>From version 2.35, binutils can be configured with
--enable-mips-fix-loongson3-llsc=yes, which means it defaults to
-mfix-loongson3-llsc. This breaks labels which might then point at the
wrong instruction.
The workaround to explicitly pass -mno-fix-loongson3-llsc has been
added in Linux version 5.1, but is only enabled when building a Loongson
64 kernel. As vendors might use a common toolchain for building Loongson
and non-Loongson kernels, just move that workaround to
arch/mips/Makefile. At the same time update the comments to reflect the
current status.
Cc: stable(a)vger.kernel.org # 5.1+
Cc: YunQiang Su <syq(a)debian.org>
Signed-off-by: Aurelien Jarno <aurelien(a)aurel32.net>
---
arch/mips/Makefile | 19 +++++++++++++++++++
arch/mips/loongson64/Platform | 22 ----------------------
2 files changed, 19 insertions(+), 22 deletions(-)
diff --git a/arch/mips/Makefile b/arch/mips/Makefile
index cd4343edeb11..5ffdd67093bc 100644
--- a/arch/mips/Makefile
+++ b/arch/mips/Makefile
@@ -136,6 +136,25 @@ cflags-$(CONFIG_SB1XXX_CORELIS) += $(call cc-option,-mno-sched-prolog) \
#
cflags-y += -fno-stack-check
+# binutils from v2.35 when built with --enable-mips-fix-loongson3-llsc=yes,
+# supports an -mfix-loongson3-llsc flag which emits a sync prior to each ll
+# instruction to work around a CPU bug (see __SYNC_loongson3_war in asm/sync.h
+# for a description).
+#
+# We disable this in order to prevent the assembler meddling with the
+# instruction that labels refer to, ie. if we label an ll instruction:
+#
+# 1: ll v0, 0(a0)
+#
+# ...then with the assembler fix applied the label may actually point at a sync
+# instruction inserted by the assembler, and if we were using the label in an
+# exception table the table would no longer contain the address of the ll
+# instruction.
+#
+# Avoid this by explicitly disabling that assembler behaviour.
+#
+cflags-y += $(call as-option,-Wa$(comma)-mno-fix-loongson3-llsc,)
+
#
# CPU-dependent compiler/assembler options for optimization.
#
diff --git a/arch/mips/loongson64/Platform b/arch/mips/loongson64/Platform
index ec42c5085905..e2354e128d9a 100644
--- a/arch/mips/loongson64/Platform
+++ b/arch/mips/loongson64/Platform
@@ -5,28 +5,6 @@
cflags-$(CONFIG_CPU_LOONGSON64) += -Wa,--trap
-#
-# Some versions of binutils, not currently mainline as of 2019/02/04, support
-# an -mfix-loongson3-llsc flag which emits a sync prior to each ll instruction
-# to work around a CPU bug (see __SYNC_loongson3_war in asm/sync.h for a
-# description).
-#
-# We disable this in order to prevent the assembler meddling with the
-# instruction that labels refer to, ie. if we label an ll instruction:
-#
-# 1: ll v0, 0(a0)
-#
-# ...then with the assembler fix applied the label may actually point at a sync
-# instruction inserted by the assembler, and if we were using the label in an
-# exception table the table would no longer contain the address of the ll
-# instruction.
-#
-# Avoid this by explicitly disabling that assembler behaviour. If upstream
-# binutils does not merge support for the flag then we can revisit & remove
-# this later - for now it ensures vendor toolchains don't cause problems.
-#
-cflags-$(CONFIG_CPU_LOONGSON64) += $(call as-option,-Wa$(comma)-mno-fix-loongson3-llsc,)
-
#
# binutils from v2.25 on and gcc starting from v4.9.0 treat -march=loongson3a
# as MIPS64 R2; older versions as just R1. This leaves the possibility open
--
2.29.2
This series hunts the problems discovered after manual enabling of
ARCH_WANT_LD_ORPHAN_WARN. Notably:
- adds the missing PAGE_ALIGNED_DATA() section affecting VDSO
placement (marked for stable);
- stops blind catching of orphan text sections with .text.*
directive;
- properly stops .eh_frame section generation.
Compile and runtime tested on MIPS32R2 CPS board with no issues
using two different toolkits:
- Binutils 2.35.1, GCC 10.2.1 (with Alpine patches);
- LLVM stack: 11.0.0 and from latest Git snapshot.
Since v4 [3]:
- new: drop redundant .text.cps-vec creation and blind inclusion
of orphan text sections via .text.* directive in vmlinux.lds.S;
- don't assert SIZEOF(.rel.dyn) as it's reported that it may be not
empty on certain machines and compilers (Thomas);
- align GOT table like it's done for ARM64;
- new: catch UBSAN's "unnamed data" sections in generic definitions
when building with LD_DEAD_CODE_DATA_ELIMINATION;
- collect Reviewed-bys (Kees, Nathan).
Since v3 [2]:
- fix the third patch as GNU stack emits .rel.dyn into VDSO for
some reason if .cfi_sections is specified.
Since v2 [1]:
- stop discarding .eh_frame and just prevent it from generating
(Kees);
- drop redundant sections assertions (Fangrui);
- place GOT table in .text instead of asserting as it's not empty
when building with LLVM (Nathan);
- catch compound literals in generic definitions when building with
LD_DEAD_CODE_DATA_ELIMINATION (Kees);
- collect two Reviewed-bys (Kees).
Since v1 [0]:
- catch .got entries too as LLD may produce it (Nathan);
- check for unwanted sections to be zero-sized instead of
discarding (Fangrui).
[0] https://lore.kernel.org/linux-mips/20210104121729.46981-1-alobakin@pm.me
[1] https://lore.kernel.org/linux-mips/20210106200713.31840-1-alobakin@pm.me
[2] https://lore.kernel.org/linux-mips/20210107115120.281008-1-alobakin@pm.me
[3] https://lore.kernel.org/linux-mips/20210107123331.354075-1-alobakin@pm.me
Alexander Lobakin (9):
MIPS: vmlinux.lds.S: add missing PAGE_ALIGNED_DATA() section
MIPS: CPS: don't create redundant .text.cps-vec section
MIPS: vmlinux.lds.S: add ".gnu.attributes" to DISCARDS
MIPS: properly stop .eh_frame generation
MIPS: vmlinux.lds.S: explicitly catch .rel.dyn symbols
MIPS: vmlinux.lds.S: explicitly declare .got table
vmlinux.lds.h: catch compound literals into data and BSS
vmlinux.lds.h: catch UBSAN's "unnamed data" into data
MIPS: select ARCH_WANT_LD_ORPHAN_WARN
arch/mips/Kconfig | 1 +
arch/mips/include/asm/asm.h | 18 ++++++++++++++++++
arch/mips/kernel/cps-vec.S | 1 -
arch/mips/kernel/vmlinux.lds.S | 11 +++++++++--
include/asm-generic/vmlinux.lds.h | 6 +++---
5 files changed, 31 insertions(+), 6 deletions(-)
--
2.30.0