From: Ard Biesheuvel ardb@kernel.org
commit bc310baf2ba381c648983c7f4748327f17324562 upstream.
The final build stage of the x86 kernel captures some symbol addresses from the decompressor binary and copies them into zoffset.h. It uses sed with a regular expression that matches the address, symbol type and symbol name, and mangles the captured addresses and the names of symbols of interest into #define directives that are added to zoffset.h
The symbol type is indicated by a single letter, which we match strictly: only letters in the set 'ABCDGRSTVW' are matched, even though the actual symbol type is relevant and therefore ignored.
Commit bc7c9d620 ("efi/libstub/x86: Force 'hidden' visibility for extern declarations") made a change to the way external symbol references are classified, resulting in 'startup_32' now being emitted as a hidden symbol. This prevents the use of GOT entries to refer to this symbol via its absolute address, which recent toolchains (including Clang based ones) already avoid by default, making this change a no-op in the majority of cases.
However, as it turns out, the LLVM linker classifies such hidden symbols as symbols with static linkage in fully linked ELF binaries, causing tools such as NM to output a lowercase 't' rather than an upper case 'T' for the type of such symbols. Since our sed expression only matches upper case letters for the symbol type, the line describing startup_32 is disregarded, resulting in a build error like the following
arch/x86/boot/header.S:568:18: error: symbol 'ZO_startup_32' can not be undefined in a subtraction expression init_size: .long (0x00000000008fd000 - ZO_startup_32 + (((0x0000000001f6361c + ((0x0000000001f6361c >> 8) + 65536) - 0x00000000008c32e5) + 4095) & ~4095)) # kernel initialization size
Given that we are only interested in the value of the symbol, let's match any character in the set 'a-zA-Z' instead.
Signed-off-by: Ard Biesheuvel ardb@kernel.org Signed-off-by: Ingo Molnar mingo@kernel.org Tested-by: Nathan Chancellor natechancellor@gmail.com Signed-off-by: Nathan Chancellor natechancellor@gmail.com ---
Hi all,
Please apply this patch to 5.4 (and older releases if you feel it necessary), as it fixes a build error that I see when linking with ld.lld on certain distribution configurations after upstream commit 5214028dd89e ("x86/boot: Correct relocation destination on old linkers") was applied in 5.4.48.
$ make -skj"$(nproc)" CC=clang LD=ld.lld O=out/x86_64 olddefconfig bzImage ... ld.lld: error: undefined symbol: ZO__end
referenced by arch/x86/boot/header.o:(.header+0x71)
...
While the commit message references bc7c9d620 as the first problematic commit, I see the same behavior of capital versus lowercase letters from nm here too. I assume this is not seen in mainline because this commit was already in the tree when 5214028dd89e was applied.
v5.4.47:
$ nm -S out/x86_64/arch/x86/boot/compressed/vmlinux | grep " _end" 000000000094b000 B _end
$ cat out/x86_64/arch/x86/boot/zoffset.h #define ZO__ehead 0x00000000000003b1 #define ZO__end 0x000000000094b000 #define ZO__text 0x000000000090ce50 #define ZO_efi32_stub_entry 0x0000000000000190 #define ZO_efi64_stub_entry 0x0000000000000390 #define ZO_efi_pe_entry 0x00000000000002f0 #define ZO_input_data 0x00000000000003b1 #define ZO_startup_32 0x0000000000000000 #define ZO_startup_64 0x0000000000000200 #define ZO_z_input_len 0x000000000090ca9e #define ZO_z_output_len 0x0000000002eeb42c
v5.4.48:
$ nm -S out/x86_64/arch/x86/boot/compressed/vmlinux | grep " _end" 000000000094b000 b _end
$ cat out/x86_64/arch/x86/boot/zoffset.h #define ZO__ehead 0x00000000000003b1 #define ZO__text 0x000000000090ccf0 #define ZO_efi32_stub_entry 0x0000000000000190 #define ZO_efi64_stub_entry 0x0000000000000390 #define ZO_efi_pe_entry 0x00000000000002f0 #define ZO_input_data 0x00000000000003b1 #define ZO_startup_32 0x0000000000000000 #define ZO_startup_64 0x0000000000000200 #define ZO_z_input_len 0x000000000090c93b #define ZO_z_output_len 0x0000000002eeb4c8
v5.4.48 with this patch:
$ nm -S out/x86_64/arch/x86/boot/compressed/vmlinux | grep " _end" 000000000094b000 b _end
$ cat out/x86_64/arch/x86/boot/zoffset.h #define ZO__ehead 0x00000000000003b1 #define ZO__end 0x000000000094b000 #define ZO__text 0x000000000090cd60 #define ZO_efi32_stub_entry 0x0000000000000190 #define ZO_efi64_stub_entry 0x0000000000000390 #define ZO_efi_pe_entry 0x00000000000002f0 #define ZO_input_data 0x00000000000003b1 #define ZO_startup_32 0x0000000000000000 #define ZO_startup_64 0x0000000000000200 #define ZO_z_input_len 0x000000000090c9af #define ZO_z_output_len 0x0000000002eeb4c8
Hopefully this clears things up.
Cheers, Nathan
arch/x86/boot/Makefile | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/x86/boot/Makefile b/arch/x86/boot/Makefile index e2839b5c246c..6539c50fb9aa 100644 --- a/arch/x86/boot/Makefile +++ b/arch/x86/boot/Makefile @@ -87,7 +87,7 @@ $(obj)/vmlinux.bin: $(obj)/compressed/vmlinux FORCE
SETUP_OBJS = $(addprefix $(obj)/,$(setup-y))
-sed-zoffset := -e 's/^([0-9a-fA-F]*) [ABCDGRSTVW] (startup_32|startup_64|efi32_stub_entry|efi64_stub_entry|efi_pe_entry|input_data|_end|_ehead|_text|z_.*)$$/#define ZO_\2 0x\1/p' +sed-zoffset := -e 's/^([0-9a-fA-F]*) [a-zA-Z] (startup_32|startup_64|efi32_stub_entry|efi64_stub_entry|efi_pe_entry|input_data|_end|_ehead|_text|z_.*)$$/#define ZO_\2 0x\1/p'
quiet_cmd_zoffset = ZOFFSET $@ cmd_zoffset = $(NM) $< | sed -n $(sed-zoffset) > $@
base-commit: 67cb016870e2fa9ffc8d34cf20db5331e6f2cf4d
On Mon, Jun 22, 2020 at 07:56:39PM +0000, Nathan Chancellor wrote:
From: Ard Biesheuvel ardb@kernel.org
commit bc310baf2ba381c648983c7f4748327f17324562 upstream.
The final build stage of the x86 kernel captures some symbol addresses from the decompressor binary and copies them into zoffset.h. It uses sed with a regular expression that matches the address, symbol type and symbol name, and mangles the captured addresses and the names of symbols of interest into #define directives that are added to zoffset.h
The symbol type is indicated by a single letter, which we match strictly: only letters in the set 'ABCDGRSTVW' are matched, even though the actual symbol type is relevant and therefore ignored.
Commit bc7c9d620 ("efi/libstub/x86: Force 'hidden' visibility for extern declarations") made a change to the way external symbol references are classified, resulting in 'startup_32' now being emitted as a hidden symbol. This prevents the use of GOT entries to refer to this symbol via its absolute address, which recent toolchains (including Clang based ones) already avoid by default, making this change a no-op in the majority of cases.
However, as it turns out, the LLVM linker classifies such hidden symbols as symbols with static linkage in fully linked ELF binaries, causing tools such as NM to output a lowercase 't' rather than an upper case 'T' for the type of such symbols. Since our sed expression only matches upper case letters for the symbol type, the line describing startup_32 is disregarded, resulting in a build error like the following
arch/x86/boot/header.S:568:18: error: symbol 'ZO_startup_32' can not be undefined in a subtraction expression init_size: .long (0x00000000008fd000 - ZO_startup_32 + (((0x0000000001f6361c + ((0x0000000001f6361c >> 8) + 65536) - 0x00000000008c32e5) + 4095) & ~4095)) # kernel initialization size
Given that we are only interested in the value of the symbol, let's match any character in the set 'a-zA-Z' instead.
Signed-off-by: Ard Biesheuvel ardb@kernel.org Signed-off-by: Ingo Molnar mingo@kernel.org Tested-by: Nathan Chancellor natechancellor@gmail.com Signed-off-by: Nathan Chancellor natechancellor@gmail.com
Hi all,
Please apply this patch to 5.4 (and older releases if you feel it necessary), as it fixes a build error that I see when linking with ld.lld on certain distribution configurations after upstream commit 5214028dd89e ("x86/boot: Correct relocation destination on old linkers") was applied in 5.4.48.
$ make -skj"$(nproc)" CC=clang LD=ld.lld O=out/x86_64 olddefconfig bzImage ... ld.lld: error: undefined symbol: ZO__end
referenced by arch/x86/boot/header.o:(.header+0x71)
...
While the commit message references bc7c9d620 as the first problematic commit, I see the same behavior of capital versus lowercase letters from nm here too. I assume this is not seen in mainline because this commit was already in the tree when 5214028dd89e was applied.
I've queued this for 5.4-4.9, thanks!
linux-stable-mirror@lists.linaro.org
-
Nathan Chancellor -
Sasha Levin