */
- while (!isprime(var_num))
var_num += 2;
- /* We need table[var_num] for hsearch_runtime algo */
- table_size = sizeof(*efi_variable_table)
+ sizeof(_ENTRY) * (var_num + 1) + var_data_size;
- ret = efi_allocate_pool(EFI_RUNTIME_SERVICES_DATA,
table_size, (void **)&efi_variable_table);
- if (ret != EFI_SUCCESS)
return ret;
- efi_variable_table->size = var_num;
- efi_variable_table->table = (void *)efi_variable_table
+ sizeof(*efi_variable_table);
- mem_pool = (u8 *)efi_variable_table->table
+ sizeof(_ENTRY) * (var_num + 1);
- var_buf_size = 128;
- var_buf = malloc(var_buf_size);
- if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
- }
- /* phase-2 loop */
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = EFI_CALL(efi_get_next_variable_name(&name_len, name,
&vendor));
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = EFI_CALL(efi_get_variable(name, &vendor, &attributes,
&var_size, var_buf));
if (ret == EFI_BUFFER_TOO_SMALL) {
free(var_buf);
var_buf_size = var_size;
var_buf = malloc(var_buf_size);
if (!var_buf) {
ret = EFI_OUT_OF_RESOURCES;
goto err;
}
ret = EFI_CALL(efi_get_variable(name, &vendor,
&attributes,
&var_size, var_buf));
}
if (ret != EFI_SUCCESS)
goto err;
if (!(attributes & EFI_VARIABLE_RUNTIME_ACCESS))
continue;
if (hsearch_runtime(name, &vendor, ENTER, &new,
efi_variable_table) <= 0) {
/* This should not happen */
ret = EFI_INVALID_PARAMETER;
goto err;
}
/* allocate space from RUNTIME DATA */
name_len = (u16_strlen_runtime(name) + 1) * sizeof(u16);
memcpy_runtime(mem_pool, name, name_len);
new->name = mem_pool - (u8 *)new; /* offset */
mem_pool += name_len;
memcpy_runtime(&new->vendor.b, &vendor.b, sizeof(vendor));
new->attributes = attributes;
memcpy_runtime(mem_pool, var_buf, var_size);
new->data = mem_pool - (u8 *)new; /* offset */
new->data_size = var_size;
mem_pool += var_size;
/* mem_pool must 2-byte aligned for u16 variable name */
if ((uintptr_t)mem_pool & 0x1)
mem_pool++;
- }
+#ifdef DEBUG
- name[0] = 0;
- name_len = name_buf_len;
- for (;;) {
name_len = name_buf_len;
ret = efi_get_next_variable_name_runtime(&name_len, name,
&vendor);
if (ret == EFI_NOT_FOUND)
break;
else if (ret != EFI_SUCCESS)
goto err;
var_size = var_buf_size;
ret = efi_get_variable_runtime(name, &vendor, &attributes,
&var_size, var_buf);
if (ret != EFI_SUCCESS)
goto err;
printf("%ls_%pUl:\n", name, &vendor);
printf(" attributes: 0x%x\n", attributes);
printf(" value (size: 0x%lx)\n", var_size);
- }
+#endif
- ret = EFI_SUCCESS;
+err:
- free(name);
- free(var_buf);
- if (ret != EFI_SUCCESS && efi_variable_table) {
efi_free_pool(efi_variable_table);
efi_variable_table = NULL;
- }
- return ret;
+}
+efi_status_t +__efi_runtime EFIAPI efi_get_variable_runtime(u16 *variable_name,
const efi_guid_t *vendor,
u32 *attributes,
efi_uintn_t *data_size,
void *data)
+{
- _ENTRY *new;
- if (!variable_name || !vendor || !data_size)
return EFI_EXIT(EFI_INVALID_PARAMETER);
- if (hsearch_runtime(variable_name, vendor, FIND, &new,
efi_variable_table) <= 0)
return EFI_NOT_FOUND;
- if (attributes)
*attributes = new->attributes;
- if (*data_size < new->data_size) {
*data_size = new->data_size;
return EFI_BUFFER_TOO_SMALL;
- }
- *data_size = new->data_size;
- memcpy_runtime(data, entry_data(new), new->data_size);
- return EFI_SUCCESS;
+}
+static int prev_idx __efi_runtime_data;
+efi_status_t +__efi_runtime EFIAPI efi_get_next_variable_name_runtime(
efi_uintn_t *variable_name_size,
u16 *variable_name,
const efi_guid_t *vendor)
+{
- _ENTRY *e;
- u16 *name;
- efi_uintn_t name_size;
- if (!variable_name_size || !variable_name || !vendor)
return EFI_INVALID_PARAMETER;
- if (variable_name[0]) {
/* sanity check for previous variable */
if (prev_idx < 0)
return EFI_INVALID_PARAMETER;
e = &efi_variable_table->table[prev_idx];
if (!e->used || efi_cmpkey(e, variable_name, vendor))
return EFI_INVALID_PARAMETER;
- } else {
prev_idx = -1;
- }
- /* next variable */
- while (++prev_idx <= efi_variable_table->size) {
e = &efi_variable_table->table[prev_idx];
if (e->used)
break;
- }
- if (prev_idx > efi_variable_table->size)
return EFI_NOT_FOUND;
- name = entry_name(e);
- name_size = (u16_strlen_runtime(name) + 1)
* sizeof(u16);
- if (*variable_name_size < name_size) {
*variable_name_size = name_size;
return EFI_BUFFER_TOO_SMALL;
- }
- memcpy_runtime(variable_name, name, name_size);
- memcpy_runtime((void *)&vendor->b, &e->vendor.b, sizeof(vendor));
- return EFI_SUCCESS;
+} +#endif /* CONFIG_EFI_RUNTIME_GET_VARIABLE_CACHING */

Hi all,
[...]
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Do we really need a multilevel hash table? I would not expect hundreds of variables.
Please don't change your point suddenly. Here we are discussing whether "The cache should be used both at runtime and at boottime" or not.
Heinrich, the idea here is to present a copy of the variables on the OS and totally disable RT variable updating from the OS. If someone wants to update UEFI variables, we can pack them to a Capsule (using FIT image format) and apply them on next reboot. Given the fact that UEFI variables are not updated that often, isn't this a viable option? If it is, then we need to keep the access separated (as Akashi-san suggests) allowing bootime to change the variables.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
You could start the cache implementation with a less complicated data structure like a linked list.
This is totally a different issue. I listed this issue in my cover letter.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables.
We are already working towards having StandAloneMM as an early OP-TEE TA. This will provide us with a secure variable storage for armv7/v8.
In case of write you could do a write-through in your cache if needed.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
Thanks /Ilias

On 6/18/19 12:34 PM, Ilias Apalodimas wrote:
Hi all,
[...]
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Do we really need a multilevel hash table? I would not expect hundreds of variables.
Please don't change your point suddenly. Here we are discussing whether "The cache should be used both at runtime and at boottime" or not.
Heinrich, the idea here is to present a copy of the variables on the OS and totally disable RT variable updating from the OS. If someone wants to update UEFI variables, we can pack them to a Capsule (using FIT image format) and apply them on next reboot. Given the fact that UEFI variables are not updated that often, isn't this a viable option? If it is, then we need to keep the access separated (as Akashi-san suggests) allowing bootime to change the variables.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
You could start the cache implementation with a less complicated data structure like a linked list.
This is totally a different issue. I listed this issue in my cover letter.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables.
We are already working towards having StandAloneMM as an early OP-TEE TA. This will provide us with a secure variable storage for armv7/v8.
What would this OP-TEE binary do? - This seems to be a source of misunderstanding when reviewing this patch.
My guess is that OP-TEE is used to read non-volatile variables only once when starting U-Boot and to write non-volatile variables whenever they are changed.
All further reading of non-volatile variables and all access to volatile variables will be handled by the U-Boot internal variable cache.
For volatile variables I would assume OP-TEE to never see them. This requires that the U-Boot variable cachek supports reading from and writing to the cache at runtime.
StandaloneMmPkg seems to be the hardware independent part of the solution. Where will the hardware driver reside in your OP-TEE solution?
Is the EDK2 hardware store for variables of the MACCHIATObin here: edk2-platforms/Silicon/Marvell/Drivers/Spi/MvFvbDxe/MvFvbDxe.c?
Which hardware platform will you use for testing the U-Boot development of you OP-TEE driver?
Best regards
Heinrich
In case of write you could do a write-through in your cache if needed.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
Thanks /Ilias

Heinrich,
On Tue, Jun 18, 2019 at 10:45:26PM +0200, Heinrich Schuchardt wrote:
On 6/18/19 12:34 PM, Ilias Apalodimas wrote:
Hi all,
[...]
I think one version of the functions serving at runtime and boottime is enough.
The cache should be used both at runtime and at boottime.
So do you mean that we should replace the existing "boottime" version of get/set_variable with my code (algorithm)?
This is a bit complicated work because we should be able to *udpate* UEFI variables at boottime, but my version of hsearch_runtime() is a stripped (and modified) version and doesn't support it.
Do we really need a multilevel hash table? I would not expect hundreds of variables.
Please don't change your point suddenly. Here we are discussing whether "The cache should be used both at runtime and at boottime" or not.
Heinrich, the idea here is to present a copy of the variables on the OS and totally disable RT variable updating from the OS. If someone wants to update UEFI variables, we can pack them to a Capsule (using FIT image format) and apply them on next reboot. Given the fact that UEFI variables are not updated that often, isn't this a viable option? If it is, then we need to keep the access separated (as Akashi-san suggests) allowing bootime to change the variables.
Making the existing hsearch_r() executable at UEFI runtime is, as I said before, quite painful.
You could start the cache implementation with a less complicated data structure like a linked list.
This is totally a different issue. I listed this issue in my cover letter.
Essentially I expect three modules working together:
UEFI API implementation <-> Cache <-> Persistence driver
I would suggest to put each of these into a separate file.
Both the API implementation and the Cache have to be available at Boottime and at Runtime. A first version of the persistence driver may only be working at boottime.
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables.
We are already working towards having StandAloneMM as an early OP-TEE TA. This will provide us with a secure variable storage for armv7/v8.
What would this OP-TEE binary do? - This seems to be a source of misunderstanding when reviewing this patch.
I and Ilias will give you more details offline, here's a short(?) answer:
Standalone MM services here means a SPD entity which provides [Get|Set]Variable APIs to non-secure side firmware, that is currently EDK2. So the source code of Standalone MM services is included in EDK2 repository as a matter of fact.
Here is one drawback: It won't allow for other entities running concurrently on secure side. One example of useful secure feature is (software-based) TPM. So Linaro is working on modifying/transforming Standalone MM to one OP-TEE application, which Ilias mentioned above.
My guess is that OP-TEE is used to read non-volatile variables only once when starting U-Boot and to write non-volatile variables whenever they are changed.
So OP-TEE version of StMM is still on-going project and I assume that this OP-TEE app will support the same set of functionality/APIs as StMM does.
All further reading of non-volatile variables and all access to volatile variables will be handled by the U-Boot internal variable cache.
For volatile variables I would assume OP-TEE to never see them. This requires that the U-Boot variable cachek supports reading from and writing to the cache at runtime.
No. As far as I correctly understand, StMM handles volatile variables as well as non-volatile variables. EDK2 on non-secure side will redirect user's request directly to secure side even without *caching* variable's values.
StandaloneMmPkg seems to be the hardware independent part of the solution. Where will the hardware driver reside in your OP-TEE solution?
Is the EDK2 hardware store for variables of the MACCHIATObin here: edk2-platforms/Silicon/Marvell/Drivers/Spi/MvFvbDxe/MvFvbDxe.c?
Which hardware platform will you use for testing the U-Boot development of you OP-TEE driver?
Ilias will be able to answer those questions.
Thanks, -Takahiro Akashi
Best regards
Heinrich
In case of write you could do a write-through in your cache if needed.
The NV-cache content should be written to non-volatile memory on Reset() and on ExitBootServices() and if possible when updating variables at runtime.
I'm not sure your intent here, but are you going to write back the cache only once? It won't work as every change of UEFI variable must be flushed to persistent storage instantly.
Thanks /Ilias

Heinrich,
[...]
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables. In the case, all the cache would be bypassed. In my old prototype, I utilized the cache but dropped that feature for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
Unfortunately, this is not practical right now because there is already some sort of assumption (and consensus) that we would re-use "Standalone MM services", which is already there in EDK2, as secure storage for UEFI variables.
We are already working towards having StandAloneMM as an early OP-TEE TA. This will provide us with a secure variable storage for armv7/v8.
What would this OP-TEE binary do? - This seems to be a source of misunderstanding when reviewing this patch.
I and Ilias will give you more details offline, here's a short(?) answer:
Standalone MM services here means a SPD entity which provides [Get|Set]Variable APIs to non-secure side firmware, that is currently EDK2. So the source code of Standalone MM services is included in EDK2 repository as a matter of fact.
Here is one drawback: It won't allow for other entities running concurrently on secure side. One example of useful secure feature is (software-based) TPM. So Linaro is working on modifying/transforming Standalone MM to one OP-TEE application, which Ilias mentioned above.
Exactly. The current StMM implementation exists for Armv8 *only* in SPM (Secure Partition Manager). The idea is to make it an OP-TEE application, so we can run it on on Armv7s as well. As Akashi-san mentions SPD (Secure Payload Dispatcher) and SPM are mutually exclusive so having everything as OP-TEE trusted applications gives us a number of advantages at the moment.
My guess is that OP-TEE is used to read non-volatile variables only once when starting U-Boot and to write non-volatile variables whenever they are changed.
So OP-TEE version of StMM is still on-going project and I assume that this OP-TEE app will support the same set of functionality/APIs as StMM does.
Yes that's the goal
All further reading of non-volatile variables and all access to volatile variables will be handled by the U-Boot internal variable cache.
For volatile variables I would assume OP-TEE to never see them. This requires that the U-Boot variable cachek supports reading from and writing to the cache at runtime.
No. As far as I correctly understand, StMM handles volatile variables as well as non-volatile variables. EDK2 on non-secure side will redirect user's request directly to secure side even without *caching* variable's values.
Similar understanding here. The question is, will we have to think of something for non-arm architectures?
StandaloneMmPkg seems to be the hardware independent part of the solution. Where will the hardware driver reside in your OP-TEE solution?
It depends on where your hardware is. If you have a NOR flash directly connected to the secure world the answer is yes. For starters we are going to use RPMB + U-Boot supplicant.
Is the EDK2 hardware store for variables of the MACCHIATObin here: edk2-platforms/Silicon/Marvell/Drivers/Spi/MvFvbDxe/MvFvbDxe.c?
No idea, i can ask around.
Which hardware platform will you use for testing the U-Boot development of you OP-TEE driver?
Ilias will be able to answer those questions.
- stm32mp1 ST board based on armv7 [1] - Socionext DeveloperBox for armv8 [2]. This has a running EDKII implementation of StMM in SPM. The underlying firmware should be irrelevant though since the whole functionality is contained within an OP-TEE TA (trusted application). If i remember correctly that will need an extra driver in OP-TEE (if we port U-Boot on that) - TI AM6 board [3]. I don't have the board in my hands yet, so no details on it
[1] https://www.st.com/en/evaluation-tools/stm32mp157c-dk2.html [2] https://www.96boards.org/product/developerbox/ [3] http://www.ti.com/tool/PROCESSOR-SDK-AM65X
Regards /Ilias

On 6/19/19 7:13 AM, Ilias Apalodimas wrote:
Heinrich,
[...]
> Unfortunately, this is not practical right now because there is > already some sort of assumption (and consensus) that we would re-use > "Standalone MM services", which is already there in EDK2, as > secure storage for UEFI variables. > In the case, all the cache would be bypassed. > In my old prototype, I utilized the cache but dropped that feature > for several reasons.
What has EDK2 code to do with it?
Did you follow my comment below?
> Unfortunately, this is not practical right now because there is > already some sort of assumption (and consensus) that we would re-use > "Standalone MM services", which is already there in EDK2, as > secure storage for UEFI variables.
We are already working towards having StandAloneMM as an early OP-TEE TA. This will provide us with a secure variable storage for armv7/v8.
What would this OP-TEE binary do? - This seems to be a source of misunderstanding when reviewing this patch.
I and Ilias will give you more details offline, here's a short(?) answer:
Standalone MM services here means a SPD entity which provides [Get|Set]Variable APIs to non-secure side firmware, that is currently EDK2. So the source code of Standalone MM services is included in EDK2 repository as a matter of fact.
Here is one drawback: It won't allow for other entities running concurrently on secure side. One example of useful secure feature is (software-based) TPM. So Linaro is working on modifying/transforming Standalone MM to one OP-TEE application, which Ilias mentioned above.
Exactly. The current StMM implementation exists for Armv8 *only* in SPM (Secure Partition Manager). The idea is to make it an OP-TEE application, so we can run it on on Armv7s as well. As Akashi-san mentions SPD (Secure Payload Dispatcher) and SPM are mutually exclusive so having everything as OP-TEE trusted applications gives us a number of advantages at the moment.
My guess is that OP-TEE is used to read non-volatile variables only once when starting U-Boot and to write non-volatile variables whenever they are changed.
So OP-TEE version of StMM is still on-going project and I assume that this OP-TEE app will support the same set of functionality/APIs as StMM does.
Yes that's the goal
Do I understand it write:
In U-Boot we would have code that essentially provides the functionality of EDK2's EFI_SMM_VARIABLE_PROTOCOL. Like MdeModulePkg/Universal/Variable/RuntimeDxe/VariableSmm.c this would talk via SMI calls with the hardware drivers, in this case the OP-TEE app.
This would allow the OP-TEE app to be used both in U-Boot and in EDK2?
All further reading of non-volatile variables and all access to volatile variables will be handled by the U-Boot internal variable cache.
For volatile variables I would assume OP-TEE to never see them. This requires that the U-Boot variable cachek supports reading from and writing to the cache at runtime.
No. As far as I correctly understand, StMM handles volatile variables as well as non-volatile variables. EDK2 on non-secure side will redirect user's request directly to secure side even without *caching* variable's values.
Similar understanding here. The question is, will we have to think of something for non-arm architectures?
The design should be open for other architectures. If we use the EFI_SMM_VARIABLE_PROTOCOL as the interface we should be able to support other architectures.
I am just wondering why does the OP-TEE module handle all the logic of EFI_SMM_VARIABLE_PROTOCOL. Wouldn't something like the EFI_FIRMWARE_VOLUME_BLOCK_PROTOCOL make more sense?
This protocol could also be used to implement CapsuleUpdate().
StandaloneMmPkg seems to be the hardware independent part of the solution. Where will the hardware driver reside in your OP-TEE solution?
It depends on where your hardware is. If you have a NOR flash directly connected to the secure world the answer is yes. For starters we are going to use RPMB + U-Boot supplicant.
Is the EDK2 hardware store for variables of the MACCHIATObin here: edk2-platforms/Silicon/Marvell/Drivers/Spi/MvFvbDxe/MvFvbDxe.c?
No idea, i can ask around.
Which hardware platform will you use for testing the U-Boot development of you OP-TEE driver?
Ilias will be able to answer those questions.
- stm32mp1 ST board based on armv7 [1]
- Socionext DeveloperBox for armv8 [2]. This has a running EDKII implementation
of StMM in SPM. The underlying firmware should be irrelevant though since the whole functionality is contained within an OP-TEE TA (trusted application). If i remember correctly that will need an extra driver in OP-TEE (if we port U-Boot on that)
- TI AM6 board [3]. I don't have the board in my hands yet, so no details on it
I have a MACCHIATObin. Would this also be usable for the testing?
Regards
Heinrich
[1] https://www.st.com/en/evaluation-tools/stm32mp157c-dk2.html [2] https://www.96boards.org/product/developerbox/ [3] http://www.ti.com/tool/PROCESSOR-SDK-AM65X
Regards /Ilias

On Wed, Jun 19, 2019 at 08:24:50AM +0200, Heinrich Schuchardt wrote:
On 6/19/19 7:13 AM, Ilias Apalodimas wrote:
Heinrich,
[...]
>>Unfortunately, this is not practical right now because there is >>already some sort of assumption (and consensus) that we would re-use >>"Standalone MM services", which is already there in EDK2, as >>secure storage for UEFI variables. >>In the case, all the cache would be bypassed. >>In my old prototype, I utilized the cache but dropped that feature >>for several reasons. > >What has EDK2 code to do with it?
Did you follow my comment below? >>Unfortunately, this is not practical right now because there is >>already some sort of assumption (and consensus) that we would re-use >>"Standalone MM services", which is already there in EDK2, as >>secure storage for UEFI variables.
We are already working towards having StandAloneMM as an early OP-TEE TA. This will provide us with a secure variable storage for armv7/v8.
What would this OP-TEE binary do? - This seems to be a source of misunderstanding when reviewing this patch.
I and Ilias will give you more details offline, here's a short(?) answer:
Standalone MM services here means a SPD entity which provides [Get|Set]Variable APIs to non-secure side firmware, that is currently EDK2. So the source code of Standalone MM services is included in EDK2 repository as a matter of fact.
Here is one drawback: It won't allow for other entities running concurrently on secure side. One example of useful secure feature is (software-based) TPM. So Linaro is working on modifying/transforming Standalone MM to one OP-TEE application, which Ilias mentioned above.
Exactly. The current StMM implementation exists for Armv8 *only* in SPM (Secure Partition Manager). The idea is to make it an OP-TEE application, so we can run it on on Armv7s as well. As Akashi-san mentions SPD (Secure Payload Dispatcher) and SPM are mutually exclusive so having everything as OP-TEE trusted applications gives us a number of advantages at the moment.
My guess is that OP-TEE is used to read non-volatile variables only once when starting U-Boot and to write non-volatile variables whenever they are changed.
So OP-TEE version of StMM is still on-going project and I assume that this OP-TEE app will support the same set of functionality/APIs as StMM does.
Yes that's the goal
Do I understand it write:
In U-Boot we would have code that essentially provides the functionality of EDK2's EFI_SMM_VARIABLE_PROTOCOL. Like MdeModulePkg/Universal/Variable/RuntimeDxe/VariableSmm.c this would talk via SMI calls with the hardware drivers, in this case the OP-TEE app.
This would allow the OP-TEE app to be used both in U-Boot and in EDK2?
I think so.
All further reading of non-volatile variables and all access to volatile variables will be handled by the U-Boot internal variable cache.
For volatile variables I would assume OP-TEE to never see them. This requires that the U-Boot variable cachek supports reading from and writing to the cache at runtime.
No. As far as I correctly understand, StMM handles volatile variables as well as non-volatile variables. EDK2 on non-secure side will redirect user's request directly to secure side even without *caching* variable's values.
Similar understanding here. The question is, will we have to think of something for non-arm architectures?
The design should be open for other architectures. If we use the EFI_SMM_VARIABLE_PROTOCOL as the interface we should be able to support other architectures.
Yeah, but please note that EFI_SMM_VARIABLE_PROTOCOL is not part of UEFI specification.
I am just wondering why does the OP-TEE module handle all the logic of EFI_SMM_VARIABLE_PROTOCOL. Wouldn't something like the EFI_FIRMWARE_VOLUME_BLOCK_PROTOCOL make more sense?
It provides only read/write interfaces *without* protection. As far as SetVariable API is concerned, you cannot compromise any authenticated variables unless you can sign a given variable with a trusted private key even if you can make a SMI call.
-Takahiro Akashi
This protocol could also be used to implement CapsuleUpdate().
StandaloneMmPkg seems to be the hardware independent part of the solution. Where will the hardware driver reside in your OP-TEE solution?
It depends on where your hardware is. If you have a NOR flash directly connected to the secure world the answer is yes. For starters we are going to use RPMB + U-Boot supplicant.
Is the EDK2 hardware store for variables of the MACCHIATObin here: edk2-platforms/Silicon/Marvell/Drivers/Spi/MvFvbDxe/MvFvbDxe.c?
No idea, i can ask around.
Which hardware platform will you use for testing the U-Boot development of you OP-TEE driver?
Ilias will be able to answer those questions.
- stm32mp1 ST board based on armv7 [1]
- Socionext DeveloperBox for armv8 [2]. This has a running EDKII implementation
of StMM in SPM. The underlying firmware should be irrelevant though since the whole functionality is contained within an OP-TEE TA (trusted application). If i remember correctly that will need an extra driver in OP-TEE (if we port U-Boot on that)
- TI AM6 board [3]. I don't have the board in my hands yet, so no details on it
I have a MACCHIATObin. Would this also be usable for the testing?
I think the answer will depend on what (feature) you want to test.
-Takahiro Akashi
Regards
Heinrich
[1] https://www.st.com/en/evaluation-tools/stm32mp157c-dk2.html [2] https://www.96boards.org/product/developerbox/ [3] http://www.ti.com/tool/PROCESSOR-SDK-AM65X
Regards /Ilias

On 6/5/19 6:21 AM, AKASHI Takahiro wrote:
EBBR v1.0 section 2.5.3 says Even when SetVariable() is not supported during runtime services, firmware should cache variable names and values in EfiRuntimeServicesData memory so that GetVariable() and GetNextVeriableName() can behave as specified.
This is an experimental patch set and the aim is to enable this feature.
Hello Takahiro,
thanks a lot for all the effort you put in this.
Reusing parts of env is a valid possibility but I am not sure if this is the right approach for something that has to work at runtime.
I will need some time to layout my ideas of the design. Unfortunately I am on a sea kayaking trip this weekend. So I may need until June 16th for a thorough suggestion.
I have already picked the trivial patches from this patch series and will put them into the next pull request.
http://git.denx.de/?p=u-boot-efi.git;a=shortlog;h=refs/tags/efi-2019-07-rc4-...
Best regards
Heinich
Cache buffer is in the same format as U-Boot environment hash table, but we cannot use functions in hashtable.c partly because most of U-Boot code are not available at UEFI runtime and partly because entries in a table are allocated by malloc(), which are again not available at UEFI runtime. It is quite painful to modify exiting U-Boot code so as to make it executable at UEFI runtime.
So I implemented a limited version of hsearch_r(). This may be a discussion. Given that there are not so many UEFI variables, we may want to use a simpler table format for caching.
Known issues:
- Currently I test this feature with a test which is temporarily embedded in ExitBootServices.
- After SetVirtualAddressMap, it won't work (TODO). ConvertPointer was implemented here just for this future work.
- So how can we test the feature? We can't use "printenv" command as its relocation won't take place.
- I see some Travis CI errors. This is probably due to no storage configured for UEFI variables.
Patch#1 to #4 are preparatory patches. Patch#5 is mainly for testing. Patch#6 is core part of this patch set.
AKASHI Takahiro (5): efi_loader: runtime: make SetVirtualAddressMap configurable efi: add RuntimeServicesSupported variable efi_loader: support convert_pointer at runtime cmd: efidebug: add "boot exit" sub-command efi_loader: variable: support runtime variable access via cache
Alexander Graf (1): efi_loader: Patch non-runtime code out at ExitBootServices already
cmd/efidebug.c | 63 ++++ include/efi_api.h | 15 + include/efi_loader.h | 22 ++ lib/efi_loader/Kconfig | 24 ++ lib/efi_loader/efi_boottime.c | 11 +- lib/efi_loader/efi_runtime.c | 155 ++++++++-- lib/efi_loader/efi_setup.c | 5 + lib/efi_loader/efi_variable.c | 467 +++++++++++++++++++++++++++++ test/py/tests/test_efi_selftest.py | 4 +- 9 files changed, 735 insertions(+), 31 deletions(-)
participants (4)
-
AKASHI Takahiro
-
Heinrich Schuchardt
-
Ilias Apalodimas
-
Mark Kettenis