Skip to contentSkip to conversation

efi: libstub: Use relocated version of kernel's struct screen_info

PATCH Superseded Ard Biesheuvel 2 participants 5 messages Mar 13, 2023

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Ard Biesheuvel

author maintainer
#0 lore raw

PATCH 1/1 efi: libstub: Use relocated version of kernel's struct screen_info

In some cases, we expose the kernel's struct screen_info to the EFI stub directly, so it gets populated before even entering the kernel. This means the early console is available as soon as the early param parsing happens, which is nice. It also means we need two different ways to pass this information, as this trick only works if the EFI stub is baked into the core kernel image, which is not always the case.

Huacai reports that the preparatory refactoring that was needed to implement this alternative method for zboot resulted in a non-functional efifb earlycon for other cases as well, due to the reordering of the kernel image relocation with the population of the screen_info struct, and the latter now takes place after copying the image to its new location, which means we copy the old, uninitialized state.

So let's ensure that alloc_screen_info() produces the correct screen_info pointer, by keeping its relocated address in a global variable.

Cc Xuefeng Li <[email protected]>
Cc Xuerui Wang <[email protected]>
Fixes 42c8ea3dca094ab8 ("efi: libstub: Factor out EFI stub entrypoint into separate file")
---
 drivers/firmware/efi/libstub/efi-stub-entry.c | 14 ++++++++++++++
 drivers/firmware/efi/libstub/efi-stub.c       |  9 ---------
 drivers/firmware/efi/libstub/screen_info.c    |  7 -------
 3 files changed, 14 insertions(+), 16 deletions(-)
drivers/firmware/efi/libstub/efi-stub-entry.c +14 -0
index 5245c4f031c0a70a..883ef2c88f522c4b 100644
@@ -5,6 +5,17 @@
 
 #include "efistub.h"
 
+static struct screen_info *si;
+
+struct screen_info *alloc_screen_info(void)
+{
+	return si;
+}
+
+void free_screen_info(struct screen_info *si)
+{
+}
+
 /*
  * EFI entry point for the generic EFI stub used by ARM, arm64, RISC-V and
  * LoongArch. This is the entrypoint that is described in the PE/COFF header
@@ -56,6 +67,9 @@ efi_status_t __efiapi efi_pe_entry(efi_handle_t handle,
 		return status;
 	}
 
+	/* point si to the relocated copy of struct screen_info */
+	si = (void *)&screen_info + image_addr - (unsigned long)image->image_base;
+
 	status = efi_stub_common(handle, image, image_addr, cmdline_ptr);
 
 	efi_free(image_size, image_addr);
drivers/firmware/efi/libstub/efi-stub.c +0 -9
index 2955c1ac6a36ee00..c4b9eccad0f103dc 100644
@@ -47,15 +47,6 @@
 static u64 virtmap_base = EFI_RT_VIRTUAL_BASE;
 static bool flat_va_mapping = (EFI_RT_VIRTUAL_OFFSET != 0);
 
-struct screen_info * __weak alloc_screen_info(void)
-{
-	return &screen_info;
-}
-
-void __weak free_screen_info(struct screen_info *si)
-{
-}
-
 static struct screen_info *setup_graphics(void)
 {
 	efi_guid_t gop_proto = EFI_GRAPHICS_OUTPUT_PROTOCOL_GUID;
drivers/firmware/efi/libstub/screen_info.c +0 -7
index 8e76a8b384ba142d..b2637420fe53911e 100644
@@ -15,14 +15,7 @@
  * early, but it only works if the EFI stub is part of the core kernel image
  * itself. The zboot decompressor can only use the configuration table
  * approach.
- *
- * In order to support both methods from the same build of the EFI stub
- * library, provide this dummy global definition of struct screen_info. If it
- * is required to satisfy a link dependency, it means we need to override the
- * __weak alloc and free methods with the ones below, and those will be pulled
- * in as well.
  */
-struct screen_info screen_info;
 
 static efi_guid_t screen_info_guid = LINUX_EFI_SCREEN_INFO_TABLE_GUID;
 
Signature · 1 line
2.39.2
author maintainer
#1 lore raw
Quoted 19 lines · Ard Biesheuvel

In some cases, we expose the kernel's struct screen_info to the EFI stub directly, so it gets populated before even entering the kernel. This means the early console is available as soon as the early param parsing happens, which is nice. It also means we need two different ways to pass this information, as this trick only works if the EFI stub is baked into the core kernel image, which is not always the case.

Huacai reports that the preparatory refactoring that was needed to implement this alternative method for zboot resulted in a non-functional efifb earlycon for other cases as well, due to the reordering of the kernel image relocation with the population of the screen_info struct, and the latter now takes place after copying the image to its new location, which means we copy the old, uninitialized state.

So let's ensure that alloc_screen_info() produces the correct screen_info pointer, by keeping its relocated address in a global variable.

Quoted 8 lines
Cc Xuefeng Li <[email protected]>
Cc Xuerui Wang <[email protected]>
Fixes 42c8ea3dca094ab8 ("efi: libstub: Factor out EFI stub entrypoint into separate file")

This is still not working :-(

Quoted 80 lines
---
 drivers/firmware/efi/libstub/efi-stub-entry.c | 14 ++++++++++++++
 drivers/firmware/efi/libstub/efi-stub.c       |  9 ---------
 drivers/firmware/efi/libstub/screen_info.c    |  7 -------
 3 files changed, 14 insertions(+), 16 deletions(-)
drivers/firmware/efi/libstub/efi-stub-entry.c +14 -0
index 5245c4f031c0a70a..883ef2c88f522c4b 100644
@@ -5,6 +5,17 @@
 
 #include "efistub.h"
 
+static struct screen_info *si;
+
+struct screen_info *alloc_screen_info(void)
+{
+       return si;
+}
+
+void free_screen_info(struct screen_info *si)
+{
+}
+
 /*
  * EFI entry point for the generic EFI stub used by ARM, arm64, RISC-V and
  * LoongArch. This is the entrypoint that is described in the PE/COFF header
@@ -56,6 +67,9 @@ efi_status_t __efiapi efi_pe_entry(efi_handle_t handle,
                return status;
        }
 
+       /* point si to the relocated copy of struct screen_info */
+       si = (void *)&screen_info + image_addr - (unsigned long)image->image_base;
+
        status = efi_stub_common(handle, image, image_addr, cmdline_ptr);
 
        efi_free(image_size, image_addr);
drivers/firmware/efi/libstub/efi-stub.c +0 -9
index 2955c1ac6a36ee00..c4b9eccad0f103dc 100644
@@ -47,15 +47,6 @@
 static u64 virtmap_base = EFI_RT_VIRTUAL_BASE;
 static bool flat_va_mapping = (EFI_RT_VIRTUAL_OFFSET != 0);
 
-struct screen_info * __weak alloc_screen_info(void)
-{
-       return &screen_info;
-}
-
-void __weak free_screen_info(struct screen_info *si)
-{
-}
-
 static struct screen_info *setup_graphics(void)
 {
        efi_guid_t gop_proto = EFI_GRAPHICS_OUTPUT_PROTOCOL_GUID;
drivers/firmware/efi/libstub/screen_info.c +0 -8
index 8e76a8b384ba142d..b2637420fe53911e 100644
@@ -15,14 +15,7 @@
  * early, but it only works if the EFI stub is part of the core kernel image
  * itself. The zboot decompressor can only use the configuration table
  * approach.
- *
- * In order to support both methods from the same build of the EFI stub
- * library, provide this dummy global definition of struct screen_info. If it
- * is required to satisfy a link dependency, it means we need to override the
- * __weak alloc and free methods with the ones below, and those will be pulled
- * in as well.
  */
-struct screen_info screen_info;
 
 static efi_guid_t screen_info_guid = LINUX_EFI_SCREEN_INFO_TABLE_GUID;
 
--

2.39.2

author maintainer
#2 lore raw
Quoted 30 lines · Ard Biesheuvel
Quoted 19 lines · Ard Biesheuvel

In some cases, we expose the kernel's struct screen_info to the EFI stub directly, so it gets populated before even entering the kernel. This means the early console is available as soon as the early param parsing happens, which is nice. It also means we need two different ways to pass this information, as this trick only works if the EFI stub is baked into the core kernel image, which is not always the case.

Huacai reports that the preparatory refactoring that was needed to implement this alternative method for zboot resulted in a non-functional efifb earlycon for other cases as well, due to the reordering of the kernel image relocation with the population of the screen_info struct, and the latter now takes place after copying the image to its new location, which means we copy the old, uninitialized state.

So let's ensure that alloc_screen_info() produces the correct screen_info pointer, by keeping its relocated address in a global variable.

Quoted 8 lines
Cc Xuefeng Li <[email protected]>
Cc Xuerui Wang <[email protected]>
Fixes 42c8ea3dca094ab8 ("efi: libstub: Factor out EFI stub entrypoint into separate file")
Quoted 2 lines

This is still not working :-(

Huacai, could you try this patch please? It is working now - I don't know what I did wrong before when testing it.

maintainer Tested-by
#3 lore raw

Hi, Ard,

Quoted 32 lines · Ard Biesheuvel
Quoted 30 lines · Ard Biesheuvel
Quoted 27 lines · Ard Biesheuvel

In some cases, we expose the kernel's struct screen_info to the EFI stub directly, so it gets populated before even entering the kernel. This means the early console is available as soon as the early param parsing happens, which is nice. It also means we need two different ways to pass this information, as this trick only works if the EFI stub is baked into the core kernel image, which is not always the case.

Huacai reports that the preparatory refactoring that was needed to implement this alternative method for zboot resulted in a non-functional efifb earlycon for other cases as well, due to the reordering of the kernel image relocation with the population of the screen_info struct, and the latter now takes place after copying the image to its new location, which means we copy the old, uninitialized state.

So let's ensure that alloc_screen_info() produces the correct screen_info pointer, by keeping its relocated address in a global variable.

Cc Xuefeng Li <[email protected]>
Cc Xuerui Wang <[email protected]>
Fixes 42c8ea3dca094ab8 ("efi: libstub: Factor out EFI stub entrypoint into separate file")
Quoted 5 lines
Quoted 2 lines

This is still not working :-(

Huacai, could you try this patch please? It is working now - I don't know what I did wrong before when testing it.

I've tested, it works for me.

But I think my original patch is also OK, unless screen_info is needed by zboot (but the earliest zboot doesn't need it).

Huacai

Quoted 1 line
author maintainer
#4 lore raw
Quoted 39 lines · Huacai Chen

Hi, Ard,

Quoted 35 lines · Ard Biesheuvel
Quoted 32 lines · Ard Biesheuvel
Quoted 19 lines · Ard Biesheuvel

In some cases, we expose the kernel's struct screen_info to the EFI stub directly, so it gets populated before even entering the kernel. This means the early console is available as soon as the early param parsing happens, which is nice. It also means we need two different ways to pass this information, as this trick only works if the EFI stub is baked into the core kernel image, which is not always the case.

Huacai reports that the preparatory refactoring that was needed to implement this alternative method for zboot resulted in a non-functional efifb earlycon for other cases as well, due to the reordering of the kernel image relocation with the population of the screen_info struct, and the latter now takes place after copying the image to its new location, which means we copy the old, uninitialized state.

So let's ensure that alloc_screen_info() produces the correct screen_info pointer, by keeping its relocated address in a global variable.

Quoted 8 lines
Cc Xuefeng Li <[email protected]>
Cc Xuerui Wang <[email protected]>
Fixes 42c8ea3dca094ab8 ("efi: libstub: Factor out EFI stub entrypoint into separate file")

This is still not working :-(

Quoted 9 lines
Quoted 2 lines

Huacai, could you try this patch please? It is working now - I don't know what I did wrong before when testing it.

I've tested, it works for me.

But I think my original patch is also OK, unless screen_info is needed by zboot (but the earliest zboot doesn't need it).

Thanks.

zboot does need it, otherwise the EFI framebuffer does not work at all.