ARM: efi: PE/COFF cleanup/hardening

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Cc Russell King <[email protected]>

Signed-off-by Ard Biesheuvel <[email protected]>

---
 arch/arm/boot/compressed/efi-header.S  | 32 ++++++++++++++------
 arch/arm/boot/compressed/vmlinux.lds.S | 30 +++++++++++++-----
 2 files changed, 46 insertions(+), 16 deletions(-)

arch/arm/boot/compressed/efi-header.S +23 -9

index 542e1ad432ae..c94a88ae834d 100644
@@ -54,20 +54,22 @@ coff_header:
 			IMAGE_FILE_EXECUTABLE_IMAGE | \
 			IMAGE_FILE_LINE_NUMS_STRIPPED	@ Characteristics
 
+#define __pecoff_code_size (__pecoff_data_start - __efi_start)
+
 optional_header:
 		.short	PE_OPT_MAGIC_PE32		@ PE32 format
 		.byte	0x02				@ MajorLinkerVersion
 		.byte	0x14				@ MinorLinkerVersion
-		.long	_end - __efi_start		@ SizeOfCode
-		.long	0				@ SizeOfInitializedData
+		.long	__pecoff_code_size		@ SizeOfCode
+		.long	__pecoff_data_size		@ SizeOfInitializedData
 		.long	0				@ SizeOfUninitializedData
 		.long	efi_stub_entry - start		@ AddressOfEntryPoint
 		.long	start_offset			@ BaseOfCode
-		.long	0				@ BaseOfData
+		.long	__pecoff_data_start - start	@ BaseOfData
 
 extra_header_fields:
 		.long	0				@ ImageBase
-		.long	SZ_512				@ SectionAlignment
+		.long	SZ_4K				@ SectionAlignment
 		.long	SZ_512				@ FileAlignment
 		.short	0				@ MajorOsVersion
 		.short	0				@ MinorOsVersion

@@ -77,7 +79,7 @@ extra_header_fields:
 		.short	0				@ MinorSubsystemVersion
 		.long	0				@ Win32VersionValue
 
-		.long	_end - start			@ SizeOfImage
+		.long	__pecoff_end - start		@ SizeOfImage
 		.long	start_offset			@ SizeOfHeaders
 		.long	0				@ CheckSum
 		.short	IMAGE_SUBSYSTEM_EFI_APPLICATION	@ Subsystem

@@ -98,9 +100,9 @@ extra_header_fields:
 
 section_table:
 		.ascii	".text\0\0\0"
-		.long	_end - __efi_start		@ VirtualSize
+		.long	__pecoff_code_size		@ VirtualSize
 		.long	__efi_start			@ VirtualAddress
-		.long	_edata - __efi_start		@ SizeOfRawData
+		.long	__pecoff_code_size		@ SizeOfRawData
 		.long	__efi_start			@ PointerToRawData
 		.long	0				@ PointerToRelocations
 		.long	0				@ PointerToLineNumbers

@@ -108,12 +110,24 @@ section_table:
 		.short	0				@ NumberOfLineNumbers
 		.long	IMAGE_SCN_CNT_CODE | \
 			IMAGE_SCN_MEM_READ | \
-			IMAGE_SCN_MEM_WRITE | \
 			IMAGE_SCN_MEM_EXECUTE		@ Characteristics
 
+		.ascii	".data\0\0\0"
+		.long	__pecoff_data_size		@ VirtualSize
+		.long	__pecoff_data_start - start	@ VirtualAddress
+		.long	__pecoff_data_rawsize		@ SizeOfRawData
+		.long	__pecoff_data_start - start	@ PointerToRawData
+		.long	0				@ PointerToRelocations
+		.long	0				@ PointerToLineNumbers
+		.short	0				@ NumberOfRelocations
+		.short	0				@ NumberOfLineNumbers
+		.long	IMAGE_SCN_CNT_INITIALIZED_DATA | \
+			IMAGE_SCN_MEM_READ | \
+			IMAGE_SCN_MEM_WRITE		@ Characteristics
+
 		.set	section_count, (. - section_table) / 40
 
-		.align	9
+		.align	12
 __efi_start:
 #endif
 		.endm

arch/arm/boot/compressed/vmlinux.lds.S +23 -7

index 1fa62432e283..dfcc2baa0077 100644
@@ -53,13 +53,6 @@ SECTIONS
     *(.rodata)
     *(.rodata.*)
   }
-  .data : {
-    /*
-     * The EFI stub always executes from RAM, and runs strictly before the
-     * decompressor, so we can make an exception for its r/w data, and keep it
-     */
-    *(.data.efistub)
-  }
   .piggydata : {
     *(.piggydata)
   }

@@ -75,6 +68,26 @@ SECTIONS
   /* ensure the zImage file size is always a multiple of 64 bits */
   /* (without a dummy byte, ld just ignores the empty section) */
   .pad			: { BYTE(0); . = ALIGN(8); }
+
+#ifdef CONFIG_EFI_STUB
+  .data : ALIGN(4096) {
+    __pecoff_data_start = .;
+    /*
+     * The EFI stub always executes from RAM, and runs strictly before the
+     * decompressor, so we can make an exception for its r/w data, and keep it
+     */
+    *(.data.efistub)
+    __pecoff_data_end = .;
+
+    /*
+     * PE/COFF mandates a file size which is a multiple of 512 bytes if the
+     * section size equals or exceeds 4 KB
+     */
+    . = ALIGN(512);
+  }
+  __pecoff_data_rawsize = . - ADDR(.data);
+#endif
+
   _edata = .;
 
   _magic_sig = ZIMAGE_MAGIC(0x016f2818);

@@ -89,6 +102,9 @@ SECTIONS
   . = ALIGN(8);		/* the stack must be 64-bit aligned */
   .stack		: { *(.stack) }
 
+  PROVIDE(__pecoff_data_size = ALIGN(512) - ADDR(.data));
+  PROVIDE(__pecoff_end = ALIGN(512));
+
   .stab 0		: { *(.stab) }
   .stabstr 0		: { *(.stabstr) }
   .stab.excl 0		: { *(.stab.excl) }

Signature · 7 lines

2.9.3


_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 6 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Quoted 4 lines

Currently the solution is to revert this patch.

Have you a better option?

I will investigate.

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

arch/arm/boot/compressed/efi-header.S +1 -1

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 
                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

Hi Ard,

Quoted 39 lines · Ard Biesheuvel

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 7 lines

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

I've just tested this change and it didn't fix anything.

Gregory

Quoted 9 lines

arch/arm/boot/compressed/efi-header.S +0 -0

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 

arch/arm/boot/compressed/efi-header.S section_table: · line 133

Quoted 7 lines

arch/arm/boot/compressed/efi-header.S +1 -1

                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

Quoted 64 lines · Gregory CLEMENT

Hi Ard,

Quoted 39 lines · Ard Biesheuvel

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 7 lines

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

I've just tested this change and it didn't fix anything.

Gregory

Quoted 9 lines

arch/arm/boot/compressed/efi-header.S +0 -0

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 

Quoted 8 lines

arch/arm/boot/compressed/efi-header.S section_table: · line 133

Quoted 7 lines

arch/arm/boot/compressed/efi-header.S +1 -1

                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

How about this?

arch/arm/boot/compressed/vmlinux.lds.S +5 -0

index 7a4c59154361..dfcc2baa0077 100644
@@ -29,6 +29,11 @@ SECTIONS
      * of the text/got segments.
      */
     *(.data)
+    /*
+     * C code that is shared with the kernel proper (but rebuilt for the
+     * decompressor) may contain exports that we have no use for here.
+     */
+    *(*ksymtab* *kcrctab*)
   }
 
   . = TEXT_START;

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

Hi Ard,

Quoted 76 lines · Ard Biesheuvel

Quoted 64 lines · Gregory CLEMENT

Hi Ard,

Quoted 39 lines · Ard Biesheuvel

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 7 lines

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

I've just tested this change and it didn't fix anything.

Gregory

Quoted 9 lines

arch/arm/boot/compressed/efi-header.S +0 -0

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 

Quoted 8 lines

arch/arm/boot/compressed/efi-header.S section_table: · line 133

Quoted 7 lines

arch/arm/boot/compressed/efi-header.S +1 -1

                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

Quoted 1 line

How about this?

It fixed the bug! (I tested with and without your previous patch and it worked in both case)

When you will send your patch, you can add my:

Tested-by Gregory CLEMENT <[email protected]>

Thanks,

Gregory

Quoted 80 lines · Gregory CLEMENT

Hi Ard,

Quoted 76 lines · Ard Biesheuvel

Quoted 72 lines · Gregory CLEMENT

Hi Ard,

Quoted 39 lines · Ard Biesheuvel

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 7 lines

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

I've just tested this change and it didn't fix anything.

Gregory

Quoted 16 lines

arch/arm/boot/compressed/efi-header.S +1 -1

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 
                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

Quoted 8 lines

Quoted 1 line

How about this?

It fixed the bug! (I tested with and without your previous patch and it worked in both case)

When you will send your patch, you can add my:

Tested-by Gregory CLEMENT <[email protected]>

Quoted 82 lines · Ard Biesheuvel

Quoted 80 lines · Gregory CLEMENT

Hi Ard,

Quoted 76 lines · Ard Biesheuvel

Quoted 72 lines · Gregory CLEMENT

Hi Ard,

Quoted 39 lines · Ard Biesheuvel

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 7 lines

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

I've just tested this change and it didn't fix anything.

Gregory

Quoted 16 lines

arch/arm/boot/compressed/efi-header.S +1 -1

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 
                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

Quoted 8 lines

Quoted 8 lines

Quoted 1 line

How about this?

It fixed the bug! (I tested with and without your previous patch and it worked in both case)

When you will send your patch, you can add my:

Tested-by Gregory CLEMENT <[email protected]>

Would you mind checking whether this fixes the issue as well?

arch/arm/boot/compressed/piggy.S +1 -0

index f72088495f43..5d52c556dd32 100644
@@ -1,5 +1,6 @@
        .section .piggydata,#alloc
        .globl  input_data
+       .align  2
 input_data:
        .incbin "arch/arm/boot/compressed/piggy_data"
        .globl  input_data_end

There may be other reasons than ksymtab entries that could result in piggydata ending up unaligned in the decompressor (which is what caused the issue before)
This is a better fix, because it addresses the root cause.

Hi Ard,

Quoted 106 lines · Ard Biesheuvel

Quoted 82 lines · Ard Biesheuvel

Quoted 80 lines · Gregory CLEMENT

Hi Ard,

Quoted 76 lines · Ard Biesheuvel

Quoted 72 lines · Gregory CLEMENT

Hi Ard,

Quoted 39 lines · Ard Biesheuvel

Quoted 38 lines · Ard Biesheuvel

Quoted 16 lines · Gregory CLEMENT

Hi Ard,

Quoted 11 lines · Ard Biesheuvel

To prevent unintended modifications to the kernel text (malicious or otherwise) while running the EFI stub, describe the kernel image as two separate sections: a .text section with read-execute permissions, covering .text, .rodata, .piggytext and the GOT sections (which the stub does not care about anyway), and a .data section with read-write permissions, covering .data and .bss.

This relies on the firmware to actually take the section permission flags into account, but this is something that is currently being implemented in EDK2, which means we will likely start seeing it in the wild between one and two years from now.

Quoted 6 lines

This patch had been merged in mainline yesterday and now prevent the Marvell Armada 370 and the Armada XP based SoC to boot. I also suspect that more Socs are impacted because the number of boot fail exploded according to kci:
https://kernelci.org/boot/all/job/mainline/branch/master/kernel/v4.13-8899-g8dc5b3a6cb2f/

Ouch.

Quoted 10 lines

I found this patch after bisecting (I can provide the bisect log if needed).

The kernel failed to boot only if CONFIG_EFI is enabled so it occurs in multi_v7_defconfig but not with mvebu_v7_defconfig.

Currently the solution is to revert this patch.

Have you a better option?

Quoted 7 lines

Quoted 1 line

I will investigate.

I cannot reproduce this on QEMU or my Beaglebone white. I have tried a locally built zImage as well as the one built by kernelci.

Could you please try whether this fixes things? It does not explain anything but it will help me figure out what is going on (hopefully)

I've just tested this change and it didn't fix anything.

Gregory

Quoted 16 lines

arch/arm/boot/compressed/efi-header.S +1 -1

index c94a88ae834d..671a6e5b7b99 100644
@@ -127,7 +127,7 @@ section_table:
 
                .set    section_count, (. - section_table) / 40
 
-               .align  12
+               .align  9
 __efi_start:
 #endif
                .endm

Quoted 8 lines

Quoted 8 lines

Quoted 1 line

How about this?

It fixed the bug! (I tested with and without your previous patch and it worked in both case)

When you will send your patch, you can add my:

Tested-by Gregory CLEMENT <[email protected]>

Would you mind checking whether this fixes the issue as well?

arch/arm/boot/compressed/piggy.S +1 -0

index f72088495f43..5d52c556dd32 100644
@@ -1,5 +1,6 @@
        .section .piggydata,#alloc
        .globl  input_data
+       .align  2
 input_data:
        .incbin "arch/arm/boot/compressed/piggy_data"
        .globl  input_data_end

Quoted 4 lines

There may be other reasons than ksymtab entries that could result in piggydata ending up unaligned in the decompressor (which is what caused the issue before)
This is a better fix, because it addresses the root cause.

I've tested this single patch and it does not fix the boot issue :(