By default GCC uses "simple" coverage counter update
mechanism. It is perfectly fine for single-threaded (or single CPU in
our case) setups, but will cause race conditions on SMP systems.
For example, I observed that counters are going backwards when running
Xen inside QEMU.
GCC starting from version 7 and LLVM/Clang starting from version 11
support -fprofile-update=atomic option, which forces coverage counter
updates to be atomic, which resolves the issue. As Xen runs mostly on
SMP systems, force use this option if it is supported by a compiler.
Signed-off-by: Volodymyr Babchuk <volodymyr_babchuk@epam.com>
---
xen/Kconfig | 4 ++++
xen/Rules.mk | 4 ++++
2 files changed, 8 insertions(+)
diff --git a/xen/Kconfig b/xen/Kconfig
index 07c4accf88..3a6635ac74 100644
--- a/xen/Kconfig
+++ b/xen/Kconfig
@@ -55,6 +55,10 @@ config CC_HAS_ASM_GOTO_OUTPUT
config CC_HAS_MCDC
def_bool $(cc-option,-fcondition-coverage)
+# Compiler supports -fprofile-update=atomic for correct SMP handling
+config CC_HAS_ATOMIC_PROFILE
+ def_bool $(cc-option,-fprofile-update=atomic)
+
# Set code alignment.
#
# Allow setting on a boolean basis, and then convert such selection to an
diff --git a/xen/Rules.mk b/xen/Rules.mk
index da21850926..7219a3033f 100644
--- a/xen/Rules.mk
+++ b/xen/Rules.mk
@@ -141,6 +141,10 @@ else
cov-cflags-$(CONFIG_CONDITION_COVERAGE) += -fcondition-coverage
endif
+ifeq ($(CONFIG_CC_HAS_ATOMIC_PROFILE),y)
+ cov-cflags-$(CONFIG_COVERAGE) += -fprofile-update=atomic
+endif
+
# Reset cov-cflags-y in cases where an objects has another one as prerequisite
$(nocov-y) $(filter %.init.o, $(obj-y) $(obj-bin-y) $(extra-y)): \
cov-cflags-y :=
--
2.50.0