Message ID | 20240126161330.1483-1-nafeabd@amazon.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | [RFC] arm64: use different compiler inlining options for arm64 kernel builds | expand |
On Fri, Jan 26, 2024 at 04:13:30PM +0000, Nafees Ahmed Abdul wrote: > It was observed that compiling the kernel with following compiler inlining > options has showed performance improvement up to 4% across tested > workloads. > > Below is the table with performance gains compared to default kernel > on few of the workloads we tested. > > | Workload | Improvement | > |-----------|-------------| > | Nginx | 4% | > | MongoDB | 3% | > | Memcached | 3% | > | Redis | 2% | > > The selected options increase inlining of the kernel functions which is > positively impacting workload performance. We have arrived at these values > by reviewing the options used with the gcc compiler at various optimization > levels such as O1, O2 and O3. > > Benchmarking was performed on AWS Graviton 2/3 generation instances > on 5.10 and 6.1 longterm supported kernels with GCC versions 10 and 11. These values will be applied to every single arm64 kernel build, regardless of hardware and GCC version (the kernel supports GCC 5.1.0 as a minimum version), are you sure that these values won't regress other people's workloads? Furthermore, this basically ties the arm64 kernel build to GCC internals, which the documentation notes "The names of specific parameters, and the meaning of the values, are tied to the internals of the compiler, and are subject to change without notice in future releases.". This does not really feel like something the upstream kernel should take but I am not an arm64 maintainer so I can't comment further. > Patch also adds noinline attribute to couple of functions for existing > kselftests to pass on ftrace target. That seems like a bandaid to me and probably warrants further investigation/explanation if adding these flags is the only thing that it takes to expose that, as it means there is either a latent compiler bug or the kernel has some dodgy code that happens to work now but some future internal compiler change could expose. > Signed-off-by: Nafees Ahmed Abdul <nafeabd@amazon.com> > --- > arch/arm64/Makefile | 6 ++++++ > fs/namei.c | 2 +- > fs/open.c | 2 +- > 3 files changed, 8 insertions(+), 2 deletions(-) > > diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile > index 47ecc4cff9d2..4ad2c7d0008d 100644 > --- a/arch/arm64/Makefile > +++ b/arch/arm64/Makefile > @@ -101,6 +101,12 @@ KBUILD_CFLAGS += -Wa,-march=$(asm-arch) \ > -DARM64_ASM_ARCH='"$(asm-arch)"' > endif > > +ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE > +KBUILD_CFLAGS += --param=early-inlining-insns=14 --param=inline-heuristics-hint-percent=600 \ > + --param=inline-min-speedup=15 --param=max-inline-insns-auto=30 \ > + --param=max-inline-insns-single=200 > +endif Regardless, the Linux kernel can also be built with clang, so this block needs to have a dependency on CONFIG_CC_IS_GCC as well, but I'd wait to send any follow up before others have a chance to comment on that. > + > ifeq ($(CONFIG_SHADOW_CALL_STACK), y) > KBUILD_CFLAGS += -ffixed-x18 > endif > diff --git a/fs/namei.c b/fs/namei.c > index 71c13b2990b4..58e8f462c137 100644 > --- a/fs/namei.c > +++ b/fs/namei.c > @@ -4445,7 +4445,7 @@ SYSCALL_DEFINE1(unlink, const char __user *, pathname) > * On non-idmapped mounts or if permission checking is to be performed on the > * raw inode simply passs @nop_mnt_idmap. > */ > -int vfs_symlink(struct mnt_idmap *idmap, struct inode *dir, > +noinline vfs_symlink(struct mnt_idmap *idmap, struct inode *dir, > struct dentry *dentry, const char *oldname) > { > int error; > diff --git a/fs/open.c b/fs/open.c > index 02dc608d40d8..afec5d0e30d3 100644 > --- a/fs/open.c > +++ b/fs/open.c > @@ -1421,7 +1421,7 @@ struct file *file_open_root(const struct path *root, > } > EXPORT_SYMBOL(file_open_root); > > -static long do_sys_openat2(int dfd, const char __user *filename, > +static noinline long do_sys_openat2(int dfd, const char __user *filename, > struct open_how *how) > { > struct open_flags op; > -- > 2.40.1 > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile index 47ecc4cff9d2..4ad2c7d0008d 100644 --- a/arch/arm64/Makefile +++ b/arch/arm64/Makefile @@ -101,6 +101,12 @@ KBUILD_CFLAGS += -Wa,-march=$(asm-arch) \ -DARM64_ASM_ARCH='"$(asm-arch)"' endif +ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE +KBUILD_CFLAGS += --param=early-inlining-insns=14 --param=inline-heuristics-hint-percent=600 \ + --param=inline-min-speedup=15 --param=max-inline-insns-auto=30 \ + --param=max-inline-insns-single=200 +endif + ifeq ($(CONFIG_SHADOW_CALL_STACK), y) KBUILD_CFLAGS += -ffixed-x18 endif diff --git a/fs/namei.c b/fs/namei.c index 71c13b2990b4..58e8f462c137 100644 --- a/fs/namei.c +++ b/fs/namei.c @@ -4445,7 +4445,7 @@ SYSCALL_DEFINE1(unlink, const char __user *, pathname) * On non-idmapped mounts or if permission checking is to be performed on the * raw inode simply passs @nop_mnt_idmap. */ -int vfs_symlink(struct mnt_idmap *idmap, struct inode *dir, +noinline vfs_symlink(struct mnt_idmap *idmap, struct inode *dir, struct dentry *dentry, const char *oldname) { int error; diff --git a/fs/open.c b/fs/open.c index 02dc608d40d8..afec5d0e30d3 100644 --- a/fs/open.c +++ b/fs/open.c @@ -1421,7 +1421,7 @@ struct file *file_open_root(const struct path *root, } EXPORT_SYMBOL(file_open_root); -static long do_sys_openat2(int dfd, const char __user *filename, +static noinline long do_sys_openat2(int dfd, const char __user *filename, struct open_how *how) { struct open_flags op;
It was observed that compiling the kernel with following compiler inlining options has showed performance improvement up to 4% across tested workloads. Below is the table with performance gains compared to default kernel on few of the workloads we tested. | Workload | Improvement | |-----------|-------------| | Nginx | 4% | | MongoDB | 3% | | Memcached | 3% | | Redis | 2% | The selected options increase inlining of the kernel functions which is positively impacting workload performance. We have arrived at these values by reviewing the options used with the gcc compiler at various optimization levels such as O1, O2 and O3. Benchmarking was performed on AWS Graviton 2/3 generation instances on 5.10 and 6.1 longterm supported kernels with GCC versions 10 and 11. Patch also adds noinline attribute to couple of functions for existing kselftests to pass on ftrace target. Signed-off-by: Nafees Ahmed Abdul <nafeabd@amazon.com> --- arch/arm64/Makefile | 6 ++++++ fs/namei.c | 2 +- fs/open.c | 2 +- 3 files changed, 8 insertions(+), 2 deletions(-)